ICT Systems and Sustainability: Milan Tuba Shyam Akashe Amit Joshi Editors

Advances in Intelligent Systems and Computing 1270
Milan Tuba
Shyam Akashe
Amit Joshi Editors
ICT Systems and

Sustainability
Proceedings of ICT4SD 2020, Volume 1
Advances in Intelligent Systems and Computing
Volume 1270
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, School of Computer Science and Electronic Engineering,
University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen , Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
Indexed by SCOPUS, DBLP, EI Compendex, INSPEC, WTI Frankfurt eG,
zbMATH, Japanese Science and Technology Agency (JST), SCImago.
All books published in the series are submitted for consideration in Web of
Science.
More information about this series at http://www.springer.com/series/11156

Milan Tuba Shyam Akashe Amit Joshi
• •
Editors
ICT Systems
and Sustainability
Proceedings of ICT4SD 2020, Volume 1
123
Editors
Milan Tuba Shyam Akashe
Singidunum University ITM University
Belgrade, Serbia Gwalior, Madhya Pradesh, India
Amit Joshi
Global Knowledge Research Foundation
Ahmedabad, Gujarat, India
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-981-15-8288-2 ISBN 978-981-15-8289-9 (eBook)
https://doi.org/10.1007/978-981-15-8289-9
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
The Fifth International Conference on ICT for Sustainable Development (ICT4SD

2020) targets theory, development, applications, experiences and evaluation of
interaction sciences with fellow students, researchers and practitioners. The con-
ference is devoted to increasing the understanding role of technology issues, how
engineering has day by day evolved to prepare human-friendly technology. The
conference provided a platform for bringing forth significant research and literature
across the field of ICT for Sustainable Development and provided an overview
of the technologies awaiting unveiling. This interaction will be the focal point for
leading experts to share their insights, provide guidance and address participant’s
questions and concerns. The conference was to be held during July 23–24, 2020, at
Hotel Vivanta by Taj, Panaji, Goa, India, but due to the pandemic, this year it was
organized through digital mode. The conference was organized by Global
Knowledge Research Foundation, Supporting Partner InterYIT, International
Federation for Information Processing, State Chamber Partner Goa Chamber of
Commerce & Industry, National Chamber Partner as Knowledge Chamber of
Commerce & Industry. Research submissions in various advanced technology areas
were received after a rigorous peer review process with the help of program
committee members and 187 external reviewers for 1000+ papers from 19 different
countries including Algeria, USA, United Arab Emirates, Serbia, Qatar, Mauritius,
Egypt, Saudi Arabia, Ethiopia, Oman out of which 160 were accepted with an
acceptance ratio of 0.15. Technology is the driving force of progress in this era of
globalization. Information and communication technology (ICT) has become a
functional requirement for the socioeconomic growth and sustained development of
any country. The influence of information communications technology (ICT) in
shaping the process of globalization, particularly in productivity, commercial and
financial spheres, is widely recognized. The ICT sector is undergoing a revolution
that has momentous implications for the current and future social and economic
situation of all the countries in the world. ICT plays a pivotal role in empowering
people for self-efficacy and how it can facilitate this mission to reach out to
grassroots level. Finally, it is concluded that ICT is a significant contributor to the
success of the ongoing initiative of Startup India. In order to recognize and reward
v
vi Preface
the extraordinary performance and achievements by ICT and allied sectors and
promote universities, researchers and students through their research work adapting
new scientific technologies and innovations. The two-day conference had presen-
tations from the researchers, scientists, academia and students on the research
work carried out by them in different sectors. ICT4SD Summit is a flagship event
of the G R Foundation. This is the fourth edition. The summit was addressed by
eminent dignitaries including Shri Manguirsh Pai Raikar, Chairperson,
ASSOCHAM MSME National Council; Shri. Prajyot Mainkar, Chairman, IT
Committee of Goa Chamber of Commerce and Industry; Mike Hinchey, President,
IFIP and Chair IEEE, UK and Ireland; Milan Tuba, Vice-Rector for International
Relations, Singidunum University, Serbia; Prof. Lance Fung, Australia; Prof.
Jagdish Bansal, India; Mr. Aninda Bose, Springer; Dr. Amit Joshi, Director, G R
Foundation. The overall conference had one inaugural session, one keynote session
and 18 technical sessions during two days.
Belgrade, Serbia Milan Tuba

Gwalior, India Shyam Akashe
Ahmedabad, India Amit Joshi
Contents
Challenges in Adoption of Secure Digitization of Graduate

Assessments: Scope for Blockchain Technologies . . . . . . . . . . . . . . . . . . 1
S. K. Shankar, Rajeswari Mukesh, and M. K. Badrinarayanan
Non-performing Asset Analysis Using Machine Learning . . . . . . . . . . . 11
Rohan Ninan Jacob
Volatile Memory Disk Forensics: Investigate the Criminal Activity
of RAMDisk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Nilay Mistry, Annan Christian, and Bansari Bhavsar
Modem Functionality Diagnostic Tool with Graphical User
Interface for Cellular Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Arghya Biswas, Sudakar Singh Chauhan, and Antara Borwankar
Regression Model to Estimate Effort for Software-Oriented
Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
V. Vignaraj Ananth, S. Aditya, S. Ragul Kesavan, and B. Keerthiga
Innovative Technology-Based Social Enterprises for Inclusive
Sustainable Healthcare in India . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Pradnya Vishwas Chitrao, Pravin Kumar Bhoyar, and Rajiv Divekar
Classification of Leaves Using Convolutional Neural Network
and Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Sapan Naik and Hinal Shah
A Low-Cost IoT-Enabled Device for Real-Time Air Quality
Monitoring System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Sai Surya Kiran Pokala, V. Panchami, Kelavath Jaisingh,
and Sandeep Kumar Yedla
Fuzzy-Based Predictive Analytics for Early Detection of
Disease—A Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . 89
V. Kakulapati, R. Sai Sandeep, V. Kranthi kumar, and R. Ramanjinailu
vii
viii Contents
A Framework for Remote Health Monitoring . . . . . . . . . . . . . . . . . . . . 101

K. Viswavardhan Reddy and Navin Kumar
GenNext—An Extractive Graph-Based Text Summarization Model
for User Reviews . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Komal Kothari, Aagam Shah, Satvik Khara, and Himanshu Prajapati
An Adaptive Control for Surrogate Assisted Multi-objective
Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Duc Dinh Nguyen and Long Nguyen
Study and Analysis of NOMA Downlink Transmission Model
in 5G Wireless System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Sanjeev Kumar Srivastava
Network Structural Analysis Based on Graphical Measures and
Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Atul Kumar Verma, Mahipal Jadeja, and Rahul Saxena
Lung Tumor Classification Using CNN- and GLCM-Based
Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Amrita Naik and Damodar Reddy Edla
A Distributed Implementation of Signature Matching Algorithms
for IoT Intrusion Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Menachem Domb
Graph-Based Multi-document Text Summarization Using NLP . . . . . . . 177
Abhilasha More and Vipul Dalal
Fault Exploratory Data Analysis of Real-Time Marine Diesel
Engine Data Using R Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Cheng Siong Chin and Nursyafiqah Binte Nazli
A Hybrid Key Management System Based on ECC and Data
Classification to Secure Data in the Cloud . . . . . . . . . . . . . . . . . . . . . . . 199
Hinddeep Purohit and Ravirajsinh Vaghela
Life Cycle Management for Indian Naval Warship Equipment . . . . . . . 209
Alok Bhagwat and Pradnya Vishwas Chitrao
Framework for Securing IoT Ecosystem Using Blockchain:
Use Cases Suggesting Theoretical Architecture . . . . . . . . . . . . . . . . . . . 223
Anshul Jain, Tanya Singh, and Nitesh Jain
Feature Selection for Handwritten Signature Recognition Using
Neighborhood Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Kamlesh Kumari and Sanjeev Rana
Blockchain-Based E-Voting Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Shreya Shailendra Chafe, Divya Ashok Bangad, and Harsha Sonune
Contents ix
SOLID: A Web System to Restore the Control of Users’

Personal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Manishkumar R. Solanki
Enhanced Three-Dimensional DV-Hop Algorithm . . . . . . . . . . . . . . . . . 269
Abhinesh Kaushik and D. K. Lobiyal
Synthesis of VLSI Structural Cell Partitioning Using Genetic
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
P. Rajeswari, Theodore S. Chandra, and Amith Kiran Kumar
Identification and Classification of Foreign Bodies from Rice Grains
Using Digital Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Rajlakshmi Ghatkamble
Improving Collaboration Between Government and Citizens for
Environmental Issues: Lessons Learned from a Case in Sri Lanka . . . . 297
Mohamed Sapraz and Shengnan Han
A Study on Blockchain Scalability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Manjula K. Pawar, Prakashgoud Patil, and P. S. Hiremath
What Influences Consumer Behavior Toward Information and
Communication Technology Applications: A Systematic Literature
Review of UTAUT2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Jick Castanha, Subhash Kizhakanveatil Bhaskaran Pillai, and Indrawati
Secure and Energy-Efficient Remote Monitoring Technique
(SERMT) for Smart Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Sohini Roy and Arunabha Sen
Balanced Accuracy of Collaborative Recommender System . . . . . . . . . . 341
Akanksha Bansal Chopra and Veer Sain Dixit
Using Transfer Learning for Detecting Drug Mentions in Tweets . . . . . 357
Laiba Mehnaz and Rajni Jindal
A Comprehensive and Critical Analysis of Cross-Domain Federated
Identity Management Deployments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Tejaswini Apte and Jatinderkumar R. Saini
Concentric Morphology Model in Detection of Masses from
Mammogram: A Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Madhavi Pingili, M. Shankar Lingam, and E. G. Rajan
A Survey on DevOps Techniques Used in Cloud-Based
IOT Mashups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
M. Ganeshan and P. Vigneshwaran
Digital Assistant with Augmented Reality . . . . . . . . . . . . . . . . . . . . . . . 395
Vijay Verma, Aditya Sharma, Ketan Kumar Jain, and Sushant Adlakha
x Contents
An Experimental Approach Toward Type 2 Diabetes Diagnosis

Using Cultural Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Ratna Patil, Sharvari Tamane, and Kanishk Patil
Predicting Students Performance Through Behavior and
Computational Thinking in Programming . . . . . . . . . . . . . . . . . . . . . . . 417
Vinayak Hegde, H. N. Meghana, R. Spandana, and M. S. Pallavi
Deployment of Ambulance with Maximum Area Coverage by Using
Particle Swarm Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 431
Amandeep Kaur
Detection of DDOS Attacks Using Machine Learning Techniques:
A Hybrid Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Datla Anurag Varma, Ravi Ashish, V. Venkata Sai Sandeep, B. Venkatesh,
and R. Kannadasan
Revisiting Algorithms for Resource Discovery in Unstructured
Peer to Peer Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Abhijith Jayan, Subi George Samuel, and P. Jayakumar
802.11 Frame-level Network IDS for Public Wireless Networks . . . . . . . 453
Anish Sujanani and Shashidhar Pai
Machine Learning for Malicious URL Detection . . . . . . . . . . . . . . . . . . 463
Gold Wejinya and Sajal Bhatia
Implementation of OSPFv3 in IPv4 and IPv6 for Establishing
a Wide Area Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Fatema Zaman Sinthia, Afsana Nasir, Bijan Paul,
Mohammad Rifat Ahmmad Rashid, and Md Nasim Adnan
Obstacle Avoiding and Fire Extinguishing Robot for Everyday
Fire Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Moumita Kabir, Fatima Noor Popy, Bijan Paul,
Mohammad Rifat Ahmmad Rashid, and Khan Raqib Mahmud
A Survey on Internet of Things (IoT): Communication Model,
Open Challenges, Security and Privacy Issues . . . . . . . . . . . . . . . . . . . . 493
R. Venkadeshan and M. Jegatha
Violence Detection Through Surveillance System . . . . . . . . . . . . . . . . . . 503
Abhishek Deshmukh, Kshitij Warang, Yash Pente, and Nilesh Marathe
Influence of Big Data Capabilities in Knowledge
Management—MSMEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Ravi Shankar Jha and Priti Ranjan Sahoo
Logistically Supervised Aspect Category Detection Using Data
Co-occurrence for Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Anice John and Reshma Sheik
Contents xi
Automated Dietary Monitoring System: A Novel Framework . . . . . . . . 535

Samiul Mamud, Punyasha Chatterjee, Saubhik Bandyopadhyay,
and Suchandra Bhandari
Intruder Insinuation and Smart Surveillance Using Face Detection
and Mask Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
M. Rajesh Khanna, G. Prakash Raj, S. Prem Kumar,
and N. S. Vignesh Raaj
Misinformation Analysis During Covid-19 Pandemic . . . . . . . . . . . . . . . 553
Shubhangi Rastogi and Divya Bansal
An Agent-Based Cloud Service Negotiation in Hybrid Cloud
Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
Saurabh Deochake and Debajyoti Mukhopadhyay
IoT-Based Smart Farming Application for Sustainable Agriculture . . . . 573
Sanjeevakumar M. Hatture and Pallavi V. Yankati
Follow Me: A Human Following Robot Using Wi-Fi Received
Signal Strength Indicator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
V. Geetha, Sanket Salvi, Gurdeep Saini, Naveen Yadav,
and Rudra Pratap Singh Tomar
Implementation of Arithmetic Unit for RNS Using 2n+ 3 as Base . . . . . 595
Nagaraj R. Aiholli, Uday V. Wali, and Rashmi Rachh
Hybridization of Artificial Bee Colony Algorithm with Estimation
of Distribution Algorithm for Minimum Weight Dominating
Set Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Sima Shetgaonkar and Alok Singh
Privacy Preservation for Data Sharing with Multi-party Access
Control in Cloud Computing: A Review Paper . . . . . . . . . . . . . . . . . . . 621
Shubhangi Tanaji Khot, Amrita A. Manjrekar, and Rashmi J. Deshmukh
Pixel-Based Attack on ODENet Classifiers . . . . . . . . . . . . . . . . . . . . . . . 629
Nitheesh Chandra Yaratapalli, Reethesh Venkataraman,
Abhishek Dinesan, Ashni Manish Bhagvandas, and Padmamala Sriram
Survey on Advanced Spectrum Sharing Using Cognitive Radio
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639
Mohamed Hassan, Manwinder Singh, and Khaild Hamid
An AI-Based Pedagogical Tool for Creating Sketched Representation
of Emotive Product Forms in the Conceptual Design Stages . . . . . . . . . 649
Sunny Prakash Prajapati, Rahul Bhaumik, Tarun Kumar, and Unais Sait
xii Contents
Social Interaction-Enabled Industrial Internet of Things for Predictive

Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
M. S. Roopa, B. Pallavi, Rajkumar Buyya, K. R. Venugopal, S. S. Iyengar,
and L. M. Patnaik
Emerging Trends in the Marketing of Financially Engineered
Insurance Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Sunitha Ratnakaram, Venkamaraju Chakravaram, Nitin Simha Vihari,
and G. Vidyasagar Rao
Attitude of the International Students Towards Integrating
ICT in Foreign Language Learning: A Case Study . . . . . . . . . . . . . . . . 685
Khushboo Kuddus and Nafis Mahmud Khan
Limiting Check Algorithm: Primality Test . . . . . . . . . . . . . . . . . . . . . . . 697
Gunjan Sethi and Harsh
Factors Affecting the Local Governance . . . . . . . . . . . . . . . . . . . . . . . . . 707
Ngo Sy Trung, Do Huu Hai, Pham Van Tuan, and Nguyen Tien Long
A Multi-factorial Code Coverage Based Test Case Selection
and Prioritization for Object Oriented Programs . . . . . . . . . . . . . . . . . . 721
Prashant Vats, Zunaid Aalam, Satnam Kaur, Amritpal Kaur,
and Saravjit Kaur
New Horizons for the Application of Microalgae in the National
Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
L. N. Medvedeva and O. Roiss
Context of the Ghanaian Government: Social Media and Access
to Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741
Adasa Nkrumah Kofi Frimpong, Ping Li, Samuel Adu-Gyamfi,
and Sandra Chukwudumebi Obiora
Convolutional Neural Network and Data Augmentation for
Behavioral-Based Biometric User Identification . . . . . . . . . . . . . . . . . . . 753
Sakorn Mekruksavanich and Anuchit Jitpattanakul
Cybersecurity Attacks: Analysis of “WannaCry” Attack and
Proposing Methods for Reducing or Preventing Such Attacks
in Future . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
Sumaiah Algarni
Enhancing Trust and Immutability in Cloud Forensics . . . . . . . . . . . . . 771
Pranay Chauhan and Pratosh Bansal
A Fuzzy Time Series Prediction Model of the COVID-19 Epidemic . . . . 779
Mohammad Minhazul Alam, S. M. Shahadat Hossain,
Md. Romman Riyadh Shishir, Sadman Hasan, Eumna Huda,
Sabrina Yeasmin, Abdul Motaleb, and Rashedur M. Rahman
Contents xiii
Modelling of Generation and Utilization Control for Isolated

Hybrid Microgrid for Rural Electrification . . . . . . . . . . . . . . . . . . . . . . 791
Kuldip Singh, Satyasis Mishra, Davinder Singh Rathee,
Demissie Jobir Gemecha, and R. C. Mohanty
Toyama de Walking: Practice of Location-Based Data Collection
Using Augmented Reality Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 805
Yuya Ieiri, Hung-Ya Tsai, and Reiko Hishiyama
Graphene-Based Wireless Power Transmission:
Charge as You Drive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
P. C. Nissimagoudar, H. M. Gireesha, R. M. Shet, Nalini C. Iyer,
Ajit Bijapur, and H. R. Aishwarya
Vision-Based Driver Authentication and Alertness Detection
Using HOG Feature Descriptor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
P. C. Nissimagoudar, A. V. Nandi, H. M. Gireesha, R. M. Shet,
and Nalini C. Iyer
ICT4D in Channelling Welfare Services in India: Case Study
of Haqdarshak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835
Komal Ramdey and Hasnain Bokhari
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845

About the Editors
Milan Tuba is the Vice Rector for International Relations, Singidunum

University, Belgrade, Serbia and was the Head of the Department for Mathematical
Sciences at State University of Novi Pazar and the Dean of the Graduate School of
Computer Science at John Naisbitt University. He received B.S. in Mathematics,
M.S. in Mathematics, M.S. in Computer Science, M.Ph. in Computer Science,
Ph.D. in Computer Science from University of Belgrade and New York University.
From 1983 to 1994 he was in the U.S.A. first at Vanderbilt University in Nashville
and Courant Institute of Mathematical Sciences, New York University and later as
Assistant Professor of Electrical Engineering at Cooper Union School of
Engineering, New York. During that time he was the founder and director of
Microprocessor Lab and VLSI Lab, leader of scientific projects and theses super-
visor. From 1994 he was Assistant Professor of Computer Science and Director of
Computer Center at University of Belgrade, from 2001 Associate Professor, Faculty
of Mathematics, University of Belgrade, from 2004 also a Professor of Computer
Science and Dean of the College of Computer Science, Megatrend University
Belgrade. He was teaching more than 20 graduate and undergraduate courses, from
VLSI Design and Computer Architecture to Computer Networks, Operating
Systems, Image Processing, Calculus and Queuing Theory. His research interest
includes nature-inspired optimizations applied to computer networks, image pro-
cessing and combinatorial problems. Prof. Tuba is the author or coauthor of more
than 200 scientific papers and co editor or member of the editorial board or sci-
entific committee of a number of scientific journals and conferences. He was invited
and delivered around 60 keynote and plenary lectures at international conferences.
Member of the ACM, IEEE, AMS, SIAM, IFNA.
Shyam Akashe is a Professor at ITM University, Gwalior, Madhya Pradesh, India.

He completed his Ph.D. at Thapar University, Punjab, and his M.Tech. in
Electronics and Communication Engineering at the Institute of Technology &
Management, Gwalior. He has authored 190 publications, including more than 50
papers in SCI-indexed journals. His main research focus is low-power system on
chip (SoC) applications in which static random access memories (SRAMs) are
xv
xvi About the Editors
omnipresent. He has authored two books entitled Moore’s Law Alive:

Gate-All-Around (GAA) Next Generation Transistor published by LAMBERT
Academic Publishing, Germany, and Low Power High Speed CMOS Multiplexer
Design published by Nova Science Publishers, Inc., New York, USA. He has also
published over 120 papers in leading national and international refereed journals of
repute. Dr. Akashe has participated in numerous national and international
conferences and presented over 100 papers.
Amit Joshi is currently the Director of Global Knowledge Research Foundation,

also an Entrepreneur & Researcher who has completed his Masters and research in
the areas of cloud computing and cryptography in medical imaging. Dr. Joshi has
an experience of around 10 years in academic and industry in prestigious organi-
zations. Dr. Joshi is an active member of ACM, IEEE, CSI, AMIE, IACSIT,
Singapore, IDES, ACEEE, NPA, and many other professional societies. Currently,
Dr. Joshi is the International Chair of InterYIT at International Federation of
Information Processing (IFIP, Austria). He has presented and published more than
50 papers in national and international journals/conferences of IEEE and ACM. Dr.
Joshi has also edited more than 40 books which are published by Springer, ACM,
and other reputed publishers. Dr. Joshi has also organized more than 50 national
and international conferences and programs in association with ACM, Springer, and
IEEE to name a few across different countries including India, UK, Europe, USA,
Canada, Thailand, Egypt, and many more.
Challenges in Adoption of Secure
Digitization of Graduate Assessments:
Scope for Blockchain Technologies
S. K. Shankar, Rajeswari Mukesh, and M. K. Badrinarayanan
Abstract Blockchain technology finds variety of applications in everyday life. It

is also capable of disrupting the implications on activities which were founded
on time-stamped record keeping of titles of ownership. In education, blockchain
technology shall disrupt activities such as the qualifications, licensing and accred-
itation, management of student records, IPR management and payments. Hence,
blockchain technology-based products or services are significantly different from
previous Internet-based commercial developments. Also blockchain is catching up
interest in the education sector. However, the stakeholders within education are
largely unaware of the social advantages and potential of blockchain technology.
Within the educational sphere, the adoption of blockchain technology in the replace-
ment of paper-based assessments has both advantages and challenges. This paper
attempts to explore the challenges faced by faculty and graduate students in the
adoption of digitized assessments and technical issues in achieving secured system.
By identifying the challenges in implementation, it will be easy to achieve successful
adoption of digitized assessments in a large scale.
Keywords Blockchain · Assessments · Accreditation
S. K. Shankar (B) · R. Mukesh · M. K. Badrinarayanan

Hindustan Institute of Technology & Science, Padur, Chennai, India
e-mail: skshankar@hindustanuniv.ac.in
R. Mukesh
e-mail: rajeswarim@hindustanuniv.ac.in
M. K. Badrinarayanan
e-mail: mkbadri@hindustanuniv.ac.in
© The Editor(s) (if applicable) and The Author(s), under exclusive license 1
to Springer Nature Singapore Pte Ltd. 2021
M. Tuba et al. (eds.), ICT Systems and Sustainability, Advances in Intelligent Systems
and Computing 1270, https://doi.org/10.1007/978-981-15-8289-9_1
2 S. K. Shankar et al.
1 Introduction
1.1 Blockchain Technology
The blockchain technology has provided fresh opportunities to deliver new business
models in the consolidated markets. The implementation of blockchain technology
in the education sector is in its nascent phase, and more tangible results can be
achieved in the future [1–5]. With the widespread usage of cryptocurrencies such as
Bitcoins, Ethereum or Swarm which offer a stable public blockchain that can be used
for secondary purposes such as document verification, it can be used for verification
of university degrees and can provide a timely and solid solution. The blockchain
technology has proven itself to be easy, trustworthy and cost-effective; hence, it can
be adopted with ease [1].
Blockchain [1] is a new technology which is based on new forms of distributed
software architectures, which withers the dependence on centralized integration
points, decentralizes transactional data sharing across large networks of untrusted
participants where components find agreements on shared states.
Blockcerts, an open-source platform, are mainly used for issuing and verifying
the certificates which have been issued digitally using blockchain.
In this paper, we analyse the challenges faced by faculty and graduate students in
adoption of digitized assessments and technical issues in achieving a secured system.
By identifying the challenges in implementation, it will be easy to achieve successful
adoption of digitized assessments in a large scale.
1.2 Digitized Assessments
With the changing industry requirements, evolving learner aspirations and increasing
expectations of the society, it becomes essential to infuse new models of curriculum
with components of technology adoption by the institutions of higher education. The
changes in existing higher education scenario are essential in order to improve the
quality of education, make it suitable to the young learners and increase its ability to
trigger innovation, build research interests and contribute actively to the development
of industry and the society. Innovations in higher education should be reflected in
terms of freedom to develop new curriculum, designing industry-relevant courses,
introduction of new assessment methods and greater flexibility for the students to
have plethora of course choices matching their interests, needs and long-term career
goals [1, 2].
Digitization of assessments shall be an effective starting point for bringing about
changes in higher education. However, the most challenging aspects in adoption of
digitization of assessments are access, quality, values, relevance and ethics. This
study focuses on challenges in digitization of assessments from the perspective of
faculty members who are the implementers of this change in the front end.
Challenges in Adoption of Secure Digitization of Graduate … 3
2 Study Methodology
The study is descriptive. It was conducted with selected faculty members who have
developed and implemented digitized assessments through platforms like Google
Classrooms, Moodle, etc.
A list of benefits and challenges in digitization of assessments were collected
from literature, which were reorganized as suitable to the present study under the
guidance of expert faculty members in the domain of blockchain technologies.
The sampling method was purposive sampling, where that faculties who were
enthusiastic about digitization of assessment and have implemented the same for
the courses they handle were specifically selected for the study. The data from the
filled in questionnaire were processed using Google analytical tools to get descriptive
statistics, and the results were interpreted.
3 Analysis
In the new era of higher education, the focus has shifted from teacher-centric to
learner-centric education. The emphasis is on learning which has placed the learner
as the focal point of all academic transactions. Teachers being the important stake-
holder, it is essential to understand their perspective and opinion about the digi-
tized assessment systems. The present study draws upon the perspectives of teachers
problems faced and recommended suggestions for better implementation of secure
digitization of graduate assessments.
The data for the study were collected by administering a four-part questionnaire
with a five-point grading scale.
The digitization of assessments comes with both benefits and challenges. The
perception of faculty members employing digitized assessment procedures is impor-
tant for repeated use and successful implementation of digitized assessments. The
perceived benefits and challenges of digitized assessments are given below.
3.1 Perceived Benefits
They are low cost, high security, enhancing student’s assessments, better control
of data access, enhancing accountability, identity authentication, improving the
management of student’s records and enhancing learner’s interactivity.
3.2 Perceived Challenges
They are preparatory issues, scalability, immutability, cost, question settings, fitness
to all domains (creative subjects), typing ability of students, privacy and security,
trust building and familiarity with the interface.
From the analysis presented above, it could be clearly observed that a majority of
the faculty members feel that digitized assessments are useful. But we cannot ignore
the fact that a sizeable number of faculty members, a little less than half of them, are
still not convinced that digitized assessments are useful.
While taking a closer look at the perceived benefits of digitized assessments
Fig. 1), better control of data access, enhanced students assessments and improved
management of student records emerge to be the top three in the perceived bene-
fits. There are concerns about the learner’s interactivity, security and cost dimen-
sions, which can be attributed to the limited platforms which are being used by the
respondents.
Among the perceived challenges, Fig. 2, the top three issues mentioned by the
respondents are typing ability of the students, structural limitations in question
Fig. 1 Perceived benefits of digitized assessments
Fig. 2 Perceived challenges in digitized assessments

Fig. 3 Overall usefulness of digitized assessments
setting and familiarity with the interface. While typing ability and familiarity can
be built through training of students over a period of time, the platform support
needs to be worked out for supporting diverse formats of questions sometimes
formula-based/pictorial to overcome the challenge (Fig. 3).
4 Multiple Regression Analysis
Multiple regression analysis is a powerful technique used to predict the value of

a variable based on the values of one or more variables of others. To understand
the influence of the perceived challenges on the perception of overall usefulness of
digitized assessments, a multiple regression analysis was performed, and the results
are presented (Tables 1 and 2).
It can be observed from the results, Table 3, that only two variables, viz. cost
and scalability have significant p-values and can be inferred to be having a stronger
impact on the perception of overall usefulness of the digitized assessments (Tables 4
and 5).
Among the perceived benefits, Table 6, enhancing learners’ interactivity,
enhancing accountability and transparency and better control of data access emerge
Table 1 Regression
Regression statistics
statistics—challenges
Multiple R 0.900101097
R square 0.810181985
Adjusted R square 0.710277767
Standard error 0.631098104
Observations 30
Table 2 Multiple regression analysis—significance statistics

df SS MS F Significance F
Regression 10.00 32.30 3.23 8.11 0.00
Residual 19.00 7.57 0.40
Total 29.00 39.87
Table 3 Multiple regression analysis—challenges

Coefficients Standard t stat P-value Lower Upper Lower Upper
error 95% 95% 95.0% 95.0%
Intercept 0.21 0.58 0.37 0.72 −1.01 1.44 −1.01 1.44
Familiarity 0.24 0.19 1.23 0.23 −0.16 0.64 −0.16 0.64
with the
interface
Trust 0.07 0.17 0.43 0.67 −0.28 0.43 −0.28 0.43
building
Privacy and −0.10 0.16 −0.58 0.57 −0.44 0.25 −0.44 0.25
security
Cost 0.35 0.14 2.53 0.02 0.06 0.64 0.06 0.64
Immutability 0.10 0.18 0.54 0.60 −0.29 0.48 −0.29 0.48
Scalability 0.24 0.14 1.75 0.10 −0.05 0.54 −0.05 0.54
Preparatory −0.17 0.19 0.89 0.39 −0.56 0.23 −0.56 0.23
issues
Structural −0.10 0.20 0.52 0.61 −0.52 0.32 −0.52 0.32
limitations in
question
settings
Fitness to all 0.17 0.13 1.31 0.21 −0.10 0.43 −0.10 0.43
domains
(creative
subjects)
Typing ability 0.18 0.20 0.91 0.38 −0.24 0.60 −0.24 0.60
of students
Table 4 Regression
Regression statistics
statistics—benefits
Multiple R 0.90
R square 0.80
Adjusted R square 0.73
Standard error 0.61
Observations 30.00
Table 5 Multiple regression analysis—significance benefits statistics

df SS MS F Significance F
Regression 8 31.94 3.99 10.57 0
Residual 21 7.93 0.38
Total 29 39.87
Table 6 Multiple regression analysis—benefits

Coefficients Standard t stat P-value Lower Upper Lower Upper
error 95% 95% 95.0% 95.0%
Intercept 1.14 0.63 1.81 0.08 −0.17 2.46 −0.17 2.46
Enhancing 0.63 0.16 3.91 0.00 0.29 0.96 0.29 0.96
learners
interactivity
Improving the −0.32 0.23 −1.36 0.19 −0.80 0.17 −0.80 0.17
management
of students
records
Identity −0.09 0.18 −0.47 0.64 −0.46 0.29 −0.46 0.29
authentication
Enhancing 0.48 0.17 2.74 0.01 0.11 0.84 0.11 0.84
accountability
and
transparency
Better control −0.42 0.23 −1.86 0.08 −0.89 0.05 −0.89 0.05
of data access
Enhancing 0.35 0.22 1.61 0.12 −0.10 0.79 −0.10 0.79
students
assessments
High security −0.05 0.21 −0.25 0.81 −0.49 0.39 −0.49 0.39
Low cost 0.15 0.16 0.98 0.34 −0.17 0.48 −0.17 0.48
as the significant variables in influencing the overall perceived usefulness of the

digitized assessments.
It is important to indicate at this point that it is essential to integrate the digital
assessment into the entire learning process from the course design. It is also for
essential assess what they learn and not just what they have learnt. Considering
the fact that learning is an ongoing process that assessment should also be done
continuously, and instructional design can be helpful for planning and implementing
the same, and we have to select the right methodologies (i.e. assessment), strategies
and learning technology to enhance it.
One more strategy that should be considered is to use different assessment methods
for measuring different processes and outcomes. Apart from the popular ones like
multiple choice questions, online role-plays, scenario-based activities, judged mathe-

matical expression and online discussions, one can also utilize techniques like ePort-
folios assessment, concept maps or personal response systems. Usage of ePortfolios,
scenario-based activities, role-paying and simulations can make the process authentic
since they provide evidence of learning. These also help in assessing the skills like
critical thinking, decision-making, team work, etc.
For several generations, the idea of feedback has been associated with the last
mile wrap up activity. It is essential to acknowledge that for effective learning to take
place, learners need substantial, regular and meaningful ongoing feedback. Digitized
assessments offer an appropriate format for providing such regular feedback. The
suggestions can be personalized and be given by different stakeholders including self,
peers, teacher and professionals. Such a feedback shall be used to improve learning
in the process and not kept aside as a final activity. Providing continuous feedback
will encourage learners to make course corrections and enhance their learning expe-
rience. In order to scale up digitized assessments, self-regulated learning should be
promoted.
5 Conclusion
Digitization of assessments is the order of the day, and the higher education system
has to adopt the digital way at a rapid pace. The digital learning platforms are
attracting users across the globe and redefining the assessment and learning experi-
ence altogether. Though the penetration into the market like India is currently low,
it is not a far future that these complimentors shall replace the brick and mortar
higher education system. In this scenario, to implement digitized assessments in
a mass scale, it is essential to understand the perspectives of faculty who are the
opinion leaders and implementers at the ground level. This study is an attempt in that
direction and has found that there are few operational challenges, but the perceived
benefits outweigh the challenges. With proper training to both faculty and students,
building secure platforms with blockchain technology will help to overcome the
challenges identified, and the digitization of assessments will be soon a reality in
higher education sector.
References
1. S. Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System (2008). Available online https://
bitcoin.org/bitcoin.pdf. Accessed on 12 June 2019
2. A. Grech, A.F. Camilleri, Blockchain in Education (Publications Once of the European Union,
Luxembourg, 2017)
3. R. Collins, Blockchain: a new architecture for digital content. E-Content 39, 22–23 (2016)
4. M. Memon, U.A. Bajwa, A. Ikhlas, Y. Memon, S. Memon, M. Malani, Blockchain beyond

Bitcoin: block maturity level consensus protocol, in Proceedings of the 2018 IEEE 5th Inter-
national Conference on Engineering Technologies and Applied Sciences (ICETAS), Bangkok,
Thailand, 22–23 November 2018, pp. 1–5
5. V. Gatteschi, F. Lamberti, C. Demartini, C. Pranteda, V. Santamaría, Blockchain and smart
contracts for insurance: is the technology mature enough? Future Internet 10, 20
Non-performing Asset Analysis Using
Machine Learning
Rohan Ninan Jacob
Abstract Loans are given by the banks to the customer, according to their needs,
and are being monitored by various entities within the organization, so that adequate
amount is received to the customers and assuring that they would not default or certain
financial advisories can help them in letting know whether a non-performing assets
(NPA) case can be triggered or not. Though there are many softwares in market which
can do this job quite well, but for small financial institutions and banks, it becomes a
big hurdle to deploy such solution. Unfortunately, there has not been a cost-effective
method of identifying and predicting these cases and has also led to some failure in
monitoring of NPA activities, in organizations. Therefore, an effective approach is to
use a machine learning algorithm, with the use of open-source softwares, that could
help in predicting the NPAs by using the essential parameters and thereby giving
a cost-effective solution to such an issue, which many of the banking and financial
organizations are facing.
Keywords NPA · Parameters · XGBoost · Gradient boosting · Feature importance
1 Introduction
The international monetary fund devises the NPA standards for the financial institu-
tions, based on which the loan is standardized to be NPA [1] if the criteria are met.
Of which some of them are as follows:
• Payments that are related to interest as well as principal that can be up to 90 days
or more
• A minimum 90 days of payments comprising of interest which is capitalized,
delayed or refinanced because of agreement clause by the organization
R. N. Jacob (B)
Department of Electronics and Telecommunications Engineering, Fr. Conceicao Rodrigues
Institute of Technology, Vashi, Navi Mumbai, India
e-mail: rohanninanjacob@gmail.com
12 R. N. Jacob
• Payments which are less than overdue for 90 days, but there are other reasons of
doubt whether those payments would be made in full or not.
Conventional collection and recovery process of the financial institution have
always been manual-driven and not data-driven. These are involved by a large number
of people who adopt various techniques (mostly manual) for the collection process
and hence lack professional approach. This generic process of collection and recovery
which is lacking focused approach has been resulting in accumulated non-performing
assets, the ever-increasing cost of collection, and dissatisfied customers.
With the advent of credit risk management solutions or other NPA detection soft-
wares, it is possible to track fraud cases resulting in NPA and makes tracking of these
quite easy. However, such software costs a lot and hence an efficient cost-effective
method could help the organization to monitor and track such activities, thereby
relying on the power of open-source softwares. Due to increase in the development
and a great community support of open source, it was possible to implement an
algorithm that helped in resolving such cases.
The paper presents an analysis which can be done by various financial institutions
and organizations, in making sure that NPAs can be in control implying to which
will give them better decisions to monitor such incidents and help them in saving a
ton of money.
The main algorithm that is used here is the XGBoost algorithm that would help
in the execution speed and the model performance. The essential part of the algo-
rithm is the scalability that helps in the fast learning by parallel and distributed
[2] computing, offering adequate memory usage. It is one of the gradient boosting
methods which falls in a supervised learning technique and aims in predicting target
variable. XGBoost is an ensemble-based learning which is composed of several base
learners.
2 Algorithm and Program Flow
The gradient boosting, when used in regression, the regression tree, helps in mapping
the input data point to one of the leaves which comprises a continuous score. Here,
the XGBoost algorithm helps in minimizing a regularized objective function which
results in the combining of a convex loss function (based on the difference between
the predicted and target outputs) as well as a penalty term for model complexity (in
other words, the regression tree functions). The training proceeds iteratively, adding
new trees and predicting the residuals or errors of prior trees that are then combined
with previous trees to make the final prediction. The fact that requires adding new
tress is to aggregate the number of trees or predictors which are individually weak
classifiers, so aggregation is required in order that prediction can improve. It is called
gradient boosting because it uses a gradient descent algorithm to minimize the loss
when adding new models.
Non-Performing Asset Analysis Using Machine Learning 13
While gradient boosting has an approach of negative gradients, in order to optimize

loss function, this makes use of Taylor series which is to calculate value of this loss
function by means of distinguished base learners [3].
XGBoost need not explore all possible tree structures but helps in the building up
of a tree greedily, whereby the regularization term is helping to reduce the building
up of complex tree by means of several leaf nodes [4].
Apart from that, XGBoost consists of other techniques which include column
subsampling, shrinkage, splitting criteria, etc.
Details of the work flow are as follows:
• Input data was provided by Hacker Earth, which was later on then converted to
test and train data to check and validate the accuracy as well as the precision and
the parameters which had to be monitored along with its description was also
provided by the same.
• With the large data which was received, there were many changes which needed
to be done in data, like in fields: ‘income’ and ‘undreamt’ were missing and hence
needed to be filled in with the mean values. Such values had to be filled by taking
upon mean values and then filling up of the result at those blank values or ‘Nan’,
which is the process of imputing missing values. Some parameters need not need
to be a part of this analysis and hence were dropped out of the consideration.
Apart from that, for better understanding of the terms, renaming was done.
• Splitting of both training and test set was performed in which 70% of data was
used in training and 30% used in the test set. It is important to make sure that
the splitting ratio should never be the same, which will alter the efficiency and
reliability of the model. It even helps in gaining deeper insights into reliable
decisions.
• For increasing the efficiency of the model, we performed cross-validation. It is
used to help in enhancing the predictive as well as accuracy performance of the
developed models, helping in seeing how they work outside the sample [5] to a
new data set, i.e., test data. The reasons come in for a cross-validation technique
are when we fit a model, we are fitting it to a training data set. Without cross-
validation, we only have information which gives us insights into how will the
model perform with respect to the sample data. Ideally, we would like to see
how the model performs when we have a new data in terms of accuracy of its
predictions [6]. In science, theories are judged by its predictive performance.
There two types of cross-validation you can perform: leave one out and k fold.
• After training the model, the best iteration is selected, due to which, we came to
know the accuracy of the model, giving us the prediction. Also, we came to know
the ROC-AUC.
Accuracy is one of the essential factors for evaluating classification models. It is
nothing but the fraction of predictions which represents the correctness of the model,
which is given by [7] (Fig. 1).
14 R. N. Jacob
Fig. 1 Accuracy equation
Fig. 2 AUC equation
The reason why predictions are important in machine learning is because machine
learning model predictions help the businesses to make highly accurate guesses in
order that they are likely outcomes of a question based on historical data [8].
The area under a curve (AUC) for a classifier on a given data set can be represented
as the probability of P(f (X+) > f (X−)), where f (X+)corresponds to a random variable
which pertains to distribution of the outputs (scores) for the classifier, belonging to the
positive observations and f (X−) the one corresponding to the negative observations.
Mathematically, represented ⎧ by [9] (Fig. 2).
⎨ 0 if x < 0,
where g() is defined as 0.5 if x = 0,
⎩
1 if x > 0.
The ROC-AUC curve gives us an indication on how well the model created is
performing with the data. This gives also an insight whether the model which is
deployed has underfit or overfit the training and test set.
A graphical workflow of program is illustrated as follows (Fig. 3).
3 Results
The test set results were compared with respect to the training set, for making sure
that the model worked well and has a good efficiency in its performance. We even
determined the accuracy in proving it so. The feature importance was evaluated,
which gives us an insight into every feature in the data and more its score, the more
it signifies to decision making. Some of the important features during the analysis of
this problem was total interest fees, home ownership, employee length, loan amount,
batch enrolled, etc.
The training process was initiated, where the best iterations were determined.
Below figure represents the start of the iteration (Fig. 4).
While the training set of iterations were ongoing and later on completed, the best
iteration was selected, thus evaluated the model accuracy and ROC-AUC (Fig. 5).
The feature importance of all features was evaluated, which are represented in the
plot as shown below, which includes all the features used during our analysis. The
y-axis represents the feature importance values, and the x-axis shows the features
used in the model building (Fig. 6).
Fig. 3 Program flow

Input Data
Data Cleaning
Imputing
Missing
Values
Splitting and
Training
Cross
Validation
Prediction
4 Summary
With the use of XGBoost algorithm, we were able to train the existing data and get
insights into the parameters that help us in the governing as well as determination
of NPA. These parameters are taken into consideration that the bank is responsible
in the legit keeping of this data. With respect to this bank, they were not performing
well for the last 3 quarters and usually depend on their group of investors to see
whether the loan can be given to particular customers, based on their entire record
and wants to make sure that whether they will default or not. Feature importance is
one such important criteria enabling us to see the necessary parameters enabling the
16 R. N. Jacob
Fig. 4 Training using XGBoost classifier
Fig. 5 Results obtained after selecting the best iteration
Fig. 6 Feature importance of variables, found after training and the best iteration selected during
the analysis
organization to monitor, and each of their customers and along with XGBoost gives
them an indication of the NPA case going to get triggered. Taking the highest scores,
which were obtained during feature importance, we were able to come to derive our
results.
Also, the model that was developed had an accuracy of 94.11%, and ROC-AUC
was 92.48%.
5 Future Scope
• With organization seeking help in need of open-source solutions, the program

when integrated to live databases would be able to track such cases which will
help real-time risk management.
• A business intelligence tool, with support of Python scripts, can help in creation
of dashboards that will help the particular user/organization to keep a record and
monitoring of NPA’s with ease.
6 Conclusion
This paper has taken into account to focus upon the NPA and other similar solu-
tions that exist in the present world, where cost has always been a major factor, for
financial organizations, especially small and medium-sized enterprises (SMEs) to
monitor their NPAs, but along with that the prediction and accuracy were always
a doubtful case when thinking of open-source platforms. This era has a vital role
in the development of variety of many developed or developing Python libraries,
enabling the use of modern and optimized solutions to exist in open sources and
thereby with use of appropriate machine/deep learning algorithms would help in
obtaining more accuracy as well as gaining more insightful analysis, which was not
the case 10–20 years back due to reasons like lack of data, software availability and
other technical constraints, etc. With the advent of such high powered and optimize
algorithm has made it possible to strengthen decision making as well as opening
variety of capabilities in different domain.
References
1. Non-performing loan, https://en.wikipedia.org/wiki/Non-performing_loan. Last accessed on 4

May 2020
2. An End-to-End Guide to understand the Math behind XGBoost, https://www.analyticsvidhya.
com/blog/2018/09/an-end-to-end-guide-to-understand-the-math-behind-xgboost/understand-
the-math-behind-xgboost/. Last accessed on 1 Apr 2020
18 R. N. Jacob
3. XGBoost Algorithm: Long May She Reign, https://www.kdnuggets.com/2019/05/xgboost-alg

orithm.html,last. Accessed on 21 Mar 2020
4. Unveiling Mathematics Behind XGBoost, https://www.kdnuggets.com/2018/08/unveiling-mat
hematics-behind-xgboost.html. Last accessed on 8 Apr 2020
5. What is the purpose of performing cross-validation? https://www.researchgate.net/post/
What_is_the_purpose_of_performing_cross-validation. Last accessed on 21 Mar 2020
6. Purpose of Cross validation, https://towardsdatascience.com/5-reasons-why-you-should-use-
cross-validation-in-your-data-science-project-8163311a1e79. Last accessed on 4 May 2020
7. Classification:Accuracy, https://developers.google.com/machine-learning/crash-course/classi
fication/accuracy. Last accessed on 14 Jan 2020
8. This is how computers predict the future, https://qz.com/1261817/predictive-algorithms-are-
not-all-that-complicated/. Last accessed on 20 Mar 2020
9. Cristiano Leite Castro, Antonio Padua Braga: Optimization of the Area Under the ROC Curve,
in 10th Brazilian Symposium on Neural Networks. IEEE computer society, Washington DC
(2008), pp. 142
Volatile Memory Disk Forensics:
Investigate the Criminal Activity
of RAMDisk
Nilay Mistry, Annan Christian, and Bansari Bhavsar
Abstract In this ever-expanding cyber space, computer is a very valuable asset, but
if used with malicious intentions it can create an environment of chaos. Computer
forensics which is a part of digital forensics aims at collecting, analyzing, and docu-
menting digital evidences such that they are admissible in court of law. But integrity
of digital evidences comes under scrutiny due to anti-forensics tools and techniques
that are exercised by cybercriminals to evade detection. RAMDisk is one of such
technique used by cybercriminals to avoid detection and perform clean up after the
crime that has been committed. Cybercriminals use this technique to store sensitive
information and data in RAMDisk and discard it once the purpose has been served.
This paper discusses the working of three popular RAMDisk softwares namely AMD
Radeon RAMDisk, ImDisk Toolkit. and Miray RAM Drive. We have also identified
proprietary drivers and unique signatures for each software and what approach can
be adopted by a forensics investigator to investigate them.
Keywords RAMDisk · Random access memory · RAM · Anti-forensics ·

Counter-forensics · Cybercrime · Volatile memory
1 Introduction
Locard’s principle for digital forensics can be defined as, any perpetrator of crime
leaves a trace, and it can be either physical or virtual. In digital crimes, there is no
physical contact between the perpetrator and the crime scene, but the perpetrator
N. Mistry (B) · A. Christian · B. Bhavsar

Institute of Forensic Science, Gujarat Forensic Sciences University,
Sector 9, Gandhinagar, Gujarat 382007, India
e-mail: nilaymistry30@gmail.com
URL: https://www.gfsu.edu.in/
A. Christian
e-mail: archieannan@gmail.com
B. Bhavsar
e-mail: bansari0712@gmail.com
20 N. Mistry et al.
would leave a digital trace behind. As the chance of leaving a trace increases, there
is always a concern about the integrity of the evidences because of the anti-forensics
exercised by perpetrator to manipulate or delete the evidence which would misguide
the forensic analyst [1]. Anti-forensics is a set of tools, techniques, methods, and
procedures that hinder the scientific analysis of evidences by affecting its quality,
quantity, and integrity.
Marcus Roger categorized anti-forensics into four categories: data hiding, artifact
wiping, trail obfuscation, and attack against computer forensic tools [3]. In data
hiding category, data is hidden into unallocated or slack space of a hard disk. Meta-
data of files, Master Boot Record (MBR), and hidden partitions on hard drive can also
be used to hide secret data. Another category is artifact wiping, and in this category
different tools are used to destroy the data by overwriting it multiple times, making
the data irretrievable. Tool like Eraser is very popular, but it also leaves indicators
of anti-forensics in the form of registry entries [2]. The worst are prophylactic tools,
they not only delete the artifacts but also wipe out any indicators of anti-forensics
or its remnants. The method of trail obfuscation has been prevalent since the era
of 70s [3]. It tries to misguide or divert the forensic investigator in a completely
different direction. Lastly, attack against computer forensics tools is a relatively new
technique as compared to the other three. In this categories, the flaws and zero-days
in the forensics tools are exploited to manipulate and forge fake digital evidences,
thus questioning the integrity and reliability of the evidences collected using those
tools.
RAMDisk (also called as RAM drive) is a block of random-access memory (pri-
mary storage or volatile memory) that a computer’s software treats as if the memory
were a disk drive (secondary storage). RAMDisk is a program that takes a portion
of the system memory and uses it as a disk drive. RAMDisk is used for both benign
and malicious purposes. RAMDisk is extensively used among gamers because of
its fast input/output (I/O) throughput, but it is equally popular among cybercrim-
inals to avoid data recovery. RAMDisk cannot be categorized into any one of the
anti-forensics categories defined by Rogers. It is a combo of both data hiding and
artifacts wiping category. RAMDisk can be used to hide data in partition created
from RAM, and once the motive is accomplished, it can be again merged into the
system memory. RAM being a highly volatile memory, the data stored in the RAM
either gets lost or overwritten by other data, and hence, artifact wiping is achieved.
In this paper, the research work is concentrated on anti-forensic that is achieved with
the use of RAMDisk and also identified different freeware software available in the
market that are used to create RAMDisk and the drivers that they use. Also identified
the signatures that they leave behind on their usage. In addition to that, how an inves-
tigator can identify whether RAMDisk was employed as an anti-forensic technique
and what approach an investigator can take to perform digital forensic if such is the
case.
Volatile Memory Disk Forensics: Investigate the Criminal Activity of RAMDisk 21
2 Different Software Available for RAMDisk
As discussed, a plethora of softwares are available for creating a RAMDisk. However,

here, we will discuss working of three popular software available for creating a
RAMDisk.
2.1 AMD Radeon RAMDisk
AMD Radeon RAMDisk provides a freeware version of the software which can be
used to create RAMDisk. An important feature of AMD Radeon RAMDisk from
security perspective is that RAMDisk has been designed to permanently delete any
information that has not been backed up or auto-saved making the information com-
pletely unrecoverable.
When the RAMDisk is successfully started, a kernel level driver named RAM-
DiskVE.sys is loaded into Windows\System32\drivers folder as shown in Fig. 1. This
driver is available to Windows whenever RAMDisk is started. Once the RAMDisk
has been stopped, the driver is removed from the folder.
2.2 ImDisk
ImDisk is a free and open-source software. Few key features of ImDisk is that it
gives an option to allocate memory dynamically which makes the data carving more
difficult.
ImDisk utilizes service named ImDisk Virtual Disk Driver Helper which starts
in the background when the computer boots, so user can create a RAM based file
Fig. 1 AMD Radeon RAMDisk driver

22 N. Mistry et al.
Fig. 2 ImDisk driver
Fig. 3 Miray RAM Drive driver
system which is automatically erased as soon as the computer is powered off. ImDisk
driver named imdisk.sys can be found in the Windows\System32\drivers folder as
shown in Fig. 2.
2.3 Miray RAM Drive
Miray RAM Drive is another tool for creating RAMDisk. Key feature of Miray RAM
Drive include that it provides persistence storage which facilitates saving the image
of RAMDisk on a hard disk.
Miray Ram Drive uses service named MRDService. The service then runs Win-
dows Kernel mode device driver mrdo.sys which initializes the RAM drive. The
driver is located at C:\Program Files (×86)\Miray RAM Drive\driver\×64 (Fig. 3).
3 Signature Creation and Identification
When a software is installed, it leaves behind some kind of traces. These traces can be
either in form of unique files or folder that are created, starting up a specific service
or modification made to the Windows registry. These traces can be used to create
signatures that can uniquely identify a software. In this work, we have used registry
snapshot analysis, and updates made to the Windows registry are used to create
signatures. Signatures were identified by comparing the snapshots of registry taken
before installing and after uninstalling the software using utility named Regshot.
3.1 Signature Creation
For identification of signature, following steps where performed chronologically

(Table 1).
1. Create test system

• Windows 10 machine
• i7 64 bit processor with 8 GB RAM
2. Take and save a snapshot of the registry (using Regshot)
3. Execute one of the following desired activity:
• Install RAMDisk creation software
• Uninstall RAMDisk creation software
4. Take and save another snapshot after the desired activity has been performed
(using Regshot)
5. Compare the snapshots
6. Identify and extract unique signatures.
3.2 Identification of Signature
Regshot allows comparison of registry snapshots giving a detailed analysis of any

keys or subkeys created, deleted, or modified in the Windows registry. Unique sig-
natures were identified by manually filtering the report generated by Regshot. The
unique signature identified for installation and un-installation of each vendors are
listed in the tables.
Table 1 Driver location

Vendor name Driver name Driver location
AMD Radeon RAMDisk RAMDiskVE.sys Windows\system32\drivers
ImDisk imdisk.sys Windows\system32\drivers
Miray RAM Drive mrdo.sys C:\Program files (×86)\Miray
RAM Drive\driver\×64
24 N. Mistry et al.
Table 2 AMD Radeon RAMDisk signatures

Action Signature
Install NA
Uninstall HKLM\SYSTEM\CurrentControlSet\Services\Ramdisk\
Table 3 ImDisk signatures

Action Signature
Install HKLM\SOFTWARE\Classes\*\shell\ImDiskMountFile
Uninstall HKLM\SYSTEM\ControlSet001\Services\bam\UserSettings\S-1-5-21-
209205877-1768948508-1567415345-1001\\Device\HarddiskVolume3\Program
Files\ImDisk\
Table 4 Miray RAM Drive signatures

Action Signature
Install HKLM\SOFTWARE\Miray\Miray RAM Drive
1.1.0-1033-604626AB-787D-49C4-9C55-280C5AFA0ED0
Uninstall HKLM\SYSTEM\CurrentControlSet\Services\MirayRAMDrive
Along with installation signatures, registry key denoting the date of installation
for few vendors are identified. These dates can be useful for creating the time-line
based analysis of the crime. The location of registry key for AMD Radeon RAMDisk
is HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\Uninstall\BBC956B0-
DD9-4A48-ACAC-DC6AC0FE10D5\InstallDate (Table 2).
The value of key denotes the date when AMD Radeon RAMDisk software
was installed. Similar registry key has been identified for Miray RAM Drive, and
the key location is HKLM\SYSTEM\ControlSet001\Control\Class\4d36e97b-e325-
11ce-bfc1-08002be10318\0001\DriverDate with value suggesting the date when the
driver was installed (Tables 3 and 4).
4 RAMDisk Data Carving and Artifacts Analysis

Methodology
4.1 Test RAMDisk Creation Methodology
To analyze RAMDisk for data carving and related artifacts, a test RAMdisk was
created of 200 MB in size using AMD Radeon RAMDisk software in Windows
10 machine. Few files were saved in the RAMDisk later to be recovered if they
were deleted or the RAMDisk was unmounted. In total seven files were saved in the
RAMDisk, out of which five files were pictures, and two files were text documents.
4.2 Volatile Memory Capturing Methodology
RAM being a volatile memory, an extra care must be taken while acquiring the
memory dump. If right tools and techniques are not used, then important artifacts
might get contaminated or sometimes even get lost. In this work, we have used
FTK Imager which is a free tool and widely prevalent in the industry for capturing
volatile memory. FTK Imager creates a bit-by-bit image of the memory including
the unallocated and slack space. We have captured the RAM for both the scenarios,
i.e., while the RAMDisk is still mounted and while it is unmounted.
4.3 Artifact that a Device Was Mounted
One of the key artifact for analyzing a RAMDisk is to find unusual entries in
mounted devices registry key. The list of mounted devices can be find out by analyz-
ing HKLM\SYSTEM\MountedDevices registry. As shown in Fig. 4, an unknown
mounted device which is a DOSDevice type has been highlighted. Such unknown and
unusual entry depicts that some sort of storage device was mounted as disk. Hence,
it could be concluded that a RAMDisk was used as an anti-forensics technique, and
further analysis of RAM dump is needed to find more potential artifacts.
4.4 Analyzing Memory Dump for Artifacts
The memory dumps obtained from FTK Imager are in raw form. To analyze this
dumps, we have to use a hex editor. Here, we have use wxHexEditor to analyze the
RAM dumps. wxHexEditor provides a user-friendly interface which is divided into
two parts. In the left part, the bit-by-bit representation of the RAM dump is shown
in a hexadecimal form, and the right side denotes the values in string of hexadecimal
values.
26 N. Mistry et al.
Fig. 4 Unusual mounted device
4.5 Analyzing Image of RAM While the RAMDisk Is Still

Mounted
4.5.1 Recovering Data from the RAMDisk
Analyzing the RAM dump taken while the RAMDisk was still mounted, it was found
that all the data can be recovered from the RAM dump. Even if files were deleted on
RAMDisk, they can be recovered from the dump. All the files saved on the RAMDisk
can be carved from the RAM image using any data carving tool. All the files which
were not deleted can be found in their original form. One of such JPEG file is shown
in Fig. 5.
4.5.2 Analyzing Image of RAM While the RAMDisk Is Unmounted
When the RAMDisk is unmounted, the RAMDiskVE.sys driver is removed, and the
RAM blocks allocated to the RAMDisk are again merged into the RAM. Partial files
can be found by analyzing RAM image. When the RAM blocks are dis-allocated and
again merged into the RAM, few blocks get overwritten by other processes, and only
few partial files can be recovered from the image of the RAM. Partial parts of the
Fig. 5 File in RAM image

Fig. 6 Partial text found in the RAM image
Table 5 Default location of .img file

Vendor Default location
AMD Radeon RAMDisk C:\RAMDisk.img
ImDisk User Defined(window opens in Documents)
Miray RAM Drive C:\Users\Miray\pers.img
text file can also be found at random locations of the RAM, hence proving that the
blocks of RAM allocated to the RAMDisk were not consecutive (Fig. 6). Merging
those parts of the file a large proportion of the original text file can be recovered.
All the data in their original form can be recovered from the .img file that the
RAMDisk software creates to restore the RAMDisk if it gets corrupted or it has to be
reinstated in case of power off. This .img file can be mounted as an external media
drive and all the data saved in it can be viewed. Different vendor has a different
default location where this .img file is stored. Default location of each vendor is
given in Table 5.
5 Conclusion
This work has concentrated on RAMDisk used as an anti-forensics technique and

identified signatures for three popular softwares that are prevalent in the market. This
signatures can be used to identify which software was used to create a RAMDisk,
and based on the software identified, the investigator can plan out the approach to
perform forensic investigation on the system. Installation and running of a driver
identified for each vendor confirms that RAMDisk was used, and so forensic analyst
should look for files and forensic artifacts in the RAM image.
From the above analysis, it was concluded that deleted files can be recovered
from the RAM image while the RAMDisk is mounted. But once the RAMDisk is
unmounted, the blocks of RAM get overwritten by other processes, and recovering
the file in its original form gets difficult. So image of RAM should be taken as soon
as possible to avoid overwriting of the blocks allocated to RAMDisk. While in the
28 N. Mistry et al.
scenario when RAMDisk is unmounted long before and data cannot be recovered
from RAM dump, in such situation, forensic analyst should look for .img file stored
at the default location to recover all the data in its original form.
References
1. R. Harris, Arriving at an anti-forensics consensus: Examining how to define and control the
anti-forensics problem, Digit. Invest. (2006). https://www.sciencedirect.com/science/article/
pii/S1742287616300378
2. K.J. Park, J.M. Park, E. Kim, C.G. Cheon, J. James, Anti-forensic trace detection in digital
forensic triage investigations. J. Digit. Forensics Secur. Law (2017). https://commons.erau.
edu/jdfsl/vol12/iss1/8/
3. M. Rogers, Anti-Forensics (Lockheed Martin, San Diego, 2005). http://www.cyberforensics.
purdue.edu/docs/Lockheed.ppt
4. K. Conlon, I. Baggili, F. Breitinger Anti-forensics: furthering digital forensic science through
a new extended, granular taxonomy. Digit. Invest. (2016). https://www.sciencedirect.com/
science/article/pii/S174228761630037
5. J. Zheng, Y.A. Tan, X. Zhang, C. Liang, C. Zhang, An anti-forensics method against memory
acquiring for android devices, in Computational Science and Engineering (CSE) and Embed-
ded and Ubiquitous Computing (EUC) (2017). https://ieeexplore.ieee.org/abstract/document/
8005797/
6. T. Rochmadi, I. Riadi, Y. Prayudi, Live forensics for anti-forensics analysis on private
portable web browser. Int. J. Comput. Appl. (2017). https://pdfs.semanticscholar.org/9974/
3a16bd4c294cc6410fb1c15ba2f3c8229c05.pdf
7. K. Lee, H. Hwang, K. Kim, B. Noh, Robust bootstrapping memory analysis
against anti-forensics. Digit. Invest. (2016). https://www.sciencedirect.com/science/article/pii/
S1742287616300408
8. N. Zhang, R. Zhang, K. Sun, W. Lou, Y.T. Hou, S. Jajodia, Memory forensic challenges under
misused architectural features. IEEE Trans. Inf. Forensics Secur. (2018). https://ieeexplore.
ieee.org/abstract/document/8323390/
9. H.K. Mann, G.S. Chhabra, Volatile memory forensics: a legal perspective. Int. J. Comput. Appl.
(2016). https://pdfs.semanticscholar.org/f140/f3bdb657f4dcfbfd4bf0183524bfa925a872.pdf
10. P. Feng et al., Private data acquisition method based on system-level data migration and volatile
memory forensics for android applications. IEEE Access 7, 16695–16703 (2019). https://
ieeexplore.ieee.org/iel7/6287639/6514899/08625420.pdf
11. A. Chetry, U. Sharma, Memory forensics analysis for investigation of online crime-a review,
in 2019 6th International Conference on Computing for Sustainable Global Development
(INDIACom) (IEEE, 2019)
12. A. Kazim et al., Memory forensics: recovering chat messages and encryption master key,
in 2019 10th International Conference on Information and Communication Systems (ICICS)
(IEEE, 2019)
13. R.A. Awad et al., Volatile Memory Extraction-Based Approach for Level 0–1 CPS Forensics
(Oak Ridge National Lab (ORNL), Oak Ridge, TN (United States), 2019). https://www.osti.
gov/servlets/purl/1606814
Modem Functionality Diagnostic Tool
with Graphical User Interface
for Cellular Network
Arghya Biswas, Sudakar Singh Chauhan, and Antara Borwankar
Abstract In both Internets of things and conventional cellular networks, the modem
is a very important component. So it should be highly stable and trustable. Any kind of
error or misbehavior must be diagnosed and need to resolve properly. In this paper,
we are proposing a very useful tool that can diagnose the workflow of a cellular
modem. The tool can observe the modem very closely and diagnose each step from
the basic power on a feature to the registration on the network with the establishment
of the data connection. If the modem faced any kind of error in between this, then
the tool is capable of detecting the error, and it will solve the problem or recommend
the possible solution to get rid of the error. We are also proposing a remote modem
diagnosis feature with this tool. If the modem is deployed in some other device on the
same network, then our tool is also able to diagnose the modem. With this diagnosis
tool, users can save a lot of time to diagnose and fix the modem error.
Keywords Cellular modem · Diagnose tool · AT command · Internet of things ·

GSM and LTE network · Yocto
1 Introduction
Cellular technology is one of the largest technologies in our daily life. This technology
was introduced only for voice communication. In the latter days, the same technology
was started to use for the data connection. Nowadays, cellular technology is being
used in various fields. One of the most important fields is the Internet of things
A. Biswas (B)
School of VLSI & Embedded System Design, NIT Kurukshetra, Kurukshetra 136119, India
e-mail: arghyabiswas05@gmail.com
S. S. Chauhan
ECE Department, NIT Kurukshetra, Kurukshetra 136119, India
e-mail: sudakarnith@gmail.com
A. Borwankar
Intel Technology India Private Limited, Bangalore 560103, India
e-mail: antara.borwankar@intel.com
30 A. Biswas et al.
(IoT). The most important part of the cellular networks is the modem (modulation–
demodulation) [1]. A proper fault diagnostic tool is required to implement, and bug
solves the modem-related issues. In this paper, we are proposing a functionality
diagnostic tool of a modem for the cellular network with a graphical user interface.
2 Literature Survey
In [2], the authors proposed an on-board diagnostic (OBD) system for euro standard
vehicles’ engine. The OBD takes the data related to the state of the engine, the exact
timing of when a spark happens, etc., and sends the data to a remote server. They
used the GSM technology with the AT command to communicate with the server. A
graphical user interface (GUI) on the server side shows the coming data and suggests
some required action if any error occurs.
The authors proposed in [3] a multi-node mesh network to analyze the data from
DTE. They grab the data from one node and send it to the other nodes, and then the
system dials up to initiate the cellular connection. In other nodes, the modem must
send the ASCII or string response according to the command that fired. They chose
WSN topology for communication and some shell to troubleshoot the issues.
In [4], authors used to send a 2.5 GHz signal after 15 min gap over a corn and
soybean field for 7 months. After analyzing the RSSI, rainfall, temperature, etc., they
concluded that the RSSI is affected by the preset and absent of the crops.
Authors proposed to share the lower layer ID IMSI for several IoT devices in [5]
to optimize the state information. It will be recognized by the upper layer devices.
They also proposed to use time division concept in machine-type communication to
upload the data in different time to enhance the speed of the system.
Authors proposed in [6], machine to machine services to implement smart
resemble metering in the application for the vehicle. They also measure the quality
of the signal overhead and analyze the data. With the proposed system, they mainly
focused on reducing the signaling load on the cellular network sever.
In [7], authors proposed a tool to collect the signal information and control plane
message to analyze the performance degradation due to the misconfiguration of
cellular network. In these analyzing steps like time threshold, control flow sequence
and signaling failure, authors used more than 3.1 million control plane messages
from 13 main cellular operators to analyze the circuit-switched fallback technology
in 3G and LTE system.
In this paper [8], authors summarized the main issue in the 5G technology evolu-
tion for fault management. To overcome those issues, the authors proposed an appli-
cation from an operational perspective based on machine learning. They describe
the details workflow of the application. They use the deep learning process to survey
the issues and based on that they also build the block diagram for the typical fault
management system.
According to the authors [9], their proposed system is able to diagnose the fault
in the cellular network for the IoT devices. They proposed a data-driven indoor
Modem Functionality Diagnostic Tool with Graphical User … 31
diagnostic framework for this application. This framework uses the machine learning
technique and trained some prediction models. Those models are able to construct
a global view of the radio environment which is well known by its name radio
environment map (REM). Later, this REM will be used for diagnosing and managing
the faults in the cellular network at the location.
In [10], authors proposed to combine the self-organizing networks (SON) and
traces. According to them, it will help to detect the issued related to the radio interface
and easy to solve the issues. As the call traces functions have the user information
so it will be easier to detect and resolve the issue automatically with this proposed
model. This model also helps to assess the RF condition based on the location data.
3 Motivation
All the projects that are based on the cellular network have a common component,
called. This modem is an essential part of GSM comm. In the IoT applications, the
modem keeps also a big contribution. So, we should concentrate on the functionality
of the modem. And we should make sure that the modem should work perfectly under
all circumstances. But if it faces any error, then we should also be prepared to fix the
issues. So, for this point of view, we are motivated to make a proper functionality
diagnostic tool with a graphical user interface for modem cum cellular networks.
4 Research Gaps
Conventionally, a cellular modem is diagnosed by manual method. The test modem

needs to be connected with a computer via a carrier board. Then AT commands are
serially sent to the modem from the computer one by one. The person responsible
for the diagnose observer the response from the modem and decide the reason of
error. So this procedure is too much time consuming, and sometime the person needs
to remember all AT command and the command sending sequence. So, we feel a
need for an automated tool with a graphical user interface that can send all those AT
commands and grab the response from the modem and sometime the tool can try to
solve the error by its own.
Another big gap that we found is remote access to the modem. If the modem is
deployed as part of a product in a remote location, then it is a bit difficult to access
and diagnose the modem if any issue occurs. So, in this project, we are also trying
to focus this remote access part also.
32 A. Biswas et al.
5 System Overview
5.1 System Architecture
The modem functionality diagnostic tool is designed for the UNIX-based operating
system only. The GUI part will be installed in the local Linux system. But for the
hardware, there are two possible options. If the modem is in the local location, then
it can be connected via a carrier board through the USB port. But if the modem is
deployed in the remote location, then one client and one host script must execute to
connect and communicate over the server.
In our setup, we were using the Intel XMM7560 modem device. One USB type
carrier board was used to communicate with this modem. If the modem is in the
remote location, then the system must connect over the server to communicate with
the modem. For this setup, we used the Intel IA board. And the modem is connected
via the PCIe interface with the Intel IA board. The board itself is connected with
the server. In our case, we test it over the LAN server. For the WAN, we need some
network side changes.
5.2 System Installation
The entire modem diagnosis tool package contains the diagnostic tool binary itself
and the dependency libraries and an auto-installation script. The package has different
binaries for the different mode of modem. For the remote mode, the target must be
set up before installing the tool. Firstly, the UNIX system and the Intel IA board must
be connected on the same local server and able to ping each other. In the installation
process, the installer will ask for the mode selection. According to the mode, it will
install the correct binary, the dependency scripts and libraries. Figure 1 shows the
mode selection process.
Fig. 1 Installation mode

selection
Fig. 2 Log for host socket and remote socket binary
For the remote or PCIe mode, we must go through some extra steps after instal-
lation. After successful ping in both host and platform, a couple of binary needs to
run in both the system like Fig. 2 to initiate the socket [11] connection.
The next step is common for the both mode of operation. In this step, the tool
binary is being executed and opens the GUI. For the PCIe mode, the tool checks
some extra things like device’s presence, device drivers, network interface, AT ports
and modem state before sending the first AT command. If the tool found any problem
here, then it will report the error.
Now the tool will send AT command, grab the response for the command and
show in the GUI. First, few commands will grab the information about the modem,
SIM, PIN, network, etc. Then, it registers the SIM with the live network and reads the
network type, signal strength, operator name, SPN, context-related data and location
info. Figure 3 is the AT command flow.
If all the response is okay and does not have any issue, then the tool down the
network interface for the modem and put the IP address, subnet mask and default
gateway on that network interface and finally up the same. Users can verify the
network interface by ping any global IP.
After every step, the tool grabs the response. If any error came in the response,
then it sends the error response to the error checking block and tries to resolve the
error. If the error is resolved, then the tool came back to the next AT command else,
and it will suggest the possible step to resolve the error.
6 Implementation Details
The hardware implementation is different for both USB and PCIe mode. In the
USB mode, a carrier board is needed. So, for this configuration UNIX is PC able to
communicate with the modem directly over the AT ports which automatically come
up when the board is connected to the PC.
34 A. Biswas et al.
Fig. 3 Diagram for the full system
For the PCIe mode, the Intel IA board must be connected with the UNIX PC via
LAN with the successful socket communication. As the modem is connected in the
PCIe bus so the PCIe drivers are also needed.
Figure 4 is a screenshot of the tool, contains basic information like device port,
IMEI, IMSI, roaming status, operator name, registration status, registration mode
and network signal strength 3GPP type location, MCC, MNC, LAC, TAC and CID.
Fig. 4 Info tab of the tool
Some elements may change the value WRT time and location. So, these values are
continuously updated in the GUI.
Figure 5 is the SIM tab that contains the SIM details like SIM presence, IMSI,
service provider name, MCC, MNC, subscriber number, PIN required or not, PIN
entry retries SIM card identifier, card slot count and the current card slot number.
This tab also has the features to unlock the PIN.
The diagnose tab in Fig. 6 contains the PCIe device name, drivers needed for the
modem, network interfaces, AT ports and modem status. This tab also contains the
Fig. 5 SIM tab of the tool

36 A. Biswas et al.
Fig. 6 Diagnose tab of the tool
context for the modem and two buttons to refresh the data and resolve the current
issue. Figure 7 is also showing the diagnose tab with some error messages. Somehow
the modem has been timed out, and the AT ports are also not available in it. In this
case, the issue may solve by clicking the resolve button.
Most of the cases, the tool automatically resolves the problem by clicking the
resolve button. But if it is beyond the scope of the tool, then it suggests the steps and
hopes for the user interface.
Fig. 7 Diagnose tab for the tool with error message

7 Conclusion and Future Work Scope
Modem functionality diagnostic tools can be used to diagnose the modem-related

issue in different types of hardware platforms. This will reduce the time to manually
diagnose the modem by sending AT command one by one. This tool is designed in
such a way that it can work with other plugins also. So, this plugin feature gives the
versatility of the software. Till the date, this tool supports three main plugins like
modem manager, Ofono and AT manager. The diagnose feature is available only in
the AT manager plugin.
This tool is totally free and open source. So, anyone can download and modify
it. This tool still has some scope of upgradation or future work. For the PCIe type
modem, we need an extra Linux-based computer. The tool needs to run in the host
Linux PC. User needs to create a network socket connection in the both Linux PC
and the Intel IA system. If the socket connection is successfully built, then the tool
can communicate with the modem through the socket connection over the IP address.
But in the future, it is also possible to embed the tool itself on the Yocto [12] recipe
with the dependencies and build the OS for the Intel IA platform. Then the extra
socket connection is not further needed.
In this diagnose tool, some modem manager feature is discarded and only focuses
on the data session diagnose. So, it is also possible to add the modem manager
features like messages, contacts, radio frequency, etc. And it is also possible to add
the diagnose feature for these features also. Then, it will be a full feature modem
diagnosis tool.
References
1. H. Kwon, K. Kim, C. Lee, The unified UE baseband modem hardware platform architecture for
3GPP specifications. J. Commun. Netw. 13, 70–76 (2011). https://doi.org/10.1109/jcn.2011.
6157254
2. M.N. Iqbal, L.Y. Xin, W.U. Rehman et al., Diagnostic tool and remote online diagnostic
system for Euro standard vehicles, in 2017 IEEE 3rd Information Technology and Mechatronics
Engineering Conference (ITOEC) (2017). https://doi.org/10.1109/itoec.2017.8122328
3. A. Alyazidi, M.B.H. Frej, N. Gupta, Multi-node partial-mesh network communication via AT
commands, in 2018 IEEE Long Island Systems, Applications and Technology Conference
(LISAT) (2018). https://doi.org/10.1109/lisat.2018.8378013
4. K.P. Hunt, J.J. Niemeier, L.K.D. Cunha, A. Kruger, Using cellular network signal strength
to monitor vegetation characteristics. IEEE Geosci. Remote Sens. Lett. 8, 346–349 (2011).
https://doi.org/10.1109/lgrs.2010.2073677
5. M. Ito, N. Nishinaga, Y. Kitatsuji, M. Murata, Reducing state information by sharing IMSI
for cellular IoT devices. IEEE Int. Things J. 3, 1297–1309 (2016). https://doi.org/10.1109/jiot.
2016.2587823
6. N. Kouzayha, M. Jaber, Z. Dawy, Measurement-based signaling management strategies for
Cellular IoT. IEEE Int. Things J. 4, 1434–1444 (2017). https://doi.org/10.1109/jiot.2017.273
6528
38 A. Biswas et al.
7. B. Hong, S. Park, H. Kim et al., Peeking over the cellular walled gardens—a method for closed
network diagnosis. IEEE Trans. Mob. Comput. 17, 2366–2380 (2018). https://doi.org/10.1109/
tmc.2018.2804913
8. S. Basangar, B. Tripathi, Literature review on fault detection of equipment using machine
learning techniques, in 2020 International Conference on Computation, Automation and
Knowledge Management (ICCAKM) (2020). https://doi.org/10.1109/iccakm46823.2020.905
1543
9. S.-F. Chou, H.-W. Yen, A.-C. Pang, A REM-enabled diagnostic framework in cellular-based
IoT networks. IEEE Internet Things J. 6, 5273–5284 (2019). https://doi.org/10.1109/jiot.2019.
2900093
10. A. Gomez-Andrades, R. Barco, P. Munoz, I. Serrano, Data analytics for diagnosing the RF
condition in self-organizing networks. IEEE Trans. Mob. Comput. 16, 1587–1600 (2017).
https://doi.org/10.1109/tmc.2016.2601919
11. M. Xue, C. Zhu, The socket programming and software design for communication based
on client/server, in 2009 Pacific-Asia Conference on Circuits, Communications and Systems
(2009). https://doi.org/10.1109/paccs.2009.89
12. A.P. Navik, D. Muthuswamy, Dual band WLAN gateway solutions in Yocto Linux for IoT
platforms, in 2017 International Conference on Internet of Things for the Global Community
(IoTGC) (2017). https://doi.org/10.1109/iotgc.2017.8008968
Regression Model to Estimate Effort
for Software-Oriented Projects
V. Vignaraj Ananth, S. Aditya, S. Ragul Kesavan, and B. Keerthiga
Abstract The assessment of the product development initiative remains a compli-

cated issue requiring additional analysis. The efficiency of the development in soft-
ware depends on an appropriate estimation of the effort needed for software develop-
ment and enables project managers to design software life cycle operations correctly
with the efficient software cost estimating techniques. The major experimental work
conducted throughout this paper is to accurately forecast the problems in developing
program-related projects. The parallelization of the behavior factors is achieved using
the M5P approach to maintain statistical precision. Besides, performance differences
with random forest are offered with the models acquired using the M5P Method to
demonstrate the efficiency attained with each method.
Keywords Machine learning · M5P technique · Random forest technique
1 Introduction
Presently used techniques of effort calculation such as feature point analysis and
COCOMO failed to effectively measure the value and energy needed for software
development. These approaches are not sufficient for calculating cost and effort
This work was supported in part by the Department of Computer Science and Engineering of
Thiagarajar College of Engineering.
V. V. Ananth (B) · S. Aditya · S. R. Kesavan · B. Keerthiga

Computer Science and Engineering, Thiagarajar College of Engineering, Madurai, Tamil Nadu,
India
e-mail: vignaraj@tce.edu
S. Aditya
e-mail: aditya77388@gmail.com
S. R. Kesavan
e-mail: srk7official1998@gmail.com
B. Keerthiga
e-mail: keershika@gmail.com
40 V. V. Ananth et al.
since they are equipped to create procedurally based applications. The computational
and object-oriented approach contrasts as the previous distinguishes records and
procedure.
The use case point (UCP) is based on the use case diagram for calculating the
effort of a provided software product. It aims to provide precise measurement of
effort from the testing process.
Case and actor are graded as ordinary, mediocre, and advanced into one of the
three levels by use. The number of operations per use case helps complexity to be
calculated in terms of value. Significantly over the past periods, the UCP model has
been regularly established. This means that the software’s size and performance are
purely corresponding with one another.
Regardless of this factor, UCP’s computing contribution equation is not widely
recognized by software enterprises. The M5P technique is used in this analysis to
tackle the constraints of the UCP model and to improve the predictive efficiency of
estimating program effort. The results obtained using the M5P-based effort evaluation
method are then computed by using random forest (RF) to determine its efficiency.
Section 1 describes the introduction to effort estimation. Section 2 provides the
literature survey. Section 3 describes the algorithm involved in this paper. Section 4
shows the experimental design. Section 5 describes the experimental results, and
Section 6 provides the conclusion and future.
2 Literature Review
(A) Early-stage software effort estimates based on case point use using random
forest technology:
Software effort estimation was done by the techniques random forest, log-linear
regression, and stochastic gradient boosting. Out of which random forest yielded
more accurate results as it generated trees in parallel and averaged the outcome of
the results [1].
(B) A summary and evaluation for the description of machine learning strategies
of automated soil mapping:
This research compared ten machine learning techniques designed to map broad
classes of soil and soil orders. With the greatest precision, support vector machine
and KNN achieved performance.
(C) Models for the urban air quality control system:
Based on comprehensive studies, M5P outperforms other algorithms due to efficiency

of the tree structure and efficient generalization capabilities. On the other hand, the
artificial neural network achieved the worst results because of its poor generalization
ability when working on a small dataset with many attributes that leads to a complex
Regression Model to Estimate Effort for Software-Oriented … 41
network that overfits the data, while having support vector machine better than arti-
ficial neural network in our case due to its adaptability with high-dimensional data
[2].
3 Basic Concepts
(A) Overfitting:
Instead of the inherent relation, a mathematical formalism of overfitting represents the

random error. A overfit happens because the configuration becomes too complicated,
as in the number of sightings and many parameters. An overfitted model has low
predictive efficiency because it responds with small variations within the training
results.
(B) Tree pruning:
Pruning is a method which automatically resizes the decision trees through elim-
inating portions of the tree that really have no potential for particular grouping.
Pruning minimizes the difficulty of overall classifiers, thereby increasing statistical
precision while decreasing overfitting [3].
(C) Tree smoothing:
It helps to address the strong divergences at the leaves induced all through the pruning
among neighboring static models.
(D) Weka Tool:
Data processing applications for clustering or regression analysis of datasets: This

is open-source software. Methodologies mentioned here are performed using this
device [4].
4 Algorithm Description
(A) M5P Algorithm:
The M5 algorithm generates a regression tree by testing repetitively dividing the

dataset into a single variable [5, 6, 7], which decreases ambiguity about the dependent
(target) variable (Fig. 1).
Fig. 1 M5P flowchart
(B) Random Forest Algorithm:
The algorithm of Breiman is commonly used to apply the RF technique [8]. The
measures discussed below are assessed in the construction of an RF method for effort
[9–11]. These phases help to construct any tree with the random forest method.
(1) S denotes total the number of trees and dataset of R points (h1, g1) (h2, g2).
(hR, gR) is considered.
(2) Repeat the steps from 3 to create a T number of trees.
(3) T represents the number of training cases, and V refers to the number of
variables in the classifier.
(4) Split the dataset as training data and testing set. Training data is represented
as the number of variables in the classifier.
(5) To the loaded data and RF tree, Zq is built by repeatedly rehabilitating the
corresponding phases for each terminal nodes of the tree until the minimum
node size n minimum. The identifiable dataset is generated for each tree.
(6) To evaluate the preference at a tree node, the no. of input variables v is chosen.
(7) For every tree, randomly choose the number of input variables.
(8) Within the training, the set is determined as the best division based on these v
variables. The value of m should be kept constant all over the forest’s growth.
Almost every tree must be expanded to the maximum and not pruned.
9) Therefore, the outcomes are obtained from a group of trees J1, J2, …, Jq.
(10) For every one of the trees in the forest, the input factor must be set aside.
5 Experiments and Results
(A) Dataset Collection
Barada Prasanna Acharya [12] using the proposed solution obtained datasets
containing 149 UCP-based results. The datasets have been taken from different sepa-
rate Web sites, fifty datasets taken from the ISBSG [13], and thirty-four datasets
have been taken from medium-sized organization and remaining taken from Western
University. In the dataset, the initial column reveals the software dimension. The
remaining columns represent the rate of productivity, project’s complexity, and actual
effort, respectively [14, 15].
(B) Results
See Figs. 2, 3, 4, and 5.
(C) Performance Analysis

prediction accuracy PRED = 1 − (ei − f i ) s ∗ 100%
where
ei actual effort
f i predicted effort
s number of projects in the dataset
PRED for random forest = 97.52% and PRED for M5P = 98.5%
Fig. 2 M5P output in Weka tool

Fig. 3 Random forest output in Weka tool
Fig. 4 M5P actual effort versus predicted effort chart

Fig. 5 Random forest actual effort versus predicted effort chart
6 Conclusion
In this paper, among different estimation techniques, we choose M5P and random
forest methods. M5P technique performed better than the RF method and also elim-
inates overfitting. Hence, it improves predictive accuracy by the reduction in over-
fitting. M5P models a single tree and predicts the output; whereas, random forest
creates many parallel trees and averages its output.
In the future, M5P can be contrasted with other approximate effort techniques
and to analyze a generalization of results and apply M5P to software project datasets
from various domains.
References
1. B. Heung, H.C. Ho, J. Zhang, A. Knudby, C.E. Bulmer, M.G. Schmidt, An overview and
comparison of the machine- learning techniques for classification purposes in digital soil
mapping. Geoderma 265, 62–77 (2016)
2. C. Lopez-Martina, A. Abran, Neural networks for predicting the duration of new software
projects. J. Syst. Softw. 101, 127–135 (2015)
3. K. B. Shaban, A. Kadri, E. Rezk, Urban air pollution monitoring system with forecasting
models. IEEE Sens. 6, 101–115 (2016)
4. A. Panda, S.M. Satapathy, S.K. Rath, Empirical validation of neural network models for agile
software effort estimation based on story points. Procedia Comput. Sci. 57, 772–781 (2015)
5. F. Sarro, A. Petrozziello, M. Harman, Multi-objective software effort estimation, in 2016

IEEE/ACM38th International Conference on Software Engineering (ICSE) (IEEE, 2016),
pp. 619–630
6. P. Rijwani, S. Jain, Enhanced software effort estimation using multi layered feed forward
artificial neural network technique. Procedia Comput. Sci. 89, 307–312 (2016)
7. A. Idri, M. Hosni, A. Abran, Improved estimation of software development effort using classical
and fuzzy analogy ensembles. Appl. Soft Comput. 49, 990–1019 (2016)
8. L. Breiman, Random forests. Mach. Learn. 45(1), 5–32 (2001)
9. P. Pospieszny, B. Czarnacka-Chrobo, A. Kobylinski, An effective approach for software project
effort and duration estimation with machine learning algorithms. J. Syst. Softw. 137, 184–196
(2018)
10. P. Phannachitta, K. Matsumoto, Model-based software effort estimation—a robust comparison
of 14 algorithms widely used in the data science community. Int. J. Innov. Comput. Inf. Control
15, 569–589 (2019)
11. H. Mustapha, N. Abdelwahed, Investigating the use of random forest in software effort
estimation. Procedia Comput. Sci. 148, 343–352 (2019)
12. S.M. Satapathy, B.P. Acharya, S.K. Rath, Early-stage software effort estimation using random
forest technique based on use case points. Research Article. Instrum. Eng. Technol. J. 18, 63–70
(2017)
13. O. Hidmi, B.E. Sakar, Software development effort estimation using ensemble machine
learning. Int. J. Comput. Commun. Instrum. Eng. 4(1), 143–147 (2017)
14. L.L. Minku, X. Yao, Ensembles and locality: Insight on improving software effort estimation.
Inf. Softw. Technol. 55, 1512–1528 (2013)
15. ISBSG: The international software benchmarking standards group, http://www.isbsg.org
(2011)
Dr. V. Vignaraj Ananth

is Assistant Professor, Computer Science and Engineering
in Thiagarajar College of Engineering, Madurai, Tamil Nadu,
India. His research interests in Software Effort Estimation,
Internet of Things, and Data Analytics. He has published
Research Article in 15 International Journals and 11 Interna-
tional Conferences.
S. Aditya
is an undergraduate student, Computer Science and Engi-
neering in Thiagarajar College of Engineering, Madurai, Tamil
Nadu, India. His research area includes Machine Learning and
Internet of Things. He has published Articles in 04 International
Journals and 02 International Conferences.
Innovative Technology-Based Social
Enterprises for Inclusive Sustainable
Healthcare in India
Pradnya Vishwas Chitrao, Pravin Kumar Bhoyar, and Rajiv Divekar
Abstract India is witnessing rapid progress in the twenty-first century. The progress
unfortunately is very uneven and non-inclusive. Many people are below the poverty
line and often have to do without sufficient food. Many children do get the benefit of
primary education. Girls and women do not get adequate nutrition or equal oppor-
tunities of education and employment. In healthcare, in particular, we see a major
disparity in the facilities available to different sections of society in India. A privi-
leged few get access to sophisticated, advanced healthcare facilities while the poor
in urban slum areas and especially in rural India either have no access to good
quality basic healthcare or cannot afford them. A lot of social enterprises in India
are using smart innovative technology to solve this problem of making available
quality healthcare to the urban poor and the needy rural in India at affordable rates.
Objective The paper studies two social enterprises, the first being AAarogyaVedh
Healthcare InfoTech Pvt. Ltd. which provides cardiopulmonary resuscitation (CPR)
and the second being a family owned social enterprise that offers a juggad ambulance
cum medical unit called AmbuPod that uses innovative technology to make available
quality and economical healthcare in a timely manner. Research Methodology The
researchers have conducted personal interviews by way of primary research. They
have also availed of secondary resources. Significance of the Study The study will
encourage enterprises to use technology in order to provide for high grade healthcare
available to everyone in India. It will inspire decision makers, NGOs, and business
entities to help such social enterprises. The study also brings out the fact that these
social enterprises not only provide crucial services like cost-effective measures for
creating physical, mental, or emotional well-being (health care), but also create job
opportunities.
P. V. Chitrao (B)
Symbiosis Institute of Management Studies (SIMS), Constituent of Symbiosis International
University, Pune, India
e-mail: pradnyac@sims.edu
P. K. Bhoyar · R. Divekar
SIMS, Pune, India
50 P. V. Chitrao et al.
Keywords Social enterprises · Innovative technology · Cost-effective healthcare ·

Job opportunities
1 Introduction
India has the distinction of having the second largest population in the world. It is
also has the fifth largest GDP. But around three fifths of Indian population is rural
[4]. India faces a lot of issues in terms of development which are related to major
areas like healthcare, employment, water, sanitation, clean environment, and means
of livelihood. The economic growth being witnessed in India does not include many
people who nonetheless pay for this growth [1]. On the one hand, life expectancy in
India has increased by around ten years. But, on the other hand, infant mortality in
India in 2015 was 47.7 per 1000 live births [14]. In June 2019, the infant mortality
rate has gone down to 36.7 per thousand births, but India’s rank in the world is still
52nd worldwide in this respect [7]. Hospitals in India, especially government ones,
are overcrowded; consequently, people do not get proper treatment. Good quality
and cost-effective healthcare facilities in semi-urban and rural regions are rare, and
if available are very costly as a result of which healthy life is enjoyed by very few.
Often, cultural and social rules restrict usually women from receiving quality health
care administered by male doctors.
Social entrepreneurship pinpoints a major societal concern and then utilizes
entrepreneurial parameters to begin and operate a business for social benefit that
will bring about a positive difference in society. Many social entrepreneurs in India
thought of novel ideas to resolve social issues in a sustainable manner and thus proved
to be great value creators for society. They used disruptive technological innovations
to bring about quality healthcare that can be afforded by all. Technology today helps
doctors give consultations through video sessions to persons facing health issues in
rural areas that face chronic low network and makes available cost-effective life-
saving gases for treatments and techniques performed in hospitals in an emergency
in order to support life after the failure of one or more vital organs.
Social entrepreneurship is the buzz word in India in the twenty-first century as
the country is attempting to attain an evenly spaced out economic growth across all
sections of in hospitals society as against a mere growing GDP statistic. It attempts
to tackle various issues like education, quality health care for all, climate changes,
and energy efficiency. Social enterprises’ work in the field of healthcare related
services is especially important. There are no hospitals in many rural parts of India
or for that matter even ambulances. The urban poor also often cannot pay for these
facilities. Precious time is spent in an attempt to reach the person requiring immediate
therapeutic intervention to the medical center. Some exurban areas even lack good
road connectivity. These sections of society consequently do not get proper treatment
for what are seen as standard cough and cold, and resort to grandmother’s cures.
In remote backward areas, pregnant women, who are about to deliver, often are
dependent on a midwife who may occasionally be caught up in an emergency at
Innovative Technology-Based Social Enterprises for Inclusive … 51
some other place and therefore may not be in a position to reach the woman who
has gone into labor on time. As a matter of fact, reaching medical services to India’s
remote areas is a major problem [12]. Women who migrate to metro cities are often
afraid of approaching a doctor on account of cultural up bringing or sheer lack of
knowledge of from where to get the required information.
In the case of cardiopulmonary resuscitation (CPR), many Indians do not even
know about it. Often, CPR is used as an excuse or a prank to have mouth-to-mouth
breathing, and thus introduce humor or romance in films. While mouth-to-mouth
breathing is often mistaken to be CPR, it is actually not as effective as putting hard
pressure on the chest. CPR is a scientifically accepted medical intervention and can
be conducted with ease and efficacy through imparting of basic training, until medical
help arrives.
The physical differences between a man and woman deter men from administering
CPR on women.
2 Literature Review
India is going through a very uneven growth phase and is facing many develop-
mental issues like poor healthcare, unemployment and poverty on account of very
fat population growth and the consequent environmental and social needs created by
limited resources. Despite the immense growth of industries and service sectors post-
Independence, India is still basically an agriculture-based economy [10]. Poverty
forces people to live in surroundings which do not have decent shelter, potable water
or proper sanitation [13] and which attract illnesses. Rapid unsystematic and non-
catered for development and rigid adherence to unconventional business processes
to generate business has caused the deterioration of the current physical fitness care
frame work in rural and semi-pastoral areas besides major cities [9]. Women in India
have a tendency of ignoring their ailments because either they have no time to spend
a lot of time waiting in clinics or because do not like being examined and exposed to
male doctors and attendants in clinics or because they are worried that their illness
will become known to someone who in turn may gossip about it in their social circles.
They therefore unnecessarily put up with a lot of discomfort and run the danger of
their illness getting aggravated.
2.1 Public Healthcare Facilities
The government-run health services in India are bad and not quality driven. Govern-
ment basically has very inadequate funding per person which in turn leads to hiring
of substandard healthcare experts and infrastructure. The poor have to rely on such
government facilities especially if they have no medical insurance. At this stage,
India cannot cater for the citizens’ social insurance that will facilitate access to
quality public as well as private healthcare facilities to the poor. The law guarantees
every individual access to quality healthcare. But approximately, 60% of the popula-
tion in India cannot use this right on account of various reasons like absence of good
facilities, too less number of qualified healthcare personnel, and unequal access to
primary drugs and medical amenities. Government is carrying out many schemes
which are not useful as they are not implemented properly.
2.2 Private Healthcare
On account of government healthcare systems being extremely substandard in terms

of quality, many Indians take recourse to private medical attention. In India, around
92% of the time, people opt for private medical services. In India, about 70% of the
population is from the country side. But, personal medical treatment is costly and is
often not checked in terms of adherence to certain accepted national or international
standards; its quality also varies. Uneducated and poor rural folks cannot afford it
nor rely on it. Good quality health care is often not availed by women who migrate
to cities on account of lack of reliable sources of information.
Cardiopulmonary resuscitation (CPR) is a technique used to save lives and is
used in emergencies where a person stops breathing in situations where the patient
is almost drowning or suffers a heart attack. CPR combines chest compressions and
mouth-to-mouth rescue breathing. Both these actions facilitate the flow of oxygenated
blood to the brain and other vital organs until the patient gets professional medical
treatment. But in India, many a times no trained person or a paramedic is nearby to
administer this life-saving technique. Precious moments are lost by the time a person
can be reached to a hospital. Some times in rural areas, either there is no hospital
nearby or no ambulance is either available or cannot reach the patient through the
narrow pathways. Many a times, the family cannot even afford an ambulance. Also,
even if a trained person is available on the spot, if the patient is a female, social norms
prevent this life-saving first aid being administered.
2.3 Social Enterprises and Technology for Better Healthcare
“Social entrepreneurship” is the twenty-first-century global phenomenon that

shows hope of bridging the ever widening difference between the pressure for
fulfilling social and habitat requirements and the correspondent reservoir of means.
Entrepreneurship that results in economic growth is today attracting academic atten-
tion. However, it is only lately that social enterprises are being considered as a step-
ping stone to social progress [11]. Social entrepreneurship is viewed as the solution to
the market and government’s inability to fulfill the requirements of a given commu-
nity [15]. Social entrepreneurs in the twenty-first century are considered today as
instruments of transformation in society; these persons use systematic solutions to
issues in society, and thereby try to change things in society in a useful manner [6].
Many of them employ innovative technology to make their enterprise a workable
business model.
Technology can bring about such differences. A lot of enterprises and non-
government organizations are utilizing mobile technology for aiding government
lessen the load on the government health services. India has more than 900 million
people who use mobile phones. Today at a global level, the mobile cellular subscrip-
tions have reached the 6.8 billion mark (Union IT 2013), which has helped tech-
nology in covering up the communication loopholes and increasing globally growth
in economy and progress [2]. Cell phones in India are popularly used because they
charge less fee rates. Studies conducted only a short period earlier indicate that the
expenditure on mobile phone bills in Indian rural families is around Rs. 5 per month
whereas in the city for poor families is around Rs. 37 on a monthly basis (Department
of Health and Family Welfare, Govt. of India MCTS, 2013). Research has indicated
that m-Health mediations and prescriptions are in vogue both in southern India (Shet
et al. 2010) as also in the country areas in northern parts of India [8].
Telemedicine and mobile health clinics thus are the new innovative and adapt-
able solutions that can answer India’s rural healthcare issues. They are not mere
“helplines.” It is a fact that telemedicine provides the basic technology infras-
tructure that will bring rural areas the benefits of medical consultancy, diagnostic
services and treatment guidelines through Internet and telephone connections. These
centers employ teleconferencing technology and remote diagnostic devices for under-
standing a patient’s basic parameters and conveying it to a doctor in an urban center
for diagnosis. The doctor then prescribes treatment that is supervised by the local
primary care worker in the rural region. The technology is an effective answer to
the healthcare issue of a lack of trained doctors in India’s rural areas. Most trained
practitioners are usually trained in the cities and therefore do not wish to relocate to
rural areas on account of an absence of well-paid jobs and well-equipped hospitals
in rural areas
In the case of emergency first aid, wherein patient needs artificial resuscitation, the
first 5 to 6 min are critical for the patient’s recovery. Globally, only 30% people are
saved, and in India, the number is less than 1%. This is because even in metropolitan
areas, the response time of an ambulance is forty minutes. As a result, people suffering
from cardiac arrest do not get timely medical aid. Cardiac arrest is basically “electrical
disturbance which causes the heart to stop beating.” The patient collapses, becomes
unresponsive, and usually stops breathing. This deprives vital organs of blood flow
and hence of oxygen. This can cause severe and irreversible damage to brain in nine
minutes. As a result, there is death in 90% of cases in out-of-hospital settings. In
fact, cardiac arrest is the most important cause of sudden cardiac death.
The answer to this is giving mouth-to-mouth resuscitation (or cardiopulmonary
resuscitation, i.e., CPR). CPR is “a procedure that is performed in an emergency
when the heart ceases to beat” [5]. It consists of 100–120 chest compressions/min
of 5–6 cm depth each. The chest compressions: breaths ratio is 30:2. Chances of
survival improve by 2–3 times when CPR is provided within 4–6 min after cardiac
arrest.
There is low level of public awareness of CPR because of scarcity of trained
healthcare workers (nursing/paramedic staff), and poor doctor-to-patient ratio. At
present, it is 1:1700. Hence, more than 99% patients do not receive appropriate CPR
especially in rural areas. This is because of underdeveloped emergency response
system. Medical assistance arrives extremely late, and so mortality increases by 10%
with each passing minute. Consequently, there is poor survival rate. There is also
no mechanical CPR-providing device currently available in India which is designed
specifically to fulfill the requirements in Indian setup. There is therefore a need to
have cost efficient life device that is easy to use and which is available in all prominent
public places.
Ultimately, technological innovations are helping promote quality healthcare at
affordable rates. Steven Steinhubl, Director (Digital Medicine) of Scripps Transla-
tional Science Institute, has accurately predicted that healthcare delivery in future
is going to undergo a massive transformation on account of the convergence of
genomics, digital devices and precision medicine [3].
Most villagers in India, who suffer from heart disease, do not have access to
consistent and quality medical facilities. The George Institute for Global Health has
initiated in India low cost-effective strategies that will provide good medical care
for people. It uses sophisticated wireless technologies including mobile phones for
giving medical care that facilitate the healthcare provider in giving case specific
computer-based medical decision support for generating the Systematic Medical
Appraisal Referral and Treatment (SMART) of a patient in that area. Every rural
area in India is given an Accredited Social Health Activist or ASHA, who is usually
an experienced female appointed to operate in her own milieu. She is thus not an
unknown person and is in fact someone whom the villagers trust; she is usually
expected to have cleared junior college. The group has started this system in over
fifty villages in Andhra Pradesh, and given training to ASHAs to find out persons
with heart ailments, and to provide the required medical care.
A good healthcare system allows sufficient universal access that will not overload
the system. This is because the monetary costs for making available correct healthcare
and consistent improvement is periodically provided. Its prominent features include
competence training, accountability, good quality medical facilities, and profitable,
efficient utilization of the latest pertinent investigation. It serves vulnerable sections
of society like kids, women, disabled and the elderly.
3 Objective
The paper tries to understand how a social enterprise can use innovative technology
to make feasible reasonably priced good healthcare to poor and to villagers as also
provide emergency CPR in small towns, rural areas, as well as cities. It also tries to
find out how such an organization can train and give good jobs to people.
4 Research Methodology
The researchers interviewed two business owners. The first enterprise is AAaro-
gyaVedh Healthcare InfoTech Pvt. Ltd. Healthcare Info Tech Pvt. Ltd. The second
was the co-owner and co-founder of Lynkambupod Pvt. Ltd, namely Ms. Yamini
Lavanian. They also availed of secondary sources of information.
4.1 Genesis of AarogyaVedh Healthcare InfoTech Pvt. Ltd.
AarogyaVedh Healthcare InfoTech Pvt. Ltd. Healthcare Info Tech Pvt. Ltd. Was
started in April 2019 by Mr. Shreekar Mangeshkar and Dr. Akshara Jogi. Mr.
Mangeshkar with qualifications in postgraduation certificate in physics and nuclear
technology and Dr. Akshara Jogi with a medical degree and an M.Sc. in molecular
neuroscience found that unlike in UK, hardly any attention is paid to healthcare
subcenters, and more attention is paid to metropolitan city centers. The two first
focused on cardiac arrest cases. They found that while globally only 30% people are
saved after a cardiac arrest, in India, the figure is less than 1%. This is because no
one helps the patient especially if it is a case of road accident. In metropolitan cities,
the average response time of any ambulance service is around forty minutes. But in
cardiac arrest, first aid has to be given in the first nine minutes. Also, in India, CPR is
given manually and requires patience and strength. So they first have developed an
automatic cardiopulmonary resuscitation (CPR) integrated with automated external
defibrillator (AED).
4.2 Designing the CPR Machine
The founders saw that in rural areas, people are not aware of medical procedures.
They think that the doctor, who is giving CPR manually, is killing the patient, and
they often beat up the person. But their perception toward machines is more positive.
So they developed an automated CPR. To check the response time of the ambulance
to the cardiac arrest, they developed the emergency procedure which will intimate
the nearest police station, fire brigade, and emergency medical hospital capable of
handling CPR. They plan to install these CPR machines in medical boxes at any public
place in malls, retail shops, petrol pumps, and even cars (they are in discussion with
automobile manufacturers). Here, only a push of a button will inform all emergency
facilities nearby.
One just has to put the CPR machine on the patient, and the machine will work on
its own. So any school going child or even a minimally literate person can operate
it without any special knowledge. In CPR, one has to put pressure on chest surface
area corresponding to lower half of the sternum (breast bone). So the founders of the
enterprise designed the machine in such a way that it will fit on a person of any age,
gender, weight, and will work in all weather conditions.
The process of international patenting takes around five to six years. For Indian
patent, one first has to file a provisional patent wherein one checks the patentability
of one’s idea within a period of one month. If this yields positive results, then one
gets a year’s time frame in which one can develop one’s product or making changes
in it. Then one can file for a long-term patent after giving the feasibility as well
as financial feasibility reports. After this, one can apply for international patent
wherein the International Committee checks and validates the idea and then gives
the recognition certificate.
4.3 Business Model and Differentiating Features
From January 2019, the two owners have so far invested around Rs. 2–3 lakhs. For
background research, they took around four years. There are thirty other similar
devices which have existed in markets globally as of now out of which there are only
four in production and all are only sold in USA. These machines are calibrated for
heart problems over there. The owners have tie-ups with different doctors and so,
on the basis of the feedback from them, they prepared a device that caters to Indian
requirements. They have priced their device at Rs. 30,000/– whereas all other devices
cost between Rs. 3 lakhs to Rs. 7 lakhs. Also, the other devices require moving the
patient around as also the presence of three para medics to fit the machine on to the
patient. This therefore takes around 2–3 min of the crucial time frame of first aid to a
cardiac arrest patient. Also, the two entrepreneurs have come up with different kinds
of pneumatic pumps and lithium batteries integrated with portable solar cells for
powering their own CPR device. The portable cells are also available in the market
and in IISER, Pune, Mumbai, Delhi, as also in IISC, Bangalore. So these batteries are
also developed in markets in India and cost between Rs. 4000/– to Rs. 10,000/– per
battery. But if mass production of these batteries happens, the cost will go down to
around Rs. 3000/– per battery.
The product right now is in the cardio care category but will soon be under basic
life support, i.e., holistic emergency response system. A lay person with minimal
training will be in a position to use the CPR device. They are also planning to tie
up with a rural institute under the government scheme of earn and learn and offer
training with certification and employment with respect to assembling of the medical
instruments.
The owners plan to focus in future on heart casualties and neurocasualties and
different scanning devices like MRI, CT scan, and PED Scan. A CT scan machine
costs around Rs. 70 lakhs and so many people find it too costly. So they will focus on
the gray areas where people do not get proper treatment. The machine should ideally
be available in sub centers where first aid is expected to be given.
4.4 Strengths/Merits of the CPR Device
• Easy to use with very basic training imparted to even lay persons.
• Cheaper than all available devices.
• Time taken to fit it on to the patient is <10 s. Simply put it on the patient’s body
and start operating it.
• Simplicity of design enables only one person to operate it without the need of
more or qualified personnel.
• Works on solar powered batteries which can become cheap if mass produced.
4.5 Challenges
The CPR device is in the prototype stage. The owners are in the stage of setting up
the manufacturing unit and are in talks with various parties where this device will
prove a boon for emergency cases. They are also seeking funding for starting projects
wherein the device will be used.
4.6 Genesis of Lynkambupod
Dr. Lavanian came up with the concept of the social enterprise Lynkambupod Pvt.
Ltd. An AFMC, Pune medical doctor, as also a former Air Force officer, he became
interested in management of patient healthcare information, hospital information
systems, and started trainings in healthcare IT post-retirement. In the capacity of
VP of telemedicine at Apollo Hospital, Andhra Pradesh, he noticed that rural areas
cannot afford expensive healthcare infrastructure. He therefore started research out
of his pension money on how technology can be utilized for providing reasonably
priced healthcare infrastructure so that all can benefit from it. He ultimately devised
a jugaad vehicle namely, AmbuPod, that could serve as an ambulance cum mobile
OPD unit.
4.7 Designing AmbuPod
Dr. Lavanian’s team of engineers devised a narrow vehicle about three feet wide
(which is more spacious than an airplane passenger seat) that would travel on very
narrow or mountainous paths, and enable one therapeutic caregiver to administer
resuscitative aid to one patient who can be kept in a lying down position. This is the
prescribed anesthesiologist positioning where patient’s head is kept in the doctor’s
lap. The medical caregiver can thus access all the vital parts of the patient’s body till
the torso. If a patient is suffering from a lower appendage damage, AmbuPod will
come to the destination and first steady the patient, and then shift to AmbuPod in
order to reach the nearest hospital.
The therapeutic hardware in AmbuPod was designed after intensive research. The
equipment in the AmbuPod was designed by keeping in mind the following points:
• The heaviness and movability of the equipment
• CE certified quality
• The quantity of energy required by the machinery to function on sunlight based
energy (The batteries are charged by sun-based boards that in turn can be powered
by even an LED bulb).
AmbuPod also contains an appliance for removing substances like blood, saliva,
mucus, and vomit from a person’s airway (i.e., a suction apparatus) as also an oxygen
tank that holds adequate oxygen and still is not cumbersome. A suction machine is
needed if there is bleeding or if the patient is suffocating on own saliva. How the
machinery is positioned is important and will help one save the patient in time. A
lot of the machinery was imported. But some machines were indigenously designed
in a creative manner within the country. For example, the 12 lead ECG machine is
only as big as a visiting/credit card, and yet records precise readings onto a tablet or
a smart phone. The cost of this ECG from AmbuPod is between Rs. 30 to Rs. 50 for
rural people as against Rs. 150 of an ECG in a public medical center.
AmbuPod also has provision for care of mother and newly born child. It has
installed a hand held ultra-sound transducer for detecting fetal heart beat for prenatal
care namely a fetal Doppler. Dr. Lavanian found that many miscarriages happen in
villages because they are not in a position to learn of the health of the fetus in the
womb. AmbuPod ensures that miscarriages do not occur through timely intervention
with the help of this test. AmbuPod thus is not just a cost-effective, single patient
ambulance. It also functions as a clinic on wheels for dispensing daily OPD services.
It is a healthcare forum because it uses the technique of telemedicine.
4.8 Business Plans/Models Offered by the Enterprise
AmbuPod offers multiple business models for delivering timely, quality healthcare to
the poor and rural folks in a cost-effective manner with the help of smart technology.
One of the plans suggests that an AmbuPod be purchased jointly by village mukhiahs
or panchayat heads or NGOs. One AmbuPod can take care of three to five villages.
AmbuPod reaches out to a village and recruits and trains an individual belonging to
that village itself. This person officiates as the intermediary of the doctor and sees
and hears the patient’s symptoms. Dr. Lavanian has devised a tele medical app named
Ambu App that records all the health issues of the patient which are then recorded
and relayed by the trained person who is nominated as the Rural Manager (RM).
A mobile tent is set up where the AmbuPod stands. Foldable chairs are arranged
for patients. As soon as the RM sends the information on the app, all the details
reach the medical professionals in a city. One can also have skype interactions as the
app is video enabled and needs only 2 GB for this (as villages have poor Internet
connection). If video facility is not feasible, medical advice can be offered by urban
doctors over phone to the RM who then hands over the prescribed treatment in the
form of an e-prescription sent by the medical professional.
This kind of communication is done when Internet is not available. If Internet
connection is feasible, there is a complete well-orchestrated connect between the
doctor and the RM and the patients with the help of Ambu App. AmbuPod bears the
salaries of the RM, the doctors, nurses, or the para medics. The RM and other staff
are paid on a monthly basis and also get a commission for exceeding certain goals
every month. Normally, the trained person who operates the AmbuPod doubles up
as the RM and is pivotal to the operation of the clinic on wheels. The patients are
attended to on a fee per consultation model wherein each fee is between Rs. 90/– to
Rs. 150/–.
The owners are offering in addition other business plans. One of them is the
PPP plan, i.e., the public private partnership where the founders will collaborate
with government to provide reasonably priced healthcare. Another model is the
JV plan according to which they will partner with important entities like business
organizations for their CSR. The founders also offer the NGO model of business
according to which they will from time to time partner with Rotary Club and other
NGOs who buy the unit, and all facilities of that unit will be carried out by the owners.
The owners are also offering the donor model. Nonresident Indians (NRIs) and
affluent persons can purchase an AmbuPod and gift it to a village. Again, the owners
will run it. In other words, in all the business models offered, the owners will be
in charge of the unit. As of now, their business plan will generate jobs among the
poor and country folks. They are also offering positions to trained, retired Armed
Medical Core (AMC) personnel employment on account of their competencies and
experience.
According to all the plans offered, the founders will operate the enterprise and
ensure quality healthcare with the help of technology.
4.9 Strengths of the Enterprise
• AmbuPod is fitted with the necessary apparatus to conduct routine tests related to
blood, urine, anemia, as also check ECGs, and temperature, collect blood samples,
find out sugar levels, pulse, etc. It is a well-fitted, ultra-modern model having
medical apparatus powered by highly sophisticated technology.
• The enterprise is a private limited company recognized by the Indian Government.
AmbuPod began functioning about four years ago and up to now about Rs. 15–20
lakhs cost has been incurred.
• This year, the founders have devised a sturdier model of AmbuPod.
• So far, they have received a lot of positive response on social media by way of
over two million hits, fifteen thousand likes and around three thousand remarks
and congratulatory communications.
• They earned recognition from Amrita Technology Business Incubators at their

pitch fest in 2017 as Best Social Sector Impact Startup.
• They recently tied up with an African company and are currently nominated the
legally recognized distributor for an African country.
• The founders of AmbuPod have prepared a plan for a pilot project covering twenty-
five villages, and for this purpose, they have appointed eighty-six R.M.s.
• Once AmbuPod implements the project, they will have vast information that
will facilitate insurance companies, government, and business houses find out
in advance which area will be afflicted by which disease at what time and know
what medicines are to be administered at what time of the year. This data will
thus help plan the storing of drugs and lessen shortage of the same. It will thus
enable accurate predictive analysis.
• AmbuPod is reasonably priced. One fully set up AmbuPod that possesses the
facility of administering prescriptions via telephone Ambu, and which can double
up as a moving clinic, is priced at Rs. five lakhs. The fundamental model comes at
Rs. 3.8 lakhs. A regular, completely prepared fundamental life bolster emergency
vehicle (ambulance) is priced at around Rs. 8 to 10 lakhs, and the sophisticated
ambulances are priced at around Rs. 15 lakhs.
• The owners have applied to National Chemical Laboratory for funding. They have
procured two patents and are now seeking an international patent. The enterprise
thus has a lot of potential for being used in poor countries or in countries where
there are areas difficult to access for making available timely and quality medical
help in a cost-effective manner.
4.10 Challenges Facing the Enterprise
AmbuPod formally became operational in January 2017. But right now, it is not
making profit. The functioning of this enterprise on a major scale is yet to happen
in India. The scheme of the owners is to function for 50,000 villages out of the six
lakhs villages in India in the first five years in any state that is interested in any of
their different business models.
4.10.1 Recommendations
Government of India should back social entrepreneurs having latent capabilities

by giving the required knowledge, and guidance. It should introduce the social
entrepreneurs to the local and global funding agencies that give money for helping
social enterprises grow and develop. It should synchronize and regulate the activities
of social enterprises, and help them in their goal of self-development. It should offer
monetary incentives for healthy progress of the social enterprises. Venture capitalists
should finance and support such enterprises that employ technology to create social
value. Qualified medical experts also can now do social service by agreeing to give
their services through such technology riven platforms that will not take too of their
time or require them to travel for reaching out to the target audience.
The researchers also suggest that business houses should opt for such cheap yet
very beneficial healthcare services as offered by AarogyaVedh Healthcare InfoTech
Pvt. Ltd. and AmbuPod for achieving their CSR goals stipulated by the law. It will also
yield good, inclusive healthcare services that can be enjoyed by poor and countryside
people. NGOs also should also go in for these reasonably priced but efficient options
like the CPR device by AarogyaVedh Healthcare InfoTech Pvt. Ltd. and AmbuPod
to promote their social services in the healthcare sector.
There is a major gap in the idea of spectrum as viewed by business houses and
social enterprise sectors. Business corporations measure success in terms of the size
they attain with respect to earnings, profits, their value in the stock market, and the
wealth they bring to their owners. But the success of social enterprises must be judged
in terms of the benefit they bring to society. In fact, the more its benefit and the smaller
the enterprise itself is, the greater its success is considered to be. The basic aim of
any social entrepreneurship is to implement social change by transforming realities
at the local level. The local context thus is responsible for the kind of opportunities
that are possible for social entrepreneurship and for deciding which strategies have
to be used. These strategies are a reflection of the entrepreneurial spirit shown with
respect to the kind of resourcefulness and the capacity of recombining the prevalent
resources into innovative value adding configurations, as also an innovative way of
doing things. Technology today is helping such social enterprises create social value
in a cost-effective manner.
References
1. A. Aggarwal, India’s services sector: gateway to development? Econ. Polit. Wkly XLVII(26
& 27), 119–123 (2012)
2. J.C. Aker, I.M. Mbiti, Mobile phones and economic development (October 2010)
3. Business Line, The Hindu (2017, February)
4. Census of India, (2011, October 31). Accessed from censusindia.gov.in on
5. Cpr.heart.org. What is CPR? (2019). [Online] Available at https://cpr.heart.org/AHAECC/CPR
AndECC/AboutCPRECC/WhatIsCPR/UCM_499896_What-is-CPR.jsp. Accessed 26 May
2019
6. J. Elkington, P. Hartigan, The Power of Unreasonable People: How Social Entrepreneurs Create
Markets that Change the World (Harvard Business Press, Boston, 2008)
7. Gazetteer India 2019. http://www.geoba.se/country.php?cc=IN
8. S. Goel, S. Bali, A.J. Singh, Impact of a short term intervention (January 2007)
9. Y. S. F. Lee, Motorization in rapidly developing cities. Scaling urban environmental challenges
(2007)
10. R. Meher, Livelihood, poverty and morbidity: a study on health and socio-economic status of
the tribal population in Orissa. J. Health Manag. 9(3), 343–367 (2007)
11. J.C. Short, T.W. Moss, G.T. Lumpkin, Research in social entrepreneurship: past contributions
and future opportunities. Strateg. Entrep. J. 3, 161–194 (2009)
12. M. Shukla, Reaching Social Entrepreneurs in India’s Hinterland (2012)
13. WHO, World Health Statistics 2013: A Wealth of Information on Global Public Health (2013)
14. WHO’s World Health Statistics 2015

15. E. Yujuico, Connecting the dots in social entrepreneurship through the capabilities approach.
Socio-Econ. Rev. 6, 493–513 (2008)
Classification of Leaves Using
Convolutional Neural Network
and Logistic Regression
Sapan Naik and Hinal Shah
Abstract The dissemination of this research paper proffers the method of classi-
fying and recognizing the different leaves of plants from their image. The leaves
of plants contain different features which vary from each other considering from its
shape, size, texture, and color. The image of the leaf is captured from each curve, and
the dataset of 50,000 images has been prepared for 50 plants (1000 images/plant)
of South Gujarat. Pre-trained convolutional neural network (CNN) is utilized as a
feature extractor, and logistic regression (LR) is used as a classifier for leaf classi-
fication. CNN is visualized using deconvolutional network (DN) to get an insight
into extracted leaf’s features. The process of leaf classification includes (1) dataset
preparation and (2) features are extracted using pre-trained CNN models. This paper
outlines six CNN models which are experimented, namely as Inception v4, Xception,
ResNet, InceptionResNetV2, DenseNet, and MobileNet. (3) Stated on the extracted
features, LR is trained and used for classification. (4) Finally used DN for visual-
izing features, which provides the insight into classification results. In experiments,
Xception confers the highest Rank-1 accuracy of 93.4% and ResNet withholds with
poor performance with 36.48% Rank-5 accuracy. MobileNet nearly takes 2.17 min
for training and classification which is the fastest. DenseNet is the slowest and takes
30.48 min for leaf classification. MobileNet is accurate and is relatively close to
Xception model and pursue of exceptionally fast execution. In our experiments, DN
claims the shape and texture of a leaf (diverse orders of venation) are significant
features for classification.
Keywords Convolutional neural network · Deconvolutional network · Feature

extractions · Leaves classification · Logistic regression
S. Naik (B) · H. Shah

Babu Madhav Institute of Information Technology, Uka Tarsadia, Surat, Gujrat 394350, India
e-mail: Sapan2307@gmail.com
64 S. Naik and H. Shah
1 Introduction
We live in a diverse world where there are indistinct trees. Trees and plants are the
residual of the ecosystem and which are the necessity of our existence and sustain-
ability. There are several plants which are not categorized in urban and rural streets.
Approximately, 20% of the world population can differentiate them and the rest of
us are not aware of the species of plants [1]. Hence, it is a laborious task to distin-
guish between them; we are going to bifurcate their leaves using computer vision
technology.
Nonetheless, we know some of the species of trees and plants; still, it is quite
challenging to differentiate similar kinds of leaves with precision. To maintain the
perfection of these leaves, all the direct, as well as indirect details relevant to the
leaves, should be acknowledged. All the minor details like the shape of the leaf, the
sub-branch of the leaf, its texture, size, color are important for the identification of
leaves [2].
Latterly, deep learning and CNN have availed a lot of popularity. Deep learning
extracts the features of the image and drastically reduces the error in image recog-
nition. CNN has been dominating in the field of image classification since long [3].
The influence of deep learning has been observed, when the team of Hinton won
the ImageNet in image classification competition [4]. Plenty of work was done on
the initial CNN model, and at the present moment, there are many modern CNN
architecture models available which have outstanding performance. Some of them
are Inception, ResNet, Xception, and MobileNet [5].
2 Related Work
By using Caffe model of deep convolutional network for implementation, an average

accuracy of 96.3% was accomplished. Resulting to this, 44 different species of plants
of Royal Botanic Gardens, Kew, England, had been classified using CNN [2] and 13
different types of plant diseases were identified out of healthy leaves [6], and visual-
ization technique based on the deconvolutional networks was utilized for visualizing
features. LeafNet—a plant identification system was developed [7] based on CNNs.
LeafSnap, Flavia, and Foliage datasets were used for training and evaluating CNN.
Results showed better performance compared to handcrafted customized systems.
For classification of the leaf, Life CLEF 2015 dataset was used [8]. CNN was
used with three architectural models, namely GoogLeNet, AlexNet, and VGGNet. It
attained an overall accuracy of 80% and an inverse rank score of 0.752. With the usage
of CNN, leaf classification was implemented, where they had clinched that the shape
is not a dominant feature for leaf, and it was more preferable to opt for different
orders of venation. Deconvolutional network was used to get insight into chosen
features. Plant identification from leaf vein patterns using CNN was proposed in [9]
three different legume species: white bean, red bean, and soybean had been classified
Classification of Leaves Using Convolutional Neural Network … 65
for the experiment. Fine-grained plant classification system submitted to LifeCLEF

2014 was presented in [10] where CNN was used, and they had achieved the score
of 0.249. Forming of the complication in plant classification, digital morphome-
tric techniques were reviewed in [11]. LeafSnap—a mobile application for plant
species identification was developed using computer vision algorithms in [12]. It
used the traditional method of image capturing, segmentation, feature extraction,
and classification using available dataset. Dataset contains samples of 184 trees in
the Northeastern USA.
Dataset was assembled in a native environment for the classification of plants
where 10,000 images of 100 ornamental plant species were captured using the mobile
phone [13] in Beijing Forestry University campus. Deep learning approach was
used and achieved a recognition rate of 91.78% on the BJFU100 dataset. A tree
classification system treelogy was established [14], where CNN and SVM were used
to classify trees based on their leaves. Dataset of 57 trees with 5408 leaf images was
prepared. Diseased leaf segmentation and recognition method were proposed in [15]
using the fusion of superpixel, K-means, and pyramid of histograms of orientation
gradients (PHOG) algorithm.
To increase the dataset size, six augmentation techniques were applied on three
leaf datasets, namely Folio, AgrilPlant, and the Swedish in [16]. CNN was used to
implement leaf classification task using AlexNet and GoogleNet. Seven new invari-
ants (translation, rotation, and scaling invariants) for multi-component shapes were
proposed in [17] for leaf classification issue. New invariants were robust to noise and
mild deformations.
2.1 Contributions
A dataset with 50,000 images of leaves has been created for 50 different plants of
South Gujarat, and the approach is presented for classification of leaves. CNN is
used for feature extraction process with six architecture models, namely Inception
v4, Xception, InceptionResNetV2, DenseNet, ResNet50, and MobileNet and to get
its insight and extracted features (based on which leaves are categorized); DN is used.
LR classifier is utilized for training and classification purpose. All the outcomes of
the experiment have been compared with available work.
3 Materials and Methods
The preparation of dataset is discussed with the data augmentation method used. Here
in this division, how to train, tune, and implement the CNN model is briefed. Later,
short details of CNN architecture models are provided. At the end of this division,
LR and DN are briefly discussed.
3.1 Dataset Preparation and Data Augmentation
The dataset of 50,000 images was befallen to categorize the leaves. For preparing
the database, fifty different plants of South Gujarat were selected. Indiscriminating
leaves, 100 s of them were gathered concerning each plant. Each of them is deter-
mined and verified by the native citizens (aged above 48 years) living in the vicinity
of a particular plant. Total 1000 images of leaves are captured from 100 leaves (ten
images of each leaf with different positions). For dataset preparation, iPhone 6 s
is used to capture images. All the images are captured from top view (putting leaf
beneath the camera) in daylight. White paper is used as background for capturing
all images to reduce segmentation issue. The size (resolution) of captured images is
2448 × 2448 pixels. Images are resized to 224 × 224 and 299 × 299 based on CNN
model. The two beneficial reasons for resizing are 1. CNN needs fixed-size input
images and 2. to reduce computational time.
For reducing overfitting problem, dataset’s size is increased by applying posi-
tion shifting, rotation, and scaling. Even brightness and contrast corrections are
performed, for increasing size of dataset. As CNN does not require any prepro-
cessing on images except resizing, no other modifications have been performed on
the dataset. We have applied data augmentation to 200 randomly selected images of
each plant and created 400 new images out of it for each plant. In this way, we have
enlarged the dataset to 70,000 images.
More detail of dataset (i.e., name of each plant whose leaf is considered for
preparing dataset with one sample leaf image) is available at https://drive.google.
com/open?id=1iN4QLGFYD2AHLMWx-FlcPoLczkp__lqw.
Dataset samples (100 images of each leaf) are available at https://drive.google.
com/open?id=1No7P5M7gwaCQC–yEZuph0xpKDXazMua.
3.2 Overview of Convolution Neural Network
Basic CNN architecture contains four layers, namely convolution, nonlinearity,

pooling, and fully connected. Each layer may occur multiple times in architecture.
The number of each layer depends on the structure/model under use.
3.2.1 Convolution Layer
It is the uppermost layer of CNN which inputs an image. In this layer, filter (kernel or
neuron) is applied to the image. Normally, filter size is 3 × 3×depth, where the depth
of input image and filter is the same. Values of filter are multiplied with the pixel value
of the image, and new array is created as output. This operation is called convolving,
and output of convolving is two-dimensional array which is called feature map or
activation map [18].
Fig. 1 Operation in convolution layer a 3 * 3 filter representation b convolving operation when

feature is available c convolving operation when feature is not available
Various filters are practiced over convolution layer where individual filter repre-
sents various features of images like color, curve, or intensity. To spectacle the filters,
convolving operation and shape detector filter are shown in Fig. 1.
3.2.2 Nonlinearity Layer
Different activation functions are applied to this layer. Some of these activation
functions are Relu, sigmoid, and tanh. Relu is more preferable because it speeds up
the training process.
3.2.3 Pooling Layer
Pooling layer is used for reducing the size of the activation map. Pooling reduces
the chances of overfitting due to less parameter. Commonly pooling is performed
with the size of 2 × 2. Different forms of pooling are max, mean, and median but
max pooling is more popular. The output of pooling layer is given for flattening step
where the two-dimensional array is converted into a single vector because this vector
goes as an input to an artificial neural network.
3.2.4 Fully Connected Layer
This layer comprises neural network where each of the neurons is connected with
the neighbor layer’s neuron. Here, this layer sets the weight and bias.
3.3 Tuning the CNN Model
Tuning normally contains three phases, i.e., training, validating, and testing [3]. In
the training phase, model is trained using dataset. Later on, this model gets validated,
and finally, it gets tested in the testing phase using new input for the same dataset.
Tuning is also associated with choosing the right architecture for CNN or
designing the new one [3]. Various architectures are available nowadays like AlexNet,
GoogleNet, InceptionResNet, VGG, etc. It is preferable to utilize standard network
architecture in the beginning. In tuning CNN, one needs to decide different param-
eters of the model like weight, biases, epoch, learning rate, validation, and testing
percentage. Normally, backward propagation method is used to set these parameters.
Once training gets completed, all the parameters are saved in a binary file known as
a model.
When new image is under consideration, the saved model is loaded in the same
network architecture and prospect of the new image is calculated. The process is
termed as inference or prediction. During training, we split the dataset into batches,
and we give training in rounds or iterations, which is called epoch.
3.4 Implementation of CNN
There is a technique called transfer learning, which reduces lot of work by using a
fully trained model, already trained on dataset like ImageNet. The existing weights
for new classes will be retrained. This can be run in less time on the laptop, without
requiring a GPU. Time taken depends on the CNN model under consideration.
Though it is not as good as full training but is unexpectedly effective.
3.5 CNN Architecture Models
Many architecture models are available for implementing CNN. Here, we have briefly
discussed VGGNet, Inception, Xception, ResNet50, DenseNet, and MobileNet.
3.5.1 VGGNet
It was introduced in 2014 by Simonyan and Zisserman [19]. There are two versions
of VGGNet, namely VGG16 and VGG19. Here 16 and 19 represent weight layers
in the network. The networks are stagnant for training purpose, and in the context
of memory, the weight is wider [20]. It has ImageNet top 5 errors as 8.0%, and it
considers 138 million parameters [21].
3.5.2 Inception
It is micro-architecture and was first proposed by Szegedy et al. in [22] and later
in [23]. Initially, it was termed GoogLeNet. The weight of Inception is smaller
compared to VGGNet and requires 96 MB of memory space [20]. Inception uses
batch normalization, image distribution, and RMSProp methods. It has ImageNet
top 5 errors as 5.6%, and it uses 25 million parameters [24]. The latest version of
this model is Inception v4.
3.5.3 ResNet
It is exotic architecture which depends on micro-architecture modules [20]. It was

first proposed by He et al. in [25]. It introduces residual connections, and the model
size is 102 MB for ResNet50. ResNet has ImageNet top 5 errors as 4.5% and uses
60 million parameters [24]. It uses max as well as average pooling.
3.5.4 Xception
It was presented by [26]. Weight for this model is 91 MB. In experiments, it slightly
outperforms the Inception v3 model.
3.5.5 DenseNet
It was introduced by [27] which connects different layers in a feed-forward fashion.

The advantages of DenseNet are that it strengthens feature propagation, alleviate the
vanishing gradient problem, encourage feature reuse, and substantially reduce the
number of parameters.
3.5.6 MobileNet
As the name suggests, it is mainly used for mobile and embedded vision applications.
It was proposed by [28]. It is faster in training as well as in execution.
3.6 Logistic Regression
In simple linear regression (Eq. 1), y is replaced with sigmoid function (Eq. 2), so
LR takes the form of Eq. 3.
y = b0 + b1x (1)
p = 1/(1 + e∧ − y) (2)
ln( p/1 − p) = b0 + b1x (3)
The reason behind choosing LR is implementation is easy and the availability of

different tools. It provides probability scores for observations which are needed in
our case [5].
3.7 Deconvolutional Network
DN is used to understand the effect of each filter applied during convolution layer.
It facilitates us to visualize the internal working of CNN and how and based on
which features it actually categorizes the leaves. The multilayered deconvolutional
network was introduced by [29]. This method helps us to understand the function
of an individual neuron. Here, feature maps are projected back to its original pixels.
All the layers from starting are deconvolved and unpooled for full input pixel space.
4 Results and Discussion
Experiments are performed on the MacBook Pro (13-inch, mid-2012) machine. The
machine has 2.5 GHz Intel Core i5 processor, 10 GB 1333 MHz DDR3 memory,
and Intel HD Graphics 4000 1536 MB graphics card running on macOS High Sierra
(version 10.13.6). Keras and TensorFlow libraries are used for the implementation
of CNN, LR, and DN. Implementation needs following simple steps.
Step 1. The training image dataset is prepared with respective labels.
Step 2. Parameters are set in the configuration file.
Step 3. Features are extracted from final fully connected layers of pre-trained CNN
and are stored. Using the transfer learning technique, the top layer of the
CNN model is retrained [3]. Our dataset of leaves is used for retraining
purpose.
Step 4. Machine learning model is trained for extracted features and labels of step
3.
Step 5. The trained model is evaluated and optimized if needed.
Step 6. Visualization of CNN is performed using DN.
Fifty folders (one for each leaf category) are created and labeled with leaf’s name
for extracting features through CNN (CNN takes each category folder as input).
Weights of ImageNet dataset are used for pre-trained CNN model (as transfer learning
is used).
To analyze, we randomly selected 200 images of leaves (total 10,000) from each
category to extract their features, and 20 images of each category (total 1000) are
taken into consideration to provide training to the classifier. For experiments, epoch
value is set to 1000, the learning rate is 0.01, training batch size is 100, and validation
percentage is 10.
Six CNN architecture models, namely Inception v4, Xception, DenseNet,
ResNet50, InceptionResNetV2, and MobileNet are tested. The motive behind
choosing these six models is good accuracy is achieved by these models and for
CNN performance comparison purpose [5].
After configuration, features are extracted and stored as HDF5 format locally.
It is important to check image size while extracting features. Image size of 224 ×
224 pixels is chosen for ResNet50 and MobileNet models, and 299 × 299 pixels
are chosen for Inception v4, Xception, and InceptionResNetV2 models. Different
models require different time for feature extraction process. Table 1 shows the time
required by each model to extract features.
After the feature extraction process is completed, features and labels are loaded.
For training these features and labels, we have implemented logistic regression and
Naïve Bayes classification models. We have compared the classification results of
LR and Naïve Bayes with Inception v4 model, where LR performs better compared
to Naïve Bayes. Due to this observation, LR is used as a classifier in all experiments.
Experiments results are shown in Table 2. It represents the accuracy achieved by
each CNN model as Rank-1 and Rank-5. It also contains execution time required to
train 20 images of each leaf (total 1000 images). If the image under consideration
is classified with a perfect label with the highest probability, it is called Rank-1
accuracy; and if it is classified within top five probability (may not be first but within
first five predicted labels), then it is called Rank-5 accuracy.
Table 1 Feature extraction time by CNN models

Model Time (minutes) Model Time (minutes)
Inception v4 242.55 ResNet50 170.89
Xception 259.09 InceptionResNetV2 341.74
DenseNet 496.14 MobileNet 99.22
Table 2 Rank-1 and Rank-5 accuracy with training execution time for CNN models
Rank-1 accuracy (%) Rank-5 accuracy (%) Time (minutes)
Inception v4 91.91 97.16 5.28
Xception 93.40 97.16 6.08
ResNet50 9.35 36.48 4.51
InceptionResNetV2 88.24 93.35 19.10
DenseNet 88.48 93.35 30.48
MobileNet 91.91 97.16 2.17
Table 3 Misclassification of leaves for CNN models

Inceptionv4 Xception ResNet50 InceptionResNetV2 DenseNet MobileNet
Hibiscus ✓ ✓
Khatumada ✓ ✓ ✓ ✓ ✓
Tecoma ✓ ✓ ✓ ✓ ✓ ✓
Amli ✓ ✓ ✓ ✓
Jackfruit ✓ ✓
Taggar ✓ ✓
Bamboo ✓ ✓
Undirkani ✓ ✓ ✓ ✓
In Table 3, misclassified leaves are listed with the CNN model which was unable
to predict them correctly. Tick mark in the box represents that given model has
predicted given leaf incorrectly. Results are shown for all CNN architecture models.
By observing Table 3 and confusion matrix of all models, we have concluded
that misclassification majorly happens with tecoma, amli, khatumada, and undirkani
category of leaves (shown in Fig. 2).
We have implemented DN as described in [29]. For implementing DN, TensorFlow
library is used in the back end and output images are generated in TensorBoard [30].
The visualization results are shown in Fig. 3. Considering the visualization, we can
state that contour of the leaf is important feature for classification (Fig. 3b and 3c, the
feature map is highlighted with yellow color). At the same time, venation of leaves
is also playing an important role in the classification process (Fig. 3b, c), the feature
map is highlighted with green color).
There are some standard datasets available to identify leaf classification, but we
have not utilized these datasets. We have mainly focused on leaves of South Gujarat
and created our own dataset. We have compared our leaves classification results with
[7]. Table 4 represents the comparison results.
Fig. 2 Sample of leaf categories which are misclassified

Fig. 3 Visualization at CNN (a); sample input image (b); visualization at Conv5 layer (c);
visualization after maxpool 4 layer
Table 4 Comparisons of
Method Dataset Top-1 accuracy (%)
leaves classification methods
Pierre Barre et al. LeafSnap, Flavia, 86.3, 97.9, 95.8
(2017) Foliage
Kadir (2014) Flavia, Foliage 97.2, 95.0
Kumar et al. LeafSnap 73.0
(2012)
Wu et al. (2007) Flavia 90.3
Our approach New for South 93.40
Gujarat
5 Conclusion
From experiments of leaves classification, we have concluded that highest Rank-

1 accuracy achieved is 93.4% by Xception model. The fastest model is MobileNet
which gives classification output (training and classification) in 2.17 min. The perfor-
mance of ResNet model is dramatically poor, and DenseNet is found the slowest in
our experiments. Performance of MobileNet is very close to Xception model in
terms of accuracy, and execution is exceptionally fast. Even the Rank-5 accuracy of
MobileNet is the highest with Xception and Inception v4 models. Our experiments
show that the majority of the misclassifications occur in tecoma, amli, khatumada,
and undirkani leaves. We have observed that shape (counter) is one of the most
important features for leaves classification, while venation also plays very vital role
for classification.
Acknowledgements Special thanks to 2013–2018 batch students of Babu Madhav Institute of

Information Technology for helping in dataset preparation. Help of Yash Rana, Purva Desai, and
Vidhi Choksi in implementation and content writing work is acknowledged by authors.
Conflict of Interest
The authors declare that there are no conflicts of interest.
References
1. W.S. Jeon, S.Y. Rhee, Plant leaf recognition using a convolution neural network. Int. J. Fuzzy
Logic Intell. Syst. 17, 26–34 (2017)
2. S.H. Lee, C.S. Chan, P. Wilkin, P. Wilkin, Deep-plant: plant identification with convolutional
neural networks, in 2015 IEEE International Conference on Image Processing (ICIP) (2015,
September), pp. 452–456
3. A. Sachan, Tensor flow Tutorial 2: Image Classifier Using Convolutional Neural Network
[online]. Available cv-tricks.com/tensorflow-tutorial/training-convolutional-neural-network-
for-image-classification/. Accessed on 17 June 2018
4. A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional
neural networks, in Advances in Neural Information Processing Systems (2012), pp. 1097–1105
5. G. Ilango, Using Keras Pre-trained Deep Learning models for your own dataset. [Online]. Avail-
able https://gogul09.github.io/software/flower-recognition-deep-learning. Accessed on 15 July
2018
6. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture
for computer vision, in Proceedings of the IEEE conference on computer vision and pattern
recognition (2016), pp. 2818–2826
7. P. Barré, B.C. Stöver, K.F. Müller, V. Steinhage, LeafNet: a computer vision system for
automatic plant species identification. Ecological Inform. 40, 50–56 (2017)
8. M.M. Ghazi, B. Yanikoglu, E. Aptoula, Plant identification using deep neural networks via
optimization of transfer learning parameters. Neurocomputing 235, 228–235 (2017)
9. L.C. Uzal, G.L. Grinblat, R. Namías, M.G. Larese, J.S. Bianchi, E.N. Morandi, P.M. Granitto,
Seed-per-pod estimation for plant breeding using deep learning. Comput. Electron. Agric. 150,
196–204 (2018)
10. N. Sünderhauf, C. McCool, B. Upcroft, T. Perez, Fine-grained plant classification using
convolutional neural networks for feature extraction, in CLEF (2014), pp. 756–762
11. J.S. Cope, D. Corney, J.Y. Clark, P. Remagnino, P. Wilkin, Plant species identification using
digital morphometrics: a review. Expert Syst. Appl. 39(8), 7562–7573 (2012)
12. N. Kumar, P.N. Belhumeur, A. Biswas, D.W. Jacobs, W.J. Kress, I.C. Lopez, J.V. Soares,
Leafsnap: a computer vision system for automatic plant species identification, in Computer
vision–ECCV 2012 (2012), pp. 502–516
13. Y. Sun, Y. Liu, G. Wang, H. Wang, Deep learning for plant identification in natural environment.
Comput. Intell. Neurosci. (2017)
14. İ. Çuğu, E. Şener, Ç. Erciyes, B. Balcı, E. Akın, I. Önal, A.O. Akyüz, Treelogy: A Novel Tree
Classifier Utilizing Deep and Hand-crafted Representations (2017). arXiv:1701.08291
15. S. Zhang, H. Wang, W. Huang, Z. You, Plant diseased leaf segmentation and recognition by
fusion of superpixel, K-means and PHOG. Optik-Int. J. Light Electron Opt. 157, 866–872
(2018)
16. P. Pawara, E. Okafor, L. Schomaker, M. Wiering, Data augmentation for plant classifica-
tion, in International Conference on Advanced Concepts for Intelligent Vision Systems (2017,
September), pp. 615–626
17. M.B.H. Rhouma, J. Žunić, M.C. Younis, Moment invariants for multi-component shapes with
applications to leaf classification. Comput. Electron. Agric. 142, 326–337 (2017)
18. M.A. Nielsen, Neural networks and deep learning (2015), [Online]. Available http://neuralnet
worksanddeeplearning.Com (2016)
19. K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image
recognition (2014). arXiv preprint arXiv:1409.1556
20. A. Rosebrock, ImageNet: VGGNet, ResNet, Inception, and Xception with Keras
[online]. Available https://www.pyimagesearch.com/2017/03/20/imagenet-vggnet-resnet-inc
eption-xception-keras/. Accessed on 18 Aug 2018
21. Coursera, Overview of modern CNN architectures [online]. Available https://www.coursera.
org/lecture/intro-to-deep-learning/overview-of-modern-cnn-architectures-U0qGv. Accessed
on 18 Aug 2018
22. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke,

A. Rabinovich, Going deeper with convolutions, in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (2015), pp. 1–9
23. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture
for computer vision, in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition (2016), pp. 2818–2826
24. S.H. Lee, C.S. Chan, S.J. Mayo, P. Remagnino, How deep learning extracts and learns leaf
features for plant classification. Pattern Recogn. 71, 1–13 (2017)
25. K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition (2016), pp. 770–778
26. F. Chollet, Xception: Deep learning with depthwise separable convolutions (2017). arXiv
preprint, 1610 – 02357
27. G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected convolutional
networks, in CVPR (2017, July), pp. 1, 3
28. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H.
Adam, Mobilenets: efficient convolutional neural networks for mobile vision applications
(2017). arXiv preprint arXiv:1704.04861
29. M.D. Zeiler, R. Fergus, Visualizing and understanding convolutional networks, in European
Conference on Computer Vision (2014, September), 818–833
30. B. Vikani, F. Shah, CNN Visualization, tf_cnnvis, 2017, [online]. Available at https://github.
com/InFoCusp/tf_cnnvis. Accessed on 27 Aug 2018
A Low-Cost IoT-Enabled Device
for Real-Time Air Quality Monitoring
System
Sai Surya Kiran Pokala, V. Panchami, Kelavath Jaisingh,

and Sandeep Kumar Yedla
Abstract Air pollution is a major problem in many metropolitan areas. It causes

health problems for the residents of the affected areas. Air pollution is due to the
increase in vehicles, industries, burning of fossil fuels, and so forth. This project aims
to use the wireless sensor networks (WSN) for monitoring the gaseous pollutants such
as Sulfur dioxide, chlorine, carbon monoxide(CO), carbon dioxide(CO2 ), benzene,
alcohol, NH3 , NOx , smoke, and stannous oxide (SnO2) in metropolitan areas. The
proposed system involves the deployment of sensor nodes which detect the level of
harmful gases in the air, and the sensor data is transmitted to the cloud platform by
using the Wi-Fi module. Sensor data is stored and analyzed in the cloud platform.
The system will send email alerts to the users if the percentage of the harmful gases
is greater than the threshold value. It provides a simple, portable, fast, and user-
friendly interface for the clients to access information about air quality from the
cloud platform.
Keywords Air pollution · Wireless sensor network · Internet of things
1 Introduction
Nowadays due to the industrialization and urbanization, measuring air quality is

necessary to take preventive measures to reduce the air pollution. The levels of
harmful gases such as sulfur dioxide, carbon monoxide, stannous oxide, carbon
S. S. K. Pokala (B) · V. Panchami · K. Jaisingh · S. K. Yedla

Indian Institute of Information Technology, Kottayam, Kerala, India
e-mail: saisuryapokala6@gmail.com
V. Panchami
e-mail: panchamam036@iiitkottayam.ac.in
K. Jaisingh
e-mail: kelavath16@iiitkottayam.ac.in
S. K. Yedla
e-mail: sandeep16@iiitkottayam.ac.in
78 S. S. K. Pokala et al.
dioxide, ammonia, smoke, and nitrogen oxides and chlorine are increasing day by
day. In India, the common people [1] in these urban areas are facing major health
issues because of the air quality dips, so there is a need for having an air quality
monitoring system to assess the levels of air pollution [2]. The Internet of things is
one of the smart solutions to address this problem. In order to monitor the air quality
level in the various locations in various time intervals, this device can be used. The
levels of harmful gases can be measured by using the cost-effective wireless sensor
networks [3]. The deployment of the IoT-enabled system in local areas can be able
to sense the levels of CO, NH3, etc. [4]. This system uses the application software
for the data analytics in the cloud and a microcontroller to read the inputs from the
sensor. This project aims to create awareness to the public about the air pollution.
The main objectives of the proposed system are:
• To create an IoT device to monitor the quality of air.
• To analyze the air quality data and provide information to the clients.
• To evaluate the quality of air by comparing the level of the harmful gases with
the threshold value.
• To facilitate the users by notifying them with an email if the levels of harmful
gases are greater than the threshold.
The paper is organized in such a way that Sect. 2 deals with the literature survey;
Sect. 3 lists hardware requirements of the system; Sect. 4 describes the proposed
system; Sect. 4 shows the implementation details of the IoT-enabled system; and last
Sect. 5 deals with the analysis results of real-time data which is stored in the cloud.
2 Literature Survey
According to the survey conducted by one of the institute in New Delhi over an area
to 1 km from the source, they have found that the amount of harmful gases found that
atmosphere had a median size of 10,000 PPM in industrial areas. The percentage of
these harmful gases is too high when compared to the optimal value which is around
400–600 PPM. There are a lot of research papers which address the problem.
Kumar et al. [5] in this paper propose an IoT device for monitoring the air quality
in urban areas using the sensors such as PM 2.5, carbon monoxide, carbon dioxide,
temperature, humidity, and air pressure. Sensor data is collected from the sensors
using Raspberry pi, and the data is stored in a IBM Bluemix cloud platform. The
paper proposed by Fioccola et al. [6] provides a solution for calculating the air quality
efficiently by using wireless sensor networks. Energy-efficient approach is adapted
in this system to reduce the power consumption of the IoT device. A SQL database is
created to store the data received from the sensors. Firdhous et al. [7] have studied the
pollutant gases present in the atmosphere. The experiments resulted in an increase
of ozone concentrations near the pollutant areas. For every five minutes, sensor data
is transmitted to the cloud using Bluetooth connection. This paper Nasution et al.
[8] intends to measure the levels of ozone and pollutant gases such as sulfur dioxide,
A Low-Cost IoT-Enabled Device for Real-Time Air Quality … 79
hydrogen sulfide, nitrogen oxides, and chlorine in the air. A web page is designed to
show the data stored in ThingSpeak cloud platform. The system is not cost-effective.
Zheng et al. [9] in this paper propose a three-tier architecture for the IoT device, and
sensors that detect the levels of harmful gases transmit the data to the cloud platform
by using ESP8266. The retrieved data is stored and analyzed in the cloud platform.
3 Experimental Setup for the Proposed System
3.1 Hardware Requirements
In this paper, we propose an IoT-enabled monitoring system in which one of the

sensors used is MQ135 [10]. The sensors are able to sense ammonia, alcohol,
benzene, smoke, carbon dioxide. It is a highly sensitive sensor and has a very simple
circuit. It is able to work in both analog and digital systems with a range of 0–5 V,
since it can be connected to any type of microcontroller. In this Fig. 1, the initial
resistance of the sensor is represented by R0 where R0 is the sensor resistance in the
clear air. The value of the resistance in the clear air is high when there is an increase
in the air pollution the resistance value decreases and the varied sensitivity is noted
using a potentiometer. The formula for the resistance of the sensor denoted as Rs is
as follows:
Fig. 1 Sensitivity characteristics [17] of the sensor MQ135


Vc
Rs = − 1 ∗ RL (1)
Vout
where
R L 20 k,
Vout Analog voltage of the sensor,
Vc Circuit voltage
MQ-7 sensor detects the carbon monoxide in the atmosphere [11]. The MQ-
8 sensor is more sensitive to hydrogen in nature [12]. It is a low-cost sensor for
detecting hydrogen gas in the atmosphere. These sensors are cost-effective and can
be used in outdoor environments.
3.2 ThingSpeak
ThingSpeak is a cloud-based platform for IoT-enabled devices. It is useful to store

the real-time data from different sensors through different channels and devices
[13]. ThingSpeak is also used for real-time data processing, data visualization and
analyzing real-time data streams in the cloud. Data is collected from the sensors
using the Wi-Fi module and it is visualized in graphical form. Connection setup for
connecting ESP8266 and the Arduino follows in such a way that the grounds (GNDs)
of both Arduino and ESP8266 are connected. TX (serial Transmit pin) in ESP8266 is
used to upload the program to the Arduino, and it is connected to D3 in the Arduino
UNO. RX (serial Receive pin) in ESP8266 acts as general-purpose input/output pin,
and it is connected to D2 in Arduino UNO. VCC in EP8266 is connected to the power
supply of 3.3 V in the Arduino UNO. RESET pin in ESP8266 resets the module.
ESP8266 can be programmed by using AT commands in the Arduino IDE to connect
to ThingSpeak.
4 Proposed System
Overview of the proposed system is shown in Fig. 2 which is designed to monitor

the quality of air. The major components involved in the system are Arduino Uno,
Wi-Fi module, and sensors. Sensors connected to Arduino Uno sense the levels of
hazardous gases in the air [14]. The proposed system provides an user interface to
the clients, and it sends an alert to the users when the levels of pollutant gases are
increased. Wi-Fi module acts as a platform to send the received data to the cloud
[15]. ThingSpeak stores large amounts of data and visualizes the data in graphical
form [16]. The API key is used to connect the Wi-Fi module with the cloud platform.
Data is received every 17 s. Results are monitored and accessed remotely from web
Fig. 2 Block diagram for IoT-enabled air quality monitoring system
pages. The monitoring results can be viewed on a web page provided by ThingSpeak
cloud service. Data can also be converted into excel data, XML or Json format.
Figure 3 and algorithm show the process involved in the cloud platform to monitor
the quality of air. In the proposed system, we are using 3 sensors denoted as S1 , S2 , S3 .
Once the data gets stored in the cloud platform, then the stream data from the 3 sensors
are exported into an excel sheet which is commonly known as a datasheet. Figure 5
shows the datasheet, in which it has 4 columns. They are timestamp, entry id, and
the stream data from the 3 sensors. The stream data from the sensors S1 , S2 , S3 are
represented as C1 , C2 , C3 . The average of C1 , C2 , C3 is taken and stored as V1 , V2 , V3 .
Threshold [17] is fixed for each sensor for assessing the quality of air. The threshold
of S1 is 200 ppm, S2 is 300 ppm, and the threshold of S3 is 400 ppm. The average
of C1 , C2 , C3 each sensor, V1 , V2 , V3 for 6 h is calculated, and the delay is set as
17 s. If the calculated average is greater than the fixed threshold value, then the
email notification is sent to the users. Regular analysis of the results retrieved by
ThingSpeak allows distinguishing how bad air pollution is from day to day [18].
Algorithm: Air Quality Monitoring System

Input: The values from the 3 sensors S1 ← MQ-7, S2 ← MQ-135, S3 ←
MQ-8
Output: Email alert if the values are greater than the threshold.
1. Start
2. Read the values from the sensors S1 , S2 , S3 .
3. Print Stream Datasheet
4. Initialize the 3 columns of Stream Datasheet as C1 , C2 , C3
where C1 ← STREAM_DATA(S1 )
C2 ← STREAM_DATA(S2 )
C3 ← STREAM_DATA(S3 )
5. Take the average of all the values of C1 , C2 , C3
V1 ← average(C1 )
V2 ← average(C2 )
V3 ← average(C3 )
6. Set time as 6 h; Delay as 17 s;

Number of columns retrieved 1270
Threshold(S1 ) → 200 PPM

7. IF V1 > Threshold(S1 )
7.1. THEN GO TO 10
7.2. Else GO TO 7
8.1. THEN GO TO 10
8.2. Else GO TO 7
9.1. THEN GO TO 10
9.2. Else GO TO 7
10. SEND EMAIL ALERT TO USERS
11. IF power is ON
11.1. THEN GO TO 2. else 12
12. Device is in OFF State
13. Stop
5 Implementation
The sensors such as MQ135, MQ7, and MQ8 are connected with the Arduino Uno kit.
The sensors sense the concentration of the harmful gases. The output of the sensors
is an analog voltage in volts. The Arduino Uno will convert the analog voltage into
Fig. 3 Flowchart for

IoT-enabled air quality
monitoring system
Fig. 4 Hardware implementation of the air quality monitoring system
digital data. The ESP8266 is a Wi-Fi module which is attached to the Arduino Uno
kit to transmit the sensor data to the cloud. The Twilio [18] cloud communication
platform acts as a service for sending SMS alerts to the users. Figure 4 shows the
hardware implementation for monitoring the quality of air.
6 Results
The data from the Wi-Fi channel is stored in the Cloud. Users should connect to the
cloud to retrieve the data from the cloud. The sensor data is analyzed as a result, and a
datasheet is being generated. Figures 5 and 6 show the graphical representation, and
Fig. 7 shows the datasheet obtained from the ThingSpeak. The sensor data receiving
from the sensors is represented in graphical form by using the ThingSpeak. If the
values from all 3 sensors are optimal, then the quality of air in the atmosphere is
good.
Fig. 5 Real-time data analysis of the MQ-135 and MQ-8 sensors
Fig. 6 Real-time data

analysis of the MQ-7 sensor
7 Conclusion
The proposed system, IoT-enabled air quality monitoring system is successfully

implemented and checked in various areas. The integration of the system with the
cloud service ThingSpeak makes the system smart for doing the alert process fast,
cost-effective, real time, and flexible. ThingSspeak is used to evaluate the observed
stream data from the sensors. AT commands of ESP8266 which is used to send data
to ThingSpeak are used for client and server communication, to set up the TCP/UDP
connection with the host and port, to find the status of the connection, for setting the
length of the data, an API key for sending the data to ThingSpeak channel, to set the
Wi-Fi mode, to enable multiple connection, and to close the TCP/UDP connection.
This implemented system sends email message alerts to the users depending on the
threshold level. It also stores large amounts of data in the cloud platform. The device
can serve many users who are living near industrial areas and in high traffic areas. The
proposed system will monitor, measure, and alert the amount of harmful gases in the
air, thereby saving mankind and the next generation. The proposed system is easily
portable and scalable to any kind of monitoring system with small modifications.
Fig. 7 ThingSpeak datasheet for the 3 sensors connected to system

In the future, it is better to add artificial intelligence for predicting the air quality
using optimal machine learning algorithms or neural networks or rule-based expert
systems. Thus, we are able to diagnose, forecast, plan, and control the air pollution
especially in metropolitan cities
References
1. J. Kang, K.I. Hwang, A comprehensive real-time indoor air-quality level indicator. Sustain-
ability 8(9), 881 (2016)
2. S. Devarakonda, P. Sevusu, H. Liu, R. Liu, L. Iftode, Nath, B. (2013, August). Real-time air
quality monitoring through mobile sensing in metropolitan areas, in Proceedings of the 2nd
ACM SIGKDD International Workshop on Urban Computing (pp. 1–8)
3. V. Laijawala, M., Masurkar, R. Khandekar, Air Quality Monitoring System (2019). Available
at SSRN 3454389
4. C.V. Saikumar, M. Reji, P.C. Kishoreraja, IOT based air quality monitoring system. Int. J. Pure
Appl. Math. 117(9), 53–55 (2017)
5. S. Kumar, A. Jasuja, Air quality monitoring system based on IoT using Raspberry Pi, in 2017
International Conference on Computing, Communication and Automation (ICCCA). IEEE
(2017)
6. G.B. Fioccola, R. Sommese, I. Tufano, R. Canonico, G. Ventre, Polluino: an efficient cloud-
based management of IoT devices for air quality monitoring, in 2016 IEEE 2nd International
Forum on Research and Technologies for Society and Industry Leveraging a better tomorrow
(RTSI). IEEE (2016, September), (pp. 1–6)
7. M.F.M. Firdhous, B.H. Sudantha, P.M. Karunaratne, IoT enabled proactive indoor air quality
monitoring systems for sustainable health management, in 2017 2nd International Conference
on Computing and Communications Technologies (ICCCT). IEEE (2017, February), (pp. 216–
221)
8. T.H. Nasution, M.A. Muchtar, A. Simon, Designing an IoT-based air quality monitoring system,
in IOP Conference Series: Materials Science and Engineering (IOP Publishing, vol. 648, No.
1, 2019, October), p. 012037
9. K. Zheng, S. Zhao, Z. Yang, X. Xiong, W. Xiang, Design and implementation of LPWA-based
air quality monitoring system. IEEEAccess, 4, 3238–3245 (2016)
10. https://www.olimex.com/Products/Components/Sensors/Gas/SNS-MQ135/resources/SNS-
MQ135.pdf
11. https://www.sparkfun.com/datasheets/Sensors/Biometric/MQ-7.pdf
12. https://cdn.sparkfun.com/datasheets/Sensors/Biometric/MQ-8%20Ver1.3%20-%20Manual.
pdf
13. https://www.mathworks.com/help/thingspeak/
14. M. Benammar, A. Abdaoui, S.H. Ahmad, F. Touati, A. Kadri, A modular IoT platform for
real-time indoor air quality monitoring. Sensors 18(2), 581 (2018)
15. H. Kumar, S. Karthikeyan, G.M. Manjunath, M.S. Rakshith, Solar power environmental air
pollution & water quality monitoring system based on IoT. Pers. Commun. Embed. Syst. Sign.
Process. PiCES 2(8), 197–199 (2018)
16. K.S.E. Phala, A. Kumar, G.P. Hancke, Air quality monitoring system based on ISO/IEC/IEEE
21451 standards. IEEE Sens. J. 16(12), 5037–5045 (2016)
17. C. Xiaojun, L. Xianpeng, X. Peng, IOT-based air pollution monitoring and forecasting system,
in 2015 International Conference on Computer and Computational Sciences (ICCCS), Noida,
(2015), pp. 257–260
18. H.F. Hawari, A.A. Zainal, M.R. Ahmad, Development of real time internet of things (IoT)
based air quality monitoring system. Indonesian J. Electr. Eng. Comput. Sci. (IJEECS) 13(3),
1039–1047 (2019)
Fuzzy-Based Predictive Analytics
for Early Detection
of Disease—A Machine Learning
Approach
V. Kakulapati , R. Sai Sandeep, V. Kranthi kumar, and R. Ramanjinailu
Abstract Influenza is significant promising and re-emerging contagious disease,

causing high morbidity and mortality throughout the world. Influenza also called
swine flu, hoard flu, pig flu. The primary task of healthcare providers is providing
diagnostic product services at low costs and to diagnose patients accurately. The
use of data analytics and machine learning is a breakthrough technology that can
have a significant impact on the healthcare field. Machine learning methods can be
used for disease identification because they mainly apply to data themselves and
give priority to outcomes of specific tasks. In this work, a framework of classifica-
tion and regression trees (CART) algorithm is to create the fuzzy rules to employ
for improved disease prediction. Validate the proposed method by choosing public
medical datasets. Results on swine flu data demonstrate that our method distin-
guished the enhancement of the disease prediction accuracy. The implementation
results exhibit that the aggregation of two methods such as fuzzy rule-based CART
algorithm with irrelevant data elimination methods that could be useful in predicting
disease. The proposed system can be helpful for healthcare providers as a healthcare
analytical technique.
Keywords CART · Fuzzy · Prediction · Accuracy · Technique · Rules
V. Kakulapati (B) · R. Sai Sandeep · V. Kranthi kumar · R. Ramanjinailu

Sreenidhi Institute of Science and Technology, Yamnampet, Ghatkesar, Hyderabad, Telangana,
India 501301
e-mail: vldms@yahoo.com
R. Sai Sandeep
e-mail: sandeepu2110@gmail.com
V. Kranthi kumar
e-mail: kranthivankadaru@gmail.com
R. Ramanjinailu
e-mail: ramu.ramanjinailu@gmail.com
90 V. Kakulapati et al.
1 Introduction
Influenza, also known as swine flu, is a communicable virus that affects pigs. This
virus identified as a pandemic in 1919, and still, it is a seasonal virus though this
virus cannot affect humans and become a worldwide outbreak in 2009. This flu is
first started in pigs and caused by virus strain called as H1N1, and this virus is a
grouping of swine, bird flu, and human inheritable factors that assortment collected
in greedy pigs then transmit to the humans [1]. This H1N1 was identified as a usual
type of seasonal flu. In the year 2009, nearly 62 million humans affected by swine
flu and above 12.5 k people [2] died due to this pandemic—report by Centers for
Disease Control and Prevention (CDC) and nearly 5.8 lakh people died due to this
virus throughout the world [3].
Influenza also caused by the H1N1 virus, which also called Spanish flu and genetic
factors, illustrates that this virus advanced from swine flu or bird flu. The number of
people died due to the H1N1 pandemic throughout the world, and the fatality rate is
higher in adults [4].
Analysis of electronic health records contains patient privacy information, which
requires to be accurate for prediction, identification and diagnosis of the diseases. For
accurate prediction, identification and diagnosis in an efficient manner, to investigate
and practice suitable ML cataloguing algorithms that appropriately detect and diag-
nose disease. Nowadays, massive resolution in the area of ML methods due to the
vast usage of computational techniques in the analysis of electronic health records
precisely and efficiently. For better diagnosis methods, which help physicians appro-
priately diagnose the disease, they need to investigate reliable models to address
the clinical trials. Many healthcare datasets require pre-processing before building
a model as these contain noisy, duplicate and inappropriate data, which may reduce
the accuracy of classification. The accuracy of disease prediction is based on the
data and the classification models. One of the significant tasks in machine learning is
classification [5–7] that retrieves facts from practical problems and develop a model,
which appropriately predicts the target class.
Fuzzy logic is a processing technique that compromises with inexactness and
incorrect information. It is a procedure of the crisp rules wherein the certainty vari-
ables lie between 0 and 1. The logic applies to perception based on the degree of
truth (1.0) or degree of false (0.0), and this logic utilized for completely imprecise
perceptions. These fuzzy rules are used in healthcare applications for better accurate
and efficient diagnosis. For identification of hidden and useful rules of observations
by using a decision tree, fuzzy logic is used in healthcare application, which handles
the diagnosis ambiguity and imprecision of symptoms [8].
In this work, propose a system for the early detection of swine flu, which uses
CART and fuzzy logic for classification. CART uses an implicit feature section
using different techniques, and fuzzy logic implemented for better classification and
improved accuracy.
Fuzzy-Based Predictive Analytics for Early Detection of Disease … 91
2 Related Work
Many researchers investigated the performance analysis between individual classifies

and ensemble classifiers [9]. Automated decision-making systems [10] are useful
to cooperating clinicians rapidly and diagnose enduring correctly within a short
time. This automated system utilizes healthcare data to attain practical consequences.
The health analysis method illustrates that how the cataloguing, regression tree, and
random Forest models assimilated to form a hybrid intellectual classification is called
as the fuzzy min-max network [11].
Intelligent swine flu prediction software [12] developed by utilizing naïve Bayes
algorithm for categorizing the swine flu enduring into three types; those are the
minimum possible, likely or maximum possible, and attained 63% accuracy. The
combination of decision tree and naïve Bayes algorithms was employed on the iden-
tical inputs to discover the genuine suspicious cases and predict the credible investi-
gation of a swine flu from the suspicious zone. Appropriate suspicious cases, naïve
Bayes correctly classified when compared with decision trees [13]. In [14], naïve
Bayes classifiers were detecting swine flu ailment from its warning signs and exploit
likely consequences.
Heart disease [15] prediction using fuzzy rules and blurring was acquainted with
data to remove uncertainties, and a membership function has been adding. Classifi-
cation and regression tree are a parametric classifier and more efficiently classified
than other parametric methods and consequently, removing the uncertainty in the
data by applying a fuzzy membership function.
In [16], an expert system based on fuzzy rule is developing for asthma-related
ailment treatment. In another method, a proficient method is with a combination of
different classification techniques such as neural network (nn), C5.0 DT (decision
tree) and linear discriminate analysis (LDA) for the cataloguing of dermatology
illness [17].
3 Framework
See Fig. 1.
3.1 Fuzzy Logic
It provides an induction morphology that empowers human intelligence abilities

to apply to decision support systems and is based on mathematical calculation for
better recognition of the uncertainties related to human intelligence and perceptive.
Fig. 1 Proposed framework of fuzzy-based classification and regression prediction of swine flu
Membership function (MF) is a curvature which defines in what manner every argu-
ment in the contribution recorded to a degree of membership between the values 0
and 1 [18].
3.2 Fuzzification
The lexical variables of the uncertainty intended in the procedure of uncertain sets
and where these variables are described as membership functions. The estimation
of this method of the degree of fit in fuzzy crisp rules is known as fuzzification.
The participation capacities might be triangular, trapezoidal, Gaussian, or ringer
moulded. As data about the level of the participation utilized for further development,
a significant amount of data may get failure in the process of fuzzification. This is
happened due to the method can perceive as a nonlinear revolution of the data. The
interpretation smears the aggregation regulations, utilizing the knowledge base and
thus membership functions to the terms of yield factors are determined, lastly, after
defuzzification, the yield result is acquired.
3.3 CART
It is one of the analytical systems used in machine learning and describes how a
destination variable can predict from other values. Consequently, the decision tree
where every node is a split in an interpreter variable and every node at the termination
has an estimation for the destination variable. In this instance, the swine flu test dataset
is classified appropriately.
3.4 Rule Induction
A set of IF-THEN rules represented as a learned model is called as rule induction,

which rule antecedent—which is the If part, and rule consequent—which is the
else part. The former contains attribute tests, and the consequent contains a class
prediction.
3.5 Membership Function
The fuzzy membership function is an interpretation of the pointer function in typical

sets. On account of triangular or trapezoidal membership, data lost in the provinces of
membership where the zero slope and the membership functions are not differentiable
at these points. Consequently, systems having triangular or trapezoidal membership
functions can come across difficulties of knowledge from information. To overcome
these problems, smoother membership functions such as Gaussian or bell function
may use.
3.6 Inference
The fuzzy input is changed into the fuzzy output with help of If-Then rules.
4 Implementation
For implementation purposes, the swine flu dataset was chosen from the UCI
machinery database and then, pre-processing the dataset for the removal of unnec-
essary data and irrelevant data and then splitting this dataset into training and test
datasets. Generated fuzzy crisp rules intend with the decision tree from which the
regulations can comprehend. Selecting random samples from the given dataset then
applying the CART algorithm to generate decision tree for each sample, which can
produce predicting results. In each step of algorithm, voting is accomplishment for
each projected consequence, and the maximum voted projected consequences as the
concluding prediction outcome.
Generate the frbs framework by training the model, and the attributes prepared
with the editor of membership function in the fuzzy editor. For easy to understand
and widely used in the different application membership function is Triangular. The
CART rules are taken in the regulation editor to methods the fuzzy rule base, and
these utilized for exhibiting the consequences. The regulation observer makes the
defuzzied yield, which can easily understand whether the persistent is affected with
swine flu or not.
Statistics and Confusion Matrix

Prediction 1 2
1 133 2
2 1 80
Accurateness: 0.9861
95% CI: (0.9599, 0.9971)
No Information Rate: 0.6204
Kappa: 0.9704
Mcnemar’s Test P-Value: 1
Sensitivity: 0.9925
Specificity: 0.9756
Pos PredValue: 0.9852
Neg PredValue: 0.9877
Prevalence: 0.6204
Recognition Rate: 0.6157
Recognition Incidence: 0.6250
Stability Accurateness: 0.9841
“Positive” Class: 1 (Table 1, Fig. 2).
The lexical terms on the contribution variables: [1] “small” “medium” “large”
“small” “medium” “large” “small” [8] “medium” “large” “small” “medium” “large”.
Membership function parameter values on the contribution variable:
Small medium large small medium large small medium large small medium large
[1,] 2.0 4.00 3.0 2.0 4.00 3.0 2.0 4.00 3.0 2.0 4.00 3.0
[2,] 0.0 0.23 0.6 0.0 0.23 0.6 0.0 0.23 0.6 0.0 0.23 0.6
[3,] 0.2 0.43 0.8 0.2 0.43 0.8 0.2 0.43 0.8 0.2 0.43 0.8
[4,] 0.4 0.53 1.0 0.4 0.53 1.0 0.4 0.53 1.0 0.4 0.53 1.0
[5,] NA 0.73 NA NA 0.73 NA NA 0.73 NA NA 0.73 NA
Table 1 Probabilities of swine flu symptoms
S. No. Split n Loss yval yprob
1 Root 64 24 2 (0.3750000 0.6250000)
2 Runny nose < 1.5 24 8 1 (0.6666667 0.3333333)
3 Runny nose ≥ 1.5 40 8 2 (0.2000000 0.8000000)
4 Headache < 3.5 16 0 1 (1.0000000 0.0000000)*
5 Headache ≥ 3.5 8 0 2 (0.0000000 1.0000000)*
6 Headache ≥ 3.5 15 7 1 (0.5333333 0.4666667)*
7 Headache < 3.5 25 0 2 (0.0000000 1.0000000)*
Fuzzy-Based Predictive Analytics for Early Detection of Disease …
* denotes prolong i.e., continuation

95
Fig. 2 Classification tree of swine flu symptoms
The number of linguistic terms on each variable

Chill Runny1se Headache Fever Swine flu
[1,] 3 3 3 3 2
The fuzzy IF-THEN rules:
V1 V2 V3 V4 V5 V6 V7 V8V9 V10 V11 V12 V13 V14 V15 V16
1 IF Chill is small and Runny1se is large and Headache is medium and Fever is
large
2 IF Chill is large and Runny1se is large and Headache is small and Fever is small
3 IF Chill is large and Runny1se is large and Headache is medium and Fever is
large
4 IF Chill is large and Runny1se is small and Headache is medium and Fever is
large
5 IF Chill is small and Runny1se is large and Headache is large and Fever is small
6 IF Chill is large and Runny1se is small and Headache is large and Fever is large
7 IF Chill is small and Runny1se is small and Headache is small and Fever is small
8 IF Chill is small and Runny1se is large and Headache is large and Fever is large
V17 V18 V19 V20
1 THEN Swineflu is 2
Fig. 3 Symptoms of swine flu disease
The certainty factor:

[1,] 1.25
[2,] 1.25
[3,] 1.25
[4,] 0.75
[5,] 0.75
[6,] 1.25
[7,] 0.75
[8,] 1.25
Estimating the percentage of error percentage to distinguish the accurateness of
the model.
R > err < −100 * sum (pred! = real. swinflu/length(pred) R > err [1] 0 (accuracy
= 100) (Figs. 3 and Fig. 4).
5 Performance Analysis
The classifier performance is subject to features data to be analysed. The performance

exhibits in the form of accuracy and the error rate. Several types of validation can
implement for the evaluation of classification performance. In the previous study,
implemented a random forest algorithm and attained 97% accuracy. In this work,
Fig. 4 Plot of membership functions
utilize fuzzy rule-based classification and regression tree with 98% accuracy with 0
classification error rate. The proposed system is exhibited efficiently predicting the
disease.
6 Conclusion
In this work, swine flu ailment is cataloguing wherein the contribution features fuzzi-
fied utilizing a lexical membership function to hold imprecise and uncertainty data.
The process of fuzzification makes the prototypical more complicated due to its fuzzy
growth that receipts training the model in more consuming time which increased
values of fuzzified are permitted to the model to retrieve the appropriate features
that make a substantial input. The proposed model demonstrates the classification of
swine flu disease prediction with 99% accuracy and 0 classification error rate, which
concludes that fuzzy logic and CART together give an efficient disease prediction.
7 Future Work
In the future, automation of swine flu prediction by using a large dataset and also
develops an intelligent framework based on similarity measures of swine flu symp-
toms. Depending on the similarity measures of patients and location-wise categorize
them by utilizing the clustering techniques. Predicting these swine flu patients and
provide recommendations to avoid the disease using deep learning and AI techniques.
In the future, develop new prediction techniques to enhance classification accuracy
and efficiency of swine flu patient privacy and provide a personalized recommen-
dation to the patient for better diagnosis by applying nature-inspired optimization
techniques.
References
1. T. Jilani et al., H1N1 influenza (swine flu). Stat Pearls. Updated December 14, 2019
2. Centers for Disease Control and Prevention. CDC Estimates of 2009 H1N1 Influenza Cases,
Hospitalizations and Deaths in the United States
3. F. Dawood et al., Estimated global mortality associated with the first 12 months of 2009
pandemic influenza A H1N1 virus circulation: a modelling study. Lancet. Infect. Dis 12(9),
687–695 (2012). https://doi.org/10.1016/s1473-3099(12)70121-4
4. D. Jordan, The deadliest flu: the complete story of the discovery and reconstruction of the 1918
pandemic virus. Centers for Disease Control and Prevention. Updated December 17, 2019
5. H. Das, A.K. et al., A novel PSO based back propagation learning-MLP (PSO-BP-MLP) for
classification computational intelligence in data mining (vol. 2, Springer, New Delhi, 2015),
pp. 461–471
6. R. Sahani et al., Classification of intrusion detection using data mining techniques Progress in
computing, analytics and networking (Springer, Singapore, 2018), pp. 753–764
7. H. Das et al., Classification of diabetes mellitus disease (DMD): a data mining (DM) approach
Progress in computing, analytics and networking (Springer, Singapore, 2018), pp. 539–549
8. V. Khatibi, et al., A fuzzy-evidential hybrid inference engine for coronary heart disease risk
assessment. Expert Syst. Appl. 37, 8536–8542 (2010)
9. C.-H. Weng et al., Disease prediction with different types of neural network classifiers.
Telematics Inform. 33(2), 277–292 (2016)
10. J. Soni et al., Predictive data mining for medical diagnosis: an overview of heart disease
prediction. Int. J. Comput. Appl. 17(8), 43–48 (2011)
11. C.L. Manjeevan Seera, A hybrid intelligent system for medical data classification. Expert Syst.
Appl. 41(5), 2239–2249 (2014)
12. A. Ankita et al., Naïve Bayes classifier for prediction of swine flu disease. Int. J. Adv. Res.
Comput. Sci. Softw. Eng. 5(4), 120–123 (2015)
13. M.J. Mangesh et al., Comparative study of decision tree algorithm and Naive Bayes classifier
for swine flu prediction. Int. J. Res. Eng. Technol. 4(6), 45–50 (2015)
14. B.A. Thakkar et al., Health care decision support system for swine flu prediction using naïve
bayes classifier, in 2010 International Conference on Advances in Recent Technologies in
Communication and Computing, pp. 101–105
15. S. Suganya et al., A proficient heart disease prediction method using Fuzzy-Cart algorithm.
Int. J. Sci. Eng. Appl. Sci. (IJSEAS) 2(1) (2016). ISSN: 2395-3470
16. Z.H.M. Fazel et al., A fuzzy rule based expert system for diagnosing asthma. Trans. E Ind.
Eng. 17(2), 129–142 (2010)
17. B.V. Ramana et al., A critical study of selected classification algorithms for liver disease
diagnosis. Int. J. Database Manag. Syst. (IJDMS) 3(2) (2011)
18. Y.H. Joo et al., Fuzzy systems modelling: an introduction. IGI Glob. Distrib. (2009)
A Framework for Remote Health
Monitoring
K. Viswavardhan Reddy and Navin Kumar
Abstract Technologies like sensor and cloud networks are evolving mainly due
to their applications and importance. The main mechanism in such networks is the
deployment of sensor nodes, collection of data, sending the data to cloud for storage
purpose, processing, analyzing, and presenting the results. Nowadays, continuous
monitoring of fluctuating vital health signals of a patient or person especially from
remote has become essential. It is more important in the pandemic like COVID-
19 situation. Hence, there is a need to design, develop, and implement a device to
perform such effective and efficient monitoring system. In this paper, we designed
and developed a framework for smart health monitoring system (SHMS). In this
system, hardware device embedded with various modules like electrocardiogram
(ECG), electromyography (EMG), galvanic skin response (GSR), pulse rate, body
position, and snore detection sensors are integrated. All these sensed vital signals
are read by particle photon (PP) ARM cortex M3 microcontroller and uploaded to
PP cloud. From the cloud, the data is exported into .csv file and imported into the
developed e-Health web and mobile application.
Keywords Sensors · Body area networks · Heart rate · Vital signals · Internet of
things · Cloud computing · E-Health monitoring · Medical diagnosis
1 Introduction
Nowadays, healthcare applications have become an important area among the

researcher’s in the field of wireless body area networks (WBANs). Also, with the
rise in aging population and chronic diseases, wearable sensors have become very
K. Viswavardhan Reddy (B)

Department of ECE, Jain University, Bengaluru, India
e-mail: viswavardhanreddy@gmail.com
N. Kumar
Department of ECE, Amrita School of Engineering, Amrita Vishwa Vidyapeetham,
Bengaluru, India
e-mail: navin_kum3@yahoo.com
102 K. Viswavardhan Reddy and N. Kumar
popular among medical, sports, military, and many other commercial fields [1].
These sensors are intensively helpful in collecting systematic and decisive data of
people’s daily activities. The following are some of the parameters that have been
identified as most common vital signals found in patients [2]: detecting the motion
of a patient, heart rate, blood pressure, blood glucose, pulse rate, oxygen saturation,
muscle conditions, brain activities, body balance, body weight, and body fat. These
are collected by the sensors such as electrocardiogram (ECG), electroencephalo-
gram (EEG), EMG, physiological sensor, motion sensors, glucose sensors, and these
are remotely monitored by authorized doctors/physicians. With the advancements in
body sensors, cloud computing, wireless, Internet of things (IoT), machine learning
(ML), and decision-making technologies, it is made possible by the physician to
examine the patient health conditions before the patients/person visits the clinic.
However, such system has not become very common and still requires practical
implementation, use cases, and effective end-to-end operation.
The usage of care alarm in the context of caretaking services for senior citizens
of the Växjö City has been presented in [3]. For this, a simulation model was devel-
oped using arena software and the results are validated. Similarly, novel advances in
wearable sensors and systems that can monitor movement, physiology, and environ-
ment have been discussed in [4] with some special use cases on Parkinson’s disease,
stroke, and severe injuries in head/neck. The appeal for wearable sensors technology
has lately advanced due to the concept of smart textiles. Authors in [5] discussed the
usage of carbon nanotubes in the domain of medical field as they are admirable in
picking up weak vital signals. Moreover, it is decided to use nanotubes for design
and development of ECG sensors: possessing weightless and easy to wear. Antonio
Celesti et al. [6] designed, developed, and implemented remote patient monitoring
system in collaboration with universities and companies in Italy and South Africa.
Big health application system (BHAS) [7] using IoT and bigdata technologies has
been discussed by the authors and a strategy for optimizing the usage of sensor
devices and various computing resources in delivering advanced applications to the
users. To oversee the human biomedical signals, sensing and monitoring framework
using IoT is suggested in [8]. Furthermore, a case study was carried to validate the
proposal: monitoring the heart condition of each player during a football match [8].
However, these studies revealed the constrained approach to patient-related informa-
tion while making the decisions and communicating with care team is not effective
[9].
In this work, a small, lightweight, low cost, powerful, and integrated SHMS has
been developed and presented. The general architecture consists of sensor network,
a cloud network, and associated links (web interface and application) and terminals.
The sensors are deployed with human body and in the close proximity of patients
to collect necessary data. This data is uploaded to the cloud via Wi-Fi or subscriber
identity module (SIM) module. Moreover, a web and mobile application have been
developed. With this, patients can view their health information and can take prior
appointments. If the patient is found in serious condition, doctors get connected to
the server by proper authentication details to access data and provide the necessary
treatment support.
A Framework for Remote Health Monitoring 103
The content of rest of the paper is as follows. Section 2 discusses smart health
monitoring system architecture with sensing components. Section 3 details the design
and development of e-Health web and mobile application along with steps involved
in deploying the application in amazon web service (AWS). Section 4 presents a
conclusion and some ideas regarding the future work that can be done based on this
work.
2 Smart Health Monitoring System Architecture
2.1 System Architecture
System as shown in Fig. 1 illustrates the proposed remote healthcare monitoring

architecture. At the patient side, we have various sensors discussed below for
capturing the vital signals, placed over the body as a part of necessary accessories.
These analog signals are then sent to processor for monitoring their symptoms. The
processor further processes the collected data and is stored in the cloud. In the cloud,
a web and mobile application are developed and deployed. This web application is
a major interaction system between the patients and hospitals or doctors. Later, this
data is retrieved by the doctors or therapists through the mobile application when
patient book for appointments, understanding that patient has sensor elements, that
is, BSN.
In BSN, sensors are the major components of acquiring the vital signal infor-
mation from the human body through various modules. Some of the major sensor
components used in the entire system are worth describing here.
ECG Sensor (AD8232). This sensor measures electrical activity of the heart. ECG
is used to determine heart rate (HR), heart rhythm and helps in diagnosing heart
arrhythmias, heart attacks, and heart failure. In this work, we have used three electrode
Fig. 1 Proposed SHMS architecture

Table 1 Cable colors and

Cable color Signal
their placement
Black RA (right arm)
Blue LA (left arm)
Red RL (right leg)
Fig. 2 Normal ECG signal from the module
pad leads and are placed closer to the heart for better measurements. The three leads
are properly placed with the help of color coding based on Einthoven’s triangle as
shown in Table 1. Once the leads are placed on the body, we will get a single-color
graph in our serial monitor as shown in Fig. 2. The device helps in acquiring the
signal in the presence of noisy conditions like motion artifacts and remote electrode
placement by two-pole high-pass filter. To compute heartbeat, we detect R–R peak
intervals from Fig. 2.
In order to have dynamic HR estimation, fast Fourier transform (FFT), peak
detection, and smoothing have to be done for every 0.5 s in a moving window. Then
for smoothening the signal, we have used K-points moving average filter (4 points).
The change in the mathematical statement for a L-point discrete time moving average
filter with respect to input ‘x’ and output ‘y’ is given in (1):
1
L−1
y[n] = x[n − k] (1)
L k=0
And, a 4-point moving average filter takes present and preceding three pieces of
input and computes the average [10] using (2):
y[n] = 0.2(x[n] + x[n − 1] + x[n − 2] + x[n − 3] (2)
Later we calculate the difference of smoothed infrared (IR) signal and then apply
to bandpass hamming window filter of 3 dB for filtering particular frequencies given
in (3).

2π n
w(n) = 0.54 − 0.46 cos 0≤n ≤ M −1 (3)
M −1
where M represents number of points in window. After the above process, it is time
for locating the peaks of the signal. Peak location is actually index of the sharpest
location of raw signals since we flipped the signal. Then, we find the distance between
the two peaks and we calculate the HR using:
HR = (ss /(nextpeak - previouspeak)) ∗ 60 Body position sensor
Accelerometer Sensor (ADXL335). This sensor helps in knowing the body posi-
tion. It is extensively being used in low power, gesture, and angle-sensing systems
like mobile phones, gaming systems, sports and fitness devices. The ADXL335 is
a three-axle accelerometer with low noise and 320uA less power utilization. The
bandwidth of this sensor depends on capacitors values and can be selected to suit
for certain applications ranging from 0.5 to 1600 Hz for the X- and Y-axes, and a
range of 0.5–550 Hz for the Z-axis. Figure 3 shows the accelerometer output on serial
monitor of different body positions.
EMG Sensor (EMGSEN). It measures small electrical signals generated by the
movement in the muscles like lifting arm up, clenching fist, or even the simplest of
movements like moving a finger as shown in Fig. 4. With the help of this signal, we can
detect muscle fatigue, defined as a gradual deterioration of the muscle performance
during long period of time.
Pulse sensor (SEN 11574). It helps to read heartbeat during exercises by placing
it to earlobe or fingertip and output is seen over serial monitor. This sensor consists
of bright LED (red in color 660 nm) and a light detector. This light is absorbed by
the tissues and peripheral blood. A photodetector collects the light reflected signals
using I2C protocol which acts as a communication interface.
Body Temp (LM35). It can measure temperature from −55 to +150 °C. From the
module, we first receive adc values and then convert into equivalent voltage using
temp_val = (temp_adc_val * 0.8). This voltage output of LM35 increases 10 mV per
Fig. 3 Accelerometer attached to PP and its output from the console

Fig. 4 a Overall raw EMG signal. b The graphical view of the first muscle contraction (0–10,800
samples) [11]
degree celsius rise in temperature using temp_val = (temp_val/10), and finally, we

convert °C to °F with temp_val = (temp_val * 1.8) + 32 as shown in Fig. 5.
All the sensed information is sent to microcontroller (120 MHz Particle Photon
ARM cortex M3) which converts analog waves into digital values. Moreover, the
device has come with inbuilt tiny Wi-Fi device (Broadcom BCM43362 Wi-Fi chip)
for connecting projects and products for the IoT. All the components are chosen in
such a manner to reduce the size of the device as much as possible. The board
is programmed with compatible version of Arduino programming environment,
allowing for testing on Arduinos. Furthermore, a framework for retrieving data over
Fig. 5 Output from pulse

sensor showing BPM as 82
the Internet is provided. Typical energy consumption of the board with WiFi enabled
is approximately 80 mA. The whole device was programmed in basic embedded C,
programming of microcontroller is a part of embedded development. The program
in controller helps to read and get data from sensors and publish it over the cloud
using the command particle. Publish. The prototype is tested by implementing as
shown in Fig. 6.
3 Development of E-Health Web and Mobile App
E-Health is a web and mobile application developed using Python, Django 1.10
Framework, and Postgresql Database running on all major operating systems. The
applications are deployed on Amazon Elastic Compute Cloud (EC2) vm instance so
that the developed applications can be easily accessible from anywhere and to all
people. Below are some of the steps involved in the development of application on
amazon web services (AWS).
Fig. 6 Prototype of the

product
3.1 Steps in Deploying Application in AWS
Sign into the AWS and then click EC2 console to launch instance for creating and
configuring the virtual machine (VM). Later, click on run instance, then it will show
the instance id, running status, and public ip address allocated. Inside the Ubuntu VM,
we install packages like python3-pip, pip3 install django ==1.11, and Postgres. Later,
we create database and we move the developed application project into the server.
Inside the server, we install various software such as gunicorn which is a python
web server gateway interface HTTP server; nginx (a web server, reverse proxy, load
balancer, mail proxy and HTTP cache), supervisor (a client/server paradigm that
allows to control a number of processes). The whole application is running in AWS
with public IP address of http://18.223.149.193:8000/. After logging into ADMIN
page, navigate to patient page → Add patient by filling all the mandatory fields.
Similarly, navigate to doctor’s page and then add doctors by filling his details as
shown in Fig. 7.
Mobile application. Currently, the e-Health mobile application has been devel-
oped for android phones only. Both doctors and patients can login with registered
mobile phone number. Figure 8 displays one of the patient’s ECG 24 BPM, and this
leads to bradycardia symptoms. If the patient wants to take the appointment from the
list of suggested doctors, he/she clicks view icon on list, then select date and time
and click on fix appointment as shown in Fig. 9.
The doctors of e-Health application are registered by the admin staff of the hospital
in the system. Once the appointment has been fixed, doctors can login into their app
Fig. 7 Admin page displaying add patient, add doctors, and list of patients added so far
Fig. 8 Displays patient

having high fever
and can view the patient’s appointment history and their medical records as shown
in Fig. 10. Moreover, if doctor is not available at that particular date and time, he/she
can update their OPD timings as shown in Fig. 11. Since government of India banned
giving medication through online, giving prescriptions over online is not shown.
4 Conclusion and Future Work
An SHMS framework for remote health monitoring of patients with the integration
of Amazon cloud has been presented in this paper. The designed device is devel-
oped to operate at a very low power, so that it helps for long term monitoring. A
controller board attached with sensor modules is discussed in Sect. 2 measures body
temperature, ECG, EMG, body position, etc., then process these vital signals and
transmits wirelessly to the cloud. Moreover, we also designed and demonstrated a
secured web and mobile application for general public, where patients can view
their medical records from anywhere, and at any time. The system also consisting a
provision to add doctors, and also, it displays the list of doctors and their specialties,
who can give the service to the public. Hence, this framework is more efficient and
cost-effective way for providing quality health services to public.
In future, we would like to develop an algorithm for suggesting doctors, predicting
diseases, and evaluating the performance of algorithm through accuracy. Moreover,
Fig. 9 Doctor appointment
we would like to update and upgrade the system with incorporating the functionality
of directly uploading the images of medical reports and storing them in database.
Furthermore, we test the device with different ages and gender and fabricating the
device for better and more accurate results.
Fig. 10 Doctors application: displaying list of patients who have taken appointment and their
medical records
Fig. 11 OPD timings of

doctor
References
1. G. Aroganam, N. Manivannan, D. Harrison, Review on wearable technology sensors used in

consumer sport applications. J. Sens. 19, 1983 (2019)
2. D. Dias, J. Paulo Silva Cunha, Wearable health devices-vital sign monitoring, systems and
technologies. Sensors 18(8), 2414 (2018, July 25)
3. D. Perez, S. Memeti, S. Pllana, A simulation study of a smart living IoT solution for remote
elderly care, in Proceedings of Third International Conference on Fog and Mobile Edge
Computing (FMEC), Barcelona, (2018), pp. 227–232
4. M.M. Rodgers, V.M. Pai, R.S. Conroy, Recent advances in wearable sensors for health
monitoring. IEEE Sens. J. 15(6), 3119–3126 (2015)
5. M. Bansal, B. Gandhi, CNT based wearable ecg sensors, in Proceedings of 3rd International
Conference on Recent Developments in Control, Automation & Power Engineering (RDCAPE),
Noida, India, (2019), pp. 208–213
6. A. Celesti, M. Fazio, F. Galán Márquez, A. Glikson, H. Mauwa, A. Bagula, F. Celesti, M.
Villari, How to develop IoT cloud e-Health systems based on FIWARE: a lesson learnt. J.
Sens. Actuator Netw. 8, 7 (2019)
7. Y. Ma, Y. Wang, J. Yang, Y. Miao, W. Li, Big health application system based on health Internet
of Things and big data. IEEE Access 5, 7885–7897 (2017)
8. H. Mora et al., An IoT-based computational framework for healthcare monitoring in mobile
environments. J. Sens. 17, 2302 (2017, October)
9. D. Tantrigoda, S. Boralugoda, S. Perera, An approach for visualizing error and obtaining a
measure of central tendency regarding a set of time series using discrete haar wavelet. J.
Wavelet Theor. Appl. 10(1), 1–18 (2016)
10. Moving Average Filter in Python and Matlab, Gaussian Waves: Signal Processing simpli-
fied. Online available at https://www.gaussianwaves.com/2010/11/moving-average-filter-ma-
filter-2/
11. E.F. Shair, A.R. Abdullah, T. Zawawi, T.N. Shuhada, S. Ahmad, S. Salleh, Auto-segmentation
analysis of emg signal for lifting muscle contraction activities. J. Telecommun. Electr. Comput.
Eng. (JTEC) 8, 17–22 (2016)
GenNext—An Extractive Graph-Based
Text Summarization Model for User
Reviews
Komal Kothari, Aagam Shah, Satvik Khara, and Himanshu Prajapati
Abstract User opinions are considered as the core of any business as it helps in
reviewing the product or the service. It is the voice of the customer and every business
strives to get full benefit from it. Relatively, this research work is based on analysis
of the user reviews to provide a better approach to effectively understand the core
of the numerous customer reviews present over the Web. It is almost inhumane job
to search and read all these user reviews manually, and hence, the concepts of text
summarization are applied with the sole aim to derive the meaningful abstract of the
entire review for quicker understanding of its content. This research article peculiarly
focuses on preparing a genre independent review summary model that can be used
on any existing systems. GenNext is a unique model which provides extractive text
summary for multi-genre data and even for user review of as less as one sentence is
developed with graphical method of extractive text summarization. This model has
been proven significantly accurate in providing effective summaries as compared
to the existing summarization models in the market, with the help of sentiment
analysis and polarity classification through natural language processing techniques.
The summaries are also assigned related emoticons based on their polarity scores for
animated representation.
Keywords User review · Review analysis · Genre independent · Text

summarization · Extractive summarization · Sentiment polarity identification ·
Sentimental analysis · Natural language processing (NLP)
K. Kothari (B) · S. Khara · H. Prajapati

Silver Oak Group of Institute, Gujarat Technological University, Ahmedabad, India
e-mail: komalkothari96@gmail.com
S. Khara
e-mail: satvik@socet.edu.in
H. Prajapati
e-mail: himanshuprajapati.ce@socet.edu.in
A. Shah
Indira Gandhi National Open University, Ahmedabad, India
e-mail: aagamshah96@gmail.com
114 K. Kothari et al.
1 Introduction
With the easy availability of Internet on fingertips, it has become extremely necessary
for the companies to reach in the sleeves of the individual to understand and examine
the pros and cons of the commodity from their valued consumers.
Many Web portals provide a platform for the consumers to post their feedback and
opinion after evaluating products or services they used, with the aim to understand the
consumer’s needs. Internet has become a huge warehouse where users over the world
put in their opinions and share their views. Reportedly, there has been an effective hike
in the volume of these customer reviews over the Web. These customer opinions are
extremely necessary for the business heads, stakeholders as well as peer customers
since it represents the voice of the customer and their experiences with the product or
service. These reviews can essentially help a business to discover market requirement
and to develop new plans accordingly for the growth of the business. Web portals,
like Amazon, Yelp, Flipkart, and many more, facilitate the customers with enriched
user reviews for each product. But any human is unlikely to read and understand all
the reviews and hence the concept of text summarization is applied on these user
reviews to analyze and understand the long textual reviews in much lesser time. In
this research work, we have developed a unique model to constructively summarize
the long textual reviews and provide an abstract with similar meaning intended in
the review.
Related Work: There are various approaches and algorithms developed in the field of
text summarization. A survey article [1] presented in this field, expressively examines
all the related works conducted in text summarization and highlights the common
problems such as the over dependency and limitation of the summary model on the
genre of the user reviews provided in the dataset. Also, most of the work conducted is
limited to summarize a group of sentences in larger user reviews and cannot provide
summary of the user review containing as less as one sentence. While a study by Hu
et al. [2] was based on portrayal of effectiveness of extractive summarization and
insists to pursue it by recognizing the top-most derivative sentences acquired from
the review. And a research work conducted by Mallick et al. [3] highlight the usage of
graphical method of test summarization where textual sentences are treated as nodes
and similarity among two different sentences as the weight of the edge between them
to build a graph-based summary.
Through GenNext model, we provide conceptual evaluation of user reviews from
multiple genres using text summarization and natural language processing methods
to understand the summarized meaning and compare it with user review. GenNext is a
unique model that summarizes the user reviews independent of their genre and review
length. Graph-based extractive summarization method provides a meaningful short
summary extracted from the top-most important sentences from the elaborated user
reviews. We have enhanced this method to provide summary for shorter user reviews.
After the model provides the summary, natural language processing techniques are
used to evaluate and compare the meaning of the obtained summary with the user
review.
GenNext—An Extractive Graph-Based Text Summarization Model … 115
We started with gathering datasets from different genres, viz., restaurants, hotels,
airlines, products, and wines from different Web portals like Yelp and Amazon,
which were openly accessible for scholastic objectives. The different datasets were
then merged into one to generate a random multi-genre review dataset. This dataset
was then cleaned by data cleaning techniques including removal of noisy data and
unwanted rows.
Figure 1 is the flow diagram of the applied methodologies for the project.
Fig. 1 Flow diagram

2 Research Design and Used Methodologies to Develop

GenNext (Proposed Approach)
2.1 Data Gathering and Cleaning
Data gathering is the process of collecting data from various sources for further usage.
Various algorithms are then performed on the collected dataset to obtain desired
outcome. Here, we have gathered datasets from different open-source datasets of
Yelp and Amazon which are available for academic purposes. Datasets from different
genres such as restaurant reviews, airline reviews, product reviews, and hotel reviews
are gathered to prepare a huge multi-genre dataset. Data cleaning is done on this
dataset to remove noisy and unwanted data. A multi-genre dataset is formed as per
the required rows such as unique reference number, product, user review, star rating,
and more, while other rows are eliminated to maintain standard format throughout
the merged dataset.
2.2 Text Summarization Using GenNext Algorithm
After merging the dataset, we now have a rich source of raw data containing important
assets necessary for further process. Text summarization method is to be applied on
user reviews of different genre and indefinite length. Before moving forward, let us
understand the concept text summarization by GenNext algorithm.
Text summarization implies the strategy for shortening long briefs of textual
content. The aim is to make a sound and familiar summary having just the major
components laid out in the document. Applying text summarization lessens under-
standing time, quickens the way toward analysis of data, and expands the volume of
data that can accommodate in an area. The latest research in the field represents an
approach to gather priority sentences from the content by text summarization feature
such as word and phrase frequency in a graph-based extraction method. This approach
prioritizes the sentences of a textual review as a component of high occurrence words
by avoiding common stop words [4]. Let us understand text summarization in detail.
Basically, there are two main types to summarize text:
Extraction-based summarization: The method that includes extracting impor-
tant phrases from the textual document and merging them to make a meaningful
abstract is called extraction-based summarization. The extraction is made dependent
on the pre-characterized condition without altering anything in the indigenous text
document. Thus, the significant task which is of extreme significance in following
this methodology is to pick and concentrate the correct sentences for summarization
[1].
Input document → sentences similarity → weight sentences

→ select sentences with higher rank.
Abstraction-based summarization: This methodology utilizes propelled ML algo-

rithms and NLP strategies to acquire a new abstract from the heap of text documents.
In this methodology, it is likewise conceivable that a few parts of the summary
may not by any means present in the original document, however, they are added
utilizing NLP strategies to give shorter and significant synopsis. This strategy incor-
porates summarizing and when applied with deep learning standards, this technique
can conquer the linguistic errors from the extractive summarization. The algorithms
required to perform abstraction summarization are hard to create; and that is on the
grounds that the utilization of extraction methods is highly popular. [1]
Input document → understand context → semantics → create own summary.
Since extraction methods are more popular and widely used, and after effective
survey of existing text summarization models in the market, we have developed
GenNext on graphical and extractive method of text summarization to overcome the
problems in existing models. Let us understand GenNext in detail.
GenNext Algorithm for Text Summarization: GenNext is a unique model devel-
oped with the combination of graphical representation and keyword extraction.
GenNext is the new age general purpose, graph-based ranking algorithm that works
within the concepts of NLP to preserve the meaning of the original text. The founda-
tion of GenNext lays within deciding the value of a vertex in a graph, obtained from
a global data recursively derived from the entire graph.
• Sentences/phrases are extracted from the content by discovering series of words
and assigning them as a vertex.
• A directed graph is made to recognize the reoccurrence the words which are
related to each other.
• A score is calculated for each expression that is the sum of individual score from
the graph.
• Top N expressions are then extracted.
GenNext algorithm weighs the score of each vertex through voting and recom-
mendation system. When a vertex corresponds to other vertex, it casts a vote in the
backend for that vertex. The vertex that gains highest votes is considered highly
important and is extracted to form a summary of that review. With the help of variety
of datasets feed in the GenNext algorithm, we can say that the model is genre inde-
pendent and can summarize user reviews of any genre and of any length. GenNext
is capable of summarizing user review with length as small as one sentence, which
was earlier the limitation in Gensim and other ranking algorithms. GenNext provides
20% of the total length of the user review in the summary for quick and better under-
standing of the summaries as well as easy accessibility while reciprocating through
huge number of reviews.
2.3 Sentiment Analysis for NLP Transition Between GenNext

Summaries and User Reviews
Sentiment analysis otherwise called as opinion mining is a field inside natural

language processing (NLP) that assembles entities that attempt to recognize and
separate opinion inside a text. As a rule, other than distinguishing the sentiment,
these frameworks extract characteristics of the expression, e.g.: polarity, subject,
opinion holder [5].
Sentiment analysis is a method that extricates feelings from the crude writings.
The intention of the client after the content written in the review is comprehended
through sentiment analysis. It gives positive negative qualities of the content. This
can be extremely helpful in understanding the client surveys. The negative, positive,
or neutral aspects obtained from the sentiment of the user review are termed as
polarity. Sentiment polarity score is considered as the evaluation result of the textual
data based on the amount of positive or negative content written in the text [6].
Polarity_Scores = Positive/Negative
In this research, the sentiment analysis is performed on the GenNext summaries

and the corresponding user review to find the polarity score, check the efficiency of
the NLP transition between summary and review, and verify from the polarity scores
whether the original meaning of the review is retained in the GenNext summary or
not. The polarity scores help in identifying the overall meaning conveyed in the user
review and the obtained summary and is also used in comparing the efficiency of
GenNext with other models in the market.
In addition, word and frequency count method was applied on the GenNext
summaries to gain an overall perspective of whether the summary is missing out
on important words from the review or not. The difference of the frequently used
words in the review and GenNext summary is almost negligible. We can say that the
most common words used to describe the review are also the most common words
used in its corresponding summary (Fig. 2).
2.4 Emoticon Assignment to GenNext Summaries
Emoticons are considerably the best way to show emotions. Emoticon assignment
is done for better visualization and interpretation of the summary. We have assigned
emoticons to the corresponding GenNext summary as per polarity scores as shown
in Fig. 3.
Following are the tasks followed by GenNext model
1. Construction of vectors for intermediate representation of the input text
2. Scoring the vectors based on the importance irrespective of the length
Fig. 2 a Word cloud of review, b word cloud of GenNext summary
Fig. 3 Emoticon assignment for polarity scores ranging from (-1) to (1)
3. Selection of a summary comprising of highly important vectors

4. Maintaining the length ratio between summary and user review for any length or
genre
5. Sentiment analysis and polarity score is calculated of GenNext summaries and
user reviews
6. Emoticons are assigned to corresponding GenNext summaries for better repre-
sentation (Fig. 4).
Fig. 4 GenNext model

Fig. 5 a Average sentiment polarity of GenNext is almost equivalent with the user reviews and
b Bleu score comparison between GenNext and other algorithms
3 Result and Discussion
Through this research, we have presented a unique methodology that summarizes the
user reviews independent of their genre and contextual length through ranking and
extractive algorithms in text summarization. We have tried to overcome the existing
problems in extractive methods through GenNext, and an approach to compare senti-
ment polarity scores as well as visual representation with emoticon assignment. In
this research, GenNext is proven far more accurate than the existing systems with
much wider implication scopes in the future (Fig. 5).
4 Conclusion
It is clearly evident that GenNext is significantly accurate with the highest score of
0.686322, through Bleu method of finding accuracy, as compared to other algorithms.
GenNext is the most efficient extractive summarizer model with highest Bleu score
that can provide summary for multi-genre user reviews. Nevertheless, there is always
a chance of improvement. Huge and varied datasets can be used to train the machine
and advanced sentiment analysis methods can be used to provide more accurate and
genre independent results over the growing time.
References
1. K. Kothari, A. Shah, S. Khara, H. Prajapati, A novel approach in user reviews analysis using
text summarization and sentiment analysis: survey. Int. J. Creat. Res. Thoughts (IJCRT) 8(3),
3067–3072 (2020). ISSN:2320-2882. Available at: http://www.ijcrt.org/papers/IJCRT2003417.
pdf
2. Y. Hu, Y. Chen, H. Chou, Y.-H. Hu, Y.-L. Chen, H.-L. Chou, Opinion mining from online hotel
reviews—a text summarization approach. Inf. Process. Manag. 53(2), 436–449 (2017). https://
doi.org/10.1016/j.ipm.2016.12.002
3. C. Mallick, A. Das, M. Dutta, A. Das, A. Sarkar, Graph-based text summarization using modified
TextRank, Soft Computing in Data Analytics (Springer, Berlin, 2019), pp. 137–146. https://doi.
org/10.1007/978-981-13-0514-6_14
4. Medium, A Quick Introduction to Text Summarization in Machine Learning (2020) [online].
Available at: https://towardsdatascience.com/a-quick-introduction-to-text-summarization-in-
machine-learning-3d27ccf18a9f
5. H2oai.github.io, Natural Language Processing Tutorial—Sentiment Analysis (2020) [online].
Available at: https://h2oai.github.io/tutorials/natural-language-processing-tutorial-sentiment-
analysis/#0
6. A. Shah, K. Kothari, U. Thakkar, S. Khara, User review classification and star rating predic-
tion by sentimental analysis and machine learning classifiers, Information and Communication
Technology for Sustainable Development (Springer, Singapore, 2019), pp. 279–288. https://doi.
org/10.1007/978-981-13-7166-0_27
An Adaptive Control for Surrogate
Assisted Multi-objective Evolutionary
Algorithms
Duc Dinh Nguyen and Long Nguyen
Abstract Multi-objective problems (MOPs), a class of optimization problems in the

real-world, have multiple conflicting objectives. Multi-objective evolutionary algo-
rithms (MOEAs) are known as great potential algorithms to solve difficult MOPs.
With MOEAs, based on the principle of population, we have a set of optimal solu-
tions (feasible solution set) after the search. We often use the concept of dominance
relationship in population, and it is not difficult to find out set of Pareto optimal
solutions during generations. However, with expensive optimization problems in the
real world, it has to use a lot of fitness function evaluations during the search. To
avoid expensive physical experiments, we can use computer simulations methods
to solve the difficult MOPs. In fact, this way often costs expensive in computation
and times for the simulation. In these cases, researchers discussed on the usage of
surrogate models for evolutionary algorithms, especially for MOEAs to minimize
the number of fitness callings. There are a series of proposals which were introduced
with the usage of RBF, PRS, Kriging, SVM models. With the concept of machine
learning, these MOEAs can solve expensive MOPS effectively. However, with our
analysis, we found that using a fixed ratio of fitness functions and surrogate functions
may make the unbalance of exploitation and exploration of the evolutionary process.
In this paper, we suggest to use an adaptive control to determine the effective ratio
during the search. The proposal is confirmed though an experiment with standard
measurements on well-known benchmark sets.
Keywords K-RVEA · Surrogate · Kriging · Adaptive control
D. D. Nguyen (B)
Military Information Technology Institute,
Academy of Military Science and Technology, Hanoi, Viet Nam
e-mail: nddinh76@gmail.com
L. Nguyen (B)
Department of Information Technology, National Defense Academy, Hanoi, Viet Nam
e-mail: longit76@gmail.com
124 D. N. Duc and L. Nguyen
1 Surrogate Assisted Multi-objective Evolutionary

Algorithms
1.1 Surrogate Models
A multi-objective optimization problem involves at least two conflicting objectives,

and it has a set of Pareto optimal solutions. MOEAs use a population of solutions to
approximate the Pareto optimal set in a single run. MOEAs have attracted a lot of
research attention during the past decade. They are still one of the hottest research
areas in the field of Computational Intelligence, especially to solve difficult MOPs in
the real world. In multi-objective optimization, a problem is defined as a expensive
problem when it has fitness functions which require time-consuming experiments or
computer simulations.
Surrogate models are used to approximate in simulation way to reduce the com-
putational cost for expensive problems. The models are described as below:
If we call f (−→x ) is an origin fitness function of a MOP, then we have f (−
→
x ) is a
meta function, which is indicated as below:
f (−
→
x ) = f (−
→
x ) + e(−
→
x) (1)
function e(− →
x ) is the approximated error. In this case, the fitness function f (−→x ) is
not to be known, and the values (input or output) are cared. Based on the responses
of the simulator from a chosen dataset, a surrogate is constructed, and then the model
generates easy representations that describe the relations between preference infor-
mation of input and output variables. There are some approaches for the surrogate
models, which are divided into some kinds such as the radial basis function (RBF),
the polynomial response surface (PRS), the support vector machine (SVM), and
the Kriging (KRG). The common concepts of using surrogate models are: Instead
of using known fitness functions always, to reduce the number of calculations, the
evaluations are suggested to use surrogate functions, which are simpler. Therefore,
without loss of generality, this paper investigates the use of the original and surrogate
functions with Kriging model in the solving process of the MOEAs for expensive
problems.
In [8], the authors proposed a method named “Kriging,” which is a response
surface method bases on spatial prediction techniques. This method minimizes the
mean-squared error to build the spatial and temporal correlation among the values
of an attribute. In [9], the authors developed a parametric regression model which
design and analysis of computer experiments, called DACE. The model is an exten-
sion of Kriging approach for at least three-dimensional problems. The model is a
combination of a known function f (x) and a Gaussian random process f (x) that
is assumed to have mean zero and covariance. In [2], the authors build a procedure
which uses Kriging approximations and NSGA-II algorithm to optimize the aircraft
design. The Kriging-based genetic algorithm procedure gives shapes which are opti-
mized for both the initial peak and perceived noised level of the signature along with
An Adaptive Control for Surrogate Assisted Multi-objective … 125
minimum drag. In [6], by providing a design of experiments (DOE) capability into

the framework, the authors hybridized the desirable characteristics of EAs and sur-
rogate models such as RSM to obtain an efficient optimization system. In [7], in the
problems with two objective functions, thermal resistance and pumping power have
been selected to assess the performance of the micro-channel heat sink. In [10], the
authors deeply discussed on surrogated MOEAs and indicated that: Kriging appears
to perform well in most situations; however, it is much more computationally expen-
sive than the rest. In [5], the KRG is used, and the term Kriging points directly to
the origin of these prediction methods dating back to the sixties, when the mining
engineer Krige used Gaussian random field models (GRFMs) to predict the concen-
tration of ore in gold and uranium mines. In [11], authors proposed such a method,
MOEA/D-EGO, for dealing with expensive MOPs. The algorithm decomposes an
MOP into a number of single-objective optimization subproblems. At each iteration,
a predictive distribution model is built for each individual objective in the MOP by
using fuzzy clustering and Gaussian stochastic process modeling.
1.2 K-RVEA
To build up our research, we analyzed the usage surrogate parameters on KRVEA

[3], and the algorithm is short described as below:
Based on RVEA [1], a reference vector based MOEA named K-RVEA was pro-
posed for solving many objective problems. In the algorithm, search process is guided
by a set of predefined reference vectors inspired from the direction vectors with the
idea is to partition the objective space into a number small subspaces using a set of
reference vectors. An elitism selection strategy is employed inside each subspace and
using an angle-penalized distance (APD) to measure the distance of the solutions to
the ideal point and the closeness of the solutions to the reference vectors. For more
detail about the algorithm, see [1].
1.3 Issues of Using Fixed Ratio
In our work, we analyzed the evolutionary processes’ behavior during the search with
surrogated assisted multi-objective evolutionary algorithms. For MOEA in general,
according to the searching process, the quality of the achieved solution population
varies according to the general rule; that is, in the first generations, the convergence
quality of the algorithm is poor, the calculation quite diverse due to the fact that
random generated solutions have not really evolved over generations. After that,
the solution quality gradually got better, the diversity also gradually decreased and
towards the Pareto optimal front spread. The last generations of a process with a
specified number of generations, the quality of the population converged was good
because of the evolution process over many generations. An algorithm is assessed
well if the diversity of the population is maintained when the solutions spread evenly
over the Pareto optimization front. In many analyzes of the evolutionary trend of
evolution, in order to ensure a quality balance between convergence and diversity,
it is necessary to guide the algorithm to keep a balance between exploration and
exploitation cascade of evolution. Therefore, in the early stages, the ability to con-
verge and the end-stage need to enhance the diversity of the population. In K-RVEA,
the authors use parameters to determine whether the model update is fixed, which
can lack the algorithm’s adaptability, possibly losing the balance between exploration
and exploitation of the algorithm. We will test the K-RVEA algorithm on the class
of expensive problems, with different parameters to evaluate the above comments.
We used wmax, the number of generations before updating the Kriging model
with values: 10, 20, 30, 50, and 70 with DTLZs problems [4] (DTLZs from 1 to 9) in
20,000 calculations. We reported the results on GD and IGD, the popular metrics to
measure the performance of the algorithm. The results are shown in Tables 1 and 2.
Through the test results on DTLZs with GD, IGD measurements, we find that: At
each time, depending on the time in evolution, the quality of the population using
the values the parameter wmax, the number of generations before the model update
is carried out, has a strong influence on the ability to explore and exploit, in other
words, the convergence and diversity of the solution population. Specifically, in the
early generations, the quick update of the model to use a surrogate function seems
to produce better results (when the wmax value is small, corresponding to DTLZ1,
DTLZ2, DTLZ4, DTLZ5 problems). Of course, it also depends on the number of
non-dominated solutions in the population that each problem is achieved according
to evolution (DTLZ3, DTLZ7, DTLZ9). When in the next generation, there is a
change in the adapt for increasing generation, ignoring the update of Kriging model
(increasing wmax value, corresponding to DTLZ1, DTLZ2, DTLZ6, DTLZ7) and
sustainable as the case of DTLZ9 problem.
Applying a predetermined number of generations, which then updates the model,
can be algorithmically inefficient, reducing convergence and diversity of the algo-
rithm. The application of numerical value generation to apply the original objective
functions before using surrogate function according to Kriging model depends on
the quality of solution of the current generation. Specifically, when the number of
non-dominated solutions in the population is small, far from the Pareto optimal class,
it would be appropriate if this parameter is applied at a relative level (in the case of
experiments about 30, 50). If the ratio is high, the number of assessments depends
largely on the original objective functions, the effect of using surrogate function does
not make much sense. If a low rate is applied, when the number of individuals is not
small, updating the model will not have many changes in quality. When the number
of non-dominated solutions in the population is already in the majority, but not close
to the Pareto optimization front, it is possible to increase the number of evaluations by
the surrogate function and the frequency of updating the Kriging model (depending
on the quality, number of non-dominated solutions in the population, the value could
be 20, 30). When the number of non-dominated solutions in the obtained population
is approaching the Pareto optimal front, then the surrogate function rate and Kriging
model update frequency may be higher, respectively, 10.
Table 1 GD values on DTLZs benchmark set

DTLZ1
10 50.7441 7.4836 5.6692 4.1103 3.2282 2.7263 2.5096 2.4711 2.4463 2.3877
20 41.9546 10.0818 10.9039 3.7888 3.7041 3.7041 4.0672 4.0672 4.0097 4.0097
30 39.0192 9.4198 7.2260 5.4170 3.9909 3.9062 3.9777 4.0149 3.9320 3.7725
50 42.5498 10.4695 6.7609 5.9252 5.5349 4.6894 4.7492 4.3930 4.1622 4.0916
70 43.5096 12.6410 8.0107 6.6715 7.1815 6.2105 5.7701 5.2643 5.0035 4.7697
ADC 46.6079 4.7267 3.3725 3.3725 3.3725 3.3725 3.3725 3.3725 3.3725 3.3725
DTLZ2
10 0.1032 0.0023 0.0014 0.0011 0.0009 0.0008 0.0008 0.0007 0.0007 0.0006
20 0.1151 0.0014 0.0009 0.0007 0.0006 0.0005 0.0005 0.0005 0.0004 0.0004
30 0.0999 0.0015 0.0010 0.0008 0.0006 0.0005 0.0005 0.0004 0.0004 0.0004
50 0.1076 0.0015 0.0009 0.0006 0.0005 0.0005 0.0004 0.0004 0.0004 0.0004
70 0.1165 0.0014 0.0010 0.0007 0.0006 0.0005 0.0005 0.0004 0.0004 0.0004
ADC 0.1227 0.0019 0.0011 0.0008 0.0007 0.0006 0.0006 0.0005 0.0005 0.0004
DTLZ3
10 183.1624 43.1944 36.1737 33.9130 30.8676 29.8677 28.4940 30.4472 29.3890 30.4415
20 171.6720 65.1569 57.1465 47.8764 47.3865 42.2447 40.1520 38.8519 36.7741 36.5496
30 220.4797 74.7820 56.8731 52.7536 52.7041 48.8366 44.5964 42.4959 41.5517 40.6848
50 194.0912 85.5257 83.2413 70.2972 59.2193 50.7384 50.3928 49.8387 48.0394 44.5380
70 151.5813 77.2795 68.4175 59.9071 51.6176 47.8923 49.5671 51.3997 48.6738 48.7996
ADC 183.1624 69.5647 54.4809 48.0276 47.8869 43.2640 42.0306 40.3582 39.2478 42.3035
DTLZ4
10 0.1739 0.0148 0.0093 0.0063 0.0051 0.0043 0.0038 0.0034 0.0031 0.0029
20 0.1613 0.0113 0.0072 0.0054 0.0044 0.0039 0.0035 0.0033 0.0031 0.0029
30 0.1634 0.0108 0.0077 0.0060 0.0051 0.0044 0.0038 0.0035 0.0033 0.0031
50 0.1841 0.0115 0.0074 0.0057 0.0048 0.0042 0.0038 0.0034 0.0031 0.0028
70 0.1564 0.0129 0.0085 0.0073 0.0063 0.0055 0.0050 0.0047 0.0045 0.0042
ADC 0.1564 0.0129 0.0085 0.0073 0.0063 0.0055 0.0050 0.0047 0.0045 0.0042
DTLZ5
10 0.1407 0.0068 0.0057 0.0054 0.0050 0.0046 0.0042 0.0042 0.0042 0.0041
20 0.1571 0.0096 0.0069 0.0063 0.0056 0.0053 0.0052 0.0050 0.0049 0.0048
30 0.1207 0.0111 0.0082 0.0068 0.0063 0.0058 0.0056 0.0054 0.0051 0.0051
50 0.1584 0.0052 0.0046 0.0042 0.0041 0.0041 0.0040 0.0040 0.0039 0.0038
70 0.1368 0.0101 0.0072 0.0065 0.0062 0.0056 0.0055 0.0052 0.0050 0.0049
ADC 0.1286 0.0071 0.0055 0.0048 0.0045 0.0042 0.0039 0.0039 0.0037 0.0037
DTLZ6
10 0.9746 0.4591 0.4409 0.3822 0.3919 0.3913 0.3306 0.3261 0.3428 0.3140
20 0.9513 0.4647 0.4897 0.4926 0.4734 0.4403 0.4151 0.3610 0.3625 0.3456
30 0.9965 0.4503 0.3943 0.4227 0.3993 0.3845 0.3798 0.3657 0.3867 0.3529
50 0.9480 0.5927 0.5124 0.4898 0.4478 0.4311 0.4444 0.4308 0.4108 0.3957
70 0.9599 0.5421 0.4830 0.4446 0.4565 0.4716 0.4370 0.4339 0.4248 0.4184
ADC 1.0474 0.4951 0.5234 0.5070 0.4695 0.4490 0.4279 0.4099 0.3823 0.3597
(continued)
Table 1 (continued)
DTLZ7
10 2.0241 0.0026 0.0012 0.0008 0.0006 0.0005 0.0005 0.0004 0.0004 0.0003
20 2.6543 0.0012 0.0007 0.0005 0.0004 0.0003 0.0003 0.0003 0.0002 0.0002
30 2.6896 0.0009 0.0005 0.0004 0.0003 0.0003 0.0003 0.0002 0.0002 0.0002
50 2.3290 0.0010 0.0007 0.0005 0.0004 0.0004 0.0003 0.0003 0.0002 2.0384
70 2.0943 0.0007 0.0005 0.0005 0.0004 0.0003 0.0003 0.0003 0.0002 0.0002
ADC 2.0241 0.0009 0.0007 0.0004 0.0004 0.0003 0.0003 0.0003 0.0002 0.0002
DTLZ8
10 0.0548 0.0322 0.0322 0.0322 0.0322 0.0322 0.0322 0.0322 0.0322 0.0322
20 0.0587 0.0327 0.0327 0.0327 0.0327 0.0327 0.0327 0.0327 0.0327 0.0327
30 0.0706 0.0358 0.0358 0.0358 0.0358 0.0358 0.0358 0.0358 0.0358 0.0358
50 0.0728 0.0303 0.0303 0.0303 0.0303 0.0303 0.0303 0.0303 0.0303 0.0303
70 0.0635 0.0357 0.0357 0.0357 0.0357 0.0357 0.0357 0.0357 0.0357 0.0357
ADC 0.0635 0.0310 0.0310 0.0310 0.0310 0.0310 0.0310 0.0310 0.0310 0.0310
DTLZ9
10 3.4231 2.2998 1.5890 1.3484 1.4657 2.2714 1.4092 1.2228 1.1880 1.1553
20 6.3959 2.5242 2.3244 2.1598 1.9384 1.9384 2.3200 1.5506 1.6016 2.1090
30 4.2762 2.1689 1.9599 1.7168 1.6040 1.4022 1.5312 1.4316 1.3738 1.2709
50 3.9907 1.5387 1.1807 0.9758 2.3940 2.1892 1.8939 1.8939 1.8430 1.7425
70 4.5920 2.5457 1.9841 2.2016 2.0792 2.0511 1.9588 1.8728 1.9472 1.7881
ADC 3.7733 2.5289 2.2851 2.1059 2.0260 2.0260 1.8907 1.8280 1.2761 1.1746
In general, from the above assumptions, it is assumed that increasing or decreasing

the number of generations (or increasing/decreasing the value of wmax) is responsive
depending on key factors such as time of evolution, quality convergence of the
population, and the computational complexity of each problem.
2 Using Adaptive Control for Surrogate Assisted MOEAs
In this section, based on the analysis from the above experimental results, the assump-
tions about appropriate parameter adjustment during evolution create an effective
balance between convergence quality and diversity of the population. Specifically,
adjusting the parameter on the number of generations before updating the model
according to the assessed factors and assuming that the direct influence in algo-
rithms using a representative function is: the process time of the process evolution,
the quality of the current population and the computational characteristics of the
problem.
Building the rate of using surrogate function and frequency of updating Kriging
model flexibly, depending on the quality of the solution population in the current
generation, to determine the quality of the solution, we use two defined values at a
time t, with the test problem j as follows:
Table 2 IGD values on DTLZs benchmark set

DTLZ1
10 37.5017 15.9551 15.9551 15.9551 8.6306 8.6306 8.6306 6.4437 6.4437 6.4437
20 69.2926 12.2018 9.0805 6.0770 6.0770 6.0770 6.0770 6.0770 6.0770 6.0770
30 73.5400 18.0276 18.0276 18.0276 18.0276 18.0276 18.0276 18.0276 18.0276 18.0276
50 75.3314 21.8503 21.8503 21.8503 18.8011 18.8011 18.8011 18.8011 18.8011 18.8011
70 56.6293 17.9351 17.9351 17.9351 15.7866 15.7866 11.9698 11.9698 11.9698 11.9698
ADC 30.8156 9.9032 9.9032 9.9032 9.9032 9.9032 9.9032 9.9032 9.9032 9.9032
DTLZ2
10 0.4863 0.0523 0.0417 0.0372 0.0343 0.0326 0.0314 0.0302 0.0293 0.0284
20 0.4862 0.0461 0.0368 0.0325 0.0297 0.0287 0.0281 0.0273 0.0265 0.0259
30 0.4720 0.0467 0.0382 0.0326 0.0293 0.0283 0.0268 0.0260 0.0249 0.0244
50 0.5233 0.0476 0.0387 0.0340 0.0306 0.0295 0.0285 0.0276 0.0272 0.0267
70 0.4848 0.0502 0.0427 0.0372 0.0338 0.0324 0.0302 0.0292 0.0283 0.0272
ADC 0.5617 0.0528 0.0398 0.0354 0.0335 0.0320 0.0305 0.0295 0.0287 0.0278
DTLZ3
10 500.9795 154.0806 154.0806 154.0806 154.0806 154.0806 154.0806 154.0806 154.0806 154.0806
20 608.2876 195.9133 185.9003 185.9003 185.9003 185.9003 185.9003 185.9003 177.8553 177.8553
30 421.7119 282.9219 229.2645 229.2645 229.2645 222.9678 132.3652 132.3652 132.3652 132.3652
50 533.7593 229.4352 225.6308 225.6308 196.0629 196.0629 196.0629 196.0629 196.0629 195.8602
70 672.9475 281.1927 274.5212 246.6838 246.6838 246.6838 183.5229 183.5229 183.5229 183.5229
ADC 500.9795 235.2239 204.4000 204.4000 204.4000 182.6566 182.6566 182.6566 182.6566 182.6566
DTLZ4
10 0.9201 0.1337 0.0794 0.0595 0.0549 0.0495 0.0477 0.0458 0.0444 0.0434
20 0.9733 0.0753 0.0606 0.0543 0.0503 0.0465 0.0447 0.0418 0.0408 0.0396
30 0.8311 0.0745 0.0626 0.0568 0.0520 0.0466 0.0441 0.0421 0.0408 0.0399
50 0.8441 0.0786 0.0637 0.0577 0.0541 0.0512 0.0474 0.0442 0.0420 0.0413
70 0.8792 0.0833 0.0662 0.0629 0.0582 0.0552 0.0542 0.0524 0.0512 0.0504
ADC 0.8792 0.0833 0.0662 0.0629 0.0582 0.0552 0.0542 0.0524 0.0512 0.0504
DTLZ5
10 0.3699 0.0395 0.0397 0.0384 0.0343 0.0330 0.0329 0.0327 0.0301 0.0301
20 0.3739 0.0583 0.0525 0.0391 0.0371 0.0357 0.0355 0.0347 0.0345 0.0344
30 0.4081 0.0722 0.0597 0.0496 0.0492 0.0478 0.0457 0.0449 0.0421 0.0421
50 0.3483 0.0381 0.0337 0.0317 0.0307 0.0367 0.0364 0.0364 0.0364 0.0363
70 0.3860 0.0494 0.0426 0.0412 0.0411 0.0394 0.0393 0.0388 0.0384 0.0376
ADC 0.4247 0.0421 0.0371 0.0380 0.0373 0.0345 0.0338 0.0336 0.0333 0.0333
DTLZ6
10 8.2380 1.8545 0.8540 0.8110 0.8110 0.8110 0.8110 0.8110 0.7906 0.7724
20 8.4386 2.2229 1.3579 1.3532 1.3532 1.2986 1.2986 1.2557 1.2557 1.2557
30 8.2757 2.4312 2.4304 1.8914 1.7946 1.7637 1.6556 1.6556 1.6556 1.6437
50 8.2283 2.7022 2.5015 2.4640 2.4640 2.4640 2.2781 2.2781 2.0575 2.0575
70 8.4906 2.4451 2.4451 2.4451 2.1171 2.0384 2.0384 2.0384 2.0384 2.0384
ADC 8.4051 2.9249 1.8047 1.6488 1.6488 1.6488 1.6488 1.6334 1.5731 1.5731
(continued)
Table 2 (continued)
DTLZ7
10 8.1067 0.0611 0.0397 0.0332 0.0295 0.0276 0.0256 0.0243 0.0231 0.0222
20 7.4668 0.0492 0.0406 0.0338 0.0317 0.0296 0.0280 0.0268 0.0256 0.0246
30 8.6059 0.0514 0.0424 0.0380 0.0358 0.0341 0.0312 0.0303 0.0293 0.0282
50 8.8847 0.0707 0.0599 0.0543 0.0469 0.0440 0.0420 0.0407 0.0398 0.0477
70 8.2937 0.0725 0.0598 0.0559 0.0506 0.0487 0.0459 0.0442 0.0433 0.0424
ADC 8.1067 0.0618 0.0546 0.0456 0.0433 0.0411 0.0359 0.0343 0.0320 0.0312
DTLZ8
10 0.2309 0.2050 0.2050 0.2050 0.2050 0.2050 0.2050 0.2050 0.2050 0.2050
20 0.2590 0.1599 0.1599 0.1599 0.1599 0.1599 0.1599 0.1599 0.1599 0.1599
30 0.2190 0.1827 0.1827 0.1827 0.1827 0.1827 0.1827 0.1827 0.1827 0.1827
50 0.2516 0.2357 0.2357 0.2357 0.2357 0.2357 0.2357 0.2357 0.2357 0.2357
70 0.2397 0.1935 0.1935 0.1935 0.1935 0.1935 0.1935 0.1935 0.1935 0.1935
ADC 0.2397 0.1790 0.1790 0.1790 0.1790 0.1790 0.1790 0.1790 0.1790 0.1790
DTLZ9
10 11.3847 6.3491 6.1299 6.1299 5.6413 5.3711 4.9843 4.9843 4.9843 4.9843
20 11.1618 6.3076 6.2910 6.2910 6.2910 6.2910 5.9713 5.9713 5.9713 5.9066
30 11.1510 7.7752 7.7752 7.7563 7.7378 7.7212 7.7103 7.7103 7.7103 7.7099
50 11.2579 7.9947 7.9363 7.9294 7.5453 7.5453 7.5453 7.5453 7.5453 7.5453
70 11.1964 6.6274 6.3928 6.3837 6.3837 6.3837 6.3837 6.3837 6.3837 6.3830
ADC 11.2529 6.5146 6.4943 6.4943 6.4943 6.4943 6.4943 6.4943 6.4697 6.4697
n ND ∗ n gen
Qt = Cj (2)
qsize tsize
In which, n ND is the number of non-dominated solutions in the main population, qsize

is the size of the population, current generation n gen , tsize is the number of generations.
C j is the complexity of the problem; this is a parameter dependent on the complexity
of the test function; it depends on the size of decision spaces, number of objectives,
and number of calculations for objective functions; we suggest to choose practical
values C j ∈ [0.8, 1], and the default value is 1. The constant 60 is used, which is the
upper range of wmax; it helps the control to get the maximized value, 50. Based on
the Q t value, we will calculate the value of wmax ∈ R (the rate of using surrogate
function and the frequency of updating the Kriging model) in the range [10, 50], and
a normalize function is defined to adjust valid wmax values.
wmax = normalize(Q t ∗ 60) (3)
In the experiments, we used the origin K-RVEA and the modified version M-K-RVEA
with the usage of the above adaptive wmax. We also used the DTLZs benchmark set
with GD, IGD metrics. The results are reported in rows ADC on Tables 1 and 2. We
also reported and compared the behaviors of K-RVEA (with the default wmax is 30)
and M-K-RVEA.
Analyzing and comparing the results, we can found that M-K-RVEA with adaptive
parameter control can get more balance of convergence and diversity for obtained
population. In details, in the early stages, it gets better in 6 problems on GD and 6
problems on IGD. During the generations (the later stages), M-K-RVEA gets equal
or better on 5 problems, one is the same. It is also better in 6 problems on IGD. In
the last stages, M-K-RVEA is getting better in 7 problems on GD and 5 problems
in IGD. In other problems, K-RVEA gets the same results with M-K-RVEA or a bit
better than M-K-RVEA.
In general, through the experimental results, it is proposed to use adaptive param-
eter control that has an effective effect to implement the optimal target optimization
algorithm using general surrogate functions, algorithms based on Kriging model in
particular. The obvious meaning of using the response parameter is to assess the qual-
ity of the population at the time of determining whether to update the Kriging model
or not? The results using adaptive parameter control demonstrate the hypothesis of
the impact of the time process, the current convergence quality of the population,
and the characteristics of the problem on maintaining the equilibrium of quality of
convergence and diversity of the population, through ensuring a balance between
the exploration and exploitation process of evolution. The effectiveness of parameter
selection in the current stages also depends on the quality and effectiveness of the
previous stage, proving that a timely adjustment to maintain a balance between explo-
ration and exploitation is needed to set. However, from the results of some problems,
K-RVEA has a better quality, although not large, it also gives us issues that need to
be adequately posed such as stability, stainability and random factors course. These
factors influence the choice of parameters appropriately or guide evolution to better
efficiency.
3 Conclusion
In this paper, we have analyzed the effects of time factor, algorithm quality, and
problem characteristics on the process of maintaining the convergence and diversity
of the population over generations. The implications for maintaining the exploration
and exploitation capabilities of the evolutionary algorithms of maximally targeted
targets using surrogate functions. The paper evaluated the properties of the temporal
process through the number of calculations, the quality of the population according
to the number of non-dominated solutions in the population and the characteristics
of spatial problems, the number of calculations of the problem. Those values are
basis for calculating and determining the parameters that determine the time when
updating machine learning model, in our experiment is the Kriging model. The
results have proved the hypothesis and suggested the evaluation of factors affecting
the evolution process to determine the parameters in a flexible way, contributing to
the guidance of well-balanced algorithms, more about the quality of convergence
and diversity through the search process. The results also suggest further research
issues for the evaluation and parameter selection in the evolution process of the
multi-objective optimal evolution algorithms, especially algorithms using surrogate
functions in solving expensive problems in the field of multi-objective optimization.
References
1. R. Cheng, Y. Jin, M. Olhofer, B. Sendhoff, A reference vector guided evolutionary algorithm

for many-objective optimization. IEEE Trans. Evol. Comput. 20(5), 773–791 (2016)
2. S. Choi, J.J. Alonso, H.S. Chung, Design of a low-boom supersonic business jet using evolu-
tionary algorithms and an adaptive unstructured mesh method. AIAA Paper 1758, 2004 (2004)
3. T. Chugh, Y. Jin, K. Miettinen, J. Hakanen, K. Sindhya, K-RVEA: a kriging-assisted evolu-
tionary algorithm for many-objective optimization. Sci. Comput. B 2 (2016)
4. K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable test problems for evolutionary multi-
objective optimization. TIK-Report no. 112. Technical report, TIK, ETH, Zurich (2001)
5. M.T.M. Emmerich, K.C. Giannakoglou, B. Naujoks, Single- and multiobjective evolutionary
optimization assisted by gaussian random field metamodels. IEEE Trans. 10(4), 421–439 (2006)
6. L. Gonzalez, J. Periaux, K. Srinivas, E. Whitney, A generic framework for the design optimisa-
tion of multidisciplinary UAV intelligent systems using evolutionary computing, in 44th AIAA
(2006), p. 1475
7. A. Husain, K.-Y. Kim, Enhanced multi-objective optimization of a microchannel heat sink
through evolutionary algorithm coupled with multiple surrogate models. Appl. Therm. Eng.
30(13), 1683–1691 (2010)
8. G. Matheron, Principles of geostatistics. Econ. Geol. 58(8), 1246–1266 (1963)
9. J. Sacks, W.J. Welch, T.J. Mitchell, H.P. Wynn, Design and analysis of computer experiments.
Stat. Sci. 4, 409–423 (1989)
10. I. Voutchkov, A.J. Keane, Multiobjective optimization using surrogates (2006)
11. Q. Zhang, W. Liu, E. Tsang, B. Virginas, Expensive multiobjective optimization by MOEA/D
with Gaussian process model. IEEE Trans. 14(3), 456–474 (2010)
Study and Analysis of NOMA Downlink
Transmission Model in 5G Wireless
System
Sanjeev Kumar Srivastava
Abstract This paper examines NOMA downlink transmission model in 5G tech-

nology wireless system. In this system, a mathematical model for unordered proba-
bility outage, ordered probability outage and optimum allocation is developed. The
outcome of the model is simulated for NOMA downlink and the analysis of the same
is discussed. When channel gain is increasing, capacity of optimum power alloca-
tion and capacity of random power allocation in NOMA improve the overall system
performance. But in large context, capacity of optimum allocation in downlink trans-
mission will produce better result as compared to that of random allocation due to
fairness in order to distribute available resources equally on high priority basis to get
high net throughput.
Keywords NOMA · OMA · OFDMA · MIMO · MUD · LDM
1 Introduction
Non Orthogonal Multiple Access (NOMA) is one of the main principles of designing
radio access techniques for the 5G technology wireless network system. It is based
on the fundamental concept of wireless technology [1].
Therefore, the motivation in this technology serves more than one users in each
orthogonal resource block that is, time slot, subcarrier, spreading code etc. In contrast,
a conventional Orthogonal Multiple Access (OMA) technique includes scarce band-
width resource which is solely occupied by the user despite its poor channel condi-
tions. This impacts negatively the spectrum efficiency and in turns the degradation
of the throughput of the overall system.
Hence, the novelty of this model is that NOMA ensures that the users with better
channel conditions can simultaneously utilize the same bandwidth resources. As
a result, fairness has to be guaranteed. Therefore, the net throughput of NOMA
can be significantly larger than that of OMA. In addition to its spectrum efficiency
S. K. Srivastava (B)
PCE, Mumbai, Maharashtra, India
e-mail: sanjeevkumar.srivastava1@gmail.com
134 S. K. Srivastava
gains, NOMA can enable massive connectivity. This is key element of supporting
the Internet of Things (IoT) infrastructure in 5G technology [2].
The extension of NOMA to cellular networks is a recent development. However,
the central principles of NOMA have been studied as successive interference cancel-
lation in the context of Multi User Detection (MUD) and Multi Input and Multi
Output (MIMO) Vertical (V)—BLAST receivers which is one of the main compo-
nents of NOMA. NOMA has been shown to be compatible with existing wireless
systems and can be integrated with other communication technologies to further
improve their efficiency. It can be blended with existing OMA techniques such as
TDMA and OFDMA [3]. NOMA has also been recommended for the 3GPP LTE.
Multi User Superposition Transmission (MUST) enables two serving users simul-
taneously on the same OFDMA subcarrier. NOMA is also being proposed for the
Fifth Generation New Radio (5G NR) standard which is termed as Resource Spread
Multiple access (RSMA) which is ideally suited for asynchronous and grant-free
access. It can be used for IoT and mission critical applications. Thus, NOMA substan-
tially enhances the spectral efficiency of TV broadcasting through superposition of
multiple data streams. Hence, NOMA has immense potential towards enhancing the
spectral efficiency in 5G networks and supporting massive connectivity.
The contribution in this system model is that mathematical analysis of NOMA
probability of unordered outage, NOMA with channel ordering in downlink trans-
mission and NOMA optimal power allocation in downlink transmission have been
derived and then its simulated results have been discussed in detail.
2 NOMA Model
The downlink NOMA model is easier and simple as compared to that of uplink.
In downlink, base station decides to which user allocates what power on basis of
channel gain whereas, in uplink, it is not clear who decides the allocation of power.
Therefore, NOMA downlink model is preferred [4] (Fig. 1).
In this model, (it has been) two users are considered (two users), namely user 1 as
weakest channel and user 2 as strongest channel (respectively). (Now, in this model),
User 1 decodes signal through base station so that it can limit the effect of noise in
user 2.
Hence, in NOMA downlink, base station creates a superposed signal containing
data of both users [5],
√ √
xs = a1 ρs x1 + a2 ρs x2 (1)
√ √
where x 1 and x 2 denote symbols of user 1 and user 2 and a1 ρs and a2 ρs denote
fractions allocated user 1 and user 2 respectively.
The quantity ρ s shows the total transmitting Signal to Noise Ratio (SNR),
Study and Analysis of NOMA Downlink Transmission Model … 135
Fig. 1 NOMA Downlink transmission
Ps
ρs = (2)
σ2
where Ps represents total transmit power.
The received symbol at each user is given by,
√ √
yi = h i xs + n i = h i a1 ρs x1 + a2 ρs x2 + n i (3)
Total power constraint is satisfied with,
a1 + a2 = 1 (4)
Now, NOMA downlink with user 1,

√ √
y1 = h 1 xs + n 1 = h 1 a1 ρs x1 + a2 ρs x2 + n 1 (5)
Hence, corresponding SNR for decoding user 1,
a1 ρs |h 1 |2 a1 ρs β1
γux11 = = (6)
a2 ρs |h 1 | + 1
2 a2 ρs β1 + 1
where β1 = |h 1 |2 and n 1 = 1.
The achievable rate for user 1,

a1 ρs β1
log2 1 + γux11 = log2 1 + (7)
a2 ρs β1 + 1
Here, the desired rate must be less than the maximum rate, otherwise probability
of outage will be faulty and subsequently, no power will be generated.
In the case of NOMA downlink with user 2,
√ √
y2 = h 2 xs + n 2 = h 2 a1 ρs x1 + a2 ρs x2 + n 2 (8)
Hence, corresponding SNR for decoding user 1 at user 2,
a1 ρs |h 2 |2 a1 ρs β2
γux22 = = (9)
a2 ρs |h 2 | + 1
2 a2 ρs β2 + 1
where β2 = |h 2 |2 and n 2 = 1.
After decoding 1st user and successive cancelling interference,
√
y2 = h 2 a2 ρs x2 + n 2 (10)
SNR for decoding user 2 at source 2 is,
ysx22 = a2 ρs |h 2 |2 = a2 ρs β2 (11)
Hence, for given maximum rates of user 1 and user 2, it is therefore necessary
that,
(12)
(13)
and are the minimum guaranteed QoS rates for users 1 and 2 respectively.
3 Analysis of NOMA Model
In this section, mathematical analysis of NOMA probability of unordered outage,

NOMA with channel ordering in downlink transmission and NOMA optimal power
allocation in downlink transmission is discussed in detail.
3.1 NOMA Probability of Unordered Outage
Probability of unordered Outage at user 1 in NOMA,

For given QoS constraints,

log2 1 + γux11 < R̃1
a1 ρs β1
⇒ γux11 = 1+a 2 ρs β2
< 2 R̃1 − 1 = R1 (14)
⇒ (a1 − a2 R1 )ρs β1 < R1
For Probability of unordered Outage < 1,

a1
a1 − a2 R 1 > 0 ⇒ R 1 < (15)
a2
Further, the Probability of unordered Outage at user 1,
Pr {(a1 − a2 R1 )ρs β1 <R1 }

(16)
= Pr β1 < (a1 −aR21R1 )ρs
Let β 1 be defined as,
β1 = |h 1 |2 , E{β1 } = δ12 (17)
PDF of β 1 is given by,
1 − βδ21
f B1 (β1 ) = e , β1 ≥ 0 (18)
δ12
The probability of unordered outage at user 1 is,

Pr β1 < R1
(a1 −a2 R1 )ρs

R1 R1
(a1 −a2 R1 )ρs (a1 −a2 R1 )ρs
= ∫ f B1 (β1 )dβ1 = ∫ 1
δ12
exp − βδ21 dβ1 (19)
0 0 1
= 1 − exp − ρ δ2 (aR−a
1
s 1 1 2 R1 )
Now, Probability of unordered Outage at user 2 in NOMA,
Pr {Cux21 < R̃1 ∪ Cux22 < R̃2 }
Pr {Cux21 < R̃1 ∪ Cux22 < R̃2 } (20)


Cux21 < R̃1 ⇒ log2 1 + γux21 < R̃1
⇒ a2aρ1sρβs2β+1
2
< 2 R̃1 − 1 = R1
⇒ (a1 − a2 R1 )ρs β2 < R1
R1
⇒ β2 < (21)
(a1 − a2 R1 )ρs

Cux22 < R̃2 ⇒ log2 1 + γux22 < R̃2
⇒ a2 ρs β2 < 2 R̃2 − 1 = R2
R2
⇒ β2 < (22)
a2 ρs
Therefore, unordered outage occurs if,
R1 R2
β2 < max , (23)
(a1 − a2 R1 )ρs a2 ρs
PDF of β 2 is given by,
β2 = |h 2 |2 , E(β2 ) = δ22 (24)

1 β2
f B2 (β2 ) = 2 exp − 2 , β2 ≥ 0 (25)
δ2 δ2
The probability of unordered outage for decoding at user 2 is,
R1 R2
Pr β2 < max ,
(a1 − a2 R1 )ρs a2 ρs

R1 R2
max (a1 −a2 R1 )ρs
,a
2 ρs
= ∫ f B2 (β2 )dβ2
0

max R1 ,a R2
(a1 −a2 R1 )ρs
1 β2 2 ρs
= ∫ exp − dβ2
0 δ22 δ22

1 R1 R2
1 − exp − 2 max , (26)
δ2 (a1 − a2 R1 )ρs a2 ρs
3.2 NOMA with Channel Ordering in Downlink
Probability of outage for ordered NOMA for channel gain of user 1 [4],
2
β̃1 = h 1 = min |h 1 |2 , |h 2 |2 (27)
where |h 1 |2 and |h 2 |2 represent weaker channel user and strong channel user
respectively.
System model for ordered NOMA,
√ √
yi = h̃ i a1 ρs x̃1 + a2 ρs x̃2 + n i (28)
PDF of β i ,
βi = |h i |2 , E{βi } = δi2 (29)

1 βi
f Bi (βi ) − 2 exp − 2 , i = {1, 2} (30)
δi δi
CDF of ordered user 1,

FB̃1 β1 = Pr B̃1 ≤ β̃1

= 1 − Pr B̃1 > β̃1

= 1 − Pr min{B1 , B2 } > β̃1 (31)

= 1 − Pr B1 > β̃1 , B2 > β̃2

= 1 − Pr B1 > β̃1 Pr B2 > β̃1
Probability of Outage Ordered NOMA at user 1,

1 β1 β1
Pr B1 > β̃1 = exp − 2 dβ1 = exp − 2 (32)
δ12 δ1 δ1
Similarly,

β1
Pr B1 > β̃1 = exp − 2 (33)
δ1
Since,

β1 β1
B1 β1 = 1 − exp − 2
F × exp − 2
δ1 δ2

1 1 1 1
= + exp −β̃1 +
δ12 δ22 δ12 δ22

1 β1
= 2 exp − 2 (34)
δ3 δ3

Pout = Pr C u 1 < R̃1

a1 ρs β̃1
= Pr < 2 R1 − 1 = R 1
1 + a2 ρs β̃1

R1
= Pr β̃1 < (35)
ρs (a1 − a2 R1 )
R1
ρs (a1 −a2 R1 )
1 β̃1
Pout = ∫ exp − 2 dβ̃1
0 δ32 δ3
R1
ρs (a1 −a2 R1 )
1 β̃1
Pout = ∫ exp − 2 dβ̃1
0 δ32 δ3

R1
= 1 − exp − (36)
ρs (a1 − a2 R1 )
where, user 1 is weaker signal and probability of Outage is increasing at user 1 and
δ32 is smaller.
Channel gain at user 2,
2
β̃2 = h̃ 2 = max |h 1 |2 , |h 2 |2 (37)
CDF of ordered user 2,

FB̃2 β2 = Pr B̃1 ≤ β̃2 Pr B̃2 ≤ β̃2

β̃2 β̃2
= 1 − exp − 2 1 − exp − 2
δ1 δ2

β̃2 β̃2 β̃2
= 1 − exp − 2 − exp − 2 + exp − 2 (38)
δ1 δ2 δ3
1 1 1
= 2+ 2 (39)
δ3
2
δ1 δ2
⎛ 2
⎞
a1 ρs h̃ 2
⎜ 2 ⎟
Pout = 1 − Pr ⎝ 2
≥ R1 , a2 ρs h̃ 2 ≥ R1 ⎠
1 + a2 ρs h̃ 2

φ φ φ
= 1 − Pr β̃2 ≥ = Pr β̃2 < = FB̃2 (40)
ρs ρs ρs
R2 R2
φ= , (41)
(a1 − a2 R1 ) a2
Since, probability of outage at user 2 is decreasing,

φ
Pout = FB̃2
ρs

∅ ∅ ∅
1 − exp − 2 − exp − 2 + exp − 2 (42)
ρs δ1 ρs δ2 ρs δ3
3.3 NOMA Optimal Power Allocation with Downlink
The sum capacity of Downlink NOMA is given as,
CUX11 + cUX22

a1 ρs β1
= log2 1 + + log2 (1 + a2 ρs β2 ) (43)
1 + a2 ρs β1
Since, a1 + a2 = 1

1 + ρs β1
= log2 + log2 (1 + a2 ρs β2 )
1 + a2 ρs β1

1 + a2 ρs β2
= log2 (1 + ρs β1 ) + log2 (44)
1 + a2 ρs β1
where (1 + ρs β1 ) is constant.
2 ρs β2
Hence, for optimal power allocation, one has to maximize, 1+a
1+a2 ρs β1
where
(β2 > β1 ) (45)
The above quantity has been observing monotonically to allocate the power by
increasing the value of a2 with constrained QoS.

1 + a2 ρs β2 1 + a2 ρs (β1 + β2 − β1 )
=
1 + a2 ρs β1 1 + a2 ρs β1
a2 ρs (β2 − β1 ) (β2 − β1 )
=1+ =1+ (46)
1 + a2 ρs β1 ρs β1 + a12
Hence, max sum-rate occurs for maximum value of a2 .

However, this has to be in the feasible region (i.e., QoS constraints) [6],

log2 1 + γux11 ≥ R̃1
a1 ρs β1
⇒ γux11 = ≥ 2 R̃1 − 1 = R1
1 + a2 ρs β1
⇒ (1 − a2 )ρs β1 ≥ (1 + a2 ρs β1 )R1
⇒ ρs β1 − R1 ≥ a2 ρs β1 (1 + R1 )
ρs β1 − R1
⇒ a2 ≤
ρs β1 (1 + R1 )
ρs β1 − R1
⇒ a2 ≤ (47)
ρs β1 (1 + R1 )
Further, we must have,

log2 1 + γsx22 = log2 (1 + a2 ρs β2 ) ≥ R̃2
2 R̃2 − 1 R2
⇒ a2 ≥ = (48)
ρs β2 ρs β2
Thus, a2 lies in the interval,
R2 ρs β1 − R1
≤ a2 ≤ (49)
ρs β2 ρs β1 (1 + R1 )
Further, for above condition to hold, we must have [7],
R2 ρs β1 − R1
≤
ρs β2 ρs β1 (1 + R1 )
R2 ρs R1
⇒ ≤ −
β2 (1 + R1 ) β1 (1 + R1 )
R1 R2 R1 R2
⇒ ρs ≥ + + (50)
β1 β2 β2
4 Result of the System Model
By referring Eq. (19) for probability of unordered outage at user 1 and Eq. (26)
for probability of unordered outage at user 2 respectively, the simulation result in
Fig. 2 shows the fixed value of NOMA. channel gain and channel fading coefficient
parameters are considered. When channel gain increases, probabilities of user 1 and
2 will limit the outage of serving a single user in each orthogonal resource block.
By referring Eq. (34) for probability of ordered outage at user 1 and Eq. (42) for
probability of ordered outage at user 2 respectively, the simulation result in Fig. 3
Fig. 2 Probability of
unordered outage for NOMA
downlink transmission
Fig. 3 Probability of
ordered outage for NOMA
downlink transmission
shows the ordered value of NOMA. Again, channel gain and channel fading coef-
ficient parameters are considered. When channel gain increases, probability of user
1 will limit the outage probability of user 2 due to noise and successive interfer-
ence cancellation with reference to multiuser detection and vertical blast of MIMO
system.
By referring Eq. (50) for optimum power allocation of downlink transmission, the
simulation result in Fig. 4 compares the capacity of optimum power allocation with
capacity of random power allocation in NOMA. It has again considered channel gain
and channel fading coefficient parameters. When channel gain is increasing, capacity
of optimum power allocation and capacity of random power allocation in NOMA
improve the overall system performance.
Fig. 4 NOMA downlink

optimum allocation
5 Conclusion
Hence, capacity of optimum allocation in downlink transmission will produce better

result as compared to that of random allocation due to fairness in order to distribute
available resources equally on high priority basis to get high net throughput.
References
1. J. Ding, J. Cai, C. Yi, An improved coalition game approach for MIMO-NOMA clustering
integrating beamforming and power allocation. IEEE Trans. Veh. Technol. 68(2), 1672–1687
(2020)
2. S.M.R. Islam, N. Avazov, O.A. Dobre, K.S. Kwak, Power domain non-orthogonal multiple
access (NOMA) in 5G systems: potentials and challenges. IEEE Commun. Surv. Tutor. 19,
721–742 (2016)
3. M.F. Kader, M.B. Shahab, S.Y. Shin, Non-orthogonal multiple access for a full-duplex
cooperative network with virtually paired users. Comput. Commun. 20, 1–9 (2019)
4. Z. Wei, J. Yuan, M. Elkashlan, Z. Ding, A survey of downlink non-orthogonal multiple access
for 5G wireless communication networks. ZTE Commun. 14, 17–21 (2017)
5. L. Yu, P. Fan, X. Lei, T. Mathiopoulos, BER analysis of SCMA systems with codebooks based
on star-QAM signaling constellations. IEEE Commun. Lett. (2019)
6. L. Lei, D. Yuan, C.K. Ho, S. Sun, Power and channel allocation for non-orthogonal multiple
access in 5G systems, tractability and computation. IEEE Trans. Wirel. Commun. 15, 8580–
8594 (2018)
7. L. Anxin, L. Yang, C. Xiaohang, J. Huiling, Non-orthogonal multiple access (NOMA) for
future downlink radio access of 5G. China Commun. 12, 28–37 (2015)
8. Y. Sun, Z. Ding, X. Dai, O.A. Dobre, On the performance of network NOMA in uplink CoMP
systems. A stochastic geometry approach. IEEE Trans. Commun. (2020)
9. K. Higuchi, A. Benjebbour, Non-orthogonal multiple access (NOMA) with successive

interference cancellation. IEICE Trans. Commun. E98-B, 403–414 (2015)
10. Z. Ding, Z. Yang, P. Fan, H.V. Poor, On the performance of non-orthogonal multiple access in
5G systems with randomly deployed users. IEEE Signal Process. Lett. 21, 1501–1505 (2014)
Network Structural Analysis Based
on Graphical Measures and Metrics
Atul Kumar Verma , Mahipal Jadeja , and Rahul Saxena
Abstract In the current era of rapid information exchange within a communication

network like social media, computer networks, etc., identifying the communication
pattern and densely interconnected region and analyzing it can yield noteworthy
information. This information gathering may help in identification of suspicious
group activities, identifying any anomaly in the group behavior, etc. In this paper,
use of basic graph theory ideas and concepts has been made to identify densely
interconnected region and patterns of communication within and between different
communities. Further, this information will be utilized to identify any deviation from
the general trend which may give us scope to further investigate the region of network.
Such studies may be employed upon social media networks to identify suspicious
groups. Based on general graph metrics analysis like motifs, community detection,
average degree distribution, etc., various scenarios will be depicted to see how the
evaluated metrics can be used to look for the solutions of the problem. The graphical
models will be mimicked using SNAP (Simulation and Networks Analytics tool) and
machine learning algorithm to evaluate the results depending upon various scenarios.
Finally, the paper will conclude with the analysis of the results obtained on the aspect
that how different situations can be handled based upon the network metrics.
Keywords Communication network · Social media · Densely interconnected

region · SNAP · Community detection · Social networks · Network metrics
A. K. Verma (B) · M. Jadeja

Malviya National Institute of Technology, Jaipur, India
e-mail: 2019rcp9050@mnit.ac.in
M. Jadeja
e-mail: mahipaljadeja.cse@mnit.ac.in
R. Saxena
Manipal University Jaipur, Jaipur, India
e-mail: 2019rcp9153@mnit.ac.in; rahul.saxena@jaipur.manipal.edu
148 A. K. Verma et al.
1 Introduction
Social network means social interconnection and personal relationship through

network. Social networking occurs when one person interacts with other person
online. A graph is plotted by studying and analyzing the data and interconnection
of social network. This graph is known as social network graph where the node
represents people and the line between the nodes, called edges.
Many networks like Internet, email networks, social network and citation network
of documents show community structure which includes set of nodes that are inter-
connected [1]. Community detection displays the concealed yet significant structure
in network such as groups in social network [2].
Communities can be defined by two ways. The first one is clustered approach in
which nodes are densely interconnected as compared to the other nodes of the graph.
Second is the denotation approach in which set of nodes have same interests. So that,
the denotation clusters can be identified by both approaches of the nodes. Social
networks are classic example of graphs with dense interconnection. Communities
are set of nodes which probably share common equity within the graph. A lot of
researches is being focused on network structural analysis for noteworthy informa-
tion. The objective of this paper is the development of a network analysis-based tool
for identifying the communication pattern and densely interconnected region and
analyzing them.
Graph and network theory have various interesting metrics and results which are
handy in finding the solutions to the realistic problems. This paper mainly focuses
on depicting the applicability of basic graph metrics of the network. Support vector
machine (SVM) classifies data in linear as well as nonlinear manner. In order to
categorize in different classes, SVM determines the linear separations in the search
space. It is used to resolve the problems of classification, clustering, etc [3, 4].
In this paper, consider two scenarios. The first scenario deals with the identification
of communities. Look for the community which has dense network, i.e., larger degree
nodes than average degree of graph. Identify the most common patterns throughout
the graph using motifs. Compare the frequency of these patterns with this community.
All rare patterns in the community may be suspected one. In the second scenario,
we ensure that all rare patterns in community are suspected or not. So, we read all
chat log file of the rare patterns and apply support vector machine (SVM) classifier
to ensure whether all rare patterns are suspicious/anomaly or not.
Thus, following the ideology discussed above, the flow of the paper will be as
follows: Section 1 deals with the basic introduction and what will be the focus areas
of the manuscript. Section 2 deals with the state-of-the-art methods dealing with the
computational solutions provided to the problems of interest using various theories
and ideologies. Section 3 covers the scenarios, discussion, mapping the problems to
graphical model and evaluating the graph measures. Section 4 deals with proposed
framework that works on detection of suspicious/anomaly node or person. Finally,
based on the methods proposed and justifications made with respect to the situations,
Network Structural Analysis Based on Graphical Measures … 149
conclusions are drawn and future directions will be discussed over the area and
approach.
2 Related Work
There have been various investigations on network structural analysis using graph
theory ideologies and using lemmas and results interesting views have been
presented. Chen et al. [5] introduced detection of community algorithms that improve
the cluster quality measurements by divide and conquer of the given densely intercon-
nected structure. This algorithm can be used to optimize any densely interconnected
graph quality metric. Newman [6] shows the network, modularity matrix or modu-
larity of particular matrix can be expressed in terms of eigenvectors. This expression
tends to generate spectral algorithm for detection of a community. Yang et al. [7]
proposed communities from vertex attributes an accurate and scalable algorithm
for detecting overlapping clusters in networks with vertex attributes. Tang et al. [8]
discussed the relationship between detecting financial crimes and the social web, and
used select case studies to illustrate the potential for applying social network analysis
techniques. Kumar and Singh [9] proposed a system for detection of the sentiments
in online messages and comments exchanged over social networking and blogging
sites, which monitors communication between users and identifies the messages of
individuals who exhibits anomaly behavior over time. The proposed system is also
able to identify group of people who are discussing about same topics which may be
about a suspicious activity.
Based upon the discussion so far, the current state of the art suggests that though
a lot of works has been done especially in the area of community detection and
anomaly identification in social network, there are very few that have come up with a
generalized framework. The approaches suggested are more inclined to the text data
mining based on the user data available on web. Further, the suspicious information
identification has a lot of insights available on approaches for identification of ‘outlier
nodes’ but the area still lacks to have a framework that can be applied on various
scenarios to produce some vital information. In this paper, an attempt has been made
to propose a framework to accomplish these tasks and using the network structure
analysis and metrics as per the relevance of applicability, the paper will try to suggest
solutions to the problems in various network scenarios.
3 Social Network-Based Measures and Solutions
3.1 Community Detection for Analysis
Community structure exploration is deemed to be the basic analysis technique which

mainly centers on network’s internal structure. Sample network with illustration of
cluster is shown in Fig. 1, where the communities indicate densely interconnected
set of nodes. The capability to identify such network is important in various real-
life applications like networks of friends on social sites, computer communication
networks, etc.; finding strongly interlinked set of nodes in network would give useful
information for its evaluation. Measure of modularity computes the true cluster in a
network given by Eqs. 1 and 2.

M∝ |(no. of edges within groups) − (expected no. of edges within groups)|
s∈P
(1)
• M is the modularity;
• P is the partition of graph;
• S ∈ P is the partitioning of the graph into groups;

1 Kx K y
M(G, P) = Ax y − (2)
2n s∈P s∈P s∈P 2n
• Axy is the edge weight between nodes x and y;

• K x and K y are the sum of the weight of the edges attached to nodes x and y,
respectively;
• 2n is the sum of all the edge weight of the graph.
Fig. 1 Community detection

If the no. of edges in the groups increases from the expected number, then
modularity will be positive. Hence, one should identify those clustered in network
where exists large positive modularity. Apart from community detection using modu-
larity, there are several more methods engaging different techniques like hierarchical
clustering, Girvan–Newman method, etc.
3.2 Detecting Suspicious/Anomalies in Network/Graph Data
An object in a data set is usually called suspicious/anomaly activity if it contradicts

from the normal/known behavior of the data, or it is not connected/similar to any other
object in terms of its characteristics. All real-life application data is interconnected
to each other through networks like social networks of friends, computer communi-
cation networks, the citation networks of documents, etc. From practical outlook, it
is important to explore the rare connections and emerging possibilities. Such explo-
ration results in deliberating anomalous activities by communities in network. Pattern
discovery activity is one of the important approaches to find out anomalies in graph. It
is interesting to note how anomalous behavior of communities represented by graph
pattern and specific method is needed for detection of such patterns. For this purpose,
it will be fruitful to develop tools and techniques to detect suspicious network activ-
ities in multi-dimensional vector data. In Fig. 2, it is possible to have more than one
suspicious in a data set as depicted. There are three isolated objects that are suspected
in this figure.
Fig. 2 Various suspicious in

a data set
3.3 Graphical Analysis Using Stanford Network Analysis

Project (SNAP)
In this section, general graph metrics will be analyzed using SNAP tool. Under-
standing these graph measures and its importance will help to propose the model
for various network structures. Considering the graphical data set obtained from
Wikipedia [10], having edge list between various nodes of the network will be consid-
ered as a reference network with number of nodes equal to 7115 and 103,689 edges
between them. The graphical distribution of the average out-degree (no. of edges
directed from source node to any other node in the network) comes out to be as
shown below.
By analysis of Fig. 3 we get, the out-degree distribution is not a well-distributed
graph (more axial), so a log–log scale distribution curve has been considered to
understand the distribution properly in Fig. 4. These results have been obtained over
Fig. 3 Out-degree
distribution curve
Fig. 4 Log–log out-degree

distribution curve
the Wikipedia edge list network using SNAP library functions. The curve seems to
follow the power law distribution [11] and heavy-tailed distribution. The average
degree comes out to be an approximately four, i.e., four out connections for majority
of nodes in the graph. Similarly, other metric represents path length (distance or no.
of hop counts to reach from one node to other node). The expression for average path
length is given as:
1
h̄ = hxy
2E max x,y=x
where hxy refers to the distance from node x to node y, and E max refers to max. number
of edges [n(n − 1)/2], where n is the total no. of nodes. This measure gives an idea
of the spread of the graph, while the average degree distribution gives the detailing
about how many neighboring nodes can be reached.
4 Proposed Framework
In the proposed framework, system will undergo the following steps to detect the
accurate anomaly, Fig. 5.
Let us take one example for better understanding of this framework with the help
of Fig. 6.
(i) First, consider a social network with lots of nodes that are interconnected to
each other.
(ii) On analyzing this interconnection in networks, we found communities, i.e.,
dense connected subgraph using NMF method [12].
(iii) Now system will randomly select motifs with same pattern [13].
(iv) These motifs may or may not contain suspicious node.
(v) Reading of chat log file present in selected motifs.
(vi) Preprocess data by NLTK [14].
(vii) Extract features based on specified suspicious keywords. If found to be
suspicious ‘1’, else ‘0’. All feature value will be stored in matrix.
(viii) Apply SVM to classify suspicious (class 0) and non-suspicious (class 1)
category [4].
(ix) Based on category, the system will ensure detection and declaration of accurate
anomaly.
5 Conclusions and Future Scope
The paper proposes a framework for the detection of suspicious node in social
network. The state-of-the-art method provides an insight for enhanced and accu-
rate detection of anomalies. The methods and techniques discussed are focused on
Fig. 5 Framework to detect anomaly in network
identifying dense interconnection, thereby analyzing and detecting the suspicious

node in the social network. This analysis provides a technique that enables system
to find out randomly rare patterns present in communities. This will help to analyze
only the rare patterns of subgraphs in the network. However, this approach works on
randomized algorithm for the selection of rare patterns in the communities. So, the
probability of non-suspicious pattern selection may be there. Moreover, this analysis
could be done using multiple methods for the detection of anomaly. Further, there
exists a scope on working out different methods to enhance the detection of anomaly
in the network. The strategies suggested are more focused on identifying the correct
suspicious node, but rare pattern selection is not up to the mark.
In the paper, an attempt has been made to propose a generalized framework
building upon the idea of analysis of the social network on the basis of selection of
Fig. 6 Example of proposed framework
rare patterns or motifs which may contain suspicious node. The idea for identifying
rare nodes of the network explicit have following steps: (i) Community detection;
(ii) random selection of motifs; (iii) feature extraction; (iv) SVM classifier. In the
future, this proposed framework needs to implement using different methods like
SNAP to analyze computational aspects and challenges for detection of anomaly in
the social network.
References
1. M.E. Newman, M. Girvan, Finding and evaluating community structure in networks. Phys.
Rev. E 69(2), 026113 (2004)
2. S. Fortunato, Community detection in graphs. Phys. Rep. 486(3–5), 75–174 (2010)
3. S. Tong, D. Koller, Support vector machine active learning with applications to text
classification. J. Mach. Learn. Res. 2, 45–66 (2001)
4. C. Cortes, V. Vapnik, Support-vector networks. Mach. Learn. 20(3), 273–297 (1995)
5. M. Chen, K. Kuzmin, B.K. Szymanski, Community detection via maximization of modularity
and its variants. IEEE Trans. Comput. Soc. Syst. 1(1), 46–65 (2014)
6. M.E. Newman, Modularity and community structure in networks. Proc. Natl. Acad. Sci.
103(23), 8577–8582 (2006)
7. J. Yang, J. McAuley, J. Leskovec, Community detection in networks with node attributes, in
2013 IEEE 13th International Conference on Data Mining (IEEE, 2013, December), pp. 1151–
1156
8. L. Tang, G. Barbier, H. Liu, J. Zhang, A social network analysis approach to detecting suspi-
cious online financial activities, in International Conference on Social Computing, Behavioral
Modeling, and Prediction (Springer, Berlin, Heidelberg, 2010), pp. 390–397
9. A.S. Kumar, S. Singh, Detection of user cluster with suspicious activity in online social
networking sites, in 2013 2nd International Conference on Advanced Computing, Networking
and Security (IEEE, 2013), pp. 220–225
10. V. Indu, S.M. Thampi, A nature-inspired approach based on Forest Fire model for modeling
rumor propagation in social networks. J. Netw. Comput. Appl. 125, 28–41 (2019)
11. A. Clauset, Inference, models and simulation for complex systems. Lecture Notes CSCI,
70001003 (2011)
12. L.Y. Tang, S.N. Li, J.H. Lin, Q. Guo, J.G. Liu, Community structure detection based on the
neighbor node degree information. Int. J. Mod. Phys. C 27(04), 1650046 (2016)
13. H. Yin, A.R. Benson, J. Leskovec, D.F. Gleich, Local higher-order graph clustering, in Proceed-
ings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining (2017), pp. 555–564
14. J. Yao, Automated sentiment analysis of text data with NLTK. J. Phys.: Conf. Ser. 1187(5),
052020 (2019)
Lung Tumor Classification Using CNN-
and GLCM-Based Features
Amrita Naik and Damodar Reddy Edla
Abstract Identification of malignant tumors in lung CT in early stage is very impor-

tant for early intervention and favorable outcomes. So far, researchers have used
several machine learning architectures to identify the malignant tumors in low-dose
CT scans. In this paper, we use a deep learning architecture which combines CNN
with GLCM-based texture features to classify the lung nodule into benign and malig-
nant. Most of the deep learning model uses softmax activation function, whereas we
use SVM classifier to predict the malignancy in the nodule so as to avoid the risk
of over-fitting. Our model was validated on LUNA dataset on which it achieved an
accuracy of 93.1%, sensitivity of 95.97%, specificity of 86.8% and positive predictive
value of 94.25%.
Keywords Deep learning · CNN · Gray-level co-occurrence matrix · Haralick ·

SVM
1 Introduction
Lung cancer has caused majority of the death worldwide [1]. The survival rate of
the patient decreases as there is delay in diagnosis of malignant lung tumor [2].
Most of the lung cancers are detected at a later stage of disease. Timely screening
of lung tumors can reduce the risk of mortality due to lung cancer. Hence, low-
dose CT scans can be used as a screening tool to diagnose lung cancer at an early
stage. The radiologist has to interpret each of these scans and identify the malignant
tumors. There could be a possibility of observation error by the radiologist. Hence, a
computer-aided diagnostic tool could be used to automate the detection of lung lesion
so as to assist the radiologist in better decision making. Several machine learning-
based decision-making tools like neural network are being used to accomplish this
task. Currently, deep learning models like convolutional neural network are showing
better results in classification of medical images.
A. Naik · D. R. Edla (B)

National Institute of Technology, Goa, India
e-mail: dr.reddy@nitgoa.ac.in
158 A. Naik and D. R. Edla
A neural network requires hand-crafted features as input, whereas CNN extracts

features on its own. CNN is composed of three layers: convolution, pooling and fully
connected layer. The convolution and pooling layer extracts the CNN feature from
the image. However, CNN cannot gather spatial information between the features.
Texture is another important feature which can be used for effectively classifying the
medical images. Gray-level co-occurrence matrix is a statistical method of analyzing
texture and extracts spatial relation of pixel. The accuracy of machine learning algo-
rithm depends largely on features selected for classification. And hence, we plan
to combine both CNN feature and statistical feature like gray-level co-occurrence
matrix to provide better accuracy in detection of malignant tumors in CT. Although
different approaches are proposed for lung nodule classification, a novel approach is
needed using deep learning strategy.
2 Preliminaries
2.1 Convolutional Neural Network
Convolutional neural network (CNN) [3] uses multiple hidden layers to learn features
from image in each layer. Our CNN architecture uses three convolution layers, three
pooling layers, one fully connected layer and a softmax layer. Each convolution
layer is followed by pooling layer and uses a ReLU activation function. The CNN
architecture is explained in brief in Fig. 1.
Fig. 1 Extraction of CNN features

Lung Tumor Classification Using CNN- and GLCM-Based Features 159
The first convolution layer takes a nodule of size 50 × 50 as input. The size of the
kernel in this layer is 3 × 3 with a stride of 1. This layer produces 32 features of size
48 × 48 which is fed to the ReLU layer so that it can only retain the positive values.
The output of the ReLU layer is sent to the max-pooling layer of size 2 × 2 to get
32 features of size 24 × 24. This output is sent to second convolution layer which
uses a kernel of size 3 × 3 and generates 64 features of size 22 × 22. The output of
convolution layer is sent to ReLU layer which is then followed by the max-pooling
layer so as to produce 64 features of size 11 × 11. The output of max-pooling layer
is fed to the third convolution layer with a kernel of size 3 × 3. The output of the
third convolution layer is 128 features of size 9 × 9 and is fed to the ReLU layer.
The output is finally fed to the max-pooling layer to get 128 features of size 4 × 4.
The output of final layers is flattened to get 2048 CNN features. These features
are then fed to the fully connected layer and finally to the softmax layer to predict
the presence of malignancy in the nodule.
2.2 Gray-Level Co-occurrence Matrix
Texture is a very important feature for analyzing images. Texture-based features

are used by most of the researcher and have also provided promising results [4–
6]. Initially, a co-occurrence matrix is created in different directions of θ and with a
specific distance ‘d’ to determine relative position of a pair of pixels in the respective
direction. We then extract statistical feature of texture using gray-level co-occurrence
matrix (GLCM) obtained from the lung nodule in CT scans. In this paper, we consid-
ered the value of d = 1, 2 and θ = 0°, 45°, 90° and 135° and hence creating eight
different gray-level co-occurrence matrices. Haralick and Shanmugam [7] used these
co-occurrence matrices to define 14 statistical features. We extract six texture features
such as contrast, dissimilarity, homogeneity, angular second moment (ASM), energy
and correlation from four different co-occurrence matrices. The formula for each of
these features is given in Table 1.
where M and N are dimensions of the image

M−1 N −1
μx = P[i, j] ∗ i
i=0 j=0

M−1 N −1
μy = P[i, j] ∗ j
i=0 j=0

M−1 N −1
σx = (i − μx )2 P[i, j]
i=0 j=0
Table 1 Haralick features

Feature Formula
Contrast N
M−1 −1
(i − j)2 P[i, j]
i=0 j=0
Dissimilarity N
M−1 −1
P[i, j]|i − j|
i=0 j=0
Homogeneity N
M−1 −1
P[i, j]
1+|i− j|
i=0 j=0
Angular second moment N

M−1 −1
P 2 [i, j]
i=0 j=0

Energy N
M−1 −1
P 2 [i, j]
i=0 j=0
Correlation N
M−1 −1
(i−μx )( j−μ y ) P[i, j]
σx σ y
i=0 j=0
Since there are eight co-occurrence matrices and six features defined for each
co-occurrence matrix, 48 features are gathered from each image.
3 Proposed Work
In this paper, we propose to design and implement a model which in turn combines the
features extracted from CNN with Haralick texture features. Apart from extracting
dominant features from CNN, the model also gathers advantage from statistical
features extracted from gray-level co-occurrence matrix to get spatial information
between pixels.
A brief explanation of the proposed architecture is demonstrated in Fig. 2.
The input to the model is a segmented lung nodule. This lung nodule is then fed
to the CNN model to extract CNN features. The image is passed through 3 layers
Fig. 2 Proposed architecture

of CNN model, wherein each layer consists of convolution and pooling layer. The
output of feature extraction phase of CNN model is a feature map of size 2048. Before
the feature is extracted from the image, we also train the CNN model with training
datasets so as to update the weights. The lung nodule is also converted into eight
GLCMs for θ = 0°, 45°, 90°, 135° and d = 1 and 2. We extract 6 statistical features
such as contrast, dissimilarity, homogeneity, angular second moment, energy and
correlation from each GLCM. Thus, we extract 48 Haralick texture features from the
nodule. Both the 2048 CNN features and 48 Haralick features are concatenated and
fed to SVM classifier to predict the malignancy of the nodule which could be either
benign or malignant. SVM has proved to be a very good two-group classifier with
high-dimensional feature map [8].
4.1 Experimental Setup
The dataset was evaluated with i7 processor, NVIDIA GeForce RTX 2070 (8 GB
onboard memory). The implementation of our networks is based on Keras 2.2.4
library with TensorFlow as back end.
4.2 Data Description
We used data from publicly available dataset LUNA16 [9] which contains 888 CT
scan images collected from LIDC-IRDC. The dataset uses 551,065 annotations, out
of which, 1351 were labeled as malignant and the rest are benign. The scans of each
person contain 200 slices, each of size 512 × 512 pixels. A team of four radiologists
have evaluated these nodules, and if the average score of the four radiologist is more
than 3, the nodule is considered to be malignant.
4.3 Simulation
LUNA dataset also maintains an annotation file which keeps record of the x, y,
z coordinates of the nodule. Nodules were then cropped based on the coordinates
mentioned in the annotation. The database consists of 6131 training sets, 1903 test
sets and 1534 validation datasets. Initially, we train the CNN with 70 epochs. The
learning rate and batch size were set to 0.02 and 100, respectively. The loss function
used for training was categorical cross-entropy. In this experiment, Adam optimizers
have provided better training performance and converge faster. After every epoch, the
Table 2 Accuracy of the

Classifier Accuracy Precision
proposed work with different
classifiers SVM 93.10 86.88
Decision tree 91.32 83.27
Linear regression 91.69 82.82
Naive Bayes 90.22 80.35
K-nearest neighbor 90.90 88.69
AdaBoost 91.09 81.51
weights were updated. These updated weights were then used to generate the CNN
feature map for training and testing. During training process, we use the trained CNN
model to generate the CNN features for the nodule. We have used 80% of the dataset
for training and 20% of the data for testing.
Table 2 gives a summary of accuracy of the proposed model with different
classifiers.
It is evident from Table 1 that SVM has proved to provide better results when
compared to other classifiers. SVM classifier provides an accuracy of 93.1%, sensi-
tivity of 95.97% and specificity of 86.8% with the proposed architecture. Figure 3
gives a plot of training score and testing score of SVM classifier for different values
of gamma parameter. The training and validation score increases as the value of
gamma increases indicating that there is no over-fitting issue.
Fig. 3 Validation curve with SVM classifier

5 Conclusion
In this paper, a computer-aided detection system to classify lung nodules using

CNN and GLCM features and support vector machines was proposed. The proposed
system has provided accuracy of 93.1% and positive predictive value of 94.25%. We
also observed that SVM classifier has provided better classification accuracy when
compared to other classification algorithm.
References
1. F. Bray, J. Ferlay, I. Soerjomataram, R.L. Siegel, L.A. Torre, A. Jemal, Global cancer statistics
2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185
countries. CA: Cancer J. Clin. 68, 394–424 (2018)
2. S.C. Byrne, B. Barrett, R. Bhatia, The impact of diagnostic imaging wait times on the prognosis
of lung cancer. Can. Assoc. Radiol. J. 66(1), 53–57 (2015)
3. A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural
networks, in Advances in Neural Information Processing Systems (2012), pp. 1097–1105
4. A. Kumar, S. Mukhopadhyay, N. Khandelwal, 3d texture analysis of solitary pulmonary nodules
using co-occurrence matrix from volumetric lung CT images, in Proceedings of the SPIE Medical
Imaging Conference, Lake Buena Vista, Florida, USA, 28 Feb 2013 (SPIE, Bellingham, WA,
2013), pp. 1–4
5. J. Wang, H. Wu, T. Sun, X. Li, W. Wang, L. Tao, et al., Prediction models for solitary pulmonary
nodules based on curvelet textural features and clinical parameters. Asian Pac. J. Cancer Prev.
14, 6019–6023 (2013). https://doi.org/10.7314/apjcp.2013.14.10.6019
6. H. Wu, T. Sun, J. Wang, X. Li, W. Wang, D. Huo, et al., Combination of radiological and gray
level co-occurrence matrix textural features used to distinguish solitary pulmonary nodules by
computed tomography. J. Digit. Imaging 26, 797–802 (2013). https://doi.org/10.1007/s10278-
012-9547-6
7. R.M. Haralick, K. Shanmugam, Textural features for image classification. IEEE Trans. Syst.
Man Cyber. 6, 610–621 (1973). https://doi.org/10.1109/tsmc.1973.4309314
8. C. Cortes, V.N. Vapnik, Support-vector networks. Mach. Learn. 20(3), 273–297 (1995)
9. A.A.A. Setio, A. Traverso, T. De Bel, M.S. Berens, C. van den Bogaard, P. Cerello, H. Chen,
Q. Dou, M.E. Fantacci, B. Geurts et al., Validation, comparison, and combination of algorithms
for automatic detection of pulmonary nodules in computed tomography images: the LUNA16
challenge. Med. Image Anal. 42, 1–13 (2017)
A Distributed Implementation
of Signature Matching Algorithms
for IoT Intrusion Detection
Menachem Domb
Abstract Common software systems manage large databases and huge incoming
and outgoing data. A major security risk is the intrusion of viruses via devices
connected to the system network. Each virus has a unique signature. Early detection
of these signatures is the core of most anti-virus software packages. Each anti-virus
uses a multi-string-matching algorithm to detect the existence of a virus signature
in the incoming data or a given data file. String matching uses pre-defined signature
lists, which are growing fast and may contain a few million entries. Studies show that
anti-viruses may have an impact on the data transmission performance and propose
ways to cope with it either by enhancing the processing power, improvements to
the matching algorithms and apply selective detection with low-risk probability. The
introduction of the Internet of things (IoT) exposes the Internet to major security
risks as numerous devices are expected to be connected to the Internet. IoT devices
have poor processing resources, and it is not able to run himself proper anti-virus
software and so, allowing malicious virus attacks without any way to prevent it. Some
solutions propose incorporating in the IoT environment a powerful security gateway
processor to handle the anti-virus aspect. This solution is against the concept of IoT
as an independent entity able to cope with such challenges. This paper presents a hier-
archical distributed and parallel processing framework that accelerates multi-string
matching for IoT devices by utilizing the accumulated excess processing power of the
local IoT devices network, eliminating the need for high capacity robust computers.
Keywords Anti-virus · Multi-string-matching algorithms · Signature matching ·

NIDS · Intrusion detection · IoT · Parallel and distributed processing · Data
transmission security
M. Domb (B)
Ashkelon Academy College, Ashkelon, Israel
e-mail: dombmnc@edu.aac.ac.il
166 M. Domb
1 Introduction
The recent increase in the number of devices, the traffic volume, introduction of
IoT, cloud computing, and other Internet expansions has made our dependency on
data networks absolute and irreversible. As a side effect, it encourages intentional
attempts to penetrate the network and utilize the transmitted data for other purposes
or simply to cause malicious damage. Basic protection from such malicious attacks
is to discover the penetrator’s unique pattern character-string or signature and block
it. Assuming such a signature is discovered, it is added to a list of universal signatures
and distributed to the public via anti-virus systems. Each local network connected
to the Internet is expected to scan every incoming and outgoing data packet to
ensure it does not contain any of the signatures listed in the corresponding table.
Once such a signature is discovered, the file transmission is blocked, and the entire
incoming file is removed. The comparison is done by a matching algorithm. Some of
the popular string-matching algorithms include Aho–Corasick, Knuth–Morris–Pratt,
Rabin–Karp, Boyer–Moore, and Wu–Manber. The Aho–Corasick algorithm uses a
finite state machine that provides the best worst-case performance but requires a
significant amount of memory for the state machine. Boyer–Moore provides the best
performance when searching for a single signature but scales poorly. The Wu–Manber
algorithm is similar to Boyer–Moore but consumes less memory than Aho–Corasick
and provides a better average-case performance [1]. Practical implementations of the
above string-matching (SM) algorithms show that it may not be sufficient for current
high-speed and capacity networks and new modalities such as cloud computing and
IoT.
IoT devices have limited capacity in computer power, memory size, and storage
space and thus demonstrate poor performance when processing heavy applications.
Compromising on string-matching algorithms (SMA) for IoT puts the entire Internet
data transmission at risk, especially due to the forecast of numerous additions of IoT
devices to the Internet. This work proposes a framework implementation of string-
matching technology adjusted to the IoT environment. It is based on distributed and
parallel processing of SMA by a network of IoT devices. It is done by splitting the
matching process into micro-processes where each processes a short piece of the
input string and scans a limited number of signatures.
In the next section, we present the problem statement. The rest of this paper
outlines the related work, the proposed solution, an example, feasibility experiment
and conclusions, and future work.
2 The Problem Statement
Common implementations of existing string-matching algorithms (SMA) are costly

concerning the memory space and cache memory it consumes. Seemingly, there is no
proper string-matching mechanism applicable for IoT devices. We could solve this
A Distributed Implementation of Signature Matching Algorithms … 167
by deploying a dedicated security-server that will handle all security aspects for the
IoT network including the execution of SMA. However, such solutions complicate
the incorporation and integration of IoT on the Internet and eliminate the entire IoT
concept. Hence, there is a real need for an appropriate solution, which we propose
in this article.
3 Related Work
Data string-matching algorithms are used to detect malicious and malware files by
monitoring incoming and outgoing data through the network. The main function
matches data patterns at hand with pre-defined lists of malicious data patterns. Some
commonly used algorithms are Boyer–Moore, Aho–Corasick, Naïve String search,
Rabin–Karp String Search, and Knuth–Morris–Pratt. These algorithms are described
in [2, 3] and were compared to determine which algorithm is most efficient for
intrusion detection. Obeidat and Alzuby [4] propose a new matching approach that
uses an array for each character and stores its index and then attempts to match it
using the character index search within the indexes array created. Several works use
graphics processor units (GPU) to accelerate expression pattern matching. GPUs
have the bandwidth required to process large data volumes and heavy transmission
of data streams. A traditional deterministic finite automaton (DFA) approach and a
new state machine for extended finite automaton (XFA) are proposed to accelerate
regular expression matching on GPU. In [5], a hierarchical parallel GPU machine
is proposed to accelerate regular expression matching and resolve the problem of
its state explosion. The results show that it demonstrates significant improvement
on performance compared to other GPU implementations using bloom filtering. An
advanced proposal of parallel processing of the AC algorithm is described in [6]. It
proposes a revised version of AC which works on an integrated CPU-GPU config-
uration. Experiments show that its performance is better than when CPUs or GPUs
are used exclusively. A comprehensive review of the distributed implementation of
matching algorithms is outlined in [1], from which we have adopted some basic
ideas for our proposal. These proposals focus on running SMAs concurrently using
threading, which requires powerful CPUs and GPUs. This approach prevents SMAs
from being executed in an IoT constraint environment. The approach we present is
based on condensing SMA algorithms so that they consume much fewer computing
resources and dividing their input into small pieces so they can be processed by
microprocessers such as IoT devices.
4 The Proposed Solution
Due to the poor processing resources with which IoT devices are equipped, they
are unable to efficiently apply basic security measures such as detecting malicious
168 M. Domb
incoming data transmissions. The common detection approach discovers the signa-
ture in the malicious data packets, which is done by string-matching algorithms
given a list of known signatures. All the existing efficient string-matching algo-
rithms consume processing resources close to the maximum capacity of most IoT
devices, which may slow down the overall data transmission.
We propose a comprehensive implementation system that operates properly in
the IoT environment. It uses reference tables to manage the entire detection process,
starting with dividing the detection process into many micro-processes, which are
distributed to a network of IoT devices and simultaneously launch the SMAs to
analyze the input data packets. The implementation is divided into two main stages,
i.e., the setup and preparation stage and the execution stage. In the setup and prepa-
ration stage, several reference tables are defined and loaded to each IoT device
connected to the local network. When an IoT device is about to accept data pack-
ages, it launches the control and monitoring system, which analyzes the reference
tables and distributes each data packet to the devices that execute the corresponding
SMA. Once the SMA process is done, the results are sent to the initiated IoT device
to integrate the results received and react accordingly. Figure 1 details the steps in
the two stages of the proposed control and monitoring system.
Figure 1 depicts the five consecutive steps of the framework system. The first stage
is setting-up the parameters the system will follow when executed. The first reference
table is the IoT devices and its attributes, such as device ID, processing power,
memory size, storage space, availability, overall strength rank, SMA preferences,
maximum input string, and parallel type. The second table is the input file attributes,
such as packet size.
In the first step, the system analyzes the reference tables to determine the strategy
of how to distribute the SMA tasks to the IoT devices. This includes which devices
will participate in the detection process for the incoming file, which SMA will be
executed in each device, the input string size, the list of malicious signatures, signature
portion, and parallelism type. In step 2, The SMA processes are launched with the
above parameters. In step 3, the incoming data is manipulated and continuously sent
in step 4 to the participating IoT devices for signature detection. If the signature has
been found, the system takes the proper action in step 5.
The following sections elaborate on the use of IoT devices to run SMAs and the
improvements in addition to the above basic decryption.
4.1 Alignment of SMA to Properly Operate on IoT Devices
Based on [5, 1], it is apparent that several string-matching algorithms require

moderate processing power and resources when applying the detection process
towards a reasonable input packet size and an average signature length. Even the
Aho–Corasick matching algorithm, which normally consumes a large memory size,
has been optimized by compression of the state representation and more [5]. Now, it
SMA Monitoring & Control System

Reference Tables
a. IoT devices attributes
b. Input file attributes
1 Analyze reference tables
2 Set-up and initiate the IoT devices
3 Distribute data packets to devices
End of
4 IoT1 IoT2 IoT3 IoTn
5 Analyze input and distribute final result
Fig. 1 Framework main components and processes
seems to be able to properly operate in high-end IoT devices. Therefore, it is recom-

mended to perform a comprehensive performance test of the most popular SMAs for
all types of IoT devices used in the specific implementation and associate the best
SMA suitable to each IoT device.
This framework proposes a systematic way of optimizing the deployment of SMAs
for a given implementation by reducing its space consumption and minimizing its
execution time. The descriptions provided in the following subsections are depicted
in Fig. 2 to clearly illustrate how the framework operates and how the improvements
are implemented.
170 M. Domb
SMA Distributor
Initiate
Devices
Divide Packets & Distribute
IoT IoT IoT IoT IoT

1
Divide Signature & Distribute
2
IoT IoT IoT IoT IoT
3
IoT IoT IoT IoT IoT
5 If Sig. 4
Integrate Results & Update
Found
Fig. 2 Splitting input packets and dividing the signature file and length
4.2 Splitting Input Packets into Shorter Strings
The processing time and the associated consumed resources depend on the algo-
rithm’s underlying concept and its actual implementation, the length of the input
string, the length of the corresponding signature, and the number of signatures
searched for. The standard execution of an SMA assumes a data packet as input,
which may be too big to handle in one cycle. However, when we divide each packet
into smaller strings, it may be processed by the IoT device. The size of each packet
portion will be adjusted to the specification of the IoT device expected to process it.
The input can be handled sequentially or parallelly. In the sequential mode, all
portions of the same packet are processed sequentially by the same IoT device, and
in the parallel mode, all portions of the same packet are processed at the same time,
each portion by a different device. In the first mode, we ensure that the entire packet
is processed by the same device to avoid missing any signature, assuming it keeps
the sequence and continuity of the pieces of the same packet and between packets.
When a signature is discovered, the process is completed, the file is blocked, and
all accepted packets of this file are purged, and notification is broadcasted to the
IoT network. In the parallel mode, when a signature is discovered by one device, it
notifies the other devices working on the same packet to stop working on the piece at
hand and wait for the next file, as the current file is infected. In this mode, a signature
may go undetected when one piece has just the beginning part of the signature, and
the remaining part of the signature is processed by another device. To handle such
cases, devices should be synchronized by using an internal messaging system that
allows an instant message to be sent to the other devices processing portions of the
same packet. To ensure a match of the two edges of the signature, an overlapping
of a bit between any two pieces of the same packet can be added. If needed, the
Harris stitching algorithm [7] can be employed. Arrow number 1 in Fig. 2 depicts the
possibility of a hierarchical division of the input data file into subfiles in several levels
enabling parallel processing of signature detection in its packets. This also enables
the possibility of dividing long packets into shorter strings and forwarding them to
a preferred IoT device for processing. Arrow 5 in Fig. 2 depicts the process where
an IoT device discovers a complete or a partial [at the string end/start] signature,
notifies the results of the integration step, which analyzes the message, concludes
the results, and forwards it to the main process, which broadcasts it and stops all
activities related to the defected file.
4.3 Dividing the Signature
Aldwairi et al. [8] conducted a performance experiment of running the Wu–Manber

matching algorithm using FPGAs hardware. The achieved performance has improved
as expected. However, for long input strings, the performance decreased. Therefore,
if the signature is significantly long, it can be divided into smaller parts. The search
can be done serially or parallelly. In a serial mode, the same device sequentially
checks the pieces of the same signature. If a signature piece does not appear in the
packet, the process will skip the current signature and move to the next signature.
In the parallel mode, if one of the devices realizes that the signature piece does not
appear in the packet, it will notify the other devices to skip to the next signature and
thereby expedite the process. Arrow 3 in Fig. 2 depicts this process. The input string
is repeatedly checked against all the pieces of the same signature.
4.4 Dividing the Signatures Table
At the same experiment described in [8], it was observed that for a large signatures
database, the performance decreases, which means that smaller signature tables are
172 M. Domb
recommended. To shorten the processing time, we divide the signatures table into
pieces such that it fits the capacity of each device that is expected to process it.
Similar to the previous section, this can be executed serially or parallelly, where each
approach has its pros and cons. Arrow 2 in Fig. 2 depicts this option by repeatedly
checking the same input string against the entire list of signatures, even when it is
divided into several sub-tables. Arrow 3 in Fig. 2 provides an additional option of
checking the same input string by a series of IoT devices, where each device has a
different sub-table with a unique list of signatures.
Figure 2 depicts the entire process including its main loops. It starts with selecting
the IoT devices that will be involved in the malicious signature’s detection process.
The selection is done by scanning the devices’ attributes table. Then, it initiates the
selected devices to ensure its readiness to execute the process. Once this step is done,
the process divides the load among the IoT devices.
5 Example
To better understand the proposed system, we provide a simple example that follows
the steps illustrated in Fig. 1. Below are the tables with the setup parameters in the
example. In this example scenario, a file of 30k bytes packet length is transmitted;
the system scans each packet to discover any malicious signature. A signature of 54
bytes appears in the 56th packet in the last 23 bytes of the second piece, and the
remaining 31 bytes appear at the beginning of the third piece of the packet. IoT1
found the first part, and IoT2 found the rest of the signature. The control system is
expected to connect the two matching partial signatures into one signature and react
accordingly.
5.1 Analysis of the Reference Table
Before a stream of data is scheduled to appear, the system sets the optimized config-
uration parameters to ensure the best detection process performance. To do so, it
collects the input attributes from Table 2 and the configuration attributes from Table 1.
In this example, the system evenly distributes the input file among the three IoT
devices in a round-robin sequence.
5.2 Set up and Initiate the IoT Devices
The system triggers IoT1, IoT2, and IoT3 devices and launches the packet formatting
process and the associated SMA in each device.
Table 1 IoT devices

Attribute IoT1 IoT2 IoT3
attributes
Device ID D10 D20 D30
Processor type (mips) 200 300 400
Memory size (Mb) 50 100 150
Storage space (Mb) 300 500 700
SMA name SM1 SM2 SM3
Max string size (bytes) 3k 9k 15k
String parallelism (V/H) V H V
Signature tables ID ST1 ST2 ST3
Max signature length (b) 500 900 1300
Signature parallelism H V H
Availability A1 L2 L3
Table 2 Datastream
Attribute Input 1 Input 2 Input 3
attributes
Type File Stream File
Packet length (k) 30 NA 40
Total size (Mb) 500 NA 1000
5.3 Distribute Data Packets to the Devices
The input file fits the definition entered in Input 1 in Table 2. The system accepts
data packets [30k] and splits each into pieces according to the lengths defined in
Table 1 of the corresponding device that is expected to process it. In our example,
the packet is divided into 5k, 10k, and 15k bytes and is forwarded to IoT1, IoT2,
and IoT3, respectively. No signature is detected in 55 packets that pass through the
system and the SMAs. However, when the 56th packet arrives, it is divided into
three pieces and forwarded to the three devices, respectively. SM1 detects a partial
signature match of 23 bytes at the end of the analyzed piece. It then sends a message
to the system. At the beginning of the piece, SM2 also detects a partial signature of
21 bytes complying to the end of the signature and sends a message to the system.
The system integrates the two partial matches discovered and realizes that the file
is affected. It notifies the source device that owns the file transmission and reacts
according to the organization’s policy.
6 Feasibility Check and Experiment
The target in this experiment was to ensure the feasibility and applicability of the
proposed framework and simulate the entire process. We developed two software
174 M. Domb
modules, (1) a pseudo-SMA similar to the compressed version of the Aho–Cora-

sick algorithm as described in [5]. The similarity focuses in terms of processing
complexity and resource consumption; (2) a control and monitoring system that
manages the input file detection process. It starts with accepting packets, dividing
it into smaller pieces, and distributes the pieces to the IoT devices connected to the
same local network. Each IoT device applies the pseudo-SM algorithm to detect
if it contains a malicious signature. In case of a positive match, the system reacts
as defined. We implemented the framework using four Arduino IoT devices. All
devices were loaded with the pseudo-SM algorithm along with the control and moni-
toring system. The data packets were taken from a billing system of a cellular phone
company. The signatures have been taken from the Broadcom Web site following the
instructions in [9]. For testing purposes, we randomly incorporated signatures within
the data packets. The execution of the example went well and proved that the frame-
work can deliver the service it is expected to provide. Although the experiment’s
purpose was just to prove the framework feasibility, we ran an initial performance
test, but the results were not distinct, and hence, this is left to be conducted in the
future.
7 Conclusions and Future Work
The recent growing trend of IoT integration in the Internet broadens its security risk.
This is due to its poor capacity and resources which limits it from implementing
proper security measures, especially the detection of malicious strings mixed in
incoming or outgoing data. This paper proposes a comprehensive framework for
enabling a variety of string-matching algorithms to properly operate in the IoT envi-
ronment and thereby close the security gap created by the IoT. The framework is
based on a combination of condensed string-matching algorithms and a variety of
techniques to better control the size of the input to be processed at a time. Future work
will focus on implementing the framework by developing an open, robust, and effi-
cient framework that easily can be integrated into heterogeneous configurations. Run
a comprehensive implementation to evaluate the framework performance impact.
References
1. C.V. Kopek, E.W. Fulp, P.S. Wheeler, Distributed data parallel techniques for content-matching
intrusion detection systems, in MILCOM 2007—IEEE Military Communications Conference,
Orlando, FL, USA, 2007, pp. 1–7
2. V. Dagar, V. Prakash, T. Bhatia, Analysis of pattern matching algorithms in network intru-
sion detection systems, in 2016 2nd International Conference on Advances in Computing,
Communication, & Automation (ICACCA) (Fall), Bareilly, 2016, pp. 1–5
3. N. Tran, M. Lee, S. Hong, J. Bae, Performance optimization of Aho-Corasick algorithm on a

GPU, in 2013 12th IEEE International Conference on Trust, Security and Privacy in Computing
and Communications, Melbourne, VIC, 2013, pp. 1143–1152
4. I. Obeidat, M. Alzuby, Developing a faster pattern matching algorithms for intrusion detection
system. Int. J. Comput. (S.l.), 278–284 (2020). ISSN 2312-5381. Available at: computingonl
ine.net/computing/article/view/1520
5. C.H. Lin, et al., A novel hierarchical parallelism for accelerating NIDS using GPUs, in Proceed-
ings of the 2018 IEEE International Conference on Applied System Invention (ICASI) (IEEE,
2018), pp. 578–581
6. V. Sanz, A. Pousa, M. Naiouf, A. De Giusti, Efficient pattern matching on CPU-GPU hetero-
geneous systems, in Algorithms and Architectures for Parallel Processing. ICA3PP 2019, ed.
by S. Wen, A. Zomaya, L. Yang. Lecture Notes in Computer Science, vol. 11944 (Springer,
Cham, 2020)
7. S. Mistry, A. Patel, Image stitching using Harris feature detection. Int. Res. J. Eng. Technol.
(IRJET) 03 (2016). www.irjet.net. p-ISSN: 2395-0072, e-ISSN: 2395-0056
8. M. Aldwairi, Y. Flaifel, K. Mhaidat, Efficient Wu-Manber pattern matching hardware for intru-
sion and malware detection, in International Conference on Electrical, Electronics, Computers,
Communication, Mechanical and Computing (EECCMC), Tamil Nadu, India, 28–29 Jan 2018
9. R. Bryan, Manually add virus signatures to new installation package. Broadcom. https://
community.broadcom.com/symantecenterprise/communities/community-home/digestviewer/
viewthread?MessageKey=df36eff8-2acb-4982-86ce-f95929489213&CommunityKey=1ec
f5f55-9545-44d6-b0f4-4e4a7f5f5e68&tab=digestviewer#bmdf36eff8-2acb-4982-86ce-f95
929489213
10. C. Hung, C. Lin, P. Wu, An efficient GPU-based multiple pattern matching algorithm for
packet filtering. J. Signal Process. Syst. 86, 347–358 (2017). https://doi.org/10.1007/s11265-
016-1139-0
Graph-Based Multi-document Text
Summarization Using NLP
Abhilasha More and Vipul Dalal
Abstract Text summarization is the process of generating the condensed view of

the text by selecting useful and relevant information from the original source docu-
ments. It is a sub-topic of natural language Processing. Text summarization is a
technique for understanding the aim of any document, to visualize large text docu-
ment within short duration. Summarization provides flexibility and convenience.
Business leaders, analyst, and researchers must undergo huge number of documents
on a day-to-day basis to stay ahead, and an oversized portion of their time is spent just
deciding what document has relevancy and what is not. By finding out important and
relevant sentences and creating summaries, it is possible to quickly check whether or
not a document is worth reading. In this paper, we propose to design automatic text
summarizer to summarize the multiple text documents. The input to the system is the
multiple sources of news articles. Important sentences from the source document are
selected and arranged in the destination documents or the summarized documents.
This is called as the extraction technique in automatic text summarization. Here, one
document visited is being processed, and its summary is generated using extraction
technique through which a weighted graph is constructed using language processing.
Keywords Text summarization · Natural language processing · Weighted graph ·

Text rank algorithm · Preprocessing · Extraction
1 Introduction
Voluminous information is available to us over the Internet due to the widespread

of the World Wide Web. There is information available through various sources
pertaining information in the form of news legal documents, journals related to
A. More (B) · V. Dalal

Vidyalankar Institute of Technology, Mumbai, India
e-mail: maurya.abhilasha@gmail.com
V. Dalal
e-mail: Vipul.dalal@vit.edu.in
178 A. More and V. Dalal
various genres, e-magazines, e-books, etc. Daily e-new sources provide us with news
containing information on various topics like politics, sports, classifieds, etc. Update
of some particular topics can re-appear frequently for few days with some variations.
Usually people just want a gist of what the news is about due to lack of time to read
all the sources of news thoroughly. Here, a text summarization tool is needed that
can save a lot of time.
Automatic text summarization is condensing or summing up the original text
to a summary without being meaningless. The condensed text or summary can be
extracted from a single source document or from multiple sources of documents,
which can be further classified as single document summarizer or multi-document
summarizer, respectively. To read this, available immeasurable information form
online resources can be very tedious and time consuming for the users linked with
different sectors of interest. A text summarizer can be used by wide variety of people
to read a concise document. For example, it can be used by a person to catch up on
news on the go, or for a lawyer to get a last minute summary, a student to understand
a concept, etc. Therefore, a summarizer can be used as a tool for any person to have
a condensed view of any article within less time.
Text summarization is the sub-topic of natural language processing which helps
to encapsulate the enormous data available online. The text summarizing tool takes
as input a document that can be one independent document or a stack of documents,
which the user can select from wide variety of sources like the World Wide Web,
user’s personal documents, or e-books. Then, it performs preprocessing of the input,
removes stop words, and then tokenizes. A sentence analysis is then performed which
includes sentence scoring and sentence ranking method. After ranking the sentences,
the summary of these documents is generated that can be used as a quick access for
the user.
Text summarization can be broadly classified as extractive summarization and
abstractive summarization. Extractive summarization: A few selected words is taken
from the text and merged together to form a summary. It can be thought of as a
highlighter that selects the main information from a source text [1]. In machine
learning, extractive summarization usually involves weighing the essential sections
of sentences and using the results to get summaries. Different types of algorithms
and methods can be used to gage the weights of the sentences and then rank them
according to their relevance and similarity with one another—and further joining
them to generate a summary.
Abstractive summarization: It uses semantic understanding to select words, even
those words that do not appear in the source documents. They understand the text
using advanced natural language techniques and generate a concise text that is
completely a new text. This text aims at conveying the most critical information
from the original text.
Graph-Based Multi-document Text Summarization Using NLP 179
2 Literature Review
Dutta et al. [2] devised a methodology to summarize the tweets. Tweets are taken
as input, and a subset, i.e., the condensed view of these tweets, is generated.
The challenge in summarizing the tweets is that they very small to be provided
as input, and also the language used in tweets is informal. Here, the similarity
between the tweets is generated by the semantic similarity between the tweets.
Similar tweets are clustered using community detection techniques through which
a representative tweet is selected.
Natesh et al. [3] propose a methodology of graph-based text summarization. NLTK
is used for tokenizing, POS tagging, pronoun resolution, and sentence extraction.
A graph-based ranking algorithm is used to generate weighted graph, and then, a
shortest path algorithm is applied to create the summary of the input documents.
Similarity between the sentences is found using the method of parts of speech and
relation among the sentences. Ranking of these sentences is done using a ranking
algorithm, and a graph is created depending on the closeness of sentences.
Thakkar et al. [1] proposed the methodology of unsupervised learning on input
text to extract sentences using graph-based ranking algorithms and shortest path
algorithms for the users to view the summary on small screens like a mobile phone
within a short duration of time. Extractive summarization method is used where
parts of speech and sentences are selected and then a shortest path algorithm is
applied to it.
Gulati and Sawarkar [4] have proposed a method of summarizing multiple docu-
ments in Hindi language. They have used the summarization concept of prepro-
cessing and extraction of feature terms. The extracted summary of these docu-
ments will then generate a consolidated single summary using fuzzy logic for
ranking the sentence based on weights of the sentences. Hindi as a language is
chosen as majority of people in India use this as a language to communicate that
is a national language of the country.
Dave and Jaswal [5] propose a method of multi-document text summarization,
wherein five input documents have been considered for summarizing that can
include around 10,000 sentences. The documents can be in the form of PDF,
DOC, or TXT. Comparison of existing tools is shown in the paper. The paper
emphasizes on multi-document text summarization methods using human-based
evaluation to make the summary readable.
Naik and Gaonkar [6] propose a single document text where a rule-based summa-
rizer is used. 15 news articles are provided as an input to the system from DUC
2002 dataset. The output of this summarizer is compared to GSM summarizer to
get better results than the existing summarizer.
3 Problem Statement
Thousands of pages can be referred to find information of any type. Millions of news
sources providing thousands of different genres of news are available every day. To
remain updated with the current events in the society or the environment, one has to
read through all these news articles; however, reading through all these sources is a
tedious task, especially during any special event like a cricket or a football match,
natural disaster, a movie release, elections, etc. News related to such events can be
too much to read. We aim to design a text summarizer for news articles for a user
who wants to keep oneself updated about more than one genre of news. The user
can select the various documents through which the news is to be read and get the
summarized content that is relevant, meaningful, and precise.
3.1 Objectives
Below is the list of objectives for the text summarization tool:

• Designing a multi-document news articles summarizer.
• Design a user-friendly summarizer.
• Reducing redundancy in the generated summary.
• Generate a weighted graph to show dependency of sentences.
4 Proposed System
In this project, we aim to design automatic text summarizer to summarize the multiple
text documents. The input to the system will be multiple sources of news articles.
Firstly, preprocessing on the input text document will be performed. Then, an extrac-
tive summary will be generated by selecting words of importance from the input text
and arrange them in the output. The last step would consist of ranking the sentence
based on feature weights and thus generating the output, i.e., the summarized content.
The following steps will be included in preprocessing the input [2, 3]:
i. Tokenizing:
Tokenizing can be performed on a paragraph or on individual sentences.
Tokenizing a paragraph means to split the paragraph into independent sentences,
and tokenizing a sentence means to split every word in a sentence. Every word
that is split is an independent token.
ii. Tagging parts of speech:
Each generated token is then identified as a relevant part of speech depending on
the tags of neighboring words. Semantic analysis of the sentence is performed
in this step which identifies to be the most important part of processing the
text as the context or the meaning of the token should also be considered while
identifying the part of speech.
iii. Resolving a pronoun:
Objects or a person is considered to be a noun in the sentence. Appropriate
mapping of pronoun to the corresponding nouns is done in this step.
iv. Building a graph:
After relevant nouns and pronouns are extracted, a weight or a rank is assigned to
them. The nouns represent vertices, and the relevant edges are created depending
on the weights of the tokens or the sentences.
v. Sentence extraction:
A consolidated weight of a sentence is calculated with the nouns present in that
sentence. Highest scoring sentences are then taken to create a summary of the
input text.
These are the basic steps of extractive text summarization. Using a lexical dictio-
nary, tokenizing and parts of speech tagging can be made easy. Further, for sentence
extraction and graph building, we have used a text rank algorithm.
4.1 Flow of the System
The flow of the proposed system is shown in Fig. 1. Input will be provided in the
form of documents. Extractive summary includes removal of phrases and sentences,
from the input text and combine them together to create a summary. After processing
a single document and generating summary of individual documents, the summary
will be further combined to generate a common condensed summary of multiple
documents depending on the ranking of the sentences generated through individual
summaries. Extracting and arranging meaningful and relevant sentences to create a
summary is a challenge and is vital in extractive method of text summarization.
A text rank algorithm will be used for extractive and unsupervised technique. The
steps in the algorithm will be as follows:
1. Concatenate all the text contained in the articles.
2. Split the text into individual sentences
3. Find vector representation for each and every sentence.
4. A matrix then stores the similarities between sentence vectors calculated in the
previous step.
5. A graph is created with the help of similarity matrix with sentences as vertices
and similarity scores as edges, for sentence rank calculation.
6. A final summary is formed by few top-ranked sentences.
7. This algorithm will be an iterative process till a single summary of all documents
is generated (Fig. 2).
Fig. 1 Flow of system
Fig. 2 Text rank algorithm
5 Future Scope
i. In the future scope of this system, the articles provided as input can be in the
form of online sources or the Web sites providing news.
ii. The user can bookmark the Web site that he/she uses regularly for reading the
news. The bookmarked sources can be used for summarization.
iii. Rather than generating summary of individual pages and then merging it, a
consolidated summary can be generated by processing all the news sources
together.
iv. A different algorithm or a technique like fuzzy logic or regression models can
be used to generate the summary of multiple documents.
6 Conclusion
In this paper, we propose a methodology of automatic text summarization of more

than one document, i.e., multi-document text summarizer using a graph-based
approach. The set of multiple documents is given as an input to the summarizer.
Here, we consider multiple e-news sources available online to be given as an input
to the proposed system. The sources should be in a form of a word document. Every
sentence in the document is first processed, and every word is independently deter-
mined as a token. The tokens are then identified with correct part of speech tag. Text
rank algorithm is used for ranking the text through which, finally, the words in the
sentences are connected by a link and a graph is built for the entire text. Depending
on the factors such as highest rank of a noun and then highest rank of a sentence, a
weight is calculated. After the calculating, a graph is drawn with nouns as vertices and
sentences as links between them. Thus, we have devised a method of multi-document
text summarizer which uses concept of natural language processing and an algorithm
that makes it convenient to draw a graph of relevant nouns and sentences.
Acknowledgements It is a genuine pleasure to express my deep sense of thanks and gratitude to

my guide Dr. Vipul Dalal, for giving me all the support and guidance for the project. His dedication
and keen interest had been solely and mainly responsible for carrying out the project work. His
timely and scholarly advice helped me a lot to accomplish this task. I would like to extend my
sincere esteems to the staff of Vidyalankar Institute of Technology and my colleagues for their
timely support.
References
1. K. Thakkar, R. Dharaskar, M. Chandak, Graph-based algorithms for text summarization, in

International Conference on Emerging Trends in Engineering & Technology (2010), pp. 516–
519. https://doi.org/10.1109/icetet.2010.104
2. S. Dutta, S. Ghatak, M. Roy, A.K. Das, A graph based clustering technique for tweet summa-
rization, in 2015 4th International Conference on Reliability, Infocom Technologies and
Optimization (ICRITO) (Trends and Future Directions) (IEEE, 2015)
3. A.A. Natesh, S.T. Balekuttira, A.P. Patil, Graph based approach for automatic text summariza-
tion. Int. J. Adv. Res. Comput. Commun. Eng. 5(2) (2016)
4. A.N. Gulati, S.D. Sawarkar, A novel technique for multi document Hindi text summarization, in
2017 International Conference on Nascent Technologies in Engineering (ICNTE), Navi Mumbai,
2017, pp. 1–6
5. H. Dave, S. Jaswal, Multiple text document summarization system using hybrid summarization
technique, in 2015 1st International Conference on Next Generation Computing Technologies
(NGCT), Dehradun, 2015, pp. 804–808
6. S.S. Naik, M.N. Gaonkar, Extractive text summarization by feature-based sentence extraction
using rule-based concept, in 2017 2nd IEEE International Conference on Recent Trends in
Electronics, Information & Communication Technology (RTEICT), Bangalore, 2017, pp. 1364–
1368
Fault Exploratory Data Analysis
of Real-Time Marine Diesel Engine Data
Using R Programming
Cheng Siong Chin and Nursyafiqah Binte Nazli
Abstract The uncertainties in diesel engine parameters often result in an inaccurate

model that give rise to poor fault analysis. A typical marine diesel engine model
consists of few subsystems such as freshwater cooling system, lubrication oil cooling
system, air cooler system and exhaust gas system. Instead of modeling and validating
the simulation results with the real marine diesel engine data to determine the possible
faults occurrence, this paper describes an exploratory data analysis on the actual data
to identify the faults. R programming is used to carry out the analysis such as k-
mean clustering by generating the graphical representation of the data from each
subsystem. Exploratory data analysis is used together with the fault tree analysis to
analyze the results for possible faults. Based on the data analysis, it is found that there
are several faults in the fuel oil system, air cooler system and scavenge air system
that requires fault corrective actions that may lead to increase shipping cost.
Keywords Marine diesel engine · Exploratory data analysis · Fault analysis · R

programming
1 Introduction
Many diesels engine modeling has been studied and analyzed for past several years
and presented various results comparing with real running engines. Diesel engine
design has a few important criteria which focus on efficiency, pollutant emissions, and
fuel flexibility. Diesel engines have been modelled in different ways like filling and
emptying methods [1], relying on basic equations like ideal gas law and conservation
of mass to simulate air path dynamics of the combustion chamber [2]. The modeling
and control of the fuel injection and the turbocharger [3, 4] were carried out. With
increased computer processing power, the use of computational fluid dynamic (CFD)
to model the geometry of engines can reach to millions of cell producing super high
C. S. Chin (B) · N. B. Nazli

Faculty of Science, Agriculture and Engineering, Newcastle University Singapore, Singapore
599493, Singapore
e-mail: cheng.chin@ncl.ac.uk
186 C. S. Chin and N. B. Nazli
accuracies with a price of relatively slow calculations due to a significant amount of

finite elements used [5, 6]. It is often used to investigate and identify fault such as a
possible spot of high thermal stress concentration in the engine.
There are many toolboxes available in MATLAB for different systems design.
For example, controlling the engine timing model, the model of a piston engine and
internal combustion engine with the throttle and rotational inertia can be found in
Simscape toolboxes such as Simscape Driveline and Simscape Multibody. A two-
stroke large-bore natural gas lean burn engine with turbocharger was also proposed
[7]. The model consists of few different sub-parts with parameters such as air-fuel
ratio to final output torque. Another software named Lotus Engineering Software is
capable of modeling the combustion and gas flow processes using air-fuel ratio and
valve timing events. The solver calculates the torque and volumetric efficiencies by
considering the influence of heat transfer and friction phenomena [8]. This program
was used for vehicle simulation. Models can be generated quickly using their inbuilt
functions. The software can estimate the fuel consumption, top speed, acceleration
and power output. Lotus engine simulation can be utilized for two or four strokes
engine with great user interface. There are many different models available from its
library but designing a new model for fault analysis is not entirely possible in the
software.
Wave Ricardo [9] is another software for validating designs using available input
data. Many engine manufacturers build multiple variations of the engine. It covers
from motor cars to marine engine and power generators. Wave offers detailed design
and technology selection from a huge database consisting of combustion models
and excellent turbine physics with the clear result and reports after the simulation.
However, it can be quite complex to design from scratch before any fault analysis.
GT-Power [10] is an alternate simulation software that uses advanced computing to
simulate various on-going processes in an internal combustion engine. Logical inter-
face and components are used along with icons and connectors. Two engine operating
modes are available either for design or analysis purposes. However, the GT-Power
engine simulation requires a mathematical tool that is not easy to configure for
general engine simulation and fault analysis. OpenModelica [11] is another package
for modeling mechanical systems for car manufacturers including engine modeling
using open source codes. OpenModelica implementation can be added in a modular
manner. It has modules for equation sorting, optimization, graphical visualization,
and code generation. However, the state machine structure identification from the
basic model that requires equations and variables that often leads to frequent error
messages during simulation that leads to more uncertainties and does not have a
statistical module for data analysis.
As observed from the above software, one of the significant problems of engine
modeling includes the accuracy of the results when compared with real engine data
(even with the availability of actual engine data). As seen in the software mentioned
above, the non-graphical-based modeling of an engine is quite complex to model even
before any fault analysis. Hence, to circumvent the problem, the R programming
on a real marine diesel engine is proposed to identify the possible faults exist in
various subsystems of the marine diesel engine. It reduces the uncertainty due to
Fault Exploratory Data Analysis of Real-Time Marine Diesel … 187
modeling inaccuracies (such as fitting the curves of the actual engine data, unknown
parameters, and other modeling assumptions). Also, the commercial software could
be quite expensive and complex to use. The use of open source software environment
for statistical computing and graphics such as R for fault data analysis gives engineer
more transparency and flexibility to apply statistic analysis on the data to identify
faults on an engine (or other systems of interest). Second, the codes written in R can be
reused to allow ship owner and engineer to review their engine performance quicker
for possible faults before implementing the corrective actions. The data analytics
enable the ship operators to determine the possible faults in the marine diesel engine.
The present study can be used for ship hull performance and design that would impact
the operating cost of the ship further.
The paper is organized as follows. The real input data from B&W 10s90MEC9
engine on a container ship are shown in Sect. 2. The descriptions of each subsystem in
the engine such as exhaust system, freshwater system, fuel oil system and lubricating
oil system are shown in Sect. 3. In Sect. 4, the engine fault analysis on the real engine
is conducted using the R programming. Lastly, Sect. 5 concludes the paper.
2 Actual Data of Marine Diesel Engine
The data from a marine diesel engine on board of APL Boston containership has
dimensions of 328 m × 46 m × 9.7 m (see Fig. 1). The ship uses a compact electron-
ically controlled ten cylinders two-stroke diesel engine, MAN B&W (10S90MEC9).
The parameters of the engine are attained and collated in an excel sheet from local
monitoring and control system through sensors placed in the engine. The data used in
this research were collected from a Bluetracker onboard server (see Fig. 2) installed
on the container ship. The server is connected to the shipboard network systems. It
transfers the data to on-shore at a regular basis (e.g. 15 min interval) for subsequent
Fig. 1 APL Boston containership (containership-info, 2016)

Fig. 2 Bluetracker system architecture
data analytics such as predictive modeling. The navigation and electro-mechanical

systems data are collected through the use of several sensors pre-installed on various
machinery in the engine room and propeller shaft. As shown in Table 1, the real-time
engine data consists of several variables (24 inputs and five outputs) from a marine
diesel engine. In this study, a section of the marine engine data was extracted from
the Bluetracker on-shore server for fault data analysis. The data consist of around
732 samples, measured at an interval of 4 h for the period 1 July 2014 (1200 h) to
30 October 2014 (0800 h). Note that not all the input and output variables are used
for the analysis in this paper.
Table 1 Input and output variables for marine diesel engine (https://ieee-dataport.org/documents/
container-vessel-data)
Variable
Inputs Fuel oil inlet pressure (bar), fuel oil inlet temp (°C)
Fuel oil inlet viscosity (cst), main lubricating oil inlet pressure (bar)
Main lubricating oil inlet temp (°C), cylinder lubricating oil inlet temp (°C)
Air cooler 1 CW inlet temp (°C), air cooler 2 CW inlet temp (°C), air cooler 3, CW
inlet temp (°C)
Air cooler CW inlet pressure (bar), starting air inlet pressure (bar)
Jacket cooling fresh water inlet pressure indicator (bar), jacket cooling fresh water inlet
temp (°C), Piston cooling oil inlet pressure (bar), intermediate bearing 1 temp (°C)
Intermediate bearing 2 temp (°C), intermediate bearing 3 temp (°C)
Thrust pad temp (°C), fuel oil mass flow (kg/h), turbocharger exhaust inlet 1 temp
(°C), turbocharger exhaust inlet 2 temp (°C)
Turbocharger exhaust inlet 3 temp (°C)
Outputs Shaft power (kW), shaft RPM (rpm), shaft torque (kNm)
Engine power (% maximum continuous rating), Turbocharger RPM indicator (rpm)
3 Main Diesel Engine Subsystems
A few subsystems within the engine involving heat input from fuel to exhaust gas
and the cooling subsystem is described. The five main subsystems such as the fuel
oil system, lubrication oil system, freshwater system, charge air cooling and exhaust
system can be seen below. The fuel oil system for a diesel engine can be split into
two parts, the fuel supply, and the fuel injection systems. For heavy fuel oil, it is
stored in the tanks at the double bottom from that is pumped to the settling tank to
get heated. The oil goes through fuel oil purifier and transfers to the daily service
tank. After the mixing tank, there is a booster pump to increase the pressure into the
combustion chamber by a fuel injector.
Proper lubrication decreases the power to overcome friction and reduces wear and
tear. Lubricating oil is used as a coolant and sealant. Lubricating oil for an engine is
kept at the bottom of the sump, or in a drain tank placed at the lower part of the engine.
The oil is absorbed from the reservoir through a strainer before passing through a
set of pump and fine filters. Subsequently, it goes through a cooler before going to
the engines before circulating to several branches of the conduit. The lubricating oil
system provides a constant flow of filtered oil to all the engine bearings and other
surfaces. The use of seawater causes corrosion which the salt deposited on the cooling
surfaces reduces the heat flow. Fresh water is used to cool the cylinder jacket covers,
piston, exhaust valve, and turbochargers working under high temperature and heat.
Fresh water is cooled by using seawater. The freshwater removes the heat from the
cylinder liners, cylinder heads, exhaust valves and the turbochargers before it returns
to the engine. The temperature of the cooling water is controlled via a three-way
control valve. A fresh water generator (FWG) which is used to produce fresh water
from sea water is also incorporated.
In a most marine diesel engine, turbocharger and supercharger are used to increase
the density of the air to enable more fuel to be burnt. It, therefore, increases the power
output of an engine. However, this charged air causes an increase in the temperature
of the air due to adiabatic compression in the turbocharger. The charged air could
not be fed directly to the engine as it is out of the operating limit of the engine. The
compressed air from turbocharger goes through the charge cooler system made of
fins and tubes to cool the charge air before going back to the engine. It increases the
density of air and reduces the engine thermal loading.
4 Fault Analysis Using R Programming
Data analysis is the process of examining data using logical interpretation and trans-
forming the raw data into useful information. There are mainly four steps in data
analysis namely: problem definition and planning, data preparation, analysis, and
deployment. As shown in Fig. 3, the first step is problem definition and planning.
The problem for this case is whether there is any fault in the engine. The plan is to use
Fig. 3 Fault tree analysis for input (left) and output (right) variables
exploratory data analysis, fault tree analysis and analytical methods to analyze the
data. The second step is data preparation. The raw data parameters are obtained from
the engine and collated in an excel worksheet. The raw data is characterized, explored
and cleaned for further processing. The third step is analysis. R-Studio is used to
generate the summary and the graphical representation of the data. Exploratory data
analysis and fault tree analysis is used to analyze the graphs generated. The last
step is deployment. The results and analysis from the previous step are deployed to
achieve the project objectives before documentation or report writing.
The exploratory data analysis plays a crucial role in data analysis as it assists in the
selection of appropriate analysis method to apply on the data, detection of a fault, and
evaluation of the rationality of the assumptions. It is mainly involved in summarizing
and visualizing data before carrying out formal modeling for visual representation.
By generating visual representations of data exploration, characteristics of the data
can be interpreted, qualities and features of the data can be examined rather than a
huge data set, and new patterns or correlations can be discovered. Making sense of
the data is easier by understanding the relationships between the variable, which can
be identified and interpreted by generating summary tables and data visualizations in
R Studio. A simple summary table helps in understanding the relationship between
variables and provide information of the data which includes meaning, median, mode,
minimum value, maximum value, and standard deviation.
Visual representation aids in finding relationships between variables. The box plot
is the visual representation of the summary table. It identifies outliers which gone
beyond the data range. Scatter plot allows the trend of the variables and the type of
relationship that could be present between the two variables. Typically, a positive
linear trend should be observed in the scatter plots when the variables are plotted
against each other. Any plot inconsistencies or outliers that occur in the scatter plot
could be an indication of an error or fault occurrence. The k-means clustering is a
common clustering technique which groups the data into clusters. One condition is to
indicate how many clusters are needed. The k-means clustering technique identifies
relationships which may exist between two variables. The k-means algorithm can be
shown as follows.
Step 1: Choose the initial number of clusters, k

Step 2: Compute the cluster sum of squares difference between data point and the
mean of the cluster
Step 3: Allocate the data into the cluster where the mean is nearest to it
Step 4: Re-compute the mean of the cluster from the data to minimize the within-
cluster sum of squares. In other words, the objective is to minimize the cost
function:

k
J = arg min x − µi 2 (1)
S i=1 x∈Si
where µi is the mean of points in S i given a set of observations (x1 , x2 , . . . , xn )

where each observation is a D-dimensional real vector. The k-means clustering
aims to partition n observations into k sets S = {S 1 , S 2 , …, S k } to minimize
(1). The Euclidean distance derived from the each point as ||x − µi ||2 will be the
minimum value of all distance from the data set to the centroid.
The data are then inspected and analyzed to determine the cause of the faults.
Investigative methods such as research, brainstorming, and fault tree analysis are
used to determine the faults or malfunctions are occurring in the systems and also
rectify the faults. Fault corrective action improves the procedures taken to remove
sources of abnormalities or other undesirable conditions. A fault tree analysis (in
Fig. 3) of the marine propulsion system is used to visualize the faults that may occur
by the parameters. It is a deductive method of analysis to find fault in the respective
subsystems. Since the engine is made up of several subsystems, they are divided into
two groups namely: input variables system and output variables system (see Fig. 3).
For example, under the input variables, the faults in the fuel oil system can be due
to leakage, the low mass flow rate of fuel and incomplete combustion.
The raw engine parameters are collated, cleaned and prepared in an Excel work-
sheet before imported into R studio. The data is then separated into different months
in its various systems. By breaking down the data into smaller subsets, the data will
be simplified for analysis. Function slice() is used to separate the data into different
months in R studio, which is presented in the environment tab. The data is ready
to be analyzed by exploratory data analysis. Example of the R programming codes
written in version 3.2.4 can be downloaded from https://docs.google.com/document/
d/12FTWVUjCRvzV33pgKVBE5JCXvbWdDU-fdwltl74NEM4/pub.
4.1 Data Analysis on Fuel Oil System
A few irregularities are found in the specific fuel consumption and fuel oil mass flow
rate. The specific fuel consumption against the days for each month is plotted. The
scatterplot examines the trend and the specific fuel consumption. From Fig. 4, it is
Fig. 4 Scatter plot of specific fuel oil consumption for each month
observed that the specific fuel oil consumption is quite constant throughout the month
except for a few days where the consumption goes down to 0 t/day. It may be due
to the ship was at the port or not moving. Also, there are two spikes in September
and October that reach 163 t/day on the average consumption of 120 t/day. With
reference to Fig. 3, the sudden surge in consumption may be due to fault or leakage
in the fuel oil tank due to unsecured seals or incomplete combustion. The higher fuel
consumption may result in overcoming strong wind and wave during the voyage. The
fuel oil inlet temperature is plotted against the fuel oil mass flow using scatter plot
with regression to show the relationship between these two variables in Fig. 4. The
fuel oil inlet temperature in July and August is higher and inconsistent as compared
to September and October. It indicates that there might be a fault in the fuel oil system
in the July and August.
The fuel oil inlet viscosity and fuel oil inlet pressure for July to October are plotted
as shown in Figs. 5 and 6, respectively. Figure 5 represents the fuel oil inlet viscosity
(cst) over days in a specific month. The typical range of fuel oil viscosity lies between
10 and 12 cst for each month. One outlier is observed in September showing a zero
viscosity (not possible in practice). With reference to the fault tree analysis in Fig. 3,
this may be due to the low mass flow rate of the fuel causing incomplete combustion
and error in the viscosity measurement. In Fig. 5, a high value of viscosity of around
17.5 cst can be seen in the month of August and October. This value corresponds
to a low pressure as seen in Fig. 6. Since it appears only once without continuous
trend, it is unlikely to cause by clogged fuel filter, pipeline or intermittent sensor
fault. On the other hand, the range of the fuel oil inlet pressure falls between 7 and
8 bars except for July and August that show a higher pressure reaching 12 bars. The
results highlight a possible fault in the fuel oil system.
As a result, another parameter to study is the fuel oil inlet temperature (°C) against
the fuel oil flow meter temperature (°C) for each month. The K-means cluster plot
gives clustered view of the temperature data between the flow meter outlet and
the inlet of the main engine. The desired plots should exhibit a linear trend in the
temperature, i.e. a high temp at the flow meter outlet should accompany by high
temperature in the inlet of the main engine. As seen in Fig. 7, the fuel oil inlet
temperature for July and August is lower than September and October. Also, there
Fig. 5 Fuel oil inlet viscosity for each month
Fig. 6 Fuel oil inlet pressure for each month
is a group of the cluster in July exhibits value approximately 20 °C higher than the
rest.
In summary, it is evident that there is a possibility of a fault in the fuel oil system,
predominantly in the path between the flow meter and inlet of the main engine during
Fig. 7 Fuel oil inlet temperature versus FO flow meter for each month (by k-means)
July and August. For example, there are some installed components along the path
between the flow meter and the inlet of the main engine. For example, a fault in
the heater due to the faulty control of heating elements may lead to the abnormality
in the fuel oil temperature causing the pressure to increase and lower the output
temperature. The pipeline may have some leakages or blockages that reduce the fuel
oil temperature.
4.2 Data Analysis on Air Cooler System
The air cooler CW inlet temperature is plotted to examine the air cooler CW inlet
temperature for each month. It is desired to see the outlet temperature is much lower
than the inlet temperature after going through the air-cooling system. In Fig. 8, the air
cooler CW inlet temperature for the three months are quite consistent. Two clusters
(in cyan and blue) have a lower temperature than (green and red). As observed
in the scatter plot, there are two lower temperature points at the beginning of the
month due to lower ambient air at the start of the voyage. As shown in Fig. 9,
the air cooler CW outlet temperature exhibits the same trend across each month.
The highest cluster (red) occurs near the beginning of each month which rises to
the maximum temperature (around 40 °C) in the cooler CW outlet. Moreover, the
high-temperature phenomena are observed to be quite far from the remaining data.
The high-temperature may cause insufficient cooling of intake air, or foreign objects
blocking the cooler intake.
The data analytics and Internet technology have a great impact on the current
maritime industry as manpower cost increases and greater uncertainties in the oil
price and global politics related to oil and gas. The data collected on the board of
Fig. 8 Scatter plot of air cooler CW inlet temperature in August
Fig. 9 Scatter plot of air cooler CW outlet temperature in October
the ship and analysis performed to enable the ship owners and operators to better
understand the underlying faults as a result of the current equipment performance
such as in this study, the marine diesel engine. The current cost reduction due to
the online data collection and analysis is approximately 10% as no manual data
collecting by engineers (time are free to do other important tasks) are needed. This
proposed methodology leads to further analysis such as hull performance that can
be further optimized in the hull design (i.e. dimensions and marine paint coating of
the ship hull) to reduce the fuel consumption and subsequent environmental impact
of shipping by greenhouse gas emissions during the voyage. A rough estimation by
the shipping company is approximately 5–10% fuel consumption since frictional
resistance constitutes up to around 90% of the total drag force experienced by the
ship hull [12]. Hence, by reducing the frictional resistance generated as the hull flow
through seawater would enable cost saving in the operating cost such as less fuel
consumption.
5 Conclusions
The paper presents exploratory data analysis on the actual engine data to identify
possible fault occurrences instead of using a marine diesel engine model that is
often subjected to uncertainties and inaccuracies. The use of R-studio with the aid
of different graphical presentation of the data allows data analysis with the help of
the fault tree analysis to determine possible faults in the engine. Through conducting
data analysis, several fault occurrences were detected in the engine data. In the fuel
oil system, the abnormality in the fuel oil consumption in July and August may
occur. The results from the fuel oil inlet pressure, temperature and viscosity show a
possibility of a fault in the fuel oil system, predominantly in the path between the
flow meter and inlet of the main engine during July and August. The inconsistency
and low temperature in July and August for fuel oil inlet temperature against mass
flow rate indicate that there is a fault occurring before the fuel enters the engine. In
the air cooler system, insufficient cooling resulting in a higher temperature in the air
cooler CW outlet temperature. In summary, the exploratory data analysis and fault
tree analysis using R studio were carried out successfully on the real engine data
from a container ship. This approach can be extended to ship hull performance and
design that would impact the operating cost and performance efficiency of the ship
during the voyage. In future work, engine parameters could be extracted from the
engine systems when it is in its optimum condition with more accurate sensors so
that more analysis can be performed to refine the fault detection and to compare
with the optimum condition. An early fault detection system can be developed and
installed on the vessel.
References
1. Y.H. Zweiri, J.F. Whidborne, L.D. Seneviratne, Complete analytical model of a single-
cylinder diesel engine for nonlinear control and estimation, Report EM/99/10, Department
of Mechanical Engineering, King’s College London, Strand, London, UK, 1999
2. J. Chauvin, A. Albrecht, G. Corde, N. Petit, Modeling and control of a diesel HCCI engine, in
Fifth IFAC Symposium on Advances in Automotive Control, CA, USA (2007), pp. 1–8
3. C.W. Tan, H. Selamat, A.J. Alimin, Modeling and control of an engine fuel injection system.
Int. J. Simul. Syst. Sci. Technol. 11, 48–60 (2010)
4. J. Wahlström, L. Eriksson, Modeling diesel engines with a variable-geometry turbocharger and
exhaust gas recirculation by optimization of model parameters for capturing non-linear system
dynamics. Proc. Inst. Mech. Eng. Part D: J. Autom. Eng. 225, 960–986 (2011)
5. L. Goldsworthy, Computational fluid dynamics modelling of residual fuel oil combustion in
the context of marine diesel engines. Int. J. Engine Res. 7, 181–199 (2006)
6. H. Barths, C. Hasse, N. Peters, Computational fluid dynamics modeling of non-remixed
combustion in direct injection diesel engines. Int. J. Engine Res. 1, 249–267 (2000)
7. M. Turesson, Modelling, and simulation of a two-stroke engine, Master of Science Thesis,
Chalmers University of Technology, Göteborg, Sweden, 2009
8. B. Duleba, Simulation of automotive engine in lotus simulation tools. Transfer inovácií 30,
48–52 (2014)
9. D. Cordon, C. Dean, J. Steciak, S. Beyerlein, One-dimensional engine modeling and valida-

tion using Ricardo wave, Final Report KLK434-B (N07-09), National Institute for Advanced
Transportation Technology, University of Idaho, 2007, pp. 1–45
10. Z. Shaohua, J. Guodong, X. Tianfeng, Design and analysis method for internal combustion
engine mufflers based on GT-power software. Autom. Technol. 7, 7–10 (2003)
11. M. Martin Sjölund, P. Fritzson, An OpenModelica Java external function interface supporting
MetaProgramming, in Proceedings 7th Modelica Conference, Como, Italy (2009)
12. M. Schultz, Effects of coating roughness and biofouling on ship resistance and powering.
Biofouling 23(5), 331–341 (2007)
A Hybrid Key Management System
Based on ECC and Data Classification
to Secure Data in the Cloud
Hinddeep Purohit and Ravirajsinh Vaghela
Abstract Cloud computing is an unparalleled innovation in the field of computer

science. However, the CSPs cannot be entrusted with secret data. Data needs to be
encrypted before it is uploaded to the cloud. But again, the data is only as secure as
the keys used to encrypt or decrypt it. This gives rise to yet another problem, i.e., key
management. This paper systematically reviews the literature, identifies the issues
prevalent in key management, and presents their corresponding solutions. Post the
literature review, we propose a model that borrows the best from the past efforts and
fixes the shortcomings of the present models by combining all the solutions presented
in this paper.
Keywords Cloud security · ECC · KMS
1 Introduction
The dawn of the high-speed Internet era bought with it an onslaught of innovations
that dramatically mutated the way people conceived and operated computer systems.
Gone are the days when one was fettered by the constraints of the hardware he or she
owned. In a wink of an eye, it is now possible to dynamically reconfigure the system
by shelling out a pittance. The introduction of cloud computing, a paradigm shift
from traditional computing, has hailed a slew of transformations that have afforded an
ever more mobile and flexible work pattern to the cyber samurais of the twenty-first
century. Cloud computing is a method of leveraging the Internet to provide essential
services such as storage, database, specialized hardware (e.g., high-performance
GPUs), operating systems, applications, and development environments remotely.
The most prevalent and authoritative definition of cloud computing is outlined in
H. Purohit (B)
RK University, Rajkot, India
e-mail: hinddeep.purohit007@gmail.com
R. Vaghela
Gujarat Technological University, Ahmedabad, India
e-mail: sunstate9999@gmail.com
200 H. Purohit and R. Vaghela
a paper published by NIST [1]. High-speed and reliable Internet connection is a

pre-requisite to access the resources hosted on the cloud. Instead of hooking up
all the hardware/software to their system, users can log into their accounts from
virtually anywhere and take advantage of the same facilities as would have been
available which had the resources been connected locally. This exonerates people
from having to lug around heavy equipment and ushers in an era characterized by
increased mobility.
Models in cloud computing: The services offered by cloud providers can be
broadly classified into three main categories, namely IaaS, PaaS, and SaaS [2]. The
cloud provider purchases and hosts hardware (e.g., storage, GPUs, CPUs, etc.) in
the data center. A subscriber pays only for the resources that he/she needs. The
key concept here is virtualization. The subscriber has complete control over the
entire technology stack. PaaS subscribers get access to operating systems and pre-
installed software (usually dev tools for building applications, hosting services, etc.)
in addition to hardware and storage. In the case of SaaS, right from the infrastructure,
middleware apps (e.g., databases) to the applications that end-users use are all hosted
on the cloud. The user does not have to manage any of the resources in the stack.
However, the cybersecurity threat landscape has expanded its frontiers far and
wide, making inroads into the privacy and security of individuals and organizations
alike, and has assumed menacing forms that were hitherto unheard of—thanks to the
attack surface cloud computing has exposed! The General Data Protection Regulation
of the European Union [3] names a number of abstract, negative effects that lack
of appropriate technical and organizational measures to ensure data protection can
have on individuals [4]. Consider the source code of a yet-to-launch application,
stored on some cloud storage, falling into the hands of the contemporary. It could
mean insurmountable financial and intellectual loss. To prevent such mishaps from
spreading mayhem, data protection in the cloud is imperative. Since data contain
sensitive information, in no case the cloud providers can be trusted. There is minimal
control over the data by the customers [5]. One solution to this problem is to encrypt
the data before storing it on cloud any service that will ensure that no unauthorized
person gains access to the data. Encryption, the most common solution to prevent
data breaches [6], is offered in several cloud platforms [7]. But encryption/decryption
of data leads to the obvious question of key management. This paper briefly reviews
the problem of key management and the solutions readily available in the literature.
We have performed a systematic literature review on key management issues in cloud

security. We consulted papers addressing key management systems for securing
the cloud as well as survey papers with an impressively broad scope that covered
security misconfigurations, virtualization issues, data storage issues, Web application
vulnerabilities and other legal challenges haunting the cloud. We aimed to study the
open problems in key management, analyze the proposed solutions, and then come
up with our own model that improves upon the previous models.
A Hybrid Key Management System Based on ECC … 201
3 Structure
Section 4 reviews and identifies weaknesses in the previous work that addresses the
issue of key management in securing data in the cloud. Section 5 describes how the
limitations identified in Sect. 4 can be overcome. Finally, Sect. 6 proposes a hybrid
model that improves upon the previously proposed models, described in Sect. 4, by
putting together all the solutions presented in Sect. 5. The hybrid model is followed
by concluding remarks that illustrate how the proposed model solves the challenges
identified in Sect. 4.
4 Literature Review
According to European Network and Information Security Agency (ENISA), impor-

tance of key management is not well addressed [5]. Ideally, it is the data owner’s
(responsibility). But due to the lack of user expertise to handle the keys, they usually
hand over the key management to CSP [8]. Rafael et al. maintain that the most
common place to store this tenant key is in a cloud system itself (e.g., a cloud
provider) where the database is accessed [4]. This lends the CSP complete access to
the data which is a serious problem mission-critical data. Another point to consider
is key destruction [5]. The keys used to encrypt/decrypt data must be destroyed and
completely removed from the cloud when the contract between CSP and client is
canceled. Failure to destroy the keys invites the possibility that the CSP or other third
party may in future access data which they should not have any access to. Since data
security and key destruction are highly sensitive, the CSP must be as vocal about its
security policies as possible. This transparency allows the client to trust [4] the CSP.
Jäger et al. proposed a sealed cloud approach [9] which ensures that nobody—not
even the CSP—other than the owner of the data has access to it. In the sealed cloud
approach, the keys are not stored in the cloud. The data owner must provide the key
explicitly each time he/she wants to access the data. Since the key is only known to
the data owner, the privacy of the data stored is guaranteed. However, the trade-off in
such a setup is that the data cannot be manipulated inside the cloud. In other words,
operations such as querying the data are not possible. Had the key been managed
by the CSP itself or a third party key management server, query processing would
have been possible. Real-world systems often require the facility of querying the
data to glean insights. Moreover, key unavailability [5] poses a major issue. If the
key used to encrypt data is lost, data is rendered unrecoverable. Therefore, a sealed
cloud approach is not feasible in most cases.
Another approach is to use a key management software (will be referred to as KMS
henceforth). Key management is the set of techniques that involves the generation,
distribution, storage, revoking, and verifying keys. The KMS is primarily responsible
for securely distributing the keys [5], i.e., no unauthorized user should receive the
key, lest the data will be compromised. Since KMS generates the key, it also needs
to update them [5] after it has expired. If the keys are not updated, even legitimate
users will have a hard time accessing data. It can be handled by the client, CSP, both
client and CSP, key splitting, centralized server, or group key management [5]. KMS
should be developed by professional companies that know about encryption, control
the integrity checking, key generation, key distribution, periodic key changing, and
key destroy. Besides, KMS has to be verified by the cloud service provider [10]. It
must be noted that using a KMS involves a lot of computational overhead [10] which
in turn impedes the performance of the system as accessing or altering encrypted data
inside a cloud involves extra computation (decryption followed by re-encryption)
that introduces a delay. Techniques that reduce this delay need to be developed. That
said, using KMS has its advantages too. A user can opt to go with different KMS
providers for different files depending on the confidentiality level of the file. This
strikes a balance between confidentiality and performance as low confidentiality
files do not have to be encrypted/decrypted with a larger key size required for high
confidentiality files.
On the other hand, when the security is entirely managed by the CSP what is called
trust issue comes into the picture. Much effort has been done to satisfy customers’
demand, but the security status is lack of transparency. It means that the client is
not clear on how service provider secures and controls their data, so they still hold a
skeptical attitude toward data security [10].
Trusted third party, proposed by I. Indu et al., acts as a cryptographic server and a
KMS that uses RSA to encrypt data. Here, the TTP acts as a server where the owner of
the file and the users are registered. TTP has multiple responsibilities such as owner
management, user management, encryption, decryption, and access control [11].
The TTP generates a set of public and private keys once the data owner is registered
in the server. The file to be uploaded is encrypted with the private key generated.
The encrypted file is then uploaded to the cloud storage service. The private key is
securely deleted from the TTP by overwriting it with one half of the public key. The
data owner provides user list/access permission list (UAPL), user-specific security
questions and answers, number of file access permission to the TTP server [11]. The
data owner provides the user with the link to TTP and a portion of his/her public key.
The user then registers to access the file. After successfully being registered, the user
logs into the TTP and provides the portion of the owner’s public key. TTP verifies
if the user has permission to access the file. If the permission has been granted, the
TTP reconstructs the public key from the partial key provided by the user, decrypts
it using the owner’s public key, and downloads the file.
All the previous attempts in securing data in the cloud have tried to target one
specific problem. But, in trying to solve that one problem, they have either intro-
duced others or overlooked multiple aspects such as level of confidentiality and
performance. For instance, if the keys are managed by the CSP, the user is freed
from the onus of keeping them safe, but as a side effect, the CSP gains complete
control over the user’s data. The sealed cloud approach gives complete control to the
data owner. However, it is a very rudimentary solution that does not allow the user to
query stored data. TTP, a KMS, seems to address the shortcomings of both of CSP
managed keys and sealed cloud approach. Although robust, TTP’s performance can
be improved. In the following sections, we will prove that the performance of TTP
can be upped by classifying data before encrypting it, and by using ECC in lieu of
RSA.
5 Factors Affecting the Performance and Security of TTP

and Data Classification
Up to this point, we have described various ways to handle the keys used to
encrypt/decrypt data in the cloud. However, the security and the performance of the
algorithm used to encrypt/decrypt data have not been taken into account. The highest
level of security, when coupled with the least amount of computational overhead,
makes for the perfect cloud data security model. Researchers have been trying to
reduce the computational overhead while improving or at the very least, maintaining
the same level of security as provided by the legacy models.
Data classification was suggested by Tawalbeh et al. [12]. The present-day cloud
storage solutions use a fixed-sized key to encrypt data. They do not take into consid-
eration the sensitivity of the data being stored. If basic, medium, and high, all three
confidentiality level data are encrypted/decrypted with a fixed-sized key, then the
computation overhead will increase significantly, leading to performance degrada-
tion. The data classification model is a granular encryption technique that takes into
account the confidentiality level of the data to be encrypted. Data is classified into
three categories, namely basic level, confidential level, and highly confidential level.
The basic level uses AES 256 + TLS, confidential level uses ten rounds of AES 128,
and the highly confidential level uses AES 256 + SHA-2 to provide data integrity
along with confidentiality. According to the graph provided in [12], it has been proven
that the data classification technique is highly effective in reducing the computational
overhead.
Nevertheless, the data classification technique uses symmetric key cryptography.
Asymmetric encryption algorithms provide a higher level of security. The selling
point of such a cryptosystem is that the key itself need not be transported with
the ciphertext thereby freeing the sender from having to protect the key while in
transit. The de facto standard in public-key cryptography is the RSA algorithm. RSA
algorithm generates public and private keys and encrypts/decrypts data using those
keys. However, when the amount of data to be transmitted is massive or when using a
resource-constrained/embedded device such as a smartphone, employing RSA would
be an inefficient choice. ECC is a viable option considering the performance gain
over RSA as illustrated in both the tables below. Therefore, the data classification
model should use ECC instead of AES as it provides better security and performance.
RSA versus other asymmetric algorithms: Key generation time, memory usage,
and the amount of time taken for encryption and decryption for RSA, DSA, DH,
ElGamal, and Paillier [13, 14] have been compared in [11]. The graphs and tables
Table 1 RSA versus ECC

Security level RSA key size ECDSA key ESP-IDF
key size
(bits) size (bits) curves
80 1024 160–223 secp192r1,
secp192k1
112 2048 224–255 secp224r1,
secp224k1
128 3072 256–383 secp256r1,
secp256k1,
bp256r1
192 7680 384–511 secp384r1,
bp384r1
256 15,360 512+ secp521r1,
bp521r1
Source [17]
Table 2 RSA versus ECC

Key length Time (s)
key generation time
ECC RSA ECC RSA
163 1024 0.08 0.16
233 2240 0.18 7.47
283 3072 0.27 9.80
409 7680 0.64 133.90
571 15,360 1.44 679.06
Source [18]
indicate that RSA outperforms other asymmetric algorithms. However, ECC—an

eminent asymmetric algorithm—was missing from the comparison done in the paper.
RSA versus ECC: A key should be produced using cryptographic module with at
least FIPS 140-2 [11] compliance for memory, strength, and security requirements
[15]. While both RSA and ECC satisfy the requirement mentioned above, ECC
provides a similar level of security with lesser key size. The key generation time for
comparable key sizes for ECC is a thousand times faster than for the RSA algorithm
[16]. To put the figures into perspective, refer to the following table taken from [17]
(Tables 1 and 2).
6 Proposed Model
While the data classification strategy is laudatory, it can be further extended to

perform automatic data classification using machine learning algorithms for certain
data types, such as plaint text, in place of manual classification.
Once the raw data has been categorized into basic level, confidential level, and
highly confidential level, we fix a key size depending on the confidentiality level
of the data to be encrypted. Basic level data should be encrypted using a smaller
key size, highly confident level data with a higher key size, and confidential level
data with a key size that falls between those two extremes. When the data owner
is registered with the TTP, it generates three different key pairs, using the ECC
algorithm, each corresponding to a level of security. Each private key is used to
encrypt the corresponding category of data. The encrypted data is then uploaded to
the cloud. All the three private keys are securely deleted from the TTP by overwriting
them with one half of their corresponding public key. The data owner provides user
list/access permission list (UAPL), user-specific security questions and answers, and
number of file access permission to the TTP server [11]. The data owner provides
the user with the link to TTP and a portion of each of his/her public keys. The user
then registers to access the file. After successfully being registered, the user logs into
the TTP and provides the portion of each of the owner’s public keys that he/she has
received. TTP verifies if the user has permission to access the file. If the permission
has been granted, the TTP reconstructs the public key from the partial key provided
by the user, decrypts it using the owner’s public key, and downloads the file (Fig. 1).
Fig. 1 Diagrammatic representation of the proposed model

7 Conclusion
Albeit with a few glitches, TTP (an example of KMS) is a marked improvement over
previous attempts at key management. Neither TTP nor data classification alone
can provide the most efficient security solution for the data stored in the cloud.
After all, data classification is only a method to speed up the process of encrypting
and decrypting data that form only a part of the entire KMS, and TTP has its own
set of drawbacks mentioned previously. Toward that end, we have come up with a
hybrid model the improvises the existing solution by using asymmetric cryptography
instead of symmetric cryptography and combines data classification with the TTP
model to improve the speed of encryption and decryption. Using automated data
classification and ECC in the data classification model improves the ease of use (due
to the automation of data classification), security and performance of key generation,
encryption, and decryption. Using this new devised data classification along with the
TTP model, we achieve a highly secure, robust, scalable, and performant model to
secure data in the cloud. It is worth mentioning that the proposed model can also be
used for securing local storage and network attached storage systems. Whether the
data is stored on the cloud, on local storage device or on a network attached storage
system, it makes little difference. The underlying method to achieve efficient data
privacy remains the same.
References
1. P. Mell, T. Grance, The NIST Definition of Cloud Computing (2011)

2. R. Velumadhava Rao, K. Selvamani, Data Security Challenges and Its Solutions in Cloud
Computing (2015)
3. European Parliament and the Council, General Data Protection Regulation (EU) 2016/679 of
the European Parliament and of the Council
4. R. Dowsley, M. Gabel, G. Hübsch, G. Schiefer, A. Schwichtenberg, A Distributed Key
Management Approach (2016)
5. A.R. Buchade, R Ingle, Key Management for Cloud Data Storage: Methods and Comparisons
(2014)
6. B. Grobauer, T. Walloschek, E. Stocker, Understanding cloud computing vulnerabilities. IEEE
Secur. Privacy 9(2), 50–57 (2011)
7. C. Esposito, A. Castiglione, B. Martini, K.-K. Raymond Choo, Cloud Manufacturing: Security,
Privacy and Forensic Concerns (2016)
8. R. Latif, H. Abbas, S. Assar, Q. Ali, Cloud Computing Risk Assessment: A Systematic Literature
Review (2014)
9. H.A. Jäger, A. Monitzer, R.O. Rieken, E. Ernst, A novel set of measures against insider attacks—
sealed cloud, in Lecture Notes in Informatics—Open Identity Summit 2013, ed. by D. Hühnlein,
H. Roßnagel (Gesellschaft für Informatik, Bonn, 2013), pp. 187–197
10. D.W.K. Tse, D. Chen, Q. Liu, F. Wang, Z. Wei, Emerging Issues in Cloud Storage Security:
Encryption, Key Management, Data Redundancy, Trust Mechanism (2014)
11. I. Indu, P.M. Rubesh Anand, S.P. Shaji, Secure File Sharing Mechanism and Key Management
for Mobile Cloud Computing Environment (2016)
12. L. Tawalbeh, N.S. Darwazeh, R.S. Al-Qassas, F. AlDosari, A Secure Cloud Computing Model
based on Data Classification (2015)
13. P. Paillier, Public-key cryptosystems based on composite degree residuosity classes, in

Proceedings of Advances in Cryptology (Eurocrypt ‘99), Prague, Czech Republic (1999),
pp. 223–238
14. Z.E. Dawahdeh, S.N. Yaakob, A.M. Sagheer, Modified ElGamal elliptic curve cryptosystem
using hexadecimal representation. Indian J. Sci. Technol. 8(15), 1–8 (2015)
15. US Nat’l Inst. for Standards and Technology, Security Requirements for Cryptographic
Modules, Federal Information Processing Standards Publication (2011)
16. S. Arun, N.R. Shanker, Data security in cloud storage using elliptical curve cryptography
(2018)
17. M. Suárez-Albela, P. Fraga-Lamas, T.M. Fernández-Caramés, A Practical Evaluation on RSA
and ECC-Based Cipher Suites for IoT High-Security Energy-Efficient Fog and Mist Computing
Devices (2018)
18. N. Jansma, B. Arrendondo, Performance Comparison of Elliptic Curve and RSA Digital
Signatures (2004)
Life Cycle Management for Indian Naval
Warship Equipment
Alok Bhagwat and Pradnya Vishwas Chitrao
Abstract Indian Navy has laid out ambitious shipbuilding plan in coming decade
with large numbers of ships and submarines planned to be built in Indian ship-
yards with indigenous design. However, a number of critical equipment viz. electric
propulsion equipment are not available from Indian defence manufacturers. These
equipment are to be imported as of now causing precious foreign exchange outgo.
Further, the dependence on foreign suppliers throughout equipment life cycle of 25–
30 years costs Indian Navy dearly. Indian defence manufacturers aspire to support
‘Make-in-India’ initiative by establishing capital intensive set-up for manufacturing
electric propulsion equipment in the country, but a lacuna in bid evaluation process
has adverse impact on this indigenous drive. The lacuna being evaluation of commer-
cial bids of equipment procurement on acquisition costs alone. The operation &
maintenance (O&M) costs, a significant element over equipment life cycle, is not
considered holistically in bid evaluation. Ministry of Defence (MoD), Govt. of India
(GoI) in the recently promulgated Defence Procurement Policy (DPP-2020) has
mandated consideration of long-term product support during equipment procure-
ment. But clear guidelines for estimation of life cycle costs (LCC) does not exists,
thus making commercial evaluation on the basis of LCC during procurement very
complex. This paper aims to bring out practices followed in US and NATO Navies
in life cycle management (LCM) practices and suggest ways to implement LCM
process and LCC estimation in Indian Navy, thus providing ‘level playing field’ to
Indian manufacturers struggling to support country’s drive in attaining self-reliance
in critical defence projects.
Keywords Indigenous warship building · LCM in Indian Navy · Indigenous

manufacture of Naval equipment · LCC in Naval procurement
A. Bhagwat (B) · P. V. Chitrao

Symbiosis Institute of Management Studies (SIMS), Pune, India
e-mail: bhagwat.alok@gmail.com
P. V. Chitrao
e-mail: pradnyac@sims.edu
210 A. Bhagwat and P. V. Chitrao
1 Introduction
Indian Navy has embarked upon an aggressive and ambitious plan of building large
number of warships and submarines in the coming decade. As many as, 30 platforms
are on order on Indian shipyards with over 50 platforms in the pipeline in various pre-
tendering stages at Naval Headquarters, New Delhi. Almost all of these platforms will
be built in Indian shipyards under indigenous design. While an increasing number
of warship equipment are being taken up for manufacturing within country, some
of the high technology equipment viz. ship propulsion and weapon systems are
imported till date. Indian industry led by renowned names like Larsen & Toubro
Limited, Tata Power Group, Kalyani Group, and Mahindra Group has been rapidly
setting up indigenous capability for manufacturing these equipment; however, the
regulatory framework is not fully supportive towards Indian industry. In fact, it is
tilted favourably towards foreign suppliers as regards deciding on the lowest bidder,
or L1 bidder, as per defence procurement terminology [1]. There is a need to correct
this aberration. Ministry of Defence (MoD), Govt. of India (GoI), is fully seized of
this situation, and corrective actions have been initiated in recent past [2].
Warships over world are moving towards electric propulsion systems for variety
of reasons, not considered in the scope of this paper. Submarines have always had
electric motors for propulsion. The primary reason being non-availability of air for
combustion engines. However, electric propulsion is coming in surface warships
also. Indian Navy has plans to install forthcoming Destroyers and Landing Dock
Ships with electric propulsion equipment comprising of high-power electric motor
and associated control drive along with auxiliary support systems. Therefore, Indian
industry players are actively considering taking up manufacture of electric propulsion
equipment for these forthcoming shipbuilding projects. Indian Navy is in dialogue
with a number of Indian manufacturers towards this goal.
It is well known that setting up a new facility for manufacturing such high tech-
nology equipment is an activity involving high capital costs. The cost is eventually
to be recovered from the project. Therefore, the Indian supplier does not remain
cost effective if the lowest bidder is decided on the acquisition cost alone. A foreign
supplier with established manufacturing set-up has lower acquisition cost. But if the
life cycle costs of the equipment involving the operation & maintenance (O&M) costs
over the entire life cycle of the equipment (25–30 years) is also taken into consid-
eration in deciding the lowest bidder, the tables turn drastically in favour of Indian
supplier. The O&M costs of s foreign supplier is significantly higher than that of the
Indian supplier. Realising this, the MoD, in the newly drafted DPP-2020 (Defence
Procurement Policy), has mandated consideration of life cycle support aspects of
an equipment and platform. However, the methodology for estimation of life cycle
costs does not exists explicitly in the Indian Naval procurement process as of now.
The paper aims to introduce the concept of life cycle management for equipment of
Indian Navy and suggest way ahead for implementation of the same.
Life Cycle Management for Indian Naval Warship Equipment 211
2 Literature Review
Ameer et al. [3] explores life cycle cost relationship of an induction motor. The
purpose of the paper is to estimate the annual costs of operation and maintenance
of the motors of Water Corporation (WC), Australia, define the conditions under
which an old electric motor would benefit from an upgrade and determine whether
only the heavily used large electric motors benefit from an upgrade. A parametric
life cycle cost model is developed with analogues and predictive elements. The
parametric equation uses the technical specifications of the motors and attached
pumps to calculate the output water and electric power. Scheduled maintenance costs
are subdivided according to minor and major service tasks and calculated by taking
the average of all the major/minor service orders at a pump station. The acquisition
costs and the rewinding costs are considered while deciding upon whether to rewind
a motor or purchase a new motor by calculating the benefits that will accrue with
high efficiency and reduced maintenance costs of a new motor over a rewound motor.
A decision matrix is presented in graphical form for deciding whether to rewind a
motor or purchase a new one.
Bhagwat and Chitrao [1] brings out the flaw in the procedure of evaluation of
competitive bids in equipment of high technology areas in Indian Naval warships. A
majority of such technologies are being imported. Indigenous manufacturers striving
to manufacture such equipment in country are at a disadvantage, as life cycle cost of
the equipment is not considered in deciding the competitive bidder. The indigenous
bidder has to incur high capital costs in establishing manufacturing set-up, thus
pushing the acquisition cost higher than those for the foreign supplier who has an
established manufacturing set-up. However, the operation and maintenance costs for
the foreign supplier is much higher than those of Indian supplier. When considered
for the equipment life cycle of 25–30 years, the life cycle costs for Indian supplier in
much lesser than the costs for the foreign supplier. But this factor is not considered,
and Indian supplier loses out the contract to the foreign supplier. The paper suggests
that life cycle costs be estimated and considered for deciding the lowest bidder.
Defence Contract Management Agency (DCMAO) [4], Baltimore, in this research
study for US Naval Sea Systems Command (NAVSEA), aims to determine the life
cycle costs of US Navy equipment. The systems are divided into three categories: low
complexity, high complexity and ordnance systems. It is argued that determination of
life cycle costs in the acquisition phase results in best estimate of life cycle costs of an
equipment. The study focused on determining the integrated logistics support (ILS)
costs. ILS costs represent a large portion of life cycle costs. The report identifies
various factors that determine the ILS costs of an equipment.
Draft DPP-2020 [2] (Defence Procurement Policy), released by Govt. of India,
Ministry of Defence (MoD), in April 2020, is a revised version of the earlier DPP
released in year 2016. DPP-2020 has addressed the issue of life cycle costing for
equipment and platforms being procured for Indian Navy. In order to optimize life
cycle support, DPP mandates that the equipment procurement agency should ensure
provision of engineering support plan (ESP), annual maintenance contract (AMC),
comprehensive maintenance contract (CMC), life cycle support contract (LCSC)

and performance-based logistics (PBL). Therefore, it is evident that the Ministry of
Defence, Govt. of India, has recognised the importance of considering life cycle costs
for a Naval equipment. The DPP will soon come into force; however, there is still
a need to evolve a model for estimation of life cycle costs of a complex equipment,
especially where imports are involved.
Fajardo et al. [5] brings out importance of life cycle management (LCM) in NATO.
The paper suggests that the organizations within NATO should undergo change to
align with the principles of LCM and goes on to recommend alignment of acquisition
and logistics processes so as to achieve LCM. The paper explains various enablers
required to be in place before commencing implementation of LCM; a key enabler
being a unified business process methodology for acquisition and logistics. A life
cycle working group has been created in NATO to address issues related to poor
management, ill-defined roles, and responsibilities of stakeholders and difficulties of
technology insertion. The objectives are to establish life cycle partnership between
all stakeholders, to have a total and shared view of the project, to create seamless life
cycle management processes that extend from mission analysis to product disposal
and to ensure continuous technology insertion by stressing preference to commercial
and non-developmental solutions to mission needs.
Gualeni et al. [6] argue that the increasing complexities of ships have necessitated
consideration of life cycle costs (LCC) and life cycle assessment (LCA) during ship
design stage. An life cycle performance assessment (LCPA) tool combining LCC
and LCA is developed as part of HOLISHIP European project. The tool allows
cost/benefit assessment over the ship life cycle, considering both the design and
operation. The paper aims to develop operational and maintenance costs, defining
a structured and flexible tool to evaluate the maintenance costs for different ship
solutions. It goes on to provide a general overview of the LCPA tool developed
during HOLISHIP project. It describes the maintenance techniques commonly used
and also the maintenance prediction model.
Martin et al. [7] undertakes a research study to review and assess the life cycle
management (LCM) of equipment instaled on US Naval ships. The study reckons
that that the Navy ships are a complex combination of different equipment ranging
from hull structure to combat systems with maintenance, training and supply require-
ments. Some of the requirements are known, and some arise unexpectedly. The study
brings out that Navy processes in managing these system life cycles are not efficiently
organised and result in seams between ships in new construction and those already
in service. The study recommends management processes for improving the readi-
ness of the equipment, increasing its efficiency and reducing sustainment costs. The
analysis is carried out in three steps. First, the tends in surface ship LCM is analysed
for a class of ships, having large number of ships in service, in all phases of LCM.
The next step is to examine new construction processes to identify any challenges
that may be complicating LCM. The final step is to assess organizational issues that
may be contributing to the identified problems. The study results in recommenda-
tions in four areas, the data sets and systems covering data standards and essential
common data, the maintenance figure of merit system, the incentive structures and
the command and organizational structures.
Sokri [8] argues that accurate cost estimation of a military equipment pays a
crucial role in any military acquisition strategy, bringing out that life cycle costing
(LCC) is an economic assessment that looks beyond the initial purchase cost of
an item. LCC also includes the cost of operating and maintaining the item over its
entire operational life. The LCC can be used as an evaluation tool in two situations,
assessing the economic impact of a given investment such as procurement of new
equipment, and assist in budget allocation if the decision is already made. With this
perspective, LCC is subdivided into three categories: acquisition cost, ownership cost
and disposal cost. The ownership cost is composed of operation and maintenance
(O&M) costs. O&M cost is an important facet of the total ownership costs of a
military equipment and can exceed the cost of capital over the life of equipment.
While applying life cycle costing concepts, analysts have been thwarted by the lack
of accepted methodology to arrive at appropriate decisions. The study aims to develop
a methodology for predicting the ownership costs of a military equipment; optimise
the LCC of acquiring, operating and maintaining military assets; and enhance the
understanding of life cycle costing, both theoretically and practically. The modelling
process is summarised as three stage process. Estimate relationship between O&M
costs and equipment’s age; compute the optimal replacement age of the equipment;
and compute the O&M LCC of the equipment over its entire optimal life.
3 The Research Gap
The brief literature review has clearly brought out the importance of life cycle
management (LCM) concept for military equipment, including equipment that go
onboard an Indian Naval warship. Indian Navy has aggressive warship and submarine
induction plans. These ships and submarines are slated for construction in Indian ship-
yards and will have indigenous design. However, a number of equipment onboard in
these platforms are not manufactured in the country as of now. Indian Navy is depen-
dent upon imports for these equipment and systems. An example is electric propul-
sion system of ship and submarine that comprises of propulsion motor & control
drive, gear box, shaft line, propeller and propulsion auxiliary systems. The technolo-
gies involved in these equipment are mature in Indian industry in civil applications.
However, an indigenous electric propulsion system solution is not available for a
warship application, thus making the country dependent upon imports. A number of
Indian companies viz. Larsen & Toubro Limited (L&T), Bharat Forge, to cite a few,
are keen to support Indian Navy by manufacturing electric propulsion equipment
in the country. Setting up a design, development, manufacture and testing of elec-
tric propulsion equipment of a warship involves large capital expenditure, thereby
making an indigenous solution much more expensive than an import option, if only
acquisition costs are considered in deciding the lowest bidder. The life of a Naval
warship equipment is 25–30 years and the operation and maintenance (O&M) costs
of an electric propulsion system is expected to be significant over the entire life.

The O&M costs sometimes can be higher than the acquisition costs of a warship
equipment [8]. Therefore, it is essential that the life cycle costs (LCC) of a warship
equipment is considered in decision-making process in equipment procurement stage,
as mandated by the new policy issued by the Ministry of Defence, Govt. of India
[2]. Therefore, it is essential that Indian Navy also move towards introducing life
cycle management (LCM) so as to holistically manage and minimise the sustenance
costs of an equipment onboard its warships and submarines. A number of studies are
underway in NATO countries as well as US Navy for adapting LCM [4, 5, 7].
However, a model for estimating the LCC of a warship equipment is not available,
thus making it difficult for procurement agencies to evaluate the bidders on the
basis of LCC. This is a crucial gap in the procurement processes of Indian Navy.
Another crucial gap is in implementation of LCM processes. Indian Navy has discreet
acquisition and logistics processes as of now, and there is a need to move towards
LCM, as is the case in modern Western Navies [4, 5, 7].
4 Research Objective
Define a roadmap for implementation of life cycle management (LCM) process for
electric propulsion system for Indian Naval warships and submarines.
5 Scope and Limitations
The study focuses on formulation of a methodology for estimation of life cycle costs
for electric propulsion equipment for new shipbuilding projects of Indian Navy in
line with stipulations of the Defence Procurement Policy of MoD, GoI. The study
also analyses various methodologies for implementation of life cycle management
aspects and suggesting way ahead in selection and implementation of the relevant
process for Indian Navy.
6.1 Primary Research
It is intended to conduct a survey followed by unstructured interviews with subject

matter experts from Naval shipbuilding agencies, shipyards and Indian industry in
chalking out a roadmap for introduction of LCM into electric propulsion equipment
of the Indian Naval warships. The concept can then be generalised for warship equip-
ment. This is not considered in the scope of this paper and will be undertaken as part
of the main research study.
6.2 Secondary Research
Most of the modern Navies world over is moving towards life cycle management
as regards equipment and systems of a warship. As brought out in literature review,
US Navy has set up a high-level task force for undertaking a research study in
implementation of life cycle management (LCM) aspects. The task force has done
considerable work in areas of LCM. Some of the concepts and findings of the research
study that can be adapted for Indian Naval scenario also are appended in succeeding
paragraphs.
6.3 Organisation Structure and Processes of Indian Navy
There are two distinct organisations in Indian Navy dealing with acquisition and
O&M aspects of a warship equipment. The Staff Branch I, headed by the Vice Chief
of the Naval Staff, is responsible for ordering the equipment for newbuild projects
of the Indian Navy. The project ordering and execution is carried out by Controller
Warship Production and Acquisition (CWP&A), through Design and Ship Production
Directorates. The order for the ship or submarine is placed in the respective shipyard
and is accounted under ‘Capital Budget’. The responsibility of this set-up is limited
up to delivery of the vessels to the operational authority of the Indian Navy, i.e.,
one of the three commands of the Indian Navy, Western, Eastern and Southern Naval
Command. Thereafter, the operation and maintenance of the ship is the responsibility
of the Materiel Branch headed by the Chief of Materiel. The shipyard is responsible
for delivery of the vessel, including the warranty of the entire ship for a period of
one year. This warranty charges are accounted for in the Capital Budget. Thereafter,
the operation & maintenance of the ship along with all the equipment for the life
cycle of 25–30 years is catered for in ‘Revenue Budget’, by the Materiel Branch of
the Indian Navy.
This is a grey area and the agency for building the ships and delivering the same
is not accountable for O&M of the ship.
Even though the draft DPP-2020 has tried to remedy this lacuna of life cycle
management of the ship equipment by introducing life cycle support, but is to a
limited extent only. The draft DPP-2020 has introduced long-term product support
mandatory for 3–5 years by introducing one or more of the following as part of
equipment procurement package:
• Engineering support package (ESP)
• Annual maintenance contract (AMC)

• Comprehensive maintenance contract (CMC)
• Life cycle support contract (LCSC)
• Performance-based logistics (PBL).
The pros and cons for the above are explained in succeeding paragraphs.
It is easy and simple to incorporate ESP, AMC and CMC into the original procure-
ment contract. A number of the equipment procurement contracts include ESP as
its integral part. The ESP comprise of on-board spares (OBS), necessary tools, test
equipment, jigs and fixtures, carried by the ship for undertaking first level mainte-
nance and defect rectification. Therefore, an ESP is almost always included in equip-
ment procurement specifications and is taken as part of the Capital Budget allocated
towards equipment acquisition.
The AMC and CMC are not easy to implement as this forms part of the Revenue
Budget. Some of the tenders makes it mandatory for the bidders to mention AMC
(or CMC) costs for a number of years as a percentage of the equipment cost or in
absolute numbers. But these are optional elements of the tender and is not considered
in the bid evaluation process. Therefore, the very purpose of binding the supplier,
even partly, towards life cycle costs is defeated as there is no incentive to quote a
realistic AMC or CMC cost. Therefore, it is left to the Materiel Branch to conclude
AMC or CMC separately after expiry of the warranty period. Therefore, this lacuna
remains to be addressed.
The PBL is a right step towards binding the equipment supplier towards life cycle
cost of the equipment. But it has to be linked with an AMC or CMC to be implemented
fully.
Life cycle support contract (LCSC) is the correct way of binding the equipment
supplier throughout the O&M cycle of the equipment. This will include the following:
• Spares
• Scheduled maintenance
• Breakdown maintenance
• Shore support facility for repairs, testing and training
• Obsolescence management at least twice in the equipment life cycle of 25–
30 years. One during mid-life update (MLU) of the ship after 10 years of operation
and second after 20 years of operation.
• Disposal of the equipment
However, estimation of life cycle costs including the elements given above is a
complex process and involves active interaction with the suppliers in the pre-bid
stage and making them aware of the various aspects of the LCC. Further, the two
distinct organisations of the Indian Navy, viz, Staff Branch I and Material Branch
will have to undergo streamlining of the processes on the principles of life cycle
management (LCM) to cater for LCSC for a warship equipment.
A number of methodologies are suggested in research studies world over towards
estimation of life cycle cost and implementation of life cycle management aspects
in Naval equipment. These are briefly explained below.
7 Findings
The secondary research was carried out by extensive study of the research papers
pertaining to the advanced Navies world over. A study undertaken for the US Navy
in year 1993 [4] is most relevant to the prevailing scenario and environment in Indian
Navy. The study established that the integrated logistics support (ILS) costs represent
a very large portion of the life cycle costs. Therefore, a great emphasis was placed on
ILS costs in program acquisition. Analyses of life cycle costs and design trade-offs
have been made mandatory for all acquisition programs of US Navy. A concept of
ILS exists in Indian Navy also as part of integrated logistics management system
(ILMS). Each equipment is broken down in equipment, assemblies, sub-assemblies
and kits (EASK) categories with each element listed in the ILMS document giving the
full details of the equipment along with costs. Therefore, it is suggested to implement
the practices followed in US Navy in ILS management. The process followed in US
Navy is known as logistics support analysis (LCA) process.
The objective of the LCA process are to influence system design, identify and
specify logistic support resource requirements, avoid duplication of the analysis
efforts and assess supportability. Such a process is applicable for warship equipment
procurement also, but it is not mature and is, therefore, implemented on a piecemeal
basis. It will become evident as we study the elements of the LCA process and
establish commonality of each process in Indian scenario. Thereafter, it will be
possible to suggest a methodology for implementation of an equivalent process for
Indian Navy, that is the objective of this research paper. A block concept of various
elements of an ILS is given below [4].
While the ILS elements depicted in Fig. 1 are as applicable to US Navy, it is
pertinent to mention here that Indian Navy also has similar elements applicable
ILS Requirements
ILS Planning Design Interface
Configuration Management
Logistics Support Analysis Standardisation
Human Factors Reliability and Maintainability
Quality Assurance System Safety
Test and Evaluation Survivability
Support Supply Packaging, Handling, Computer

Maintenance Equipment Support Storage & Resource Support
Planning Transportation
Technical Data Facilities Manpower & Training and

Personnel Training Support
Fig. 1 ILS elements

for equipment acquisition and O&M. However, as the responsibility for various
elements is divided between the authorities responsible for acquisition and O&M,
the overall costs are not considered during equipment induction, thus creating an
adverse environment for Indian manufacturers, as explained above.
The research studies carried out in NATO Navies and US Navy [5, 7] have also
shown similar situation where in life cycle management processes are not followed
by all stakeholders in an integrated manner. Acquisition authorities focus on ship
delivery and do not take holistic view of the O&M costs. Other studies [6, 8] have
laid thrust in life cycle costing and evaluation of life cycle costs during evaluation of
various design alternatives as it have been conclusively found out that estimation of
life cycle costs during design stage results in most competitive alternative of sourcing
the equipment. Estimation of LCC helps in budget allocations during the entire life
cycle of equipment. The MoD, GoI has also stipulated consideration of long term
product support in the recently promulgated draft DPP [2].
With an aim to link the life cycle cost elements, as applicable to Indian Navy with
the well-defined ILS elements as brought out in Fig. 1 above, a table has been made
as given below. Table 1 links the elements of ILS with the stipulations of the DPP
and suggests methodology for adaptation of the life cycle management process in
Indian Navy.
Therefore, it is necessary to formulate a unified life cycle management (LCM)
process in the acquisition as well as maintenance organisations of the Indian Navy.
Further, the life cycle cost (LCC) of a complex equipment viz. electric propulsion
equipment is to be estimated in line with the methodology explained in Table 1. This
research paper explains the qualitative methodology of adaptation of LCM and esti-
mation of LCC as applicable to electric propulsion equipment of Indian Navy. Further
study will be carried out with primary research methodology involving interaction
with subject matter experts of Indian Navy, academia and Indian industry in defining
quantitative estimation of life cycle cost of an electric propulsion equipment.
8 Recommendation
Indian Navy should adapt life cycle management (LCM) processes in organisations
of all stakeholders responsible for acquisition as well as operation and maintenance
of a warship and its equipment. Evaluation of bidders on the basis of LCC and
not acquisition costs alone, will go a long way in providing level playing field to
Indian defence industry. Indigenous development of electric propulsion equipment
can only succeed if such practices are adapted by Indian Navy in time for procurement
of these equipment. The regulatory framework in the form of Defence Procurement
Policy-2020 is in place and provides a solid foundation towards this journey.
Table 1 Life cycle cost elements for Indian Navy

ILS element Link to Indian Navy process Remarks/suggestions for Indian
Navy
ILS planning Long-term support Loosely defined as of now. Will
get addressed once all other
elements of ILS are taken care
of during acquisition phase
Design interface Various elements required to There is a need to quantify
be considered in the design these factors with an aim to
phase of acquisition program arrive at contribution of these
viz. configuration factors towards life cycle costs.
management, standardisation, This exercise can only be
reliability & maintainability, undertaken when life cycle
safety, survivability, logistics management aspects are borne
support analysis, quality in mind in the equipment
assurance, manned systems acquisition phase. The LCM
engineering and the plans, processes are to be followed in
analysis and data required to a unified manner by the
support these factors are stakeholders viz. acquisition
considered during design authority and maintenance
stage. But these are not authority
quantified
Maintenance planning These are linked with ESP, Quantifying the maintenance
AMC/CMC, PBL, LCSC, as support package as per the DPP
stipulated under the long-term and considering the same
product support, newly during quantitative evaluation
introduced in the DPP-2020 of bids in acquisition phase will
go a long way in encouraging
indigenisation of high
technology equipment viz.
electric propulsion equipment
Support equipment and Support equipment for electric Presently, support equipment is
training equipment propulsion system will include not considered in tender
shore-based test and repair evaluation
facility for undertaking
second-level maintenance. It
will also include shore-based
training facility for imparting
effective training on the
equipment
Supply support A very elaborate supply Presently, spares for a period of
support organisation exists in two to three years of O&M are
Indian Navy under Materiel considered as part of
Branch to ensure logistics acquisition process. This needs
processes in making various to be extended for entire
spare parts available to the service life of the equipment
ship as well as dockyards,
necessary for operation and
maintenance of the equipment
(continued)
Table 1 (continued)
ILS element Link to Indian Navy process Remarks/suggestions for Indian
Navy
Packaging, handling, storage Not significant for electric Is considered in acquisition
and transportation propulsion equipment phase
Computer resource support Indian Navy maintains a Is managed well as of now and
repository of all software of needs minor streamlining
the system
Technical data A very exhaustive technical Although the exploitation
documentation is mandatory documentation is adequately
along with equipment supply captured in acquisition phase,
the maintenance and
troubleshooting documentation
are not given due importance
during acquisition phase. These
document will form the basis of
estimation of O&M costs under
spares, deputation of
specialists, etc. Indian Navy is
paying exorbitant sums in this
element of life cycle costs, and
this element is not considered
during commercial evaluation.
This is a serious lacuna
Facilities and manpower These are as are considered Not recommended to be
separately as manpower budget considered at this stage
Total ILS cost Life cycle cost (LCC) The objective of this study is to
estimate LCC for electric
propulsion equipment
9 Future Scope
The future scope includes quantitative research study with primary research method-
ology involving focused interaction with senior Indian Naval officer in shipbuilding
and equipment maintenance organisations, subject matter experts in Indian shipyards,
Indian industry and academia.
References
1. A. Bhagwat, P. Chitrao, Life cycle costing model for equipment of Indian Naval shipbuilding
programme, in Proceedings of ICT4SD 2019, vol. 1. Advances in Intelligent Systems and
Computing, vol. 1077 (Springer, Singapore, 2019). ISBN 978-981-15-0936-0)
2. Draft Defence Procurement Policy (DPP) 2020. Published by Govt. of India, Min. of Def. https://
mod.gov.in/dod/sites/default/files/draftdppnew200320c.pdf
3. I. Ameer et al., A study into the true life cycle costs of electric motors. Presented in CEED
Seminar in year 2009. https://ceed.wa.edu.au/wp-content/uploads/2017/02/10.Irfan_.Ameer_.
pdf
4. DCMAO Baltimore, Integrated Logistics Support Life Cycle Cost (1993). https://apps.dtic.mil/
dtic/tr/fulltext/u2/a264270.pdf
5. E. Fajardo et al., Life cycle management in NATO. Presented at the RTO Studies, Analysis and
Simulation Panel (SAS) Symposium held in Paris, France, 24–25 Oct 2001. https://apps.dtic.
mil/dtic/tr/fulltext/u2/a418689.pdf
6. P. Gualeni et al., Life cycle performance assessment tool development and application with
a focus on maintenance aspects. J. Mar. Sci. Eng. (2019). https://www.researchgate.net/pub
lication/335252086_Life_Cycle_Performance_Assessment_Tool_Development_and_Applic
ation_with_a_Focus_on_Maintenance_Aspects
7. B. Martin et al., An Approach to Life cycle Management to Shipboard Equipment (RAND
Corporation, Santa Monica, CA, 2018). https://www.rand.org/content/dam/rand/pubs/research_
reports/RR2500/RR2510/RAND_RR2510.pdf
8. A. Sokri, Life cycle costing of military equipment, in Proceedings of Control, Dynamic Systems,
and Robotics, Ottawa, Ontario, Canada, 15–16 May 2014. https://pdfs.semanticscholar.org/e28e/
a1db8770fcc2c96a0f80905781f1ca5ed555.pdf
Framework for Securing IoT Ecosystem
Using Blockchain: Use Cases Suggesting
Theoretical Architecture
Anshul Jain, Tanya Singh, and Nitesh Jain
Abstract Blockchain is a distributed, open record between two customers for all
trades. The decentralized nature is possible because each trade is controlled by the
understanding of a larger number of customers, who are interested in anything in
the system. The authors concentrate and present methods to address real issues of
security in IoT. The authors review use cases and organize common issues of security
with respect to IoT layered building. Despite the current plans for attacks, hazards and
cutting-edge game, we are designing essential IoT security conditions. IoT security
problems are also orchestrated against the current game plans discovered in writing
and mapped against them. Furthermore, we are trying to discuss how blockchain,
which is the fundamental development of Bitcoin, can be the primary enabling impact
to deal with immeasurable IoT security problems. Besides, the paper includes open
research issues and challenges to IoT security. In this paper, the potential informative
applications are focused, along with explorations on how blockchain advancement
can be used to address some preparation issues. In the wake of inquiring about
security issues, authors present an architecture of IoT blockchain ecosystem. This
paper also talks about the highlights and ideal states of blockchain improvement.
Keywords Blockchain · IoT · Data security · Network security · Blockchain IoT

protocols · IoT security · IoT ecosystem · Blockchain-enabled IoT
A. Jain (B)
AIIT, Amity University, Noida, Uttar Pradesh, India
e-mail: anshuljain13@gmail.com
T. Singh
ASET, Amity University, Noida, Uttar Pradesh, India
e-mail: tsingh2@amity.edu
N. Jain
JSSATE, Uttar Pradesh Technical University, Noida, Uttar Pradesh, India
e-mail: jainitesh@gmail.com
224 A. Jain et al.
1 Introduction
Blockchain progress is generally referred to as appropriate record innovation. It

allows people to efficiently check transactions performed, exchange and transfer
advantages in [1]. A model flow of cryptographic currency blockchain exchange is
technique of ease. To initiate a transaction with User B, a shared blockchain is used
by Customer A. A cryptographic character check (two or three open keys and private
key) is used to make an unusual correlation between Customer A and Customer B in
the framework. At this stage, the exchange will be transferred strongly to the memory
pool of the blockchain framework for transaction check and approval. The new block
is shaped by acquiring a specific measure of the avowed center points; this is called
arriving at a consensus. This square incorporates every one of the exchanges that
have occurred during this minute. It is “connected” through the modern imprint to
the main obstacle in the plan in [2]. The foundation of the accord is practiced utilizing
an agreement calculation. This block embodies each of the exchanges taking place at
this time. It is “connected” to the main obstacle in the plan through the modern imprint
in [2]. The platform of the consensus is practiced using a consensus algorithm. This
is referred to as mining industry. In particular, the peer-to-peer plan competes with
the present circumstance of the spread record [3]. Each center can cast a vote by
denying advances by taking increments or dismissing invalid blocks by their CPU
ability to perceive authentic blocks. Any fundamental direction and propelling powers
can be executed through this plan of understanding [4]. Any block of exchange is
distinguished with a specific time mark. The two blocks are equally interfaced by
a timestamp. The blockchain data thus demonstrates a property of time, and the
span of the chain is always developing. It implies that the blockchain is a variety of
transmitted transactions that upgrade the timestamp organization [5–8]. To create a
notable chain of SHA-256 hash limit, explicit hardware is utilized by blockchain,
and cryptographic data is used to prevent invincible client data from being adjusted
[1].
The organization of this paper is as follows. Section 2 explains different applica-
tions in IoT ecosystem using blockchain technology, and it also describes its security
requirements. Next Sect. 3 explains different benefits of blockchain which makes it
robust and reliable. Section 4 describes a combined and modular architecture of IoT-
enabled blockchain technology. A brief literature survey is done in Sect. 5 which goes
through several papers and writes the outcome of the research done. This section also
reviews several papers and presents them in a tabular format representing different
IoT blockchain methods proposed, their benefits, shortcomings and future scope.
Finally, conclusion is presented in last Sect. 6.
Framework for Securing IoT Ecosystem Using Blockchain … 225
2 Internet of Things Domain Use Cases Using Blockchain

and Their Security Requirements
This section includes various IoT applications which explains the problems faced
due to lack of security in the study.
2.1 IoT Blockchain-Enabled Financial Services
Numerous IoT frameworks are likely to be subject to smaller-scale budget exchanges

between computerized articles, requiring the IoT devices be associated in a manner
that considers the so-called machine-to-machine (M2 M) economy–which is essen-
tially the money trading between non-human devices. The number of transactions is
concerned; countless blockchain schemes have restricted execution every time they
can cope with it. This implies most blockchain executions of proof of work and
proof of stake blockchain present the limited potential for versatility, making them
unacceptable for large-scale handling of M2 M microtransactions. Nevertheless, it
is worth mentioning that countless blockchain operations are shifting toward adapt-
ability provisions, such as the Bitcoin lightning network and the Ethereum Plasma
[9].
2.2 IoT Security with Blockchain
Another safety building that dodges the use of cryptography scheme based on
the trade-in symmetric or asymmetric cryptography keys is suggested that is not
prescribed in IoT devices due to their handling, storage and energy containment
[10]. The suggested scheme relies on the multi-operator being used to guarantee the
safety of related objects. With no compelling reason to register keys, these studies
modified with specialists can speak to different items. The main goal of this job is to
maintain an abnormal state of safety by enhancing the energy usage of IoT devices
[11].
2.3 Privacy-Preserving Healthcare Blockchain for IoT
Designs are for confirming an unbound IoT-based implementation, especially

concerning applications for medicinal services where customer approval is a key
issue. The three procedural examples, i.e. Security Goal Documentation, Choose
the Right Stuff and enroll from Outsider Pattern, are examined in [12]. Five other
examples of approval are shown in this investigation as reference monitor, access
226 A. Jain et al.
matrix role approval guidelines and role-based access control, remote authenti-
cator/authorizer and file authentication patterns. For IoT-based E-healthcare applica-
tions, a protected gathering-based lightweight validation conspire has been proposed
where the proposed model will provide shared validation and energy proficiency and
calculation for IoT-based social insurance applications [13], which will use elliptic
curve cryptography (ECC) guidelines that provide mention included of the suggested
model [14].
2.4 IoT Blockchain Support to LoRaWAN Network Servers
An IoT framework model is presented that uses the LoRaWAN convention to transmit
information from sensor hubs to our cloud administration and the things network
stage that executes the backend administrations of LoRaWAN [15]. The authors have
organized a structure that can be adapted and extended to expand new advantages
just as other IoT phases are coordinated. It is also equally adaptable, which means
that a presentation can be designed simply by creating new server instances (Fig. 1).
Fig. 1 Application and benefits of IoT blockchain security

3 Benefits of Blockchain
Reliability. The databases of the entire transaction records are changed by the decen-
tralized concept. Cloud adds to the reliability where which helps data to be easily
accessed and processed with security features like asymmetric encryption [16].
Trust. The blockchain structure is happening as new trust providers with decentralized
data. These data is distributed across a system of fixed centers [17].
Security. The blockchain framework uses one-way hash work. The yield has no
unmistakable connection to the information.
Efficiency. Blockchain advancement could accelerate the clearing and repayment of
specific exchanges identified with cash by decreasing the number of middle people
included and making the trade-off procedure faster and gradually profitable [18].
4 Literature Review
Another IoT access control structure is based on the blockchain’s development. The
objective of the study is to provide a reference model for the proposed structure in the
IoT objectives, models, architecture and mechanism [19]. In addition, fair access is
presented as a full pseudonymous and security-saving executive system authorization
that empowers customers to own and regulate their data. For updating the model, the
blockchain is used and adjusted to decentralized control of access.
A decentralized trust system, called IoT passport, is suggested using innovation
in blockchain for coordinated cross-stage efforts [20]. A reference design for shared
IoT applications based on blockchain, featuring the main correspondence channels
anticipated to interface the IoT layer with the blockchain layer, was shown in [21].
By establishing frameworks and conventions for shared IoT applications based on
blockchain, the practicality of the suggested engineering has been explained.
A successful and controlled blockchain plot-based transmitted cloud architecture
is discussed in CRVANET’s biological system instead of custom cloud engineering
to protect driver safety with minimal effort and on-demand detection in [22].
In addition to providing research methodology to identify concerns and arrange-
ments, an efficient survey of existing IoT security and protection systems is provided
in [23]. It discusses IoT design, characterizes the distinct areas of safety and security
and provides a scientific classification and close investigation to plot safety goals,
risks and attacks and agreements suggested late (2015–2018) (Table 1).
228 A. Jain et al.
Table 1 Review of proposed methods

Reference No Proposed method Benefits Shortcomings Future scope
[20] IoT passport It helps in It is completely Future work calls
commodity user dependent on for collaborative
interfaces trust and is very IoT
susceptible to
security threats
[6] IoT chain It provides a The Private ethereum
flexible performance blockchain network
authorization and robustness needs upgradation
have not been to be used in the
much evaluated PoS-based version
of the ledger
[8] A blockchain-based It provides The number of Future
proxy re-encryption scalability and mined implementation
scheme security to the transactions includes
automated depends on implementation on
payments various factors different
including gas blockchain
price and limit platform, e.g.,
hyperledger
[24] An It provides a The firmware Future work
over-the-blockchain secure mechanisms are includes the
firmware update verification to too lengthy to development of a
framework the firmware operate robust and
lightweighted
protocol
[25] Unibright New templates A huge number Future work needs
are created to of risks are to focus mostly on
recognize associated with the risks associated
different it
industries
[26] A decision It shows the This framework This decision
framework to systematic still did not framework can be
support blockchain identification of provide any improved into a
platforms for IoT blockchain specific recommender tool
and edge computing characteristics
suitable for IoT
platform
5 Proposed Architecture–Blockchain in IoT Ecosystem
The architecture of IoT ecosystem as explained in [27] is fully modular, and all
the modules can be easily decoupled from one another in order to add a module
and keeping the rest of the system intact. [28] Further provides a comparative study
on the architecture, which shows security issues and solutions on application layer.
Integrity-related problems can be handled by blockchain solutions discussed in this
paper.
Fig. 2 Proposed architecture for blockchain in IoT ecosystem [27, 28]
This paper proposes in Fig. 2 a modification to the architecture discussed in [28],

where blockchain is introduced before the application layer which can address some
of the security issues addressing our concerns related to confidentiality, integrity
and accountability. Several linked devices possessing the capabilities of communi-
cation, data storage and computing are present in the IoT physical layer. The connec-
tivity layer provides services which include managing the networks, managing the
security and passing of messages. The IoT blockchain service layer provides all
services including managing identities, peer-to-peer communication and consensus.
The application layer consists of the interfaces which visualize the data from all the
physical devices for regulating and manipulating devices.
6 Future Projection of an IoT and Blockchain

Authorization Model
The Internet of things (IoT) has stretched the Web availability to reach PCs and
people, yet most of our population is missing Internet of things. The IoT may be able
to combine billions of articles at the same time, which has the effect of improving the
sharing of data needs as a result of improving our lives. Even though the advantages of
IoT are boundless, due to its concentrated server/customer model, the IoT are facing
numerous difficulties. For example, the versatility and security issues that arise as a
result of the unreasonably large number of IoT issues in the system. The server model
requires all devices to be linked and verified through the server, which is the single
point of failure. In this way, moving the IoT framework in a decentralized direction
230 A. Jain et al.
could be the right choice. One of the well-known frameworks for decentralization
is blockchain. The blockchain is a ground-breaking innovation that decentralizes
computation and board forms that can capture a vast number of IoT issues, security.
Blockchain innovation is the missing link to solve the problems of versatility, protec-
tion and reliability in the Internet of things. Blockchain advances might be the silver
slug required by the IoT business. New Innovations in Blockchain technology can
be used to track billions of devices, empower exchanges and coordination between
devices are considered as significant reserve funds for IoT industry. This decentral-
ized methodology would be used for the sole purpose of disappointment, creating
a stronger biological system for devices to keep going. Cryptographic calculations
used by blockchains would make purchaser information confidential. Blockchain-
based IoT arrangements are appropriate for streamlining business forms, improving
customer experience and achieving significant cost efficiencies.
7 Conclusion
As we envision the future growth of IoT, we must also discuss about its security as one
of the utmost requirements of IoT ecosystem. In this paper, we have addressed secu-
rity issues in IoT ecosystem using blockchain technology. We have also discussed a
few domains like finance, healthcare, network, smart devices, etc., all these domains
have their specific security requirements and challenges. This paper discusses the
probable models which can be used by these domains to address security issues using
blockchain technology. The broad overview provided in this paper also discusses
about the combined and modular architecture of IoT and blockchain, and this archi-
tecture is designed in such a flexible way that all the devices and layers can be coupled
or decoupled as per the domain’s requirement. This paper also sheds some light on
the benefits of blockchain which gives us confidence of using this technology in
our future deployments. Furthermore, we have done a brief literature survey on the
solutions discussed by several technical papers on this technology. Different solu-
tions provided by several writers are compared based on their benefits, shortcomings
and future scope on the proposed method. Before concluding this paper, we have
also discussed a future projection of an IoT and blockchain-based authorized model,
which mentions specific details required for this model. This paper is just a drop
in the ocean of blockchain technology, there is long way to go and much more to
dive into, but we expect that we are able to address some of the challenges of IoT
ecosystem using blockchain technology and we hope that paper will be helpful in
researches going to happen in near future. We are hopeful that blockchain is going
to be one of the pioneer technologies in addressing security issues.
References
1. F. Tschorsch, B. Scheuermann, Bitcoin and beyond: a technical survey on decentralized digital

currencies. IEEE Commun. Surv. Tutorials 18(3), 2084–2123 (2016)
2. J. Yli-Huumo, D. Ko, S. Choi, S. Park, K. Smolander, Where is current research on blockchain
technology?—a systematic review. PLoS ONE 11(10), e0163477 (2016)
3. D. Kraft, Difficulty control for blockchain-based consensus systems. Peer-to-Peer Networking
Appl. 9(2), 397–413 (2016)
4. S. Nakamoto, Bitcoin: A peer-to-peer electronic cash system (2008)
5. S. Haber, W. Stornetta, How to time-stamp a digital document, Crypto’90, LNCS 537 (1991)
6. O. Alphand, M. Amoretti, T. Claeys, S. Dall’Asta, A. Duda, G. Ferrari, F. Rousseau, B.
Tourancheau, L. Veltri, F. Zanichelli, IoTChain: a blockchain security architecture for the
internet of things, in 2018 IEEE Wireless Communications and Networking Conference
(WCNC) (IEEE, 2018), pp. 1–6
7. S. Mendhurwar, R. Mishra, Integration of social and IoT technologies: architectural framework
for digital transformation and cyber security challenges, in Enterprise Information Systems
(2019), 1–20
8. A. Manzoor, M. Liyanage, A. Braeke, S.S. Kanhere, M. Ylianttila, Blockchain based proxy
re-encryption scheme for secure IoT data sharing, in 2019 IEEE International Conference on
Blockchain and Cryptocurrency (ICBC) (IEEE, 2019), pp. 99–103
9. The derivative effect how financial services can make IoT technology pay off. Retrieved
from https://www2.deloitte.com/us/en/pages/financial-services/articles/the-derivative-effect-
how-financial-services-can-make-iot-technology-pay-off.html. Last accessed 15 Jan 2020
10. B.A. Abdelhakim, B.A. Mohamed, B. Mohammed, B.A.O. Ikram, New security approach for
IoT communication systems, in Proceedings of the 3rd International Conference on Smart City
Applications (ACM, 2018), pp. 2
11. How to secure the internet of things with blockchain. Retrieved from https://www.devteam.
space/blog/how-to-secure-the-internet-of-things-iot-with-blockchain/. Last accessed 15 Jan
2020
12. M. Aydar, S.C. Cetin, S. Ayvaz, B. Aygun, Private key encryption and recovery in blockchain,
arXiv preprint (2019), arXiv:1907.04156
13. A. Pazaitis, P. De Filippi, V. Kostakis, Blockchain and value systems in the sharing economy:
the illustrative case of backfeed. Technol. Forecast. Soc. Chang. 125, 105–115 (2017)
14. A.D. Dwivedi, G. Srivastava, S. Dhar, R. Singh, A decentralized privacy-preserving healthcare
blockchain for iot. Sensors 19(2), 326 (2019)
15. Network Architecture, https://www.thethingsnetwork.org/docs/network/architecture.html.
Last accessed 25 Jan 2020
16. Y. Zhang, C. Xu, J. Ni, H. Li, X.S. Shen Blockchain-assisted public-key encryption with
keyword search against keyword guessing attacks for cloud storage. IEEE Trans. Cloud
Comput. (2019)
17. S. Underwood, Blockchain beyond bitcoin. Commun. ACM 59(11), 15–17 (2016)
18. H. Wang, K. Chen, D. Xu, A maturity model for blockchain adoption. Finan. Innov. 2(1), 12
(2016)
19. S. Moin, A. Karim, Z. Safdar, K. Safdar, E. Ahmed, M. Imran, Securing IoTs in distributed
blockchain: analysis, requirements and open issues. Future Generation Comput. Syst. 100,
325–343 (2019)
20. B. Tang, H. Kang, J. Fan, Q. Li, R. Sandhu, IoT passport: a blockchain-based trust framework
for collaborative internet-of-things, in Proceedings of the 24th ACM Symposium on Access
Control Models and Technologies (ACM, 2019), pp. 83–92
21. G.S. Ramachandran, B. Krishnamachari, A reference architecture for blockchain-based peer-
to-peer IoT applications. arXiv preprint (2019), arXiv:1905.10643
22. S. Nadeem, M. Rizwan, F. Ahmad, J. Manzoor, Securing cognitive radio vehicular ad hoc
network with fog node based distributed blockchain cloud architecture. Int. J. Adv. Comput.
Sci. Appl. 10(1), 288–295 (2019)
232 A. Jain et al.
23. M. Gulzar, G. Abbas, Internet of things security: a survey and taxonomy, in 2019 International
Conference on Engineering and Emerging Technologies (ICEET) (IEEE, 2019), pp. 1–6
24. A. Yohan, N.W. Lo, An over-the-blockchain firmware update framework for IoT devices, in
2018 IEEE Conference on Dependable and Secure Computing (DSC) (IEEE, 2018), pp. 1–8
25. S. Schmidt, M. Jung, T. Schmidt, I. Sterzinger, G. Schmidt, M. Gomm, K. Tschirschke,
T. Reisinger, F., Schlarb, D. Benkenstein, B. Emig, Unibright-the unified framework for
blockchain based business integration, in Unibright Project White Paper (2018)
26. N. El Ioini, C. Pahl, S. Helmer, A decision framework for blockchain platforms for IoT and
edge computing. SCITEPRESS (2018)
27. A. Jain T. Singh, in Security challenges and solutions of IoT ecosystem, ed. by M. Tuba, S.
Akashe, A. Joshi. Information and Communication Technology for Sustainable Development.
Advances in Intelligent Systems and Computing, vol. 933 (Springer, Singapore, 2020)
28. A. Jain, T. Singh, S.K. Sharma, Threats paradigm IoT ecosystem, in 2018 7th Interna-
tional Conference on Reliability, Infocom Technologies and Optimization (Trends and Future
Directions) (ICRITO) (Noida, India, 2018), pp. 1–7
Feature Selection for Handwritten
Signature Recognition Using
Neighborhood Component Analysis
Kamlesh Kumari and Sanjeev Rana
Abstract Advance trends adopted in classification of images using deep learning

networks have paved opportunities for handwritten signature identification and veri-
fication. Convolution neural network has demonstrated to be profoundly proficient
in image classification, yet this methodology has a few weaknesses also. Notably, a
huge number of images are required for training to achieve high level of a precision
for classification. Transfer learning with pretrained deep convolution neural network
model can be utilized to conquer these issues. To bring about further enhancements
in terms of recognition rate, storage and computational time, feature selection proce-
dure plays a key role by distinguishing the most significant features. In this work, we
used a pretrained DCNN model, i.e., GoogLeNet for feature extraction using transfer
learning and further used neighborhood component analysis to reduce the irrelevant
features. Support vector machine is used for classification. Experiments have been
conducted using three signature datasets, i.e., CEDAR, SigComp2011 and UTSig.
Keywords AlexNet · Support vector machine (SVM) · GoogLeNet ·

Neighborhood component analysis (NCA) · Deep convolution neural network
(DCNN) · VGGNet · Center of excellence for document analysis and recognition
(CEDAR)
K. Kumari (B) · S. Rana

Department of Computer Science & Engineering, M.M.D.U, Mullana, Ambala, India
e-mail: savyakamlesh@gmail.com
S. Rana
e-mail: dr.sanjeevrana@mmumullana.org
234 K. Kumari and S. Rana
Filter Concatenation
1*1 3*3 5*5 3*3 Max

Convolution Convolution Convolution Pooling
Previous Layer
Fig. 1 GoogLeNet-inception module
1 Introduction
Amongst various biometric frameworks, handwritten signature recognition has got

distinct fascination throughout the most recent three decades. In recent years, deep
convolution neural network-based features show the best result for image classifi-
cation. So in our research work, we investigate the performance of signature iden-
tification and verification utilizing features based on GoogLeNet and further use
neighborhood component analysis for feature selection. In AlexNet, there is not
any fixed pattern [1]. The convolutions for each layer are chosen experimentally. In
VGGNet, you can distinguish a structure obstruct as two convolution layers followed
by pooling layers [2]. GoogLeNet has a quite different architecture than AlexNet and
VGGNet [3]. It uses inception module that performs numerous convolutions, with
various filter sizes, and just as pooling in one layer. Consequently, it makes network
wider than deeper. Inception module is limited to filter sizes 1 × 1, 5 × 5 and 3 × 3
as shown in Fig. 1.
The outline of this paper is as follows: Neighborhood component analysis is
described in Sect. 2. Section 3 shows the related work of the authors who have
considered the features extracted from pretrained CNN model. The methodology
has been elaborated in Sect. 4. Experimental results and analysis are detailed in
Sect. 5. Conclusion of the work is discussed in Sect. 6.
2 Neighborhood Component Analysis
NCA is a compelling strategy for choosing critical component that focuses on

high-dimensional information [4]. Neighborhood component analysis is used for
Alzheimer’s disease [5] and skin lesion classification [6]. The subtleties of the feature
section process in context of signature recognition are as per the following:
Let us consider signature subtype classification, multi-class classification.
Suppose c be the subtype classes and n the number of signature samples. At that
Feature Selection for Handwritten Signature Recognition … 235
point, a given training set can be depicted as follows:
S = {(x1 , y1 ), (x2 , y2 ), . . . , (xn , yn )} (1)
whereas x i stands for the r-dimensional vector and yi denotes the class corresponding
to vector.
The choice of reference point in NCA is based on some probability which is called
selection probability pij as described in Eq. (2). If the reference point of x, i.e., x i
is closer to x, then the probability pij will have higher value, as measured by the
distance function in Eq. (3) and kernel (k).
k(dw (xi , yi ))
Pi j = n (2)
j−1, j=1 k(dw (x i , yi ))

p

dw (xi , yi ) = wr2 xir − x jr , Wr is a weight assigned to the r th feature. (3)
r =1
For a given S, the probability of correct classification of signature sample x i is

defined by Eq. (4) and is called average leave one out probability pi where yij is 0
for i = y.

n
pi = Pi j yi j (4)
i=1, j=i
n
Probability of correct classification of all signature samples is F(w) = i=1 pi −
p
λ r =1 wr2 , where λ is regularization parameter.
3 Related Work
Architecture similar to inception is used for offline signature verification [7]. Perfor-
mance of proposed architecture is tested on three public datasets, i.e., UTSig, CEDAR
and BHSig260. False acceptance rate of 6.38, 3.38 and 23.34 is obtained for Hindi,
Bengali and Persian signature, respectively. Also, performance on CEDAR is 100%.
Particle swarm optimization approach to select features is proposed in the work [8].
Database used is GPDS-960, and features used are signet-2048. With this method,
55% feature give same result that was achieved with 2048 feature from signet. This
method controls the overfitting.
Experiments are performed on signature of thirty users in [9]. Transfer learning
approach using two DCNN (i.e., VGGNET and AlexNet) is used. With 50 epochs,
outcome is 100%. Three different DCNN models with three, four and five CONV
layers are also tried. Result of DCNN with five CONV layer is 99.17% with execution
time of 115s. A DCNN with 4 CONV layer and one dense layer with twin structure
is proposed in [10]. Syntactic signatures are generated to increase the samples. The
equal error rate of 4.84% for CEDAR and 3.93 for SigComp2011 are obtained.
Performance of features based on pretrained DCNN model is tested on datasets used
in signature verification competitions [11]. Investigations are performed on seven
sets, i.e., two sets for Japanese, Chinese, Italian and one set for Dutch. Four types of
signatures are used, and outcome of SVM with AlexNet-based features is better than
DT. Results are expressed in confusion matrix. Accuracy of Japanese and Chinese
signature is 99.2% and 100% for one set of each which are better than other two
languages, i.e., Dutch and Italian.
Investigations are performed on scanned signatures of three datasets using
AlexNet-based features and two machine learning algorithms in the research [12].
Two sets are formed each of seven users. Outcome of SVM is better than DT. Perfor-
mance is expressed in confusion matrix for each user. The recognition rate of 100%
and 99.1% is obtained for Bengali and Hindi signatures, respectively. Also, outcome
on CEDAR and UTSig dataset is good, i.e., above 90%. Performance of bag of
features is also compared with these features.
4 Methodology
Flow diagram of proposed approach, i.e., from input to prediction is shown in Fig. 2.
4.1 Preprocessing
GoogLeNet recognizes input size of 224 × 224 × 3, so all the scanned signatures of
dataset are resized to 224 × 224.
Scanned Feature Extraction

Preprocessing
Signature (GoogLeNet)
Feature Selection
Genuine / Forged SVM
(NCA)
Fig. 2 Flow diagram of proposed approach, i.e., from input to prediction

4.2 Feature Extraction
To extract features from any pretrained deep CNN model, it is essential to mention
from which layer of pretrained model, we want to extract feature. The length of
feature vector depends on the layer from which we are extracting features. Extracting
feature from layer named pool5-drop_7 × 7_s1 results in a set of 1024 features for a
single image. Whereas, in our work, feature representation of scanned handwritten
signature images is extracted from a layer named as loss3-classifier of GoogLeNet
that yields a length of 1000 values for each image.
4.3 Feature Selection
We used NCA for feature selection. A higher weight value of features indicates better
discrimination value. The best λ value is chosen to reduce the number of features
through cross-fold validation considering minimum loss value.
4.4 Classification
Support vector machine is used. Outcome is expressed in confusion matrix.
5 Experimental Results and Analysis
We investigate the performance on two set of seven users on CEDAR [13] and UTSig
[14] datasets and Dutch users from SigComp2011 [15] in MATLAB [16]. Table 1
shows the notations used in experiment.
Table 1 Dataset and their

CEDAR CED-I User 1–7
notations
CED-II User 8–14
UTSig P-I User 1–7
P-II User 8–14
SigComp2011 DUT Dutch users
Fig. 3 Feature weights for Dutch signatures
5.1 Experiment on DUT
Feature weights of 1000 features extracted from GoogLeNet for Dutch signature are
shown in Fig. 3. Using NCA, we got 176 features and obtained recognition rate of
96.4% as shown in Fig. 4. Plot of loss values for different lambda values is shown
in Fig. 5.
5.2 Experiment on P-I and P-II
Feature weights of 1000 features extracted from GoogLeNet for P-I are shown in
Fig. 6. Plot of loss values for different lambda values for P-I is shown in Fig. 7. Using
NCA, we got 97 features for P-I with outcome of 89.5% (as shown in Fig. 8) and
167 features for P-II with recognition rate of 91.7%.
5.3 Experiment on CED-I and CED-II
Feature weights of 1000 features extracted from GoogLeNet for CED-I are shown in
Fig. 9. Plot of loss values for different lambda values for CED-I is shown in Fig. 10.
Using NCA, we got 162 features for CED-I with outcome of 91.8% as shown in
Fig. 11 and 159 features for CED-II with recognition rate of 91.8%.
Fig. 4 Confusion matrix for Dutch signatures
Fig. 5 Average loss values versus lambda values for Dutch signatures
Fig. 6 Feature weights for P-I
Fig. 7 Plot of loss values for different lambda values for P-I
5.4 Analysis
Comparison of performance achieved through NCA is described in Table 2. Outcome

of GoogLeNet with NCA is better for Dutch and Persian signatures in terms of
Fig. 8 Outcome of 97 features selected by NCA for P-I
Fig. 9 Average loss values versus lambda values for CED-I
prediction and storage (due to reduced feature). Outcome of GoogLeNet without

NCA is better for CED-I and CED-II.
Fig. 10 Feature weights of 1000 GoogLeNet features (CED-I)
Fig. 11 Outcome of 162 features selected by NCA (CED-I)
Table 2 Comparison of
Dataset GoogLeNet (%) GoogLeNet + NCA (%)
performance with features
based on GoogLeNet CED-I 92.9 91.8
CED-II 93.9 91.8
DUT 92.9 96.4
P-I 88.7 89.5
P-II 85 91.7
6 Conclusion
Thinking about the ongoing accomplishment of deep learning, we explored a tech-

nique for the classification of handwritten signature recognition using NCA for
feature selection. In the wake of actualizing this thought, we are in a situation to
advance a couple of cases including: (i) further improvement in analyzing feature
selection and (ii) dimensionality step to be further decreased that significantly
improves the classification results in terms of storage and execution time and recog-
nition rate for both Persian and Dutch signatures. Finally, as future work, more
advanced convolutional network models can be used to acquire better results for
handwritten signature recognition problems.
References
1. A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional

neural networks, in Advances in Neural Information Processing Systems (2012)
2. C. Szegedy et al., Going deeper with convolutions, in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (2015)
3. K. He et al., Deep residual learning for image recognition, in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (2016)
4. W. Yang, K. Wang, W. Zuo, Neighborhood component feature selection for high-dimensional
data. JCP 7(1), 161–168 (2012)
5. M. Jin, W. Deng, Predication of different stages of Alzheimer’s disease using neighborhood
component analysis and ensemble decision tree. J. Neurosci. Methods 302, 35–41 (2018)
6. T. Akram, H.M.J. Lodhi, S.R. Naqvi, S. Naeem, M. Alhaisoni, M. Ali, N.N. Qadri, A multilevel
features selection framework for skin lesion classification. Human-centric Comput. Inform. Sci.
10, 1–26 (2020)
7. R.K. Mohapatra, K. Shaswat, S. Kedia, Offline handwritten signature verification using CNN
inspired by inception V1 architecture, in 2019 Fifth International Conference on Image
Information Processing (ICIIP) (IEEE, 2019), pp. 263–267)
8. V.L. Souza, A.L. Oliveira, R.M. Cruz, R. Sabourin, Improving BPSO-based feature selec-
tion applied to offline WI handwritten signature verification through overfitting control. arXiv
preprint (2020), arXiv:2004.03373
9. A. Hirunyawanakul, S. Bunrit, N. Kerdprasop, K. Kerdprasop, Deep learning technique for
improving the recognition of handwritten signature. Int. J. Inform. Electron. Eng. 9(4) (2019)
10. V. Ruiz, I. Linares, A. Sanchez, J.F. Velez, Off-line handwritten signature verification using
compositional synthetic generation of signatures and siamese neural networks. Neurocom-
puting 374, 30–41 (2020)
11. K. Kumari, S. Rana, Offline signature recognition using pretrained convolution neural network
model. Int. J. Eng. Adv. Technol. (IJEAT) 9(1), 5497–5505 (2019)
12. K. Kumari, S. Rana, S. Offline signature recognition using deep features, in International
Conference on ICT for Intelligent Systems (Springer, 2020)
13. http://www.cedar.buffalo.edu/NIJ
14. A. Soleimani, K. Fouladi, B.N. Araabi, UTSig: a persian offline signature dataset. IET
Biometrics 6(1), 1–8 (2016)
15. TC11, I. A. P. R., ICDAR 2011 Signature verification competition (SigComp2011) (2011)
16. http://www.mathworks.in/
Blockchain-Based E-Voting Protocol
Shreya Shailendra Chafe, Divya Ashok Bangad, and Harsha Sonune
Abstract Today, each individual voter expects that his vote is counted correctly,
and that the election results are not rigged. Every voting process affects voter’s
trust in the system. In other words, it is very important that the government is able
to successfully hold non-fraudulent, transparent, trustworthy and most importantly,
secure elections to affirm the faith of public. In this paper, we propose a system
that makes use of blockchain to make the elections secure, maintain anonymity
of voter and is decentralized. We propose a new e-voting protocol which adheres
to the fundamental e-voting properties and in addition offers to make the process
decentralized.
Keywords Election · Transparent · Trustworthy · Secure · Blockchain ·

Anonymity · Decentralized
1 Introduction
Election plays an important role in building a nation and determining its fate. Elec-
tions permit voters to exercise their democratic power by casting a vote for a meriting
candidate who would represent them within the government [1]. However, sadly,
voters are being bribed or the ballot machines are being tampered by the parties to
win the elections in an unethical manner. There have been many such cases of rigging
and fraud in past, across the world. Cases related to election manipulation, election
fraud or illegal intrusion with the electoral process have been seen in US Election
(2016), Malawi elections (2019), Turkish General Election (2015), Ugandan General
S. S. Chafe (B) · D. A. Bangad · H. Sonune

Cummins College of Engineering for Women, Pune, India
e-mail: shreya.chafe@cumminscollege.in
D. A. Bangad
e-mail: divya.bangad@cumminscollege.in
H. Sonune
e-mail: harsha.khedkar@cumminscollege.in
246 S. S. Chafe et al.
Election (2006) and many more. Therefore, it is the need of an hour to fabricate an
election system free from external interventions.
Every citizen’s vote should be autonomous, and he should be able to verify whether
or not his vote is counted properly. In this paper, we are proposing a system that aims
at creating the present e-voting system transparent, auditable and secure. The current
electoral system in India is as mentioned further. Constituencies are created within
which people can cast their vote or can even apply to get a ticket to stand in elections
themselves. Each constituency consists of an electoral roll that contains the lists of
voters. Any Indian national who is eighteen or more can exercise his power to vote.
On the day of the election, individuals have to head to the election constituencies
to cast their vote. They need to press a button against a candidate of their choice on the
electronic mechanical device (electronic voting machine—EVM) to cast their vote.
On the day of vote counting, EVMs are checked and unlocked. Then, the counting
of votes begins, and a candidate with maximum number of votes wins the election.
This method needs to be carried out under the presence of eminent officers.
The current electoral system is centralized, which means the system is controlled
by a central authority [1]. This could lead to tampering of votes while counting [1].
Sometimes, it might happen that the EVMs are programmed to constantly yield the
same output for different inputs. Thus, every voter of India cannot keep track of his
vote. Therefore, we can use blockchain to overcome the issues like change of state of
the votes and rigged EVMs [1] within the current election system. Online voting can
prove to be an efficient solution for this problem [1]. Using blockchain technology,
online voting could encourage voter’s involvement and help reinstate the public’s
trust in the democracy [2].
Let us take a look at what is blockchain and how it will develop a secure online
voting system. Blockchain, also known as distributed ledger technology (DLT) [3], is
a peer-to- peer network [4], open to anyone [3]. A blockchain is a chain of blocks that
contains information, where each block is linked with each other using cryptography
[3]. Each block consists of a hash, a timestamp and the data of transactions that have
occurred [3]. A very interesting property of blockchain is that once some data has
been recorded, it becomes very difficult (almost impossible) to change it, rendering
the records immutable [3, 5]. Other main features of blockchain include validating,
consensus and verifiability.
Blockchain is used to store data in addition to keeping it safe from any malicious
attack while keeping the identity of voters disclosed. Blockchain would shift the
current centralized electoral system to a decentralized voting system, giving complete
transparency. Thus, implementation of blockchain is a very efficient approach toward
a secure and transparent election process.
2 Design Considerations
The following are some design requirements that should be taken under consideration
[6–8]:
Blockchain-Based E-Voting Protocol 247
2.1 Eligibility
The system ought to be designed solely on the degree of eligibility of a person

who is exercising the power to cast a vote. Eligibility needs should embrace age
restrictions and citizenship of the individual.
2.2 Anonymity
The system ought to be fabricated in a way that the votes cannot be traced back to a
exact voter, thus revealing the identity of the voter, i.e., to possess the flexibility of
permitting a citizen to vote anonymously [9].
2.3 Unchangeable Records
The system ought to be fabricated in such a way that the votes casted cannot be
mutated or modified by anyone [9].
2.4 Single-Cast
The system ought to be designed such that a citizen will exercise the power to vote
just once and not multiple times.
2.5 Authentication
The system ought to be designed such that solely those that had registered previously
will vote. In different words, a citizen shall be supplied with some variety of identity
(election-card) which can attest them to vote [10].
2.6 Transparency
The system ought to be designed in such a way that it provides transparency to avoid
any ambiguity, within the style of a verifiable assurance to every citizen that their
vote was counted, correctly, besides maintaining anonymous standing of citizen [2].
3 Properties of Blockchain to be Used in E-Voting
3.1 Smart Contract
Smart contract is a small computer code deployed in blockchain nodes. They are
immutable and distributed [5]. Being immutable means that no one can tamper with
the code of the contract. And being distributed means the code and conditions in the
contract are publicly available on the ledger. They are initiated by the message which
is embedded in the transaction. These are used to layer the trust infrastructure of the
blockchain with logic and computation.
3.2 Boot Node
The boot node is assigned a static IP, instead of dynamic IP so that district nodes can
locate their peers quicker. A boot node is a coordination and discovery service that
helps the district nodes to learn about each other [1, 6].
3.3 District Node
Each district will have its own district node, which will have a software mediator
that interacts independently with the boot node and manages the life cycle of smart
contract and entire balloting process [1, 6].
3.4 Asymmetric Cryptography
This uses mathematically related public key and private key. Private key is kept
secret, while the public key is shared with everyone. Data encrypted by the sender’s
private key and receiver’s public key can be decrypted only with the sender’s public
key and receiver’s private key.
3.5 Hashing
It is used to verify that the information is not been tampered or modify [9]. Each
node in the blockchain has a unique identity which is a 256-bit hash placed on its
header, which is calculated using the secure hash algorithm (SHA-256) [11]. Input in
all the states of the blockchain is a plain text, i.e., is all the transactions along with the
previous block hash. The output is the current state of the blockchain encrypted in
256-bit binary value (hash). To generate the hash for the first block (genesis block),
the series of the initial transactions is used.
This is strictly a unidirectional algorithm, which means we cannot use the same
algorithmic program to convert the hash code to plain text. Besides if the data of
any transaction is changed, the hash of that block changes and conjointly of any
following block. On scrutiny with the original hashes, the change of state is caught
easily.
3.6 Consensus Protocol
A consensus algorithmic program aims at finding a common agreement among all the
nodes of blockchain that is a win for the whole network. There are various types of
consensus protocols that are based on different mechanism. Some of the consensus
protocols are Proof of Stake (PoS), Practical Byzantine Fault Tolerance (PBFT),
Proof of Work (PoW) and Delegated Proof of Stake (DPoS). We will implement DPoS
in our system, and in this, the stakeholders rather than creating a block themselves
give the power to the delegates they support. This reduces the process power of
the stakeholders. In some DPoS versions, a delegate should show a commitment
by depositing some funds into a time-locked security account. DPoS is sort of a
government system, if the delegates miss their turn of creating a block, they will not
be allowed, and the stakeholders will select new nodes to exchange their place. DPoS
is affordable and high-efficiency consensus protocol as compared to PoW and PoS
[5, 12].
4 Requirements for the Elections
4.1 Registration
Every eligible voter should register his/her name only once in lifetime. During regis-
tration, the individual must provide Aadhar card and permanent account number
(PAN) card. Biometrics like fingerprint and iris of an individual ought to be taken
and stored in the database. In case of any ambiguity, the biometrics can also be
verified with the Aadhar card database. These are unique to a person, and hence,
it ensures that someone solely votes against his own name. This would add more
security to the system, ensuring that an individual casts vote just once and will do no
frauds. Phone number and e-mail id should be provided by the voters during regis-
tration. These data fields will be very useful for sending OTP, confirmation messages
and some additional details and information.
After registration, the voters will be provided with voter’s ID card, each of which
will contain a unique QR code. This unique code will act as a username for the
concerned individual. The QR code will be used for login credentials.
4.2 Blockchains Used
Private blockchain
Every district will have a private blockchain in which each node will consist of the
votes casted by the voters Fig. 1a.
Public blockchain
Every state will have a public blockchain. This will contain the boot nodes linked to
a district node. Figure 1a, b in other words, the total vote tally of each party in each
district will constitute this blockchain.
The blockchain in every district can be activated that is the voting process in each
district will begin only when all the administrators have logged-in in the system. This
ensures decentralization. These administrators can be trusted people from NGOs who
look after the welfare of the people and run campaigns to fight against the voting
frauds [3].
Fig. 1 Representing types of blockchains used

5 The Proposed Protocol
For the actual process of casting votes, the voters are supposed to report to their
voting centers, along with their voting IDs. Let us see step-by-step as to how the
votes shall be casted.
5.1 Prerequisites
Prior to the election, every voter’s cryptocurrency wallet meant exclusively for the
election process will be set with few coins [13]. This wallet can be accessed only
with a private key that each voter has. Private key is unique to every voter.
5.2 Login
The voters have to scan the QR code from the voter’s ID card. After this, the biometrics
will be taken as input, to ensure the validity of the voter. If the biometrics and the
login details match, only then shall the process proceed further, otherwise the current
session shall be aborted.
5.3 Initiating Voting
After the identity of the voter is verified, he would be given the authority of creating
a node in the private blockchain by the administrators using the DPoS consensus
algorithm.
5.4 Displaying the Candidates
The list of candidates to be voted for will be displayed along with their party symbol.
For the voter to cast his vote, he will have to spend all his coins on the candidate of
his choice [14]. The voters can use their private key to access their wallet. While the
public key of the voter will be shared with a deserving candidate for whom the voter
is willing to spend the coins and wants to vote for.
Fig. 2 Encrypting transactions
5.5 Cast a Vote
When the voter spends his coins to cast his vote, the transaction will be added in the
ledger. A hash code will be generated using the hash of the previous block and the
voter’s current transaction. Then, this transaction along with the hash code will be
encrypted by the voter’s private key, and this is the digital signature. We can say that
digital signature is the combination of hashing and encryption of the transaction by
the private key of the voter. Digital signature can be defined as the following function
(Fig. 2):
f (hash code, current transaction) = y

encryption(y) = digital signature
5.6 Receipt Generation (not Mandatory)
After casting the vote, a receipt will be generated indicating the vote casted by the
voter. These receipts can be used to verify the results of the election if needed.
Figure 3 illustrates the working of proposed protocol.
5.7 Tally
The tally is performed using a blockchain which is a distributed ledger. Each voter
can see the entire network of blocks.
After results are declared, every individual voter can verify whether his/her vote
was counted correctly (i.e., that system was not rigged), and if the total tally matches,
the actual votes casted. This ensures that the votes were not tampered with. In case
of any tampering, the hash of the block whose vote is tampered with changes and so
does the hashes of further blocks.
Every voter shall be able to see his own block, showing the vote that he has
casted (Fig. 3). For this, the authority that is conducting the elections shall provide
Fig. 3 Working of proposed system
an interface where a voter can login and see his/her details clearly. The details may
include the voter’s name, voter ID number, Aadhar number and the vote that he had
casted. Personal details can only be seen by the voter herself/himself. He/she can
also be able to see all other votes, casted by other voters, while maintaining their
anonymity of other voters.
The details of other voters shall include only the hash of the block and the vote
casted. This way, anonymous status of other voters is maintained, and the results
(total tally of votes) can also be cross-checked by the voters [5]. Hence, the control
is not just in the hands of one specific body.
If the voter finds out that the vote he had casted is not the same as that shown in
the block, he may take legal action and register complaint of vote rigging (Fig. 4).
Fig. 4 View of voter
6 Discussion
One major inconvenience caused by the existing process of elections is its inflexi-
bility. The existing system does not allow any voter to cast his/her vote in case the
voter is not present in constituency where he/she has been registered. This rigidity of
the system results in decreased participation of the voters. We can have an additional
flexibility in the proposed system.
As discussed earlier, on the day of voting, administrators in each city must first
login into the system to begin a session. These login credentials can be shared with the
administrators using a smart contract over the state so that those voters who could not
be in their respective constituencies can vote from their current location. This helps in
motivating the population of the country to turn up at the voting centers and cast their
vote. This also eliminates any possibility of voter fraud as smart contracts cannot be
changed. To conclude, the names of the voters should be registered in a smart contract
so that these smart contracts can be exchanged between the administrators. Different
polling centers should be set up for those who do not belong to the constituency from
where they are voting. In this, the administrator in that polling center will make sure
that the voter is added to the correct boot node.
7 Result and Conclusion
India is the largest democracy in the world, yet EVM fails to provide justice to the
citizen’s right to vote [2]. Political instability also prevails as the casted votes are
rigged by hacking the electronic machines. Hence, comes the need for a more reliable
platform for casting votes and electing a political party for the welfare of country. It
is the need of an hour to design a secure and immutable voting system.
Blockchain, a technology that offers various features such as transparency, secu-
rity, decentralization and immutability, is an appropriate approach toward the imple-
mentation of such system [2]. We have proposed the use of cryptocurrency wallets
in order to invest on the favored politician. Digital signature preserves authenticity
of the voter and tampering with the vote. Smart contracts, an important feature of
blockchain, make sure that the voter does not spend more than the available amount.
Using all these features of blockchain, this paper proposes a working e-voting system.
References
1. F. Hjálmarsson, G.K. Hreiðarsson, M. Hamdaqa, G. Hjálmtýsson, Blockchain-based E-voting

system, in 2018 IEEE 11th International Conference on Cloud Computing
2. A. Menon, V. Bhagat, Blockchain based e-voting system (2020)
3. A. Khandelwal, Blockchain implementation on e-voting system, in International Conference
on Intelligent Sustainable Systems (ICISS 2019), IEEE Xplore
4. T. Moura, G. Alexandre, Blockchain Voting and its effects on election transparency and voter
confidence, in Proceedings of the 18th Annual International Conference on Digital Government
Research (2017)
5. B. Wang, J. Sun, Y. He, D. Pang, N. Lu, Large-scale election based on blockchain, in 2017 Inter-
national Conference on Identification, Information and Knowledge in the Internet of Things
(2017)
6. W. Li, Escape analysis on the confinement-escape problem of a defender against an evader
escaping from a circular region (2016)
7. N. Sharma, A. Chakrabarti, V.E. Balas, Data management, analytics and innovation, in
Proceedings of ICDMAI 2019, vol. 1 (2019)
8. I.L. Awalu, P.H. Kook, J.S. Lim, Development of a distributed blockchain e-voting system,
in Proceedings of the 2019 10th International Conference on E-Business, Management and
Economics (2019)
9. R. Hanifatunnisa, B. Rahardjo, Blockchain based e-voting recording system design (2017)
10. M. Pawlak, A. Poniszewska-Maránda, N. Kryvinska, Towards the intelligent agents for
blockchain e-voting system, in The 9th International Conference on Emerging Ubiquitous
Systems and Pervasive Networks (2018)
11. A.B. Ayed, A conceptual secure blockchain-based electronic voting system (2017)
12. S. Zhang, J.-H. Lee, Analysis of the main consensus protocols of blockchain (2019)
13. K. Patidar, B. Rahardjo, Decentralized e-voting portal using blockchain, in 2019 10th
International Conference on Computing, Communication and Networking technologies (2019)
14. F.S. Hardwick, A. Gioulis, R.N. Akram, K. Markantonakis, E-voting with blockchain: an e-
voting protocol with decentralisation and voter privacy, in 2018 IEEE Conferencess on Internet
of Things, Green Computing and Communications, Cyber, Physical and Social Computing,
Smart Data, Blockchain, Computer and Information Technology, Congress on Cybermatics
(2018)
SOLID: A Web System to Restore
the Control of Users’ Personal Data
Manishkumar R. Solanki
Abstract The concept of Web was conceived to facilitate sharing of information

between scientists and researchers in universities and institutes around the world. But
this juggernaut has given birth to personal and business data breach, cyber bullying,
harassment, hate speech and other cybercrimes. The privacy of a user is always at
stake while using Internet due to commercial exploit of private data. This paper gives
overview of a project “SOLID (SOcial LInked Data)”, initiated by the inventor of the
Web Tim Berners-Lee to give a person the ultimate control of his private information
on the Internet.
Keywords Angular · Cybercrimes · PODRDF · React · WebID
1 Introduction
The harmonization of World Wide Web (www) service and the hardware backbone
Internet has significantly changed the living standard of a human across the globe.
People can access Internet services at anytime from anywhere with the help of smart
phones. Almost every kind of business has the Web presence today. The E-Commerce
platform is increasing its user base at every moment due to ease of use and saving
of time and money. Millions of business transactions carried out on Internet store
consumers’ private data in central repository for commercial purpose. Consumers
have lost the control of their own data. Openness of Web has been compromised at
great extent. To bring back it on its track, the founder of www, Tim Berners-Lee has
started to build a system named “SOLID”. This paper gives overview of Solid and
its working.
The goal of this paper is to introduce the functioning of the Solid project, especially
how it can prevent the misuse of personal information of a user and to give him the
M. R. Solanki (B)
Shri Bhagubhai Mafatlal Polytechnic, Mumbai, India
e-mail: manish_ratilal2002@yahoo.com
258 M. R. Solanki
ultimate control of the same. Section 2 describes the exploitation of the present Web,
Sect. 3 introduces the working principles of Solid project, and Sect. 4 concludes the
paper.
2 Exploitation of the Present Web
Tim Berners-Lee, a British scientist, invented the World Wide Web (WWW) in 1989,
while working with Robert Cailliau at CERN [1]. The purpose of inventing the
Web was to enable seamless information sharing between researchers and scientists
residing across the globe. The Web began to flourish with the release of its’ open
license in 1993. It sparked a global wave of creativity, collaboration and innovation.
Tim moved from CERN to the Massachusetts Institute of Technology in 1994 to
found the World Wide Web Consortium (W3C), an international community devoted
to developing open Web standards [2].
The free, open and accessible nature of Web is being compromised due to corpo-
rate exploit of information, hate speech, state-sponsored hacking and attacks, crim-
inal behavior and online harassment and other such activities taking place on the
Internet [2]. The world has witnessed huge data breaching victims, i.e., Adobe,
Adult Friend Finder, Ebay, Linkedin, Mariott International, Yahoo, Facebook and
other business giants in the last decade. The secret and sensitive data of millions of
users have been exposed in the public area for the commercial exploit. Such breach
of valuable data has raised the questions on reputational of brand, loss of money,
legal ramifications and security of intellectual property.
During the inauguration of Web Summit, in Lisbon, Portugal, on November 5,
2018, Tim Berners-Lee launched a campaign to persuade governments, companies
and individuals to sign a “Contract for the Web” designed to defend a free and open
Internet [3].
3 Basic Components of Solid System
This section discusses about basic components of Solid system and demonstrates the
working of a demo app developed with Solid.
3.1 POD and WebID
Solid provides a distributed data service by extending the present Web standards.
Individuals and organizations keep their data in personal online data stores (PODs).
They can store photos, comments, address books, calendar events, bank details,
documents and other personal information on a POD as shown in Fig. 1. PODs are
SOLID: A Web System to Restore the Control … 259
Fig. 1 Personal data stored in POD [5]
designed to be secure, and the owner can provide fine-grained access to portions of
their data, as well as revoke it as needed [4]. PODs can be stored either in a public
cloud storage service, i.e., Solid Community or Inrupt, or self-hosted.
WebID provides globally unique decentralized identifiers, Solid uses WebID
instead of usernames which internally includes POD of a user.
The data and link types are written in standardized resource description framework
(RDF) form, multiple applications can use the same data. So, the data being stored
in the POD can be given access to all applications, i.e., Facebook, Twitter, Google,
Amazon, etc., instead of separately giving them the data and remembering to manage
the same across all apps. The application development is underway for it now. The
demo app to retrieve the user profile is developed in Angular and React and is
described on Solid URL.
260 M. R. Solanki
3.2 Setting up the Environment for the Demo Application
In this paper, I am showing the application development using Angular. The following
steps should be executed on command line in Windows OS.
1. Install Yeoman: npm install-g yo
Yeoman is a generic scaffolding system allowing the creation of any kind of app.
Yeoman is language agnostic. It can generate projects in any language, i.e., Java,
Python, C#, etc. [6] (Fig. 2).
After installation, we will get the Yeoman Doctor message as shown in Fig. 3.
2. Install solid-angular generator: Generators are basically plug-ins in the
Yeoman environment.
3. Now, navigate to the project folder where you want to create your app. In my case,
the project folder is SOLID which resides in drive d: yo @inrupt/solid-angular
(Fig. 4).
It will ask the application name or folder name, i.e., in my case d:\SOLID\solidapp
folder, so I gave the name as “solidapp” as shown in Fig. 5.
After successful installation of necessary resources in the “solidapp” application,
“welcome to solid” message will be displayed as shown in Fig. 6.
4. Install the Angular CLI: The Angular CLI is a command line interface tool that
you use to initialize, develop, scaffold and maintain Angular applications (Fig. 7).
5. Running the application: Make solidapp as your present working directory and
run the following command:
If you get an exception as shown in Fig. 8, then refer the reference [7] and employ
Approach 2 to fix the issue.
Fig. 2 Installation of Yeoman

Fig. 3 Successful installation of Yeoman
Fig. 4 Successful installation of solid-angular generator
Fig. 5 Setting application name to store necessary files and folders

262 M. R. Solanki
Fig. 6 Successful installation of necessary resources in the application
npm install -g @angular/cli
Fig. 7 Successful installation of Angular CLI (can be done in the beginning also)
Fig. 8 Receiving “could not find implementation for builder….exception”
After resolving the issue, we get “Compiled successfully” run the application on
Web browser using http://localhost:4200/.
3.3 Running the Application
The first and foremost requirement to get ownership of data, the user has to set up
his POD by registering at Solid Community, Inrupt or he can have his own setup.
After registering, the user will get his WebID which he can share to allow to the apps
wants to access user data as shown in Figs. 9 and 10, i.e., https://mrsolanki1903.inr
upt.net/profile/card#me.
When the application is being run on the browser using 4200 port (as in our case),
the user will be asked his POD provider and has to set permissions of his data to be
accessed by the app as shown in Figs. 11 and 12.
The user can see his profile and permissions given to the data at the provider site,
i.e., Inrupt by logging in his account as shown in Figs. 13 and 14.
The app can see the user data as shown in Fig. 15. As the user has given modi-
fication permissions to the app, it can modify the user data, i.e., the role of a user
from “Employee” to “Lecturer”. As the save button is pressed, “Success!” message
is displayed at the top-right corner.
Fig. 9 Getting POD from solid providers [8]
Fig. 10 User WebID with included POD

264 M. R. Solanki
Fig. 11 Selection of POD provider
Fig. 12 Setting personal data permissions
Now, if the user changes the permission “Read” only as shown in Fig. 16, then the
app cannot modify the role, i.e., from “Lecturer” to “Senior Lecturer”, and an error
message “All Required Access Modes not Granted” would be displayed at top-right
corner as shown in Fig. 17.
This demonstration shows that before accessing the users’ personal data, an app
will have to take permission from the user, and user will not sign up on the application
Web site or it will not even perform social sign in instead with his WebID at POD he
can grant permission for his personal data.
Fig. 13 User profile at Inrupt
Fig. 14 User profile at Inrupt with trusted applications
4 Conclusion
Companies and governments dip into the personal information of user data in increas-
ingly and innovative ways, specially tracking what they are, what they do, to whom
they interact and where they go. The commercial exploitation of personal data has
defeated the working philosophy of www at great extent. By proposing the Solid
project, Sir Tim Berners-Lee has initiated a revolutionary action to give back the
user to have control of his own data. As the system is in development phase, it
266 M. R. Solanki
Fig. 15 Modification of role from “Employee” to “Lecturer”
Fig. 16 Changing access modes to “Read” only
would be challenging task for Solid team to succeed in a space dominated by giant
corporations who make a living by harvesting and selling user data.
Fig. 17 “Required Access Modes Not Granted” error
References
1. The Birth of the Web, https://home.cern/science/computing/birth-web

2. A. Hern, Tim Berners-Lee on 30 years of the world wide web: ‘We can get the web we
want’, https://www.theguardian.com/technology/2019/mar/12/tim-berners-lee-on-30-years-of-
the-web-if-we-dream-a-little-we-can-get-the-web-we-want
3. Technology News November 6, 2018/ 12:23 AM/ A Year Ago, Web creator Berners-Lee
launches contract for better internet, https://www.reuters.com/article/us-portugal-websummit-
berners-lee/web-creator-berners-lee-launches-contract-for-better-internet-idUSKCN1NA2CX
4. D. Cardinal, Tim berners-lee’s solid project: can it save the web? (2018), https://www.extrem
etech.com/extreme/281334-tim-berners-lees-solid-project-can-it-save-the-web
5. https://solid.inrupt.com/how-it-works
6. Getting started with yeoman, https://yeoman.io/learning/
7. https://thecodebuzz.com/resolved-could-not-find-the-implementation-for-builder-angular-dev
kit-build-angularbrowser/
8. Solid Pod Providers, https://solid.inrupt.com/get-a-solid-pod
Enhanced Three-Dimensional DV-Hop
Algorithm
Abhinesh Kaushik and D. K. Lobiyal
Abstract Numerous applications of Wireless Sensor Networks (WSNs), need loca-

tion information for various purposes, such as identification of the location of objects,
improve the routing process, etc. The accuracy of the location information also plays a
significant role in these applications due to which localization of WSNs has emerged
as an important field of research. Among the various algorithms designed for local-
ization, DV-Hop is a favourite among the researchers. Lots of improved adaptations
of DV-Hop are suggested in the literature as a result of its simple usage. In this paper,
we have proposed an enhanced three-dimensional DV-hop algorithm to enhance its
location accuracy. The proposed algorithm uses only the beacon nodes in close prox-
imity for localization. Further, the concept of coplanarity is used to select an optimal
set of beacon nodes around an unidentified node for its location estimation. Further-
more, to reduce the propagation error, we use a new method to solve the system of
distance equations. The simulations illustrate that our proposed technique outclasses
the DV-Hop algorithm.
Keywords WSNs · DV-Hop · Localization
1 Introduction
Wireless sensor networks comprise various tiny sensor nodes appropriated over
an enormous zone to quantify the physical properties like temperature, pressure,
humidity, motion, etc. [1]. These are multifunctional, low-cost sensors having
powerful sensing capability along with a capacity to communicate over small
distances. They are capable of forming self-organizing, distributed, and multi-hop
wireless communication network [2].
WSNs have numerous applications in the field of target tracking [3], remote
sensing, smart homes, medical care, military affairs, natural disasters, intelligent
transportation, precision agriculture, and so forth [4]. Location of the place where
A. Kaushik (B) · D. K. Lobiyal

School of Computers and Systems Sciences, Jawaharlal Nehru University, New Delhi, India
e-mail: abhinesh.kaushik@gmail.com
270 A. Kaushik and D. K. Lobiyal
the event occurred is an important aspect. The data received from these sensors is
not of much use without the location information. Therefore, this makes location
estimation a primary requirement in countless applications of the WSNs. Most of
the commonly used methods for location estimation are designed based on Global
Positioning System (GPS) [5]. In GPS based methods, nodes are equipped with
GPS supporting hardware chips [6]. These GPS chips make the sensor nodes, bulky,
costly, and power-intensive. GPS also restricts the sensor nodes usability in indoors
and in an underground scenario as it requires to communicate with the satellite [7].
Therefore, designing a localization algorithm with minimal hardware, minimal cost,
and limited power is a challenging task. These localization algorithms use techniques
like Distance Vector, Approximate point in triangulation, centroid [8], received signal
strength indicator (RSSI), etc. for localization. DV-hop algorithm is among the most
frequently used algorithms for localization because of its simple usage and vast scope
for improvement in location accuracy. Consequently, researchers have proposed
scores of improved variants of DV-Hop algorithm [2, 9–11] in the literature. But
most of them are restricted to two-dimensional scenarios.
This manuscript is arranged as follows. In Sect. 2, we discuss DV-hop algorithm
and its improved version. Section 3 elaborates the proposed work. Simulation results
and functioning of the system is discussed in Sect. 4. Whereas, conclusion of the
paper is deliberated in Sect. 5.
2 Correlated Work
The DV-hop algorithm [12] is designed for a two-dimensional scenario. It consists of

three steps. The first step, involves flooding coordinates of all the beacon nodes (nodes
knowing their positions) and hop count to all the nodes (beacon and unidentified) in
the network. In second step, an average hopsize is calculated using the information
obtained from the first step. It also involves, estimation of distance between the
beacon and unidentified node. Whereas in the last step, the information from the
preceding steps is used to localize the unidentified node by solving a set of distance
equations. The biggest shortcoming of this algorithm is that it provides poor accuracy.
The algorithm is developed for a two-dimensional scenario, whereas most of the
real-life applications involve a three-dimensional scenario.
A three-dimensional DV-Hop localization algorithm based on RSSI in [13]
(3DVR) is proposed to upgrade the existing DV-Hop algorithm. The second step
of 3DVR algorithm involves the concept of RSSI to find the average hop size. An
error correction factor is also added to the distance. 3DVR does improve the location
accuracy when compared with the DV-Hop but introduces its own limitations due to
the usage of RSSI. The most common limitation of the model is shadowing, fading,
and path loss. The algorithm also consumes more power in contrast to the other
improved renditions of DV-Hop algorithm.
There are few more modified versions of DV-Hop like one in [14] where the
author uses a method of intersecting circles (of communication radius) between the
Enhanced Three-Dimensional DV-Hop Algorithm 271
unidentified node and the beacon node. Further, these intersecting circles are used
to localize the unidentified node. Although the method makes calculation easier but
compromises on the location accuracy.
3 Proposed Work
In this paper, we propose an enhanced three-dimensional DV-Hop algorithm(E3DV-

Hop) for enhancing location accuracy of the DV-Hop. The proposed algorithm is
broadly working in three steps.
In the Step 1, E3DV-Hop algorithm works in the same way as first step of the DV-
hop algorithm. In the Step 2 of the algorithm, we determine a beacon node’s average
hop size. To find the hop size, we use hop count and location information from step
1. The hopsize, thus computed, helps in computing the distance between the nearest
beacon node and the unidentified node. Before calculating the distance, we locate the
nearest beacon nodes. To get a more accurate estimation, we use the beacon nodes
within the close proximity of one to three hop count from the unidentified node.
On finding the nearest nodes, we check for degree of coplanarity [15], which helps
in selecting the best topology for location estimation. In the step 3, we restrict the
total beacon nodes used for localization, which leads to a reduction in the number
of distance equations for localization of unidentified node. Further, in the step 3, we
use a distinctive method to solve the set of distance equations meant for computing
the location of an unidentified node. There is a reduction in the propagation of
error because of the new method used for solving equations. The simulation results
also support our proposed method in terms of betterment in location accuracy. The
proposed technique works in the subsequent three steps:
Step 1 In this step, the beacon nodes present in the network broadcast their loca-
tion information in the wireless sensor network so that all other nodes get
coordinates of the beacon nodes and the hop count.
Step 2 In this step, we calculate every beacon node’s hopsize (average) by using the
coordinates information of beacon nodes and the hop count received from
step 1, using Eq. (1),
2 2 2

i= j xi − x j + yi − y j + z i − z j
Hopsizei = (1)
i= j hopsi j
After calculating the hopsize, we now find the beacon nodes in the close proximity
of the unidentified node to be localized. To decrease the network utilization, we
use beacon nodes, which are one to three hop away from the unidentified node.
There shall be a minimum of 4 beacon nodes around the unidentified node to use
the multilateration method for localization of unidentified node. After finding the
beacon nodes in close proximity, we check them for coplanarity. Beacon nodes that
are present in the close proximity to unidentified node and non-coplanar are the
only ones used for localization. The degree of coplanarity helps us in checking the
coplanarity of four neighboring beacon nodes. A tetrahedron is formed by joining the
lines connecting the four non-coplanar beacon nodes in a 3D space. The following
Eq. (2) represents the radius ratio of the tetrahedron,
3rin
ρ= (2)
rout
where ρ is the radius ratio of the tetrahedron, and rin , and rout are the radius of
the inscribed and the circumscribed sphere represented using Eq. (3), and Eq. (4),
respectively.
3σ
rin = 3 (3)
i=0 μi
√
(a + b + c)(a + b − c)(a + c − b)(b + c − a)
rout = (4)
24σ
where volume of the tetrahedron is σ , and μi represents the surface areas of the four
sides. The edges of the tetrahedron are denoted by a, b and c.
216σ 2
ρ = 3 √ (5)
i=0 μi (a + b + c)(a + b − c)(a + c − b)(b + c − a)
Therefore the radius ratio can be calculated by using Eq. (5), where the value of
ρ varies from 0 to 1. The value of ρ approaching 0 represents that the four points are
coplanar, whereas a value approaching 1 denotes that a tetrahedron is formed out of
the four points. Therefore, the degree of coplanarity is stated as:

0 coplanar
Degree of Coplanarity =
ρ otherwise
The degree of coplanarity helps us in finding the topologically optimal set of

beacon nodes around the unidentified node for localization.
After getting the beacon nodes in the close proximity and checking for the copla-
narity, the beacon nodesbroadcast their hopsize in the network. The unidentified
node u picks the hopsize hopsizei from the nearest beacon node i and evaluates the
distance between the beacon nodes and itself by using Eq. (6)
dup = hopsizei × hopsup (6)
where hopsup is the least hop count between uth unidentified node and pth beacon
node.
Step 3 In this step, to decrease the propagation of error, we introduce a new method
of solving the distance equations. In this method, we subtract all the distance
equations from the largest distance equation and then square the resulting
equations to solve further. Subtraction before squaring leads to a reduc-
tion in propagation of inherent error as compared to the squaring and then
subtracting method, which increases the error rapidly. We assume the uniden-
tified node u has (xu , yu , z u ) coordinates, and (xn , yn , z n ) are the coordinates
of the nth node, and the distance between unidentified node u and the last
beacon node n, which is not more than three hops away, is denoted by dn .
The coordinates of unidentified node are calculated using Eq. (7)
⎫
(xu − x1 )2 + (yu − y1 )2 + (z u − z 1 )2 = d1 ⎪
⎪
⎪
⎪
(xu − x2 )2 + (yu − y2 )2 + (z u − z 2 )2 = d2 ⎪
⎪
⎪
⎬
.
(7)
(xu − xm )2 + (yu − ym )2 + (z u − z m )2 = dm ⎪
⎪
⎪
⎪
⎪
. ⎪
⎪
⎭
(xu − xn ) + (yu − yn ) + (z u − z n ) = dn
2 2 2
Subtracting all the n equations from the largest distance equation (mth) having
distance dm
⎫
(x − xm )2 + (yu − ym )2 + (z u − z m )2 − (xu − x1 )2 + (yu − y1 )2 + (z u − z 1 )2 = dm − d1 ⎪
⎪
u ⎪
⎪
⎪
⎪
(xu − xm )2 + (yu − ym )2 + (z u − z m )2 − (xu − x2 )2 + (yu − y2 )2 + (z u − z 2 )2 = dm − d2 ⎪
⎬
. ⎪
(8)
⎪
⎪
. ⎪
⎪
⎪
⎪
⎭
(xu − xm )2 + (yu − ym )2 + (z u − z m )2 − (xu − xn )2 + (yu − yn )2 + (z u − z n )2 = dm − dn
Squaring both sides and solving Eq. (8) and further simplifying the set of
equations,
⎫
−2(xm + x1 )xu − 2(ym + y1 )yu − 2(z m + z 1 )z u + 2r = dm2 + d12 − (tm + t1 ) ⎪
⎪
⎪
−2(xm + x2 )xu − 2(ym + y2 )yu − 2(z m + z 2 )z u + 2r = dm2 + d22 − (tm + t2 ) ⎪
⎪
⎬
.
⎪
⎪
. ⎪
⎪
⎪
−2(xm + xn )xu − 2(ym + yn )yu − 2(z m + z n )z u + 2r = dm2 + dn2 − (tm + tn ) ⎭
(9)
where r = xu2 + yu2 + z u2 and t = xi2 + yi2 + z i2 , ∀ i = 1, 2, 3 . . . ..n where n is the

last beacon node in the close proximity.
Matrix form of the system of n − 1 equations in Eq. (9) is as follows:
AX = B (10)
where A, X, and B are defined as:

⎡ ⎤
−2(xm + x1 ) −2(ym + y1 ) −2(z m + z 1 ) 2
⎢ −2(xm + x2 ) −2(ym + y2 ) −2(z m + z 2 ) 2 ⎥
⎢ ⎥
A=⎢. .. .. .. ⎥
⎣ .. . . . ⎦
−2(xm + xn ) −2(ym + yn ) −2(z m + z n ) 2
⎡ 2 ⎤
dm + d12 − (tm + t1 )
⎢ d 2 + d 2 − (tm + t2 ) ⎥
⎢ m 2 ⎥
B = ⎢. ⎥
⎣.. ⎦
dm2 + dn2 − (tm + tn )
⎡ ⎤
xu
⎢ yu ⎥
X =⎢ ⎣ zu ⎦
⎥
Least square method is used to solve Eq. (10), and Eq. (11) is used to find the
location coordinates of the unidentified node,
−1
Z = AT A AT B (11)
where A T denotes the transpose of A.
4 Simulation Results and Performance Analysis
To assess the performance of the proposed algorithm, it is simulated on MATLAB

2018b. In this section, we examine the localization error (LE) with respect to various
parameters. We have defined localization error in Eq (12) as one of the performance
indicators of the location accuracy.
2 2 2
N
j=M+1 x uj − x aj + y uj − y aj + z uj − z aj
LE = (12)
C × (N − M)
Where (x uj , y uj , z uj ), (x aj , y aj , z aj ) are the computed and the actual position coordinates

of the jth unidentified node, respectively, N denotes the overall number of nodes in
the WSN, M is percentage of beacon nodes, and C denotes the communication radius
of a sensor node. In order to provide a uniform environment for all the algorithms we
have kept all the parameters as constant for all the algorithms. We have also extended a
third dimension to the DV-Hop algorithm to simulate it in a three-dimensional space.
Fig. 1 Communication radius versus localization error
The simulation is carried out in a random network area of 100 m × 100 m × 100 m,
and results are averaged over a thousand trials.
4.1 Localization Error Versus Communication Radius
Figure 1 shows the effect of change in communication radius of the sensor node on
the localization error. In this trial, the overall number of sensor nodes has been fixed
at 400; the percentage of beacon nodes is fixed at 20%, and the communication radius
is varied from 15 m to 45 m. The figure shows that LE for E3DV-Hop is smaller in
contrast to DV-Hop and 3DVR algorithms for same values of communication radius.
We noticed that the fall in LE for E3DV-hop is steeper in every case as compared to
that of the other two algorithms.
4.2 Localization Error Versus Percentage of Beacon Nodes
Figure 2 depicts the effect of change in the percentage of beacon nodes on the
localization error. Here also, the overall number of sensor nodes considered is 400,
communication radius is selected as 25 m, and the percentage of beacon nodes alters
from 5 to 35% of the overall number of nodes. The figure shows that LE for E3DV-
Hop is smaller in contrast to DV-Hop and 3DVR algorithms for the same value of
Fig. 2 Percentage of beacon nodes versus localization error
percentage of beacon nodes. We also observe that LE for E3DV-Hop remains lower
by about 3–10% as compared to other two algorithms.
4.3 Localization Error Versus Total Number of Nodes
Figure 3 depicts the effect of variation in the total number of sensor nodes in the
network on the localization error. The percentage of beacon nodes is kept constant
at 20% of the total number of nodes in the network, communication radius is taken
as 25 m, and the total number of sensor nodes are altered from 100 to 800. This
figure shows that LE for E3DV-Hop is smaller in contrast to DV-Hop and 3DVR
algorithms for the identical value of total number of nodes. We also observe that LE
for E3DV-Hop remains lower by about 10% as compared to the other two algorithms.
5 Conclusion
Results obtained from the simulation proves that E3DV-Hop algorithm reduces the
localization error. The selection of closed proximity beacon nodes and a check for
coplanarity selects the topologically optimal set of beacon nodes. This leads to better
localization accuracy over the other two compared algorithms. Whereas a distinctive
Fig. 3 Total number of nodes versus localization error
method used for solving the set of distance equations leads to a reduction in propaga-
tion error. The performance evaluation of E3DV-Hop, 3DVR, and DV-Hop algorithms
indicates that E3DV-Hop algorithm is appreciably superior over the other variants
of DV-Hop. To be more precise, the localization error of these three localization
algorithms DV-Hop, 3DVR, and E3DV-Hop is 55%, 42%, and 25%, respectively.
References
1. I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor networks: a survey.

Comput. Netw. 38, 393–422 (2002)
2. S. Kumar, D.K. Lobiyal, An advanced DV-Hop localization algorithm for wireless sensor
networks. Wireless Pers. Commun. 71, 1365–1385 (2013)
3. S. Biswas, S. Gupta, F. Yu, T. Wu, A networked mobile sensor test-bed for collaborative
multi-target tracking applications. Wireless Netw. 16, 1329–1344 (2010)
4. G. Han, J. Chao, C. Zhang et al., The impacts of mobility models on DV-hop based localization
in mobile wireless sensor networks. J. Network Comput. Appl. 42, 70–79 (2014)
5. S. Čapkun, M. Hamdi, J.P. Hubaux, GPS-free positioning in mobile ad-hoc networks, in
Proceedings of the Hawaii International Conference on System Sciences 255 (2001)
6. R. Nagpal, H. Shrobe, J. Bachrach, Organizing a global coordinate system from local informa-
tion on an ad hoc sensor network, in Lecture Notes in Computer Science (including subseries
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 2634,
pp. 333–348 (2003)
7. B. Hofmann-Wellenhof, HC.J.E. Lichtenegger, Global Positioning System: Theory and
Practice (Springer Science & Business Media, 2012)
8. C.C. Chen, C.Y. Chang, Y.N. Li, Range-free localization scheme in wireless sensor networks
based on bilateration. Int. J. Distrib. Sens. Netw. (2013)
9. S. Kumar, D.K. Lobiyal, Power efficient range-free localization algorithm for wireless sensor
networks. Wireless Netw. 20, 681–694 (2014)
10. O. Cheikhrouhou, G.M. Bhatti, R. Alroobaea, A hybrid DV-hop algorithm using RSSI for
localization in large-scale wireless sensor networks. Sensors 18, 1–14 (2018)
11. S.P. Singh, S.C. Sharma, A modified DV-Hop algorithm for wireless sensor network
localisation. Int. J. Commun. Netw. Distrib. Syst. 21, 187–201 (2018)
12. D. Niculescu, B. Nath, Ad hoc positioning system (APS), in GLOBECOM’01. IEEE Global
Telecommunications Conference (Cat. No.01CH37270). IEEE, pp. 2926–2931
13. P. Gou, F. Li, X. Jia, et al., A three-dimensional DV-Hop localization algorithm based on RSSI,
in Proceedings—2018 3rd International Conference on Smart City and Systems Engineering
(ICSCSE, 2018), pp. 830–833
14. X. Yu, R. Xiao, Y. Liu, et al., Modified DV-Hop localization algorithm based on intersecting
circle and hops revaluation, in Proceedings of the 3rd International Conference on Intelligent
Information Processing, pp. 33–39 (2018)
15. Y. Xu, Y. Zhuang, J.J. Gu, An improved 3D localization algorithm for the wireless sensor
network. Int. J. Distrib. Sens. Netw. (2015)
Synthesis of VLSI Structural Cell
Partitioning Using Genetic Algorithm
P. Rajeswari, Theodore S. Chandra, and Amith Kiran Kumar
Abstract The explosive growth in VLSI Design technology has been juxtaposed
with the complexities, hence requiring a more efficient and faster way of minia-
turization, with little to none human efforts. The way to go is to explore novel
methodologies to find out the best as well as the least Time-Consuming method
to find the optimum solution without compromising with quality. Here this paper
describes the use of the genetic algorithm to find the most optimum path in a smaller
number of iterations by the judicial use of controlled mutation at various stages, all
the more so with the least impact on the optimization result with a robust structure.
Hence enabling large VLSI circuits to be miniaturized with least cross over, by which
circuit efficiency, as well as Cost-Saving, can be achieved in the least time possible
with high accuracy at the expense of a very least possible trade-off between time and
cost.
Keywords Genetic algorithm · VLSI · Circuit integration · Partitioning
1 Introduction
Due to phenomenal growth experienced in the VLSI field and there are expectancies
for its exponential growth, adding on further there are no signs of saturation in the
near feature. Due to the flooding of deep Sub-micron technologies in the field of
VLSI, the Self-partitioning Algorithm has become more and more important. As
P. Rajeswari (B) · T. S. Chandra · A. K. Kumar

Dayananda Sagar Institutions, Bangalore, Karnataka, India
e-mail: prajeswarisugans@gmail.com
T. S. Chandra
e-mail: Theodore-ece@dsu.edu.in
A. K. Kumar
e-mail: amithkhul@gmail.com
280 P. Rajeswari et al.
VLSI Circuit design is hampered by a significantly large and wide range of issues,
they are design complexities, time complexity, increase in time delay, large power
consumption, large area consumption and high cost are some of the common areas
of concern.
Circuit partitioning techniques can be used to breakdown the large VLSI circuits
effectively and logically in to a large number of smaller interacting Sub-elements for
better and easy handling. It’s the most efficient way possible to address the difficulties
that have shown up due to highly dense VLSI circuit chips in unearthing the design
tradeoffs in order to make a more effective design in feature circuits we require a
better, faster, accurate, partitioning algorithms to provide a low cost high efficient
chips to market in a very short time [1]. This makes VLSI circuit partitioning of the
physical design of VLSI system an integral part of VLSI technology. Universities
and Industries are working continuously to further fortify and improve the algorithms
as well as finding new alternative algorithms to break these large VLSI design into
small Sub-parts None the less, partitioning of these highly dense chips is a tedious
as well as a challenging task. It requires a lot of time and effort in arriving at a
satisfactory solution in a reasonably fair time. From the past observation, it is noted
that partitioning is harder to a Non-deterministic polynomial level. This encourages
the advent of the use of heuristic algorithms and new methodology that would arrive
at a reasonably fair and acceptable solution in the shortest and allocated amount
of time. Even though there were a lot of other subsidiary techniques they have
some shortcomings. The algorithms have now started to be built based on nature’s
inspiration such as from neural network, bee colony, honeycomb, and now finally the
genetic algorithms. It is said there are VLSI circuit designs that if they are drawn on
a paper would cover the entire earth thrice. Now due to the use of genetic algorithms,
we can build such unimaginably bulky circuits today with less human efforts.
2 Motivation and Background
2.1 Genetic Algorithm
Due to the efforts of vigorous and deep studies and research in the area of VLSI
circuit partitioning to meet the enormous technology scaling, the advent of new
and powerful technique such as heuristic procedure has been developed for better
partitioning. Among all these algorithms, genetic algorithm has been proven as a very
promising algorithm for the partitioning of VLSI circuits. This is the base on this the
paper is being developed. Genetic algorithms are nothing but heuristic algorithms
constructed on the basis of Charles Darwin’s theory of evolution employing a natural
selection of the stronger species, and introduction of occasional mutation to combat
the attainment of local minima and arrive at a global maximum. It’s also important to
choose the level of mutation and the extent of mutation to create an ideal partitioning
algorithm a highly mutated algorithm may miss some of the local or possibly the
Synthesis of VLSI Structural Cell … 281
global minima and a very low mutated algorithm take a long time to arrive at the
global minima [2–5].
There is always a Trade-off between time and efficiency in the partitioning algo-
rithm, the important thing is how minimum is the Trade-off and robust is the algo-
rithm. Creation of very high mutation would increase the accuracy but it would also
take longer time which of course is affordable for smaller VLSI circuits but, when
it comes to practical demand there is a lot of restrains related to the time it may take
years to build a perfect partition. So, there is a requirement to restrict the number of
mutation possible in a circuit.
2.2 Fitness Parameter and Finding Global Maxima
The fitness parameter is designed to find the most optimum result or outcome of an
individual circuit and it represents how well or how efficient this floorplan or the
built circuit is, that is it describes how well a chromosome can survive. In a genetic
algorithm after each iteration, in which multiple possible combinations of the circuit
partitioning returns the chromosomes that have the local maxima value as its fitness,
fitness parameter is checked and then that result showing the level of fitness is passed
as an argument to the genetic algorithm where it forms a distributed record of the
data (as shown in Fig. 4) and would proceed further with mutations at a random
interval [6, 7].
It means to say first the different possible combination of partitioning is started
by doing random changes. There are some times when the partitioning might not
have reached the most optimum result, yet it might have reached the optimization
limit, so the mutations are introduced to avoid getting into local minima and reach
maxima.
It makes a list of a parameter such as local maxima which are then finally combined
to form a global maximum of the function, the global maxima are the final intended
result of the program (Fig. 1).
The fitness function is designed to find the optimum cost, area, and time.
Please note that the first paragraph of a section or subsection is not indented. The
first paragraphs that follow a table, figure, equation etc. do not have an indent, either.
Subsequent paragraphs, however, are indented.
Fn = −(α ∗ F(a) + β ∗ F(t) + γ ∗ F(c)) (1)
f n = ((Fn − Fn (min)) ∗ (U ))/(Fn (max) − Fn (min)) (2)
where F n the nth optimization level of VLSI circuit, F(a) is the area function, F(t)
is the time function and F(c) is the cost function, α is the weight of area function, β
is the weight of time function and γ is the weight of cost function.
Fig. 1 Distribution of the local maximums and minimums in a function
When it’s mentioned as time function it is not the time taken by the genetic
algorithm, it is the time delay experienced in the VLSI circuit.
The negative sign indicates fitness function value decreases as the cost, time and
area function values increases [8].
The Eq. (2) is used to map the fitness function to a positive value in the Scale-Up
to U.
where U is the upper limit, F(min) is the minimum optimization level required
and F(max) is the maximum optimization level noticed.
After evaluation of Eq. 2, all the negative readings are turned to positive values.
3 Flow Chart and Algorithm
The proposed design first import the VLSI circuit data from the V, .itf, or .lib files
and make them as chromosomes.
It then finds the fitness of the circuit if fitness doesn’t meet the required standard
and the iteration count limit n (max) is not reached, introduce mutation at every 1000
steps of iteration and then select a particular few chromosomes with the best fitness
to perform Crossing-Over to get new set of population, and send it again to find the
fitness [9, 10].
This is performed iteration or fitness is satisfied, once it is satisfied, it stops and
gets the global maxima.
The flow diagram of the proposed genetic algorithm is shown in Fig. 2.
Fig. 2 Flow diagram of the genetic algorithm
4 MATLAB Implementation
4.1 Selection of Chromosome Location
The VLSI circuit data is taken from the .lib file and the chromosome locations are
chosen from that and these locations are represented as a matrix and are then plotted
on a graph after that the chromosome data is imported for finding the cost, time and
area function (Figs. 3 and 4).
Fig. 3 Chromosome
location from import data
Fig. 4 Distribution graph of

maxima and minima
4.2 Data Representation
To find the global maxima the weight of each trial is represented in the distribution
graph. The graph records all the distribution of local maxima and local minima’s
which helps in determining the global maxima once the number of iteration limit is
reached, the time allotted is used up or the most possible best solution is obtained.
5 Results
The proposed solution in the genetic algorithm is implemented using MATLAB and
the results are obtained. [10] All the crossing-overs are optimized to get the better
cost area and time function and the final partitioned circuit is displayed in the GUI
display (Table 1).
From the table, it is clear that the genetic algorithm is almost four times better
than the brute force partitioning technique and as the complexity and the volume
of the circuit increases, this difference also increases it’s more obvious in the larger
circuit (Fig. 5).
Figure showing the structuring of VLSI circuit by partitioning using Genetic
algorithm the number of crossing-over is reduced to achieve the most optimum
results.
The graph has been obtained after over 3200 iterations for more complex circuit
a greater number of iterations are required (Fig. 6).
As we see from the graph as the number of Iterations increases along the x axis
the fitness is improved and has reached a constant mean to say it has reached the
global maxima.
Table 1 Comparison table

Comparision
for the efficiency of genetic
algorithm versus brute force Partitioning technique Fittness value Number of iteration
partitioning technique Brute force partitioning 401.29 4290+
Partitioning using genetic 100.36 1320
algorithm
Fig. 5 Genetic algorithm

crossing-over in matrix
Fig. 6 Cost, area, time

function on the number of
trials in the genetic algorithm
6 Conclusion
In this paper, several papers have studied, researched implemented and modified
that have given may insight for the partitioning of VLSI circuit identification. The
study briefly depicts that most of the methods concentrate on studying various way
on improving efficiency, cost, time and performances. Since VLSI has a limitless
future, more research work should evolve. Those future works should study not only
individual aspects but also minimize the tradeoffs, thereby boosting the performance
capacity of the system. Feature works have to be focused on Multi-Dimension circuits
and possible solutions to carbon nanotube technology. Hence this paper has helped
to build a robust algorithm for VLSI partitioning.
References
1. M. Shunmugathammal, C.C. Columbus, S. Anand, A nature inspired optimization algorithm

for VLSI fixed-outline floorplanning. Analog Integr. Circ. Signal Process. https://doi.org/10.
1007/s10470-020-01598-w
2. E.V. Kuliev, V.V.L. Kureichik, I.O. Kursitys, Decision making in VLSI components placement
problem based on grey wolf optimization, in 2019 IEEE East-West Design and Test Symposium
(EWDTS)
3. Y. Wang, G. Wang, Y. Shen et al., A memristor neural network using synaptic plasticity and
its associative memory. Circ. Syst. Signal Process. 39, 3496–3511 (2020). https://doi.org/10.
1007/s00034-019-01330-8
4. I.H. Shanavas, R.K. Gnanamurthy, T.S. Thangaraj, A novel approach to find the best fit for
VLSI partitioning—physical design, in 2010 International Conference on Advances in Recent
Technologies in Communication and Computing (ARTCom), pp. 330, 332, 16–17 Oct. 2010
5. R. Pavithra Guru, V. Vaithianathan, An efficient VLSI circuit partitioning algorithm based on

satin bowerbird optimization (SBO). J. Comput. Electron. (2020). https://doi.org/10.1007/s10
825-020-01491-9
6. M. Rapoport, T. Tamir, Best response dynamics for VLSI physical design placement, in 2019
Federated Conference on Computer Science and Information Systems (FedCSIS)
7. T.N. Bui, B.-R. Moon, Genetic Algorithm and Graph Partitioning
8. S.K. Roy, C. Giri, H. Rahaman, Optimization of test wrapper for TSV based 3D SOCs. J.
Electron. Test. 32, 511–529 (2016). https://doi.org/10.1007/s10836-016-5610-4
9. G.K. Chellamani, P.V. Chandramani, An optimized methodical energy management system
for residential consumers considering price-driven demand response using satin bowerbird
optimization. J. Electr. Eng. Technol. 2020, 955–967 (2020)
10. N. Arya, A.P. Singh, A fault avoidance approach with test set generation in combinational
circuits using genetic algorithm, in 2018 2nd International Conference on Inventive Systems
and Control (ICISC)
Identification and Classification
of Foreign Bodies from Rice Grains
Using Digital Image Processing
Rajlakshmi Ghatkamble
Abstract This research article presents the method for the identification and classifi-
cation of foreign bodies from rice grains using the digital image processing approach.
Any matter other than the rice grains in the rice images is considered as the foreign
bodies in the current research work. The foreign bodies can be in terms of weed,
stones, soil lumps, pieces of stems, plant leaves, or other types of grains. The amount
of the foreign bodies decides the quality of the rice grains which in turn may be
helpful to determine the quality of the rice grains. In the manual inspection of grains,
the foreign bodies are evaluated based on the manual inspection which may not be
so accurate in every inspection. Whereas the machine vision system automatically
determines the amount of the foreign bodies present in the rice grains which can
be helpful to farmers in sowing the seeds and also marketing the rice grains. The
digital image analysis algorithms are developed using machine learning techniques
using MATLAB and Python programming languages to determine the foreign bodies
present in the rice grain samples using a neural network method. The foreign bodies in
the rice grain image samples are determined based on the color texture and morpho-
logical features using a digital image processing method. All three above mentioned
features are presented to the neural network for training purpose and this trained
network is later used to identify the foreign bodies in rice grains.
Keywords Foreign bodies · Neural network · Image acquisition · Pre-processing ·

Feature extraction
R. Ghatkamble (B)
School of Research and Innovation, Department of Computer Science, C. M. R University,
Bangalore, Karnataka, India
e-mail: rajlakshmi.5samarth@gmail.com
290 R. Ghatkamble
1 Introduction
The current research article depicts the method for identification and classification
of foreign bodies from rice grain image samples using the digital image processing
approach. In the current situation, the machine vision inspection of agricultural prod-
ucts is emerging as an important area of research interest [5]. The variability in the
appearance of grains and foreign bodies makes it a challenging task for any machine
vision system to recognize and classify all the biological entities of rice grains as
well as foreign materials [5]. There is a lot of research undertaken till now in deter-
mining the potential of morphological features of rice grains as well as impurities
using pattern recognition techniques [3, 4]. In the current situation, the amount of
foreign materials present in the container of rice sacks affects the value of their
products and there is a chance for the human inspectors to inspect the rice grains to
play mischief during the evaluation of rice grains. Also, the decision-making capa-
bilities of human inspectors can be biased based on their mental health or work
pressure. So to avoid such circumstances occurring in rice grain evaluation grain
the current research work helps in estimating foreign bodies by analyzing the rice
grain image samples using machine learning techniques which can be done based on
digital image processing [3]. This paper is organized into sections mentioned below
which are proposed methodology which is concerned with image acquisition, pre-
processing, image enhancement, image segmentation, feature extraction, knowledge
base, and the last section is the conclusion of the work.
2 Proposed Methodology
The below-mentioned block diagram illustrates the procedure of identifying and

classifying foreign materials from rice grain image samples using digital image
processing as shown in Fig. 1. And the methodology is as given in algorithm 1.
Algorithm 1: Identification and Classification of Foreign Bodies from Rice
Grains Input: original 24-bit color image.
Output: Classified foreign materials.
Image Pre Image

Acquisition Processing Enhancement
Knowledge Feature Image

Base Extraction Segmentation
Fig. 1 Flow chart of the system design

Identification and Classification of Foreign Bodies … 291
Start
Step 1 Acquire rice grain images with impurities.
Step 2 Pre-processing of image samples.
Step 3 Enhance the grain image with impurities to remove noise and blurring by
filtering an image.
Step 4 Do the sample image segmentation.
Step 5 Extract morphological, color, and texture features.
Step 6 These extracted features of rice grain and impurities are analyzed and
identified using a probabilistic neural network.
Stop
(1) Image Acquisition
In this stage, the rice grain image samples along with impurities are captured under
standardized lighting conditions. The images are acquired using a color digital camera
with 4.0 megapixels fixed to a tripod stand at a fixed distance of approximately
700 mm with an orientation of 0° and a focal length of 3.2 mm. During image
acquisition the camera is placed at a location situated with a plane normal to the
path of the object so that the captured images are of good quality to further process
the imaging technique which is used to collect the data of a rice grain samples with
impurities from the image. The green simple plain background is used to capture the
image samples (Fig. 2).
The images acquired are 3264 * 2448 pixels in size. Rice images along with
impurities are captured and stored in JPEG format automatically. With the help of
data cable, the captured images are transferred and stored in the disk managing device
in a proper sequence (Figs. 3 and 4).
The proposed system determines the aspects of the ratio distribution of the image
sample. These samples are analyzed and examined based on their color, morpho-
logical as well as texture features. This technique of identifying the impurities from
rice grain image samples can be done using the machine vision technology that is
by using a probabilistic neural network algorithm. The current research work also
attempts to give the computerized results to analyze and classify the foreign bodies
Fig. 2 Block diagram of

image acquisition Rice grain image
Digital camera
with impurities
Grain image stored

Images stored in
in jpeg format
hard disk
292 R. Ghatkamble
Fig. 3 Image acquisition
Fig. 4 Rice grain image

with foreign bodies
present in rice grains and helps in determining the quality of rice grains up to some
extent, whose accuracy pertains to be higher.
(2) Pre-processing
Color image pre-processing of image sample and its segmentation are the classical
examples of the multichannel information processing concerned with digital images.
Pre-processing is a stage that is also called as data preparation step which is helpful
in contrast enhancement, noise reduction, or filtering and edge enhancements of the
given input image. Image filtering operations involve the removal of small objects
or the noise terms from an image.
(a) The grayscale image is a sample input image where each pixel in the image is
concerned with the pixel of rice and impurities holding the input image. Every
pixel in an image depicts the intensity information of the image which can be
either in terms of the black or white image after grayscale processing of the
image. It has an only black or white color in which black depicts as weakest
intensity and white as the strongest intensity.
Fig. 5 Pre-processed image
(b) RGB to gray processing is a stage where each pixel in an image comprising of
the red, green, and blue pixel which comprises of an array of single, double, or
whole pixel values specifying intensity values. This phase converts the RGB to
grayscale values.
(c) Binary image processing is a phase that converts the grayscale image into a
binary image. The output rice grain image replaces all the pixels in the input
image with the luminance greater than the level with the value as one that is
rice grain and replaces all the other pixels with the value 0 which is below the
luminance value lesser is considered as foreign bodies.
The function gray thresh is used to estimate the argument level automatically to
separate the object such as rice grains and impurities in the image from the back-
ground. The color of the rice grain whose threshold value is more than the set values
and is termed as a rice grain and the rest is separated as impurities and background
of the image (Fig. 5).
(3) Image Enhancement
This stage is one of the most appealing and simplest forms of methods used in digital
image processing. The idea behind this technique is to bring out the hidden details
from the image or highlight the features of interest from the rice sample image with
impurities. This phase is a subjective area of image processing. Image enhancement
techniques are used to improve the quality of the degraded image [1]. Grayscale
manipulation, filtering, and histogram equalization schemes are some of the various
image retrieval techniques used in the current situation. The histogram equalization
method is used to remap the gray levels of an image and hence it is considered as
the most convenient image enhancement technique.
(4) Image Segmentation
In the image segmentation process, the thresholding plays an important role in clus-
tering the objects in an image. The image segmentation process is a method of subdi-
viding an image into different objects or segments using the thresholding technique
294 R. Ghatkamble
Fig. 6 Segmented and cropped rice grains
which is considered as the first step in image processing [2]. The image is usually
subdivided until the rice grains and impurities are isolated from their background.
The image segmentation has generally two approaches in which one approach is
based on the discontinuity of gray level values and the other approach is based on
the similarities of the gray level values. The first approach uses the method of the
partition of an image based on the abrupt changes in the gray level of an image.
The second segmentation approach uses the thresholding method, region growing,
merging, and region splitting (Fig. 6).
(5) Feature Extraction
Algorithms are developed in the windows environment using python and MATLAB
programming language which is used to extract the morphological features, color,
and texture features of the rice grains as well as foreign materials.
The extracted morphological, color, and texture features in this research work are
as mentioned below:
(1) Area (mm2 ): This feature extracts the number of pixels inside the rice grain
and foreign bodies including their boundaries which is later multiplied by the
calibration factor (mm2 /pixel).
(2) Major axis length (mm): It is the distance between the endpoints of the longest
line that can be drawn through the rice grain and the foreign bodies. The major
axis endpoints are found by computing the pixel distance between every combi-
nation of border pixels of the rice grains as well as the impurities boundary and
hence finding the pair with maximum length.
(3) Minor axis length (mm): It is the distance between the endpoints of the longest
line that could be drawn through the rice grain as well as the impurities while
maintaining perpendicularity with the major axis.
(4) Aspect ratio: Major axis length/Minor axis length.
(5) Red mean: It is the average or means the value of red color or pixel in the rice
grain and the impurities present in the input image.
(6) Green means: It is the average or means the value of green color or pixel in the
rice grain and the impurities present in the input image.
(7) Blue means: It is the average or means the value of blue color or pixel in the
rice grain and impurities present in the input image.
(8) Entropy: Entropy value of red, green, and blue color in rice grain, as well as
foreign materials, is calculated.
(6) Knowledge Base
The knowledge base concerned with this research work is an organized set of data
elements or values using a model of vertical columns and horizontal rows of data and
a cell being the unit where the column and row intersect. All the extracted features
of the rice grains and foreign bodies extracted from the input image are stored in
this knowledge base during the training of the sample images and this knowledge
base is later used to classify the rice grains and impurities during the testing phase
of probability neural network.
3 Conclusion
In the current research work, the digital imaging approach is devised to analyze
and classify the foreign bodies from rice grain image samples using a digital image
processing method. The algorithm developed here helps in identifying the rice grains
and foreign bodies from the input image samples. The rice grain and impurity clas-
sification accuracies are found to be 80% and above respectively. The increased
presence of foreign materials in the rice grains samples reduces the capacity of clas-
sification accuracies of the varieties of rice grains and hence it also affects the quality
of the rice grains which is based on the purity of rice grains. According to the study
if the percentage of the foreign materials in the rice grain samples is above 50%,
then the identification and classification of the mixed rice grain samples become an
impossible task.
296 R. Ghatkamble
4 Future Scope
Future work may include rice image acquisition in bulk amount and with varying
backgrounds.
References
1. S. Shantaiya, U. Ansari, Identification of food grains and it’s quality using pattern classification.
Int. J. Commun. Comput. Technol. 2(2,3,4) (2010)
2. B.S. Anami, D.G. Savakar, Improved method for identification and classification of foreign
bodies mixed. ICGST 9(1) (2009)
3. B.S. Anami, D.G. Savakar, Effect of foreign bodies on recognition and classification of bulk
food grains. J. Appl. Comput. Sci. 6(3) (2009)
4. T. Jiang, Y. Zhang, F. Cai, J. Qian, Optical coherence tomography for identifying the variety of
rice grains. Adv. Optoelectron. Micro/nano-opt.
5. A. Mehrez, D. Ali, A. Mohamed, Hybrid method for cereal grain identification using
morphological and color features. IEEE
Improving Collaboration Between
Government and Citizens
for Environmental Issues: Lessons
Learned from a Case in Sri Lanka
Mohamed Sapraz and Shengnan Han
Abstract The collaborations between government and citizens are very crucial for
addressing environmental issues, but how such a collaboration can be developed in
practice is an open question. An interpretative case study is conducted to investigate
‘Wilpattu’ deforestation, one of the recent controversial environmental issues in Sri
Lanka. The paper studies citizens’ activities in the social media platforms and the
responses of government authorities regarding the case. The results indicate the root
cause is lack of proper collaboration and coordination between government and citi-
zens and other stakeholders. The study also reveals the societal, political factors, as
well as governing structures of government authorities, exacerbate the environmental
problem. By understanding the lessons learned from the study, a Digital Collabora-
tive Platform (DCP) is proposed for potentially improving collaborations between
government and citizens for addressing environmental issues in future research.
Keywords Digital collaborative platform · Environmental problems ·

E-government · Co-production
1 Introduction
Environmental sustainability is fundamental to humanity’s sustainable development

[1] and it is a dominant challenge at present. We have seen massive environmental
problems such as climate change including global warming, rising sea levels, ocean
acidification, dramatically expanding droughts, and loss of biodiversity [2]. The
global environmental conditions are changed at an unprecedented level, in which
developing nations like Sri Lanka are still struggling to adapt. Sustainability focuses
on the interrelatedness of social, economic, and environmental aspects, and is often
M. Sapraz (B)
Faculty of Computing, NSBM Green University Town, Homagama, Sri Lanka
e-mail: shafraz@nsbm.lk
M. Sapraz · S. Han
Department of Computer and Systems Science, Stockholm University, Kista, Sweden
e-mail: shengnan@dsv.su.se
298 M. Sapraz and S. Han
discussed in terms of the triple bottom line [3] where the key ideal is that all three
aspects need to be harmonized for something to be sustainable.
The collaboration between government and citizens is crucial in addressing envi-
ronmental problems. Collaboration is one form of citizen engagement and it is defined
as ‘citizens and administrators collaborate with one another to solve problems and
get things done’ [4]. With the recent development of new ICT technologies such
as Web 2.0 tools, mobile-related technologies, and social media platforms, improve
and provide a new avenue for the collaboration between government and citizen
in e-governance [5]. Due to the scarcity of government resources and financial
pressures, more focus appears in citizen co-production (a form of collaboration)
of public services. Co-production improves the collaboration between citizens and
government, improve efficiency and trust.
However, effective collaboration into solving environmental problems remains
uncertain and still, it is an open question [6, 7]. Citizens’ engagement plays a pivotal
role in environmental decision making [8]. Further, empowerment and citizen partic-
ipation play a key role in achieving sustainability [7]. Citizens are much eager than
ever to engage in environmental protection and they use ICT as a channel to raise
their voice against environmental issues [9]. It is required to conduct further research
to explore the possibilities of the practical application of collaboration into practice
in addressing environmental sustainability with the aid of ICT [10].
The study focuses on adopting an interpretive case study to investigate a case from
Sri Lanka. The case is based on deforestation taken place at ‘Wilpattu,’ the largest
National Park in Sri Lanka. The results indicate the root cause is lack of proper collab-
oration and coordination between government and citizens and other stakeholders.
The study also reveals the societal and political factors, as well as governing struc-
tures of government authorities, exacerbate the environmental problem. Hence, a
Digital Collaborative Platform (DCP) is proposed for potentially improving collab-
orations between government and citizens for addressing environmental issues in
future research.
2 Background: E-Government Services and Collaboration
The term ‘Electronic government (e-government)’ or ‘Digital government’ is a multi-

faceted concept and can be defined in different ways. The World Bank has defined E-
government as ‘government-owned or operated systems of information and commu-
nications technologies (ICTs) that transform relations with citizens, the private sector
and/or other government agencies to promote citizen empowerment, improve service
delivery, strengthen accountability, increase transparency, or improve government
efficiency’ [11]. For developing countries, ICT can be an effective solution among
different stakeholders including government and citizens in information interchange.
[12]. E-government can facilitate and provide better ways of collaboration and
information interchange to offer enhanced public services [13].
Improving Collaboration Between Government … 299
Three levels of citizen engagement in e-government services are identified as; 1.

The government shares the information with citizens (Government as a platform);
2. Both the government and citizens make a productive dialogue(interaction); 3.
Collaboration (Co-production/Citizen-government joint service delivery) [14]. Co-
production in this level of engagement is, ‘The public sector and citizens making
better use of each other’s assets and resources to achieve better outcomes or improved
efficiency’ [15]. In the past, governments had a constraint and limited ability to
collaborate with citizens and citizens to self-organize. But the digital platform is an
option to increase interactivity and enable co-production [16]. With the development
of DCP for co-production, citizens allowed to identify and discuss problems, propose
innovative solutions, and provide crowdsourcing solutions.
Overall, e-government services could be designed and facilitate major revolutions
in enhancing democratic citizen participation, adopt co-production, and offering
innovative solutions through crowdsourcing [17]. There are practical applications
of collaborative e-government projects (Environmental Exchange Network, Partici-
pateDB, GovLoop, MuniGov2.0, Joined-Up, etc. [13]) and Digital participatory plat-
forms [14] currently in use in different countries. Applications specially designed for
addressing the environmental problems are very few and previous empirical studies
are evidence to suggest that most of the existing e-government systems and plat-
forms remained mostly in information dissemination rather than focus on the effec-
tive use of government-citizen collaboration and citizen participation for practical
problem-solving [18, 19].
3 Research Method
The study adopts an interpretive case study approach to investigate why collabo-
rations between government and citizens cannot be set up in the case in Sri Lanka.
Interpretive case studies suggest starting with existing theories, yet allow ‘…a consid-
erable degree of openness to the field data, and a willingness to modify initial
assumptions and theories’ [20].
3.1 Case Description—‘Wilpattu’ Deforestation
The case is based on ‘Wilpattu’—a forest located in the Northwest coast lowland dry
zone of Sri Lanka and it had become controversial due to deforestation by different
stakeholders. ‘Wilpattu’ (Willu-pattu) means ‘Land of Lakes’ and the ‘Wilpattu’
Forest Complex is a land area of 528,200 acres and is the largest remaining forested
land in Sri Lanka [21]. Illegal deforestation taking place and threatens the reserved
forest since early in the year 2009. Unauthorized land settlements and encroachment
in the area have resulted in the deforested area.
4 Data Collection
4.1 Citizens Comments from ‘Twitter’ and ‘Facebook’
Citizens’ comments on ‘Wilpattu’ deforestation had been retrieved from ‘Twitter’

through a Web crawler. ‘The crawler is a computer program that will crawl a site or
set of sites, analyze the pages fetched and then report its results’ [22]. In retrieving
data, the following hashtags are used as inputs to the ‘Web crawler.’ During the
process, ethical consent is obtained from the ‘Twitter’ through submitting the details
and purpose of analyzing the data.
#wilpattu #Wilpattu #wilpaththu #Wilpaththu #wilpattu deforestation
#wilpaththu deforestation #wilpattudeforestation #save wilpattu #Save wilpattu
#savewilpattu #Savewilpattu #deforestation sri lanka #SriLanka deforestation
Web crawler generated 623 ‘Tweets’ and content is studied for the relevancy of
the case. After exclusion, 558 ‘Tweets’ are filtered for further analysis.
From Facebook, data retrieved from 15 active user profiles (8 males and 7 females
between the age of 20 and 50). The data retrieved from Facebook is limited to 15
users as we need to get the consent of the users to retrieve the data available in the
user profiles. As this case is controversial, many of the Facebook users reluctant to
share the information posted on their personal Facebook account. With the consent of
the users, in each Facebook profile, users are requested to search using the keywords
‘Wilpattu’ and ‘Wilpattu deforestation’ (or similar) for text posts and text comments.
The process of retrieving the data from profiles was completed in front of the author by
Facebook users. Initial results generated 328 posts and comments and it is reduced
to 235 different posts and comments after excluding the comments and posts not
directly related to the case.
4.2 Interviews of the Government Authorities
Semi-structured interviews had performed with 5 different government officers who

are directly or indirectly responsible for environmental issues of the country to gather
data about the case. The author visited the government authorities in the October 1st
week of the year 2019 and conducted interviews and recorded the interviews with the
consent of interview participants. Interestingly, none of the interviewees are willing
to disclose their job identity to the public, as ‘Wilpattu’ is a controversial issue in
the country. As an ethical practice, the aim of the study is well explained to the
interviewees’ and their consent to record, analyze, and publish gathered information
is obtained through a formal consent form. Four interviews were conducted as face to
face interviews and the other one is a telephone conversation. Interview questions are
broadly based on the ‘Wilpattu’ case incident, citizen and government engagement,
and collaboration in the case and what to do to improve the current situation. Three
Table 1 Profiles of interviewees’

Organizations Duration (min) Gender
Department of Forest Conservation (DFC) 40 Male
Department of Wildlife Conservation (DWC) 45 Male
Central Environmental Authority (CEA) 40 Female
Sri Lanka Army 40 Male
Information Communication Technology Agency (ICTA) 45 Male
Table 2 Results from Twitter and Facebook comments

Category Number Percentage (%)
(1) Role of the citizens in engage in environmental protection 345 42
(citizen attitude, behavior, responsibility, etc.)
(2) Blame and urge the government to take necessary actions 645 78
(3) Problem is identified as a racial issue (positive and negative) 449 55
(4) People invite others to gather and organize protests 180 22
*A single Tweet or comment may belong to more than one category
government officers are interviewed in the native language (‘Sinhala’) and the other
two are in English. Recorded interviews are transcribed for the data analysis (Table 1).
4.3 Data Analysis
The content analysis is used to analyze the collected citizens’ data (‘Twitter’ and
‘Facebook’). And, it is useful in eliciting meaningful findings by analyzing the
collected data and output can be presented in a quantifiable form [23].
All the related Tweets and comments are stored in separate excel sheets for the
analysis. The collected comments are thoroughly reviewed manually by reading the
content for the relevancy and excluded from the study if it is not directly related to
the. Another round of review carried out by an independent reviewer to validate the
categorization and the process of analysis.
The transcribed interviews were analyzed using inductive thematic analysis. It is a
qualitative study to identify, analyze, and reporting patterns from the data collection
to elicit meaning and describe the data in detail [24].
5 Results
The social media content analysis of citizen comments exhibits (Table 2), citizens’
enthusiasm, and a strong willingness to collaborate in environmental protection
(Category 1 and 4). Most of the citizens believe the government is accountable
and they blame for not taking timely actions to stop deforestation (Category 2).
Surprisingly, the case study is directed and viewed by certain citizens as a racial
issue (Category 3). Results of government interviews (Table 3: Themes and explana-
tion are provided) highlight reasons for their inability to stop deforestation and their
challenges.
5.1 Citizens Comments in Social Media
Role of the Citizens who Engage in Environmental Protection

Out of the total 558 comments analyzed 345 (42%) tweets are toward citizens’ role
in engaging in stop ‘Wilpattu’ deforestation (or similar in future). People oppose
and suggest their views toward deforestation. In such comments, people already are
speculating the root causes of the problem and emphasize the ‘role of the public.’
Ex: ‘Controlling wildlife would be easier than controlling the wild attitude of
those people who are destroying flora and fauna, not only in Sri Lanka but also in
other parts of the world’; ‘It’s not only Wilpattu, but all other jungles and forests of
Table 3 Themes and their associated meaning

Theme Explanation
(1) Issues faced by environmental authorities Drawbacks of certain policies, interferences of
in addressing issues certain politicians, and inter-governmental
communication gaps are certain obstacles for
the environmental authorities
(2) Officers accept the fact citizens made a Social media strongly impact in penetrating the
strong voice via Social media platforms message, but it could not resolve an issue
(3) Realize the need for ICT based platform for All the officer believes the presence of an ICT
communication between the stakeholders based collaborative platform could bring
consensus/solve/assist authorities and citizens
communication addressing environmental
issues
(4) Government authorities lack information Officers of environmental authorities could not
to provide solutions to the environmental access the required information to make certain
issues decisions and concerns raised by different
stakeholders
Sri Lanka should also be protected’; ‘number of social responsibilities entrusted in

the public…’
Blame and Urge the Government to Take Necessary Actions

From the dataset, 78% believe and blame the government for not taking proper
actions in addressing this issue. Citizens publicly comment about their voice against
politicians and government actions toward this deforestation issue.
Ex: ‘Not only the politicians but even the government officers are also spineless.’;
‘Useless talking about Wilpattu. Nobody takes any action. Administrators are afraid
of those Culprits’, ‘…the entire matter should go to the courts to get the matter
cleared without any doubt. Some politicians cannot survive without such issues
being raised….’
Problem is Identified as a Racial Issue (Positive and Negative) Unfortunately,

55% of the dataset proves this problem has led to unrest among ethnic groups,
and it brings misunderstanding and racism among people. It became a humungous
generator of heated arguments, rivalry, and fights among different races.
Ex: ‘In sum, the conversational domain around #Wilpattu is mired in over-
whelming toxicity, with rabid, racist, extremist content aimed at a specific commu-
nity, faith, and culture.’; ‘Have to reiterate that Wilpattu deforestation is a
NATIONAL issue. Please don’t infuse race or religion into it.’; ‘SaveWilpattu is
just pure racism-based propaganda.’
People Invite Others to Gather and Organize Protests

Another 22% comments were about inviting and gathering other citizens in orga-
nizing protests and gathering people in taking actions against ‘Wilpattu’ deforesta-
tion. Although most of the protests were on social media platforms, some were orga-
nized by different so-called environmentalists and third parties who gain benefits
from intragovernmental issues and racial issues.
Ex: ‘We will organize an island-wide trade union action against Wilpattu issue’;
‘Those who are against the deforestation of Wilpattu should get together, visit
Wilpattu and see what we can do.’; ‘So yeah let’s get together for #savewilpattu,’
‘Someone good at this, please start organizing I will gladly commit myself for that.’
5.2 Analysis to Understand Environmental Authorities’

Response
All five interviewees were aware of the reality of the ‘Wilpattu’ incident and had
varied expressions toward the case based on their professional points of view.
6 Lessons Learned from the Case Study
The interpretative case study results are vital in understanding the failure of the case.
Discovered failures are useful in finding feasible solutions in similar environmental
problems in the future. Study results show the main failure of the case is the lack
of collaboration and coordination between citizen-government, internal and inter-
government authorities in terms of providing timely actions. Citizens’ social media
analysis reveals their readiness to collaborate with the government in providing reli-
able information, evidence, and suggestions to stop deforestation. They are depressed
with the actions and decisions executed under alleged patronage of both officers and
politicians staying in Colombo and had never visited or does not aware much about
either the factual truth or societal issues. Administrative officers miserably failed
to perform legal duties and obligations mandated by the law and there are several
cases filed against authorities by environmental organizations (Including bribery and
corruption).
Some of the information published in mass media and social media is ‘bias’ and
‘false information,’ distorted reality, and published and compromised for political
survival, popularity, and monetary gains. This was leading to dire implications and
perceptions about the case among the citizens. Some citizens questioned about their
‘freedom’ to disclose ‘truth’ to the public. There is a socio-political debut between
the right of the people displaced by the war to resettle in their original villages
in ‘Wilpattu’ and the right to destroy the environment. As a result, deficiencies in
these miscommunications, the environmental issue had converted into a racism-based
ethnic issue. Controversially, it has not only affected the ‘Wilpattu’ area in the north,
but it has also disturbed the peace and harmony of the societies of the whole country.
Citizens’ comments further reveal, their ability to self-organize and execute many
endeavors to protect the forest and to assist the government in environment-related
decision making. Overwhelming inputs and campaigns of citizens through social
media platforms could not derive solutions, but they were able to create the awareness
among the majority of the citizens in the country, Sri Lankan ex-pats (citizens never
heard about even the term ‘Wilpattu’) and government including head of the state to
take actions to stop ‘Wilpattu’ deforestation. Ultimately, the government lost citizen
‘trust’ as they are unable to enforce effective measures toward the conservation of
natural resources by using administrative and legal power.
Government struggle to provide a solution to ‘Wilpattu’ catastrophe as it is a
complicated issue inextricably interwoven with political, social, and environmental
aspects. Addressing environmental sustainability is a complex societal problem as
many stakeholders and issues revolve around. But they failed to collaborate and
communicate with all these stakeholders (government, citizen, politicians, offices of
different authorities, etc.) in deriving consensus or solutions. They have faced the
challenge of decision making due to the overlap of institutional policies and political
interferences. Though laws and regulations exist for the conservation of state-owned
forest resources, officers face practical difficulties in law enforcement due to various
factors. Certain conflicts in the case could have been easily avoided if substantive
environmental information is available with officers. Due to the ‘unavailability of

information’ and government processes are ‘not transparent’ and ‘not well-defined,’
certain politicians and some other officers had misused the facts their benefits. In
the summary, failure of the case is mainly due to collaboration and coordination
problems and other different societal, political, and administrative issues that had
been affected in the case.
7 Conclusion
This paper presents the case of a serious environmental problem in Sri Lanka. The
lessons learned from this case emphasize the critical importance of improving the
collaboration between government and citizens in a developing country, e.g., Sri
Lanka. Therefore, we aim to continue the research and design a digital platform
to potentially improve this collaboration. The study had a limitation of increasing
the number of citizens and government officers’ comments as studied ‘Wilpattu’
deforestation is a controversial issue, some are reluctant to participate in the study.
A digital platform could provide co-production, democratic participation and
engagement, crowdsourced solutions and to improve transparency and account-
ability. ICT is an enabler to increase the interactions and relationships closer and more
frequent between citizens and government. Collaboration is critical to design public
services, co-production of government, and citizens to provide solutions for many
of the societal, political, and governance issues related to environmental problems.
References
1. B. Moldan, S. Janoušková, T. Hák, How to understand and measure environmental sustain-

ability: Indicators and targets. Ecol. Indic. 17, 4–13 (2012). https://doi.org/10.1016/j.ecolind.
2011.04.033
2. R. Gholami, R.T. Watson, A. Molla, H. Hasan, N. Bjørn-Andersen, Information systems solu-
tions for environmental sustainability: how can we do more? J. Assoc. Inf. Syst. 17, 521–536
(2016). https://doi.org/10.17705/1jais.00435
3. T.F. Slaper, The triple bottom line: what is it and how does it work? The triple bottom line
defined. Indiana Bus. Rev. 86, 4–8 (2011)
4. K. Callahan, Citizen participation: models and methods. Int. J. Public Adm. 30, 1179–1196
(2007). https://doi.org/10.1080/01900690701225366
5. R. Sandoval-Almazan, J.R. Gil-Garcia, Are government internet portals evolving towards more
interaction, participation, and collaboration? Revisiting the rhetoric of e-government among
municipalities. Gov. Inf. Q. 29, S72–S81 (2012). https://doi.org/10.1016/j.giq.2011.09.004
6. M. Cuthill, Exploratary research: citizen participation, local government and sustainable devel-
opment in Australia. Ann. Assoc. Econ. Geogr. 45, 167–168 (1999). https://doi.org/10.20592/
jaeg.45.2_167_2
7. T.M. Koontz, Collaboration for sustainability? A framework for analyzing government impacts
in collaborative-environmental management. Sustain. Sci. Pract. Policy 2, 15–24 (2006). https://
doi.org/10.1080/15487733.2006.11907974
8. D.M. Konisky, T.C. Beierle, Innovations in public participation and environmental decision
making: examples from the great lakes regions. Soc. Nat. Resour. 14, 815–826 (2001). https://
doi.org/10.1080/089419201753210620
9. G. He, I. Boas, A.P.J. Mol, Y. Lu, E-participation for environmental sustainability in transitional
urban China. Sustain. Sci. 12, 187–202 (2017). https://doi.org/10.1007/s11625-016-0403-3
10. C.J.A.M. Termeer, A. Bruinsma, ICT-enabled boundary spanning arrangements in collaborative
sustainability governance. Curr. Opin. Environ. Sustain. 18, 91–98 (2016). https://doi.org/10.
1016/j.cosust.2015.11.008
11. World Bank: E-government, http://documents.worldbank.org/curated/en/527061468769894
044/pdf/266390WP0E1Gov1gentina1Final1Report.pdf, last accessed 2020/03/20
12. G. Viale Pereira, M.A. Cunha, T.J. Lampoltshammer, P. Parycek, M.G. Testa, Increasing collab-
oration and participation in smart city governance: a cross-case analysis of smart city initiatives.
Inf. Technol. Dev. 23, 526–553 (2017). https://doi.org/10.1080/02681102.2017.1353946
13. S.A. Chun, L.F. Luna-Reyes, R. Sandoval-Almazán, Collaborative e-government. Transform.
Gov. People, Process Policy 6, 5–12 (2012)
14. E. Falco, R. Kleinhans, Digital participatory platforms for co-production in urban development:
a systematic review. Int. J. E-Plann. Res. 7, 1–27 (2018). https://doi.org/10.4018/IJEPR.201
8070105
15. T. Bovaird, E. Loeffler, From engagement to co-production: the contribution of users and
communities to outcomes and public value. Voluntas 23, 1119–1138 (2012). https://doi.org/
10.1007/s11266-012-9309-6
16. D. Linders, From e-government to we-government: defining a typology for citizen coproduction
in the age of social media. Gov. Inf. Q. 29, 446–454 (2012). https://doi.org/10.1016/j.giq.2012.
06.003
17. J.C. Bertot, P.T. Jaeger, J.M. Grimes, Promoting transparency and accountability through ICTs,
social media, and collaborative e-government. Transform. Gov. People Process Policy 6, 78–91
(2012)
18. L.A. Brainard, J.G. McNutt, Virtual government-citizen relations: informational, transactional,
or collaborative? Adm. Soc. 42, 836–858 (2010). https://doi.org/10.1177/0095399710386308
19. W. Olphert, L. Damodaran, Citizen participation and engagement in the design of e-government
services: the missing link in effective ICT design and delivery. J. Assoc. Inf. Syst. 8, 491–507
(2007). https://doi.org/10.17705/1jais.00137
20. G. Walsham, Interpretive case studies in IS research: nature and method. Eur. J. Inf. Syst. 4,
74–81 (1995). https://doi.org/10.1057/ejis.1995.9
21. Wipattu National Forest (2010) https://www.wilpattunationalpark.com/visiting_wilpattu_nat
ional_park.php?lId=02, last accessed 2010/02/03
22. M. Thelwall, A web crawler design for data mining. J. Inf. Sci. 27, 319–325 (2001). https://
doi.org/10.1177/016555150102700503
23. M. Bengtsson, NursingPlus open how to plan and perform a qualitative study using content
analysis. NursingPlus Open 2, 8–14 (2016). https://doi.org/10.1016/j.npls.2016.01.001
24. V. Braun, V. Clarke, Using thematic analysis in psychology. Qual. Res. Psychol. 3, 77–101
(2006). https://doi.org/10.1191/1478088706qp063oa
A Study on Blockchain Scalability
Manjula K. Pawar, Prakashgoud Patil, and P. S. Hiremath
Abstract Blockchain is an emerging technology for various business and indus-

trial applications because of its characteristics like security, immutability, and trans-
parency. But still, blockchain faces the main issue of scalability, which influences
several factors such as throughput, which is defined as the number of transactions
processed per second (tps), the limited size of the block, cost, and capacity (size
of the blockchain) and latency. In this paper, a survey of literature on the various
solutions for scalability issues, namely cashing, side chain, off chain, inter-chain,
child chain, on chain, and also deep learning methods along with their pros and cons.
Keywords Blockchain · Scalability · Cashing · Side chain · Off chain ·

Inter-chain · Child chain · On chain · Deep learning
1 Introduction
Satoshi Nakomoto developed blockchain technology in the year 2008 [1], which
gained popularity due to its application in cryptocurrency, like Bitcoin [2]. This
technology is utilized in many fields [3, 4] where immutability, transparency, and
security are the requirements. Of late, many banks do their transactions quickly with
the help of this technology. Several countries have come up with blockchain for bank
transactions [5]. Though it has several advantages, it has some disadvantages, such
as low throughput and poor scalability [6]. Ethereum and Bitcoin are two blockchain
platforms that process 20 and 7 transactions per second, respectively, which is much
M. K. Pawar (B) · P. Patil · P. S. Hiremath

KLE Technological University, Hubballi, Karnataka, India
e-mail: manjulap@kletech.ac.in
P. Patil
e-mail: prakashpatil@kletech.ac.in
P. S. Hiremath
e-mail: pshiremath@kletech.ac.in
308 M. K. Pawar et al.
less compared to common systems, like PayPal and Visa, that process 193 and 1667
transactions per second, respectively.
In this paper, the solutions for improving blockchain scalability represented in the
literature are discussed. Scalability can be expressed in terms of throughput, cost, and
capacity [7, 8]. As the size of the block is limited, transactions wait for more time for
completion, which affects throughput. As and when transactions are created, the fees
have to be paid to the minors, which generates the blocks. This is burdensome even
if the transactions are lighter since they have to pay micro values. This causes the
cost issue. As the number of blocks increases in the chain, which results in difficulty
in managing the blockchain, it causes the capacity issue. Solutions to overcome
these issues are: (1) Cashing method [5], which attempts to improve scalability and
throughput by adapting FPGA(Field Programmable Gate Array) key value for storing
in gigabit Ethernet Network Interface Controller, and this consumes low resources
and low power; (2) on chain, where all the transactions are performed in the main
blockchain only; (3) off chain, where only results are stored in main blockchain and
transactions are performed outside the main blockchain; (4) child chain, where the
results are stored in main blockchain and transactions, are performed in child chain;
(5) side chain [9], where different assets and functions from different blockchains
are exchanged; (6) inter-chain [7], which is a technique for communication between
blockchains; (7) Deep learning method, which improves the scalability by having
two layers, namely a base layer and branch layer.
The methods, namely cashing, on chain, off chain, child chain, side chain, inter-
chain, and deep learning, are described in detail in the following section.
2 Scalability Methods
2.1 Cashing
Cashing method improves the scalability and throughput by adapting the FPGA(Field
Programmable Gate Array) network interface controller (NIC). The cashing tech-
nique uses customized SHA-256 for performing actual hashing operations only when
the request parameter is not a hashed value. This saves time of hashing executions
and thus improves the performance. It adapts [10] the method having NIST (National
Institute of Standards Technology) recommendation, such as SHA-256, that helps
provide collision-free solutions. It uses the SHA-256 hash function that avoids the
collision and improves the hash table, which in turn improves the efficiency and
performance of the system. SHA-256 is a standard recommended by NIST [10],
which is used with the Bitcoin core server through remote procedure calls (RPC)
Jansson and curl libraries developed in [5]. Finally, it showed better performance
with more than 160 GB data in the lab. The results show a better improvement, i.e.,
nearly 103 times of system cache hit. It has a small area FPGA (Field Programmable
Gate Array) that uses less power consumption.
A Study on Blockchain Scalability 309
2.2 On Chain
The on chain technique involves increasing the size of the block by making size
as unlimited, wherein the transaction limit is higher, and also it reduces the cost
of transmission. But it has a disadvantage, specifically, as the size increases speed
is reduced, and the probability of generating orphan blocks will be more and thus
causes the maintenance cost to be more, and thereby it leads to throughput issues.
Due to these disadvantages, increasing the size of the block is not the right solution
[7].
The alternative solution offered is Merkelized Abstract Syntax Tree [11], which is
developed in BIP-114 by combining abstract Syntax Tree and Merkle Tree. Merkle
tree is a kind of binary tree that concatenates the two transactions of their hash values
of a block. Later it hashes the concatenated two transactions. The abstract Syntax tree
is the one that connects each function until all the dependencies in the program are
mapped. Figure 1 shows a segment of a Bitcoin script example, and its corresponding
MAST is shown in Fig. 2.
When MAST performs its operation, one branch is hashed, as shown in Fig. 3a,
b. The main advantage of this is that (i) it reduces the size of the block since it
grows with a log scale due to the binary structure, and (ii) it increases the privacy as
one branch is hidden and its information is not visible. But the disadvantage of this
method is that it doesn’t provide complete privacy as another branch is not hidden.
The next solution is Segregated Witness (Segwit) [12]. Transactions also consist
of digital signatures that occupy 70% of space. In order to solve this issue, signatures
are stored in a separate data structure called Witness, which separates them from
transactions. This also solves the problem of malleability, in which attacker creates
OP_IF
<A’s PubKey> OP_CHECKSIG
OP_ELSE
<time delay> OP_CHECKSEQUENCEVERIFY
OP_DROP
<B’s PubKey> OP_CHECKSIG
OP_ENDIF
Fig. 1 Bitcoin script example: code 1
<A’s PubKey> OP_CHECKSIG <time delay> OP_CSV OP_DROP

<B’s PubKey> OP CHECKSIG
Fig. 2 MAST of code 1 example in Fig. 1

Included in chain H
H
Not included in chain

H H
<timedelay> OP_CSV OP_DROP

<A’s PubKey>
OP CHECKSIG <B’s PubKey> OP_CHECKSIG
(a) MAST case1 (b) MAST case2
Fig. 3 MAST cases of code 1 example in Fig. 1
his transaction, which is similar to the original transaction but with different trans-
action Id (TX_Id). The attacker includes this in the next block earlier to the original
transaction to make it appear as if payment not made [12]. To solve this problem,
it removes the signature from data structure while generating TX_Id. It increases
the complexity of code. The number of transaction processing increases as capacity
usage also gets increased. This also leads to fungibility problems [12].
Another solution incorporated in this technique is Sharding, which is popular in
distributed processing systems [13]. As the chain size increases, the burden on nodes
increases as they need to store all block data, which is troublesome. Sharding is a
method that groups the nodes to form a shard, and each shard processes different
blocks instead of storing all blocks. This technique improves the throughput by
parallelizing the transaction processing. However, it becomes vulnerable to attack
(by about 1%), as an attacker may gain control over the entire shard.
2.3 Off Chain
In the off chain method, the transactions are processed outside the blockchain. This
is done by storing the main chain state and applies the last processed state that has
been done on the other channel. This method is called a state-channel chain. This
also lowers the load of the main blockchain, and hence, improves the throughput.
The general example of off chain is the Lightning Network of Bitcoin [14].
Usually, in Bitcoin, transactions need to pay the transaction fee even for micro-
payment transactions and need to wait until they enter the block. In the Lightning
Network, one opens the channel together with a payment user. A user puts the coin in
a multi-sig address so that they both can share the channel. Once the channel is open,
both of them can use it. Hence, the transaction fee is paid only in the main chain to
open the channel. After this transaction is processed, the main chain doesn’t record
these events, and the waiting time is also zero. Once the channel is closed, its state
will be maintained in main chain just as in on chain. For example, consider that John
and Merry establish a channel, and also Merry and Kelvin maintain a channel each.
Then John and Kelvin can also have a connection with this. Thus, it can enhance
scalability. The advantages of the method are: fewer burdens on the main chain,
transaction fee reduced, and standby time is reduced.
By applying this method, the solution network called Raiden Network [15] is
obtained, which is similar to Lightning Network except that, Raiden network handles
a different variety of transaction types, e.g., the Ethereum supports smart contracts
and IoT, etc.
2.4 Side Chain
The side chain method helps in exchanging the properties of multiple blockchains
among each other. This approach aims to bring the functionalities of one blockchain
into another [9]. If someone wants to use bitcoin, he can put this in another blockchain
as its cryptocurrency is of high value, limited function, and slow. Suppose one wants
to use Bitcoin on Ethereum, then he can use the Ethereum smart contract with Bitcoin.
For example, to operate the bitcoin in another blockchain, it can be implemented as
follows: first, the amount of Bitcoin that can be operated need to be transferred to a
special account and freeze it. Now it creates coin in another blockchain and operates
through it. After all the transactions are completed after the freeze, the remaining
amount is available. If this is possible, interoperability is possible; that is, it can be
used from another blockchain. But this requires controlling of cryptocurrencies since
the values of currency keep changing on and on.
2.5 Child Chain
The solution of the child chain method is parent and child structure. In this technique,
child chain processes all the transactions, and the parent chain records all the results.
It is applied in the Ethereum plasma chain [16], where the plasma chain recursively
creates child chains. Transactions are processed in child chain. All transactions in
the generated block are transformed to a Merkle hash root, as shown in Fig. 4a. The
same is recorded in the parent chain. The initial stage is the same as that of off chain.
Once the child chain is created, it is going to become a parent chain in the plasma,
and it deposits coin through a smart contract. To use this child blockchain, it receives
the token. During transactions, if fraud occurs, then the bad node (the node where
the fraud occurred) will become slower than the owner of the token to withdraw the
token. In such cases, the owner will connect to any other parent, as shown in Fig. 4b.
Hence, the deposit is lost by the bad node.
The verification process is carried out by any child chain. The benefit of the
method is that it is lighter since only Merkle hash roots of child chains are stored in
parent chains. The disadvantage of this method is that verification by child chains is
A->B A->B
2ET 2ET
Plasma Plasma Plasma Plasma

Blockchain Blockchain Blockchain Blockchain
(Ethereum) (Ethereum (Ethereum) (Ethereum)
Plasma
Plasma
Blockchain
Blockchain
(Ethereum)
(Ethereum)
Root
Root
(Ethereum)
(Ethereum)
(a) Plasma example (b) If the parent chain

is invalid
Fig. 4 Plasma chain example
expensive because it has to keep checking the honesty of all parent chains. It is also
troublesome, as transforming all the transactions of the child chain to a Merkle hash
needs to be valid in the parent chain and then recursively to the root.
2.6 Inter-chain
The inter-chain approach provides communication between blockchains, just like

the Internet. This is the same as the side chain. In another word, inter-chain provides
the infrastructure for side chain.
Let us consider the case study of this method as Atomic-swap [15]. Consider that
Harry uses Lightcoin, and John uses Bitcoin. Harry wants 30 light coins and John
wants 2 bitcoins to be exchanged among each other. Then they can use inter-chain as
follows: First, John creates a transaction of 2 bitcoins with Harry’s digital signature,
sets the number k which only he knows, or he gets back it as coin after a certain time.
Next, John creates the public key for k and broadcasts that to network. Harry, after
receiving this message, creates the transaction 30 light coins using John’s digital
signature or gets back the coin after a certain time, which is smaller than John’s
expiration time, which he had set. Usually, the expiration time for Harry is shorter
than John’s expiration time; this is because if expiration time for Harry is larger
than John’s expiration time then John receives the 30 lightcoins just before the end
expiring time and coins return to John before he takes 2 bitcoins. The advantage of
this approach is that the exchange of properties between different blockchains and
disadvantages are that all blockchains need to implement the same algorithms such
as POW hashing and so on. If one of the users is lazy among all, then exchange gets
completed almost when expiration time is reached.
2.7 Deep Learning Method
In Deep Learning method, it needs big data for training and builds an effective deep
learning model. Deep Learning [17] consists of two-layer blockchain architecture,
such as the Base Layer and Branch layer. Base Layer can be implemented with
Bitcoin or Ethereum. It can have several branch layers. Validation of each branch
layer will be done in the base layer.
Merkle trees implement branch layers. Merkle tree helps in achieving efficient
and secure verification of transactions in a distributed environment. The root of every
Merkle tree is stored in the base layer, which ensures that trees stored in branch
layers are unchangeable. If any modification happens in the tree, then its validation
fails, since it doesn’t match with root stored in the base layer. The advantage of
the approach is that it helps in providing features like POET (Privacy, Ownership,
Efficiency, Transparency). It is implemented by using the Hash graph and FBA
(Federated Byzantine agreement) protocols that improve the throughput and reduce
the latency. Further, it uses fewer resources but requires high bandwidth. The hash
graphs work only privacy granted or permissioned network.
3 Analysis
Herein, the solutions discussed in the previous section are compared by considering
factors that are influenced by the scalability, namely throughput, cost, and capacity.
Table 1 provides a performance comparison of all these solutions. It is observed that
all the methods increase throughput, except that child chain and off chain reduce the
capacity by reducing the load on main chain. These also reduce the cost by handling
the transactions outside the main blockchain. Increasing the block size does not
solve the issue of blockchain. Hence, the on chain method, such as MAST method,
slightly modifies the Merkle tree and solves the capacity issue. Segwit on the other
hand separates signature from the transaction and consumes more capacity, which
is a structural solution. Sharding, on the other hand, provides a distributed database
system and solves both capacity and cost issues. Inter-chain, which is similar to the
side chain, hence another row for side chain is not shown in the table. It achieves a
functional part of scalability by connecting different blockchains. Cashing improves
the performance by using FPGA with less power consumption, and hence cost is
reduced, and throughput is improved. Deep Learning, by using two layers (base layer
Table 1 Comparison of perforance of different methods in terms of throghput, cost and capacity
Category Method Tps Capacity Cost Advantage Disadvantage
On chain MAST – Low – Provides high Privacy is not
privacy completely
provided
Big block High High Low Transmission Increases orphan
limit is high blocks,
centralization of
mining pools
Segwit High – Low Possibility to Leads to
apply to Bitcoin Fungibility
with various problems
solutions
Sharding High Low – Parallelized It leads to 1%
processing, the attack
low burden on enhancement
the capacity
Off chain Lightning Low Low Low No cost for Only payment
network transition and channel
waiting time is
almost zero
Raiden Low Low Low General-purpose
network channel
Child chain Plasma Low Low – Parent and child Verification is
structure expensive
blockchain
Inter- chain Atomic-swap – – – Communication Bounded
among applicable
blockchains situations
Deep learning DeepLinQ High Low Low Provides privacy, Require rigorous
ownership, verification, the
accountability, hash graphs are
transparency more suitable for
private or
permissioned
chain
Cashing FPGA High – Low Less power
consumption, the
result is
improved over
103 times of
average
and branch layer), improves the throughput. Besides, by considering the Merkle tree
in Baselayer, it enhances the cost.
4 Conclusion
Blockchain is an emerging trend in the industry, as well as in business applications.

But, to use this in real-world applications, it needs better scalability, which is lacking.
The standard platforms of the blockchain are Ethereum and Bitcoin. Ethereum
processes 20 transactions per second and Bitcoin processes seven transactions per
second, whereas real-world applications need more number of transactions to be
processed per second. Throughput, latency, cost, and capacity are the factors which
affect the scalability. Hence, to improve the scalability, there are several approaches,
namely cashing, on chain, off chain, side chain, child chain, inter-chain, and deep
learning. These solutions are compared in terms of performance metrics, such as
throughput, cost, and capacity. From the present literature survey, it is observed that
to apply the blockchain for real-life applications, and it is necessary to study existing
methods and suggested new solutions in view of scalability. The future scope of the
work reported on the studies of blockchain presented in the survey points to consid-
eration of the latency factor, along with the number of nodes in the blockchain, in
the development of new efficient solutions for the implementation of blockchain
technology in real-world applications.
Acknowledgements The authors are grateful to the reviewers for their helpful comments and
suggestions, which improved the quality of the paper.
References
1. S. Nakamoto, Bitcoin: a peer-to-peer electronic cash system. Cryptography at https://met

zdowd.com (2009)
2. S.D. Lerner, RSK white paper overview (2015)
3. C. Khan, A. Lewis, E. Rutland, C. Wan, K. Rutter, C. Thompson, A distributed-ledger
consortium model for collaborative innovation. Computer 50(9), 29–37 (2017)
4. J. Ray, A next-generation smart contract and decentralized application platform. Ethereum
Wiki
5. A.I. Sanka, R.C.C Cheung, Efficient high performance FPGA based NoSQL caching system
for blockchain scalability and throughput improvement, in 2018 26th International Conference
on Systems Engineering (ICSEng) (IEEE, 2018), pp. 1–8
6. H. Kohad, S. Kumar, A. Ambhaikar, Scalability issues of blockchain technology. Int. J. Eng.
Adv. Technol. IJEAT. 9(3) (2020). ISSN: 2249-8958
7. S. Kim, Y. Kwon and S. Cho, A survey of scalability solutions on blockchain, in 2018 Interna-
tional Conference on Information and Communication Technology Convergence (ICTC), Jeju,
2018, pp. 1204–1207
8. C. Worley, A. Skjellum, Blockchain tradeoffs and challenges for current and emerging appli-
cations: generalization, fragmentation, sidechains, and scalability, in 2018 IEEE International
Conference on Internet of Things (iThings) and IEEE Green Computing and Communications
(GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart
Data (SmartData) (IEEE, 2018), pp. 1582–1587
9. A. Back, M. Corallo, L. Dashjr, M. Friedenbach, G. Maxwell, A. Miller, A. Poelstra, J. Timón,
P. Wuille, Enabling blockchain innovations with pegged sidechains, vol 72 (2014). http://www.
opensciencereview.com/papers/123/enablingblockchain-innovations-with-pegged-sidechains
10. NIST, Secure hash standard (SHS) Fed. Inf. Process. Stand. Publ. 180–184 (2015)
11. L. Johnson, Merkelized abstract syntax tree. https://github.com/bitcoin/bips/blob/master/bip-
0114.mediawiki,Github Repository. Accessed 2018-06-30 (2016)
12. L. Eric, L. Johnson, W. Pieter, Segregated witness. https://github.com/bitcoin/bips/blob/mas
ter/bip-0141.mediawiki,Github Repository. Accessed 2018-06-30 (2015)
13. S.S.M. Chow, Z. Lai, C. Liu, E. Lo, Y. Zhao, Sharding blockchain, in 2018 IEEE International
Conference on Internet of Things (iThings) and IEEE Green Computing and Communications
(GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart
Data (SmartData) (Halifax, NS, Canada, 2018), pp. 1665–1665, https://doi.org/10.1109/cyb
ermatics_2018.2018.00277
14. J. Poon, T. Dryja, The bitcoin lightning network: scalable offchain instant payments, draft
version 0.5 9: 14 (2016)
15. Komodo an advanced blockchain technology, focused on freedom. https://komodoplatform.
com/wpcontent/uploads/2018/03/2018-03-12-Komodo-White-Paper-Full.pdf (2018)
16. J. Poon, V. Buterin, Plasma: scalable autonomous smart contracts. White paper, pp. 1–47 (2017)
17. E.Y. Chang et al., DeepLinQ: distributed multi-layer ledgers for privacy-preserving data
sharing, in 2018 IEEE International Conference on Artificial Intelligence and Virtual Reality
(AIVR) (Taichung, Taiwan, 2018), pp. 173–178
What Influences Consumer Behavior
Toward Information and Communication
Technology Applications: A Systematic
Literature Review of UTAUT2 Model
Jick Castanha , Subhash Kizhakanveatil Bhaskaran Pillai ,
and Indrawati
Abstract Success of any Information and Communication Technology (ICT) appli-

cation, whether developed for industrial or individual use, depends on the level of
acceptance by the consumers and usage of the same on a regular base in the most
sustainable manner. Result of the content analysis on UTAUT2 model reveals that this
model have more predictive ability than earlier models. The variance explained on
behavior intention was of 61–80% and on use behavior was of 41–60%. The external
factors which are applied newly, namely, trust, risk, innovativeness, privacy, security,
self-efficacy and information are found to be good predictors. Various stakeholders,
especially those entrepreneurs who are developing new as well as existing owners of
ICT applications consider these constructs for adoption and continuous use among
consumers. The present scenario around the world is going to result in more and
more use of ICT applications by the people for all their needs.
Keywords UTAUT2 · Technology adoption · Behavior intention · Content analysis
1 Introduction
During the last one decade, exponential development and growth are being witnessed
around the world in the field of ICT applications, technology adoption is being one
of the major areas in almost all fields of business which completely changed the
lifestyle of people, and the success completely depended on the level of acceptance
and continued use of the ICT applications by the consumers. Both public and private
J. Castanha (B) · S. K. B. Pillai

Goa University, Goa, India
e-mail: jickcas@gmail.com
S. K. B. Pillai
e-mail: subhash@unigoa.ac.in
Indrawati
Telkom University, Bandung, Indonesia
e-mail: indrawati@telkomuniversity.ac.id
318 J. Castanha et al.
players are showing great interest and huge investments for the development and
introduction of new technologies for the use of various stakeholders in all fields of
business. Unless and until these innovations in ICT applications are not success-
fully adopted and efficiently used by the end-users, time and money spent on these
innovations become useless, hence researchers around the world started focusing on
understanding and identifying what makes a consumer adopt and use technology.
Technology adoption is a stage in consumer buying behavior where in the selection
of appropriate technology among the alternatives available for the use by an individual
or an organization [1]. Technology adoption is not only related to technology but also
related to the user’s attitude [2], personal and performance-related benefits [3], social
influence [4], facilitating conditions [5] and also other factors. This being the case,
researchers in technology adoption and diffusion are considered to be more mature
areas of exploration [6] which led to the development of various theories and models
by various scholars and tested in different systems and in a different context.
During the last 52 years period (1960–2012), researches in this domain of tech-
nology adoption identified new conceptualizing factors that influence the behavior
of users to adopt the technology which led to the development of various consumer
behavior theories and models, namely, Diffusion of Innovation Theory (DOI) by
Everett Roger in 1960, Theory of Reasoned Action (TRA) by Martin Fishbein and
Ajzen in 1975, Social Cognitive Theory (SCT) by Bandurra in 1986, Technology
Adoption Model (TAM) by Davis in 1989, Theory of Planned Behaviour (TPB) by
Icek Ajzen in 1991, Model of Personal Computer Utilization (MPCU) by Thompson
and Higgins in 1991, Motivation Model (MM) by Davis, Bagozzi and Warshaw in
1992, Extended Technology Adoption Model (TAM2) by Venkatesh and Davis in
2000, Unified Theory of Acceptance and Use of Technology (UTAUT) by Venkatesh,
Morris, Davis G and Davis F in 2003, and Model of Acceptance with Peer Support
(MAPS) by Sykes, Venkatesh and Gosain in 2009 [6–8]. The most recent model being
the Extended Unified Theory of Acceptance and Use of Technology (UTAUT2) by
Venkatesh, Thong and Xu in 2012.
The presence of many theories, each having pros and cons leads confusion
among the researchers, as they are forced to pick from a wide variety of competing
models and theories which suits the study characteristics [8]. To sort this confu-
sion, Venkatesh, Morris, Davis and Davis developed a unified model – The Unified
Theory of Acceptance and Use of Technology in 2003, popularly known as UTAUT,
by systematic review and consolidation of various constructs used in earlier models
developed during the period 1960–2000 [6]. The UTAUT considered four main
constructs, namely Performance Expectancy (PE), Effort Expectancy (EE), Social
Influence (SI) and Facilitating Condition (FC) which are direct determinants of
Behavioral Intention (BI) and ultimate Use Behavior (UB) leads to actual system
use.
However, this UTAUT model was having inherit limitations such as it does not
represent the external factors that can affect the performance of the behavior, behavior
intention has weak ability to predict behaviors that are not completely within an
individual control [9], hence Venkatesh, Thong and Xu extended the original UTAUT
model with three more main constructs [10], namely, Hedonic Motivation (HM),
What Influences Consumer Behavior Toward Information … 319
Fig. 1 Extended unified theory of acceptance and use of technology (UTAUT2). Source Venkatesh
et al. [10]
Price Value (PV), and Habit (H) which are direct determinants of behavioral intention
and this model is commonly known as UTAUT2, having three moderators, namely,
Age (A), Gender (G) and Experience (E), the detailed model is shown (see Fig. 1).
The original UTAUT Model and the modified UTAUT2 are shown together (see
Fig. 1), the shaded constructs (HM, PV and H) and the three moderators (A, G
and E) are the addition made over UTAUT to make UTAUT2 which resulted in
increased predictive capability. The orange lines indicate that all the seven constructs
or independent variables (PE, EE, SI, FC, HM, PV and H) influences the construct
or dependent variable (BI) which in turn influences the final resulting construct or
dependent variable (UB), which is shown in the blue line. Also the two constructs
or independent variable (FC and H) shown in blue line directly influences UB. The
moderating variables are shown in dotted green lines.
It was found that the UTAUT model explained the variance of 56% on behavior
intention but in the UTAUT2 it explained 74% which is a good predictor model of
behavioral intention. With respect to technology use, UTAUT explained variance of
40% whereas UTAUT2 explained 52% [10]. Hence, UTAUT2 was widely accepted
by the researchers and also is been tested in a different context and hence the scope
for exploring the applicability of UTAUT2 in assessing consumer behavior when it

comes to ICT applications in different fields of business around the world.
The goal of this research is to analyze the research works carried out on the
UTAUT2 model during the period 2012–2020 to see how best this model is applied in
different areas of ICT applications around the world so as to see those critical factors
which are considered significant and influencing consumer behavior. Subsequent
sections discuss (1) the methodology followed for the content analysis; (2) findings
and discussion on the result of the content analysis; (3) conclusion of the analysis
and the implications which provides some insights on the applicability of UTAUT2
model and finally (4) limitations of the work as well as future research directions. As
far as the authors’ knowledge, no similar study has been carried out so far addressing
the above-mentioned issues, which makes this paper unique because it throws some
light on an otherwise unexplored area which needs to be studied in detail and how
best the developers and also existing organizations of ICT application can improve
the services to have complete customer satisfaction and improved brand loyalty and,
therefore, the study fills the gap by adding valuable knowledge, new perspectives,
and presents possibilities for consideration. The paper offers valuable inputs for
academicians and researchers, developers of ICT applications, business entities using
ICT applications, the governments, and other stakeholder groups.
2 Methodology
The study aim was to conduct the systematic review of all the research works carried
out during the period 2012 till February 2020 on the application of UTAUT2 model,
mainly for assessing and identifying the most and least influencing constructs,
namely, the independent and dependent variables and also the moderating demo-
graphic variables. A comprehensive search was conducted on all research database,
namely, Emerald, Elsevier, Inderscience, JSTOR, MDPI, Sage, Springer, Taylor and
Francis, Google Scholar and other open sources by using keywords such as “Extended
Unified Theory of Acceptance and Use of Technology” or “UTAUT2.” The search
was limited to only research articles and conference proceedings. After the identi-
fication of papers, the criteria for selecting the papers were established. The papers
which used any of the extended UTAUT2 constructs (HM, PV, and H) along with the
original UTAUT constructs (PE, EE, SI, and FC) have been selected for further anal-
ysis. Finally, 204 articles were considered for carrying out content analysis. These
articles were then analyzed in series of aspects, namely, methodological details,
types of relationship found between model constructs, explained variance by model
constructs, moderating variables, and external variables, which are in line with the
earlier works [8, 11, 12].
For the purpose of assessing the inter-relationships and the predictive capability of
the constructs, weight analysis was carried out. A total of 204 articles were reviewed
where in 186 studies were quantitative in nature, of which 12 studies made use
of UTAUT2 more than once in the same study due to different time span or user
type, bringing about a total of 221 different occurrences of the relationship between
the variables. In order to better understand the predictive power of each indepen-
dent/dependent variable pairs, a weight analysis was performed as suggested by
Jeyaraj [13]. The predictor’s weight of each UTAUT2 construct (PE, EE, SI, FC,
HM, PV, H on BI and UB) was obtained by dividing the number of times each of
these constructs has been found to be significant on behavior intention and on use
behavior by the total number of times that relationship has been examined across all
studies [8]. A weight of “1” indicates that the relationship between two pairs is signif-
icant across all studies, whereas “0” indicates that this relationship is non-significant
across all the studies. The Independent variable is said to be the best predictor, if its
weight is 0.80 or more [13].
3 Findings and Discussion
Three way analysis is being carried out on the final 204 research papers, namely,
demographic characteristics of the research papers, methodological aspects used in
the research papers and the analysis of constructs used in the UTAUT2 model. The
findings of each category have been discussed in the following sections.
3.1 Demographic Characteristics
Content analysis result of the demographic characteristics of 204 papers on UTAUT2

during 2012–2020 reveals certain interesting facts about frequency of research
works carried out over the years, research collaboration, the most prolific author,
geographical concentration of the work, prominent publishers as well as application
area/industry.
Though the UTAUT2 model was developed in the year 2012, the initial three
years (2012–2014) very few researchers adopted this model in their research (only
4%) as this new model was not considered superior over the earlier models when
it comes to studying consumer behavior in ICT applications. But over the next five
years (2015–2010) researchers identified the superiority of the UTAUT2 model and
started applying it extensively (96%) and found that this model explains consumer
behavior better than earlier models. Another interesting aspect is about the collab-
oration among the researchers, where only less than 9% of the research works are
being carried out by individual authors whereas more than 91% of the research works
are the result of the collaboration of at least 2 researchers. This clearly indicates that
among the researchers also, the technology adoption and use is increasing, whereby
geographical distance is not an issue for sharing ideas. The most prolific authors
who published at least 4 research works using the UTAUT2 model in different fields
of business along with their respective institutions are listed in Table 1. These eight
Table 1 Most prolific authors

Authors (Years of sample works) University No.
Dwivedi [14] Swansea University, UK 12
Alalwan [15] King Abdulaziz University, Saudi Arabia 9
Oliveira [16] New University of Lisbon, Portugal 9
Williams [17] Swansea University, UK 7
Indrawati [18] Telkom University, Indonesia 6
Tarhini [19] Sultan Qaboos University, UK 5
Rana [20] Swansea University, UK 5
Baabdullah [21] King Abdulaziz University, Saudi Arabia 4
Source Authors Compilation
authors collectively published around 57 (28%) collaborative research works out of

the total 204 papers.
When it comes to geographical concentration, the Asia Pacific region accounts for
around 40% of the research works, followed by Europe (24%), the Middle East and
Africa (23%), North America (10%) and Central and South America (3%). When
the country-wise analysis result is considered, concentration is more seen in India
(25%) and Malaysia (22%), followed by the USA (13%), Jordan (12%) and Spain
(12%). The result of publisher-wise analysis shows that the majority of the papers
are published by Emerald (22%) and Elsevier (21%) and the remaining 57% of the
research papers are shared by Springer, Taylor & Francis, Inderscience, Wiley, Sage,
JSTOR, MDPI, etc.
The result of the application area/industry where the UTAUT2 model is applied
reveals that Education (16%), General purpose (14%) and Healthcare (10%) are the
three toppers having double-digit figures. Those with a share of between 5-9% shares
include Banking (9%), Mobile Technology (8%), Shopping (7%), Payment (5%) and
Travel (5%). The result also reveals that almost all the area/industry are being studied
by the researchers to identify the consumer behavior toward adoption and use of ICT
applications.
3.2 Methodological Aspects
Content analysis result of the methodological aspects provides insights on the

methodology adopted for identifying and collecting sample data and various statis-
tical techniques used for analyzing the sample data. For the purpose of data iden-
tification and collection, most of the researchers used exploratory research (66%)
that was based on an online survey (50%) using convenience sampling (25%) with
a sample size of 201–400 (48%) using Likert’s scale for assessing the influencing
factors of consumer behavior toward ICT applications. The majority of the studies
Table 2 Statistical
Technique name %
techniques used
Confirmatory factor analysis (CFA) 31.14
Structure equation modeling (SEM) 30.56
Descriptive analysis 27.47
Regression analysis 4.26
Exploratory factor analysis (EFA) 3.87
Correlations 1.55
Analysis of variance (ANOVA) 0.58
Meta-analysis 0.39
Chi-square test 0.19
indicated a high response rate, 13% of the research between 61 and 80% response
rate and 19% of the research between 81 and 100%, which clearly shows that the
result of the research works will have a high degree of reliability.
Statistical techniques used for analyzing the data is shown in Table 2. Since the
majority of the research identified and collected multivariate data, data analysis is
carried out mostly using confirmatory factor analysis (31%) and structural equation
modeling (31%). This clearly shows that complexity involved in identifying and
assessing the consumer behavior for the purpose of understanding whether there is
any pattern emerging out of the data which indicates the most and least influencing
constructs/variables when it comes to adoption and use of ICT applications with the
help of the UTAUT2 model.
3.3 Construct Used in UTAUT2 Model
This section provides a discussion about the different constructs/variables used in

analyzing and testing the UTAUT2 model and their inter-relationships, reliability
statistics, moderating variables used, and finally, use of external variables and their
impact. The UTAUT2 model has seven constructs (see Fig. 1) as independent vari-
ables, namely, Performance Expectancy (PE), Effort Expectancy (EE), Social Influ-
ence (SI), Facilitating Conditions (FC), Hedonic Motivation (HM), Price Value (PV),
and Habit (H). The additional two more constructs treated as dependent variables
are Behavioral Intention (BI) and Usage Behavior (UB).
The UTAUT2 model (see Fig. 1) focuses on 10 inter-relationships (shown as
orange and blue arrows), namely, PE-BI, EE-BI, SI-BI, FC-BI, HM-BI, PV-BI, H-
BI, FC-UB, H-UB, and BI-UB. Those constructs having weight above 0.80 are
considered to have the best predictors when it comes to finding inter-relationships,
remaining all constructs with weight less than 0.80 do not have good predictive
ability [13]. Weight analysis of independent variables (refer Table 3) indicates that
Table 3 UTAUT2 variables relationship and weight analysis

Relations Significant Non-significant Not tested Total Relation Weight of
examined predictors
PE-BI 151 22 48 221 173 0.87
EE-BI 79 97 45 221 176 0.45
SI-BI 89 82 50 221 171 0.52
FC-BI 82 76 63 221 158 0.52
HM-BI 124 47 50 221 171 0.73
PV-BI 84 50 87 221 134 0.63
H-BI 115 22 84 221 137 0.84
FC-UB 40 26 155 221 66 0.61
H-UB 53 9 159 221 62 0.85
BI-UB 82 5 134 221 87 0.84
only four pairs of the UTAUT2 are said to be best predictors, namely, PE-BI (0.87),
H-BI (0.84), H-UB (0.85) and BI-UB (0.84), whereas the other variables did not meet
this requirement, the fifth closest predictor being HM-BI, with weight being 0.73.
This provides a generalized result that only PE and H are the two constructs in the
UTAUT2 model which allows prediction of BI and UB when it comes to adoption
and use of ICT applications.
When it comes to the explained variations, considered as an indicator for reliability
(R2), earlier studies on UTAUT model having six main variables 56% of the variations
are explained w.r.t BI and UB of 40% [6], whereas the extension proposed in UTAUT2
produced a substantial improvement in variance explained in BI of 74% and UB of
52% [10]. Result of the content analysis of the 204 research works on UTAUT2
also reveals that the variance explained (R2) in BI is found to be between 61 and
80% of the 52% of the studies and in UB between 41 and 60% of the 42% of the
studies, which is similar to the UTAUT2 models explained variations. This is a clear
indication that the UTAUT2 model is superior over the earlier models on consumer
behavior when it comes to the adoption and use of ICT applications.
About the influence of the demographic variables on the constructs, only 29%
of the research applied and the most common moderator was gender (25% of
the research) and age (24% of the research). Few researchers also examined the
moderating role of income, nationality, employment, marital status and also cultural
factors. Apart from those seven constructs/variables, researchers also used additional
external constructs/variables. The result shown in Table 4 shows that trust is the most
frequently applied construct, followed by risk, innovativeness, privacy, security, self-
efficacy and information. Among these seven constructs, self-efficacy is having more
significance (67%), followed by trust (58%), risk and security (54%) and innovative-
ness (48%). Privacy and information have less significance. Analysis of the original
as well as the external constructs used and the moderating variables reveals that the
Table 4 Significance of external variables

Relations Significant Non-significant Not tested Total
Trust—BI 25 7 11 43
Risk—BI 13 4 7 24
Innovativeness—BI 10 4 7 21
Privacy—BI 4 3 7 14
Security—BI 7 3 3 13
Self-Efficacy—BI 8 2 2 12
Information—BI 4 3 3 10
UTAUT2 model does have superiority over the earlier consumer behavior models
developed during 1960–2012.
4 Conclusion and Implication
Though various models of consumer behavior on the adoption and use of ICT appli-
cations are developed from 1960 till 2012, there was no systematic literature review
was carried out on the UTAUT2 model developed in 2012, which was considered
superior over other previous models. Two such systematic literature was carried out
so far was on TAM [11] and UTAUT [8], which formed the basis for the present
research work.
Acceptance of the UTAUT2 model and its use in academic research was very
limited during 2012–2014 but got acceptance and witnessed an exponential increase
taken place during 2015–2020 as the model is having superior predictive power. The
Asia Pacific is the top region and India and Malaysia are the two countries where
researchers applied the UTAUT2 model for assessing the consumer behavior model
when it comes to ICT applications in various business fields, most prominent ones
being education, general purpose and health-related ICT applications. This is an
indication that business entities are trying to capture the markets in the Asia Pacific
region.
By using weight analysis, it was found that only four out of ten pairs are said
to be best predictors, namely, PE-BI, H-BI, H-UB and BI-UB. The majority of
the studies also gave a similar variance explained (R2 ) with were in line with the
findings of Venkatesh [10]. Gender, age and experience were the most commonly
used moderators. The most significant external constructs used by the researchers
were found to be self-efficacy, trust, risk, security, and innovativeness. Developers of
ICT application must consider the attributes, both from the original UTAUT2 model
as well as those which are added subsequently, to improve the existing application
and also incorporate these attributes in new upcoming applications for the purpose
of ensuring that existing and potential future consumers are adopting and using for
all future purposes. Trust is found to be one of the best external construct, and it is
related with other external constructs, namely level of risk, level of privacy and the
level security associated with the ICT application.
5 Limitations and Future Research Direction
Every research is bound to have some limitation and the present study is not an excep-
tion. First of all, efforts are taken to cover all the relevant content from the existing
research works carried out during 2012–2020, but there is always the possibility
of losing relevant content, hence scope for future research work. Second, articles
reviewed are only the research papers and conference proceedings whereas disser-
tation, thesis, unpublished work and other work being not taken into consideration,
thus providing further scope for the research. Third, the present study is not taken
into consideration the limitations faced in the studies of Lee and William [8, 11], thus
one can make a detailed study of the constraints faced by those researchers. Fourthly,
many researchers might have focused and used UTAUT2 but lack the keywords in
title, hence become scope for further research. Lastly, all studies using UTAUT2
construct were studied in general, thus, once can study the relationship between each
construct under each ICT application industry which influences the behavior of user
toward technology adoption.
References
1. J.V.H. Carr, Technology adoption and diffusion. http://www.icyte.com/system/snapshots/fs1/

9/a/5/0/9a50b695f1be57ce369534ac73785801745a8180/index.html, last accessed 2020/05/07
2. M. Fishbein, I. Ajzen, Belief, Attitude, Intention and Behavior: An Introduction to Theory and
Research (Addison-Wesle, Reading, MA, 1975)
3. A. Bamdura, Self-efficacy: towards a unifying theory of behavioral change. Psychol. Rev.
84(2), 191–215 (1977)
4. I. Ajzen, The theory of planned behavior. Organ. Behav. Hum. Decis. Process. 50(2), 179–211
(1991)
5. R.L. Thompson, C.A. Higgins, Personal computing: towards a conceptual model of utilization.
MIS Q. 15(1), 125–143 (1991)
6. V. Venkatesh, M.G. Morris, G.B. Davis, F.D. Davis, User acceptance of Information technology:
toward a unified view. MIS Q. 27(3), 425–478 (2003)
7. R. Sharma, R. Mishra, A review of evolution of theories and model of technology adoption.
Indore Manage. J. 6(2), 17–29 (2014)
8. M.D. Williams, N.P. Rana, Y.K. Dwivedi, The unified theory of acceptance and use of
technology (UTAUT): a literature review. J. Enterp. Inf. Manage. 28(3), 443–488 (2015)
9. V. Venkatesh, S.A. Brown, L.M. Maruping, H. Bala, Predicting different conceptualizations of
system use: the competing roles of behavioral intention, facilitating conditions, and behavioral
expectation. MIS Q. 32(3), 483–502 (2008)
10. V. Venkatesh, J.Y.L. Thong, X. Xu, Consumer acceptance and use of information technology:
extending the unified theory of acceptance and use of technology. MIS Q. 36(1), 157–178
(2012)
11. Y. Lee, K.A. Kozar, K.R.T. Larsen, The technology acceptance model: past, present and future.
Commun. Assoc. Inf. Syst. 12(1), 752–780 (2003)
12. V. Kumar, Z. Rahman, A.A. Kazmi, Sustainability marketing strategy: an analysis of recent
literature. Glob. Bus. Rev. 14(4), 601–625 (2013)
13. A. Jeyaraj, J.W. Rottman, M.C. Lacity, A review of the predictors, linkages, and biases in IT
innovation adoption research. J. Inf. Technol. 21(1), 1–23 (2006)
14. Y.K. Dwivedi, M.A. Shareef, A.C. Simintiras, B. Lal, V. Weerakkody, A generalised adoption
model for services: a cross-country comparison of mobile health (m-health). Gov. Inf. Q. 33,
174–187 (2016)
15. A.A. Alawan, Investigating the impact of social media advertising features on customer
purchase intention. Int. J. Inf. Manage. 42, 65–77 (2018)
16. P. Rita, T. Oliveira, A. Estorninho, S. Moro, Mobile services adoption in a hospitality consumer
context. Int. J. Cult. Tour. Hosp. Res. 12(1), 143–158 (2018)
17. H.R. Marriott, M.D. Williams, Developing a theoretical model to examine consumer acceptance
behavior of mobile shopping, in Social Media: The Good, the Bad, and the Ugly 2016, vol.
9844, ed. by Y. Dwivedi et al. (LNCS, Springer Cham, 2016), pp. 261–266
18. N.R. Indrawati, Predicting acceptance and use behavior of consumers towards indihome
services by using extended UTAUT model (a case study in Bandung). Int. J. Sci. Res. 6(5),
1860–1865 (2017)
19. A. Tarhini, R. Masadeh, K.A. Al-Busaidi, A.B. Mohammed, M. Maqableh, Factors influencing
students’ adoption of e-learning: a structural equation modeling approach. J. Int. Educ. Bus.
10(2), 164–182 (2017)
20. K. Tamilmani, N.P. Rana, N. Prakasam, Y.K. Dwivedi, The battle of brain versus heart: a
literature review and meta-analysis of hedonic motivation use in UTAUT2. Int. J. Inf. Manage.
46, 222–235 (2019)
21. A.M. Baabdullah, A.A. Alalwan, N.P. Rana, H. Kizgin, P. Patil, Consumer use of mobile
banking (M-Banking) in Saudi Arabia: towards an integrated model. Int. J. Inf. Manage. 44,
38–52 (2019)
Secure and Energy-Efficient Remote
Monitoring Technique (SERMT)
for Smart Grid
Sohini Roy and Arunabha Sen
Abstract Monitoring and automation of the critical infrastructures like the power
grid are improvised by the support of an efficient and secure communication network.
Due to the low-cost, low-power profile, dynamic nature, improved accuracy and scal-
ability, wireless sensor networks (WSN) became an attractive choice for the Infor-
mation and Communication Technology (ICT) system of the smart grid. However,
the energy efficiency and security of WSN depend highly on the network design
and routing scheme. In this paper, a WSN-based secure and energy-efficient remote
monitoring technique (SERMT) is proposed by demonstrating a WSN-based ICT
network model for pervasive monitoring of the generation and transmission part of
the power network in a smart grid system. The performance of the proposed network
model designed for a smart grid of IEEE 118-Bus system coupled with the secure
routing technique is tested during cyber-attacks by means of NS2, and the simulation
results indicate that it performs better than existing smart grid monitoring methods
like Lo-ADI (Dhunna and Al-Anbagi in IEEE Sens J 19(13), 2019 [1]) with respect
to packet drop count and throughput.
Keywords Information and Communication Technology · Wireless sensor

network · Cyber-attacks · Remote monitoring · Smart sensor nodes · Smart grid
1 Introduction
A sustainable lifestyle for human race is reliant upon critical infrastructures

like power network, communication network, water supply system, transportation
system, etc. Therefore, pervasive monitoring [2] and control over such critical
systems are very important. It is to be noted that only an improved Information
S. Roy (B) · A. Sen

Arizona State University, Tempe, AZ 85281, USA
e-mail: sroy39@asu.edu
A. Sen
e-mail: asen@asu.edu
330 S. Roy and A. Sen
and Communication Technology (ICT) can transform every such system into a smart
system and thus can form a smart city. Smart grids are an inherent part of a smart city,
and it is obtained by incorporating features like full-duplex communication between
the ICT entities, automated metering in the smart homes, power distribution automa-
tion and above all intelligent decision making by means of pervasive monitoring
of the power system and securing its stability. Therefore, it is beyond any question
that the ICT network of a smart grid must be accurate, scalable and secure enough to
instantly identify any kind of abnormal behavior in the entities of the power network,
securely communicate that information to the control center (CC) and thereby help
in taking necessary and timely action to ensure uninterrupted power supply.
Although, designing of a robust ICT network for smart grid has become a boiling
topic of research, the best-suited ICT design for a smart grid is still not very clear.
In [3], only a crude idea of the design of a joint power communication network
is given using a test system consisting of 14 buses. While the authors of [4] have
come up with a realistic design of the ICT for smart grid by taking help from a
power utility in the U.S. Southwest, their ICT system relies completely on wired
channels that either use SONET-over-Ethernet or Ethernet-over-dense wavelength
division multiplexing. A completely wired ICT system is neither cost effective nor
energy saving. Every communication entity in [4] draws power, and thus a huge
amount of power is devoted for monitoring the power network itself. Moreover,
addition/isolation of ICT entities for hardening purpose or fault tolerance/during a
failure or a security threat is extremely difficult and costly in a wired system.
Smart sensor nodes like phasor measurement units (PMUs) are popular in remote
monitoring system for smart grid. Power generation and transmission, power quality,
equipment fitness, load capacity of equipment and load balancing in the grid can
also be monitored by data sensing techniques. As a result, wireless sensor networks
(WSNs) are currently acquiring attention of the researchers for the purpose of
designing the ICT network for a smart grid system. WSNs are comprised of low-
powered sensor nodes with easy installation process, lesser maintenance requirement,
low installation cost, low power profile, high accuracy and scalability. All these have
convinced the researchers that WSNs are a very good choice for the designing of the
ICT of a smart grid. Therefore, in this paper, a new design of the ICT network for a
smart grid is proposed using WSN.
Despite all these advantages, the most common drawback of a sensor node is
that it is battery powered and battery replacement is difficult. As a result, energy
conservation becomes important. In secure and energy-efficient remote monitoring
technique (SERMT), energy efficiency is obtained by energy-aware routing and use
of more expensive rechargeable energy harvesting sensor nodes (EHSNs) for MU
and PMU data transfer, respectively. Also, the sensor nodes as well as the wire-
less channels are vulnerable to common cyber-attacks [5] like sinkhole, wormhole,
Sybil attack, packet eavesdropping, flooding attack and most importantly false data
injection attack by means of node compromise. SERMT aims at securing the data
transfer together with energy conservation by means of lightweight security protocols
Secure and Energy-Efficient Remote Monitoring Technique … 331
used in [6] like elliptic curve public key cryptography (ECC), elliptic curve Diffie–
Helman key exchange scheme (ECDH), nested hashed message authentication codes
(NH-MAC) and RC5 symmetric cypher.
The rest of the paper is structured as follows. Section 2 gives an overview of
the ICT network setup phase for SERMT. A realistic risk model for the SERMT
network design is given in Sect. 3 of this paper. Mitigation of the threats discussed
in Sect. 3 by adopting a secure and energy-efficient routing technique is explained
in Sect. 4. Section 5 does performance analysis of the proposed scheme SERMT by
means of comparing the simulation results with an existing secure remote monitoring
technique for power grid named as Lo-ADI [1]. Section 6 concludes the paper and
discusses the scope for future works.
2 Overview of the ICT Network Setup Phase for SERMT
In order to provide a reliable remote monitoring for the smart grid, a generic ICT
network design is proposed by SERMT that can be applied on any given power
network. In this section, the network setup phase for SERMT is divided into four
steps. Each step is illustrated with the help of IEEE 14 bus system as the power
network.
In the first step, a given power grid is divided into several substations. The buses
connected by transformers are placed in one substation with the assumption that
transformers remain within a substation only [7]. This process is repeated until no
transformers are left. Then each bus is placed in a unique substation. In Fig. 1, the
IEEE 14 bus system is divided into 11 substations denoted by Si where i is the
substation ID.
After the substation division, the distance between all pairs of substations
(Si & S j ) is calculated [4]. Now, starting from a substation Si with the maximum
connectivity among the substations at the network borders, all substations which are
within a given distance of it are marked as substations of a common monitoring region
Rx . Then the next substation which is the closest to Si but beyond the given distance
and which is not yet placed in a monitoring region is selected and the same process is
repeated. This process is continued till every substation is placed within a monitoring
region. For example, in the IEEE 14 bus system, substations S3 , S4 , S5 , S6 , S7 , S8
and S10 are at the borders of the smart grid area. Among them, S4 has the maximum
connectivity; therefore, the region division starts from it. Figure 1 shows that after
this process is completed, the smart grid network for IEEE 14 bus system is divided
into four regions.
SERMT considers two different types of smart sensors for monitoring the power
network entities. The first type is the measuring unit (MU)-based smart sensors [8].
The MU-based smart sensors have a sensing unit consisting of sensors like current
transformer (CT) and voltage transformer (VT) and a merging unit for performing
signal amplification, conditioning and analog–digital conversion. There is an internal
Fig. 1 Region division and sensor node placement in a smart grid of IEEE 14 bus system
clock in the merging unit that can timestamp every sample and thereby time synchro-
nize them with an external UTC source. The MU-based sensor has a network commu-
nication module which helps it to communicate with intelligent electronic devices
(IEDs) or applications. The second type of sensors used for SERMT is the PMU-
based smart sensors. The sensing module of a PMU has sensors like voltage and
current sensor which provide analog inputs for the PMU. Current and voltage phases
are sampled at a rate of 48 samples per cycle in a PMU using a common time
source for synchronization. It also has an internal clock and a processing module
responsible for signal conditioning, timestamped analog–digital signal conversion
and computing voltage and current phasors to produce a synchronized phasor and
frequency.
In this step, MU-based sensors are placed at every bus in the power network.
PMU-based sensors are placed at some of the buses using an optimal PMU placement
algorithm [7]. Low-cost, non-rechargeable battery-enabled sensor nodes denoted by
N j with ID j are randomly dispersed across the network area. These N j s can carry
the MU-based sensor data to the CCs in a power-efficient manner. A phasor data
concentrator (PDC), responsible for receiving accumulating PMU data from multiple
PMUs, is placed at each region of the smart grid and few EHSNs denoted by E S j
with j as the ID are randomly deployed across the smart grid region. The idea behind
the deployment of the two kinds of sensor nodes is that the cheaper N j s will follow
an event-driven hierarchical routing approach to send the MU data from the MU-
based sensor nodes to the CCs; and the E S j s will always be active to accumulate
synchrophasor data from the substations of each region and send the data to the local
PDCs and finally to the CCs. Due to the high volume of PMU data transfer from each
substation having a PMU, the sensor nodes carrying them should always be active.
In Fig. 1, the MU-based sensors are denoted by Mi where i is the ID of the sensor,
and it is same as the bus number it is placed at. The PMUs are represented by Pi ,
and they are numbered in the order they are placed. It is assumed that N j s and E S j s
are deployed region wise and are numbered in the order of their deployment.
SERMT assumes that there is a main CC and a backup CC. So, even if the
main CC is damaged, the backup CC can efficiently monitor the network state. The
connectivity of each substation is calculated based on the number of transmission
lines connecting that substation with other substations. The substation having the
highest connectivity is selected as the main CC and that having the second highest
connectivity is selected as the backup CC. After the CCs are selected, all substations
are equipped with a router acting as a gateway for the substation servers, and the
servers in each CC are access points for the operator. The CC gateways can wirelessly
receive data from the sensors and PDCs and send those data to server via a wired
LAN connection in control centers. In Fig. 1, S1 is the main CC and S2 is the backup
CC.
3 Risk Model and Assumptions for the SERMT Network

Setup
The WSN nodes are vulnerable to several attacks that are explored by researchers
over the ages. In SERMT, some of the common attacks on WSN are considered. The
threat model for SERMT predicts that a compromised entity of the ICT can pretend
as a CC gateway and congest the network with bogus communications to launch
a flooding attack which can result in denial of service (DoS) by the ICT entities.
Another common attack of the WSN is the Sybil attack [5] where a legitimate node
of the network gets compromised and deceits the other nodes by providing multiple
fake location information of itself. In SERMT network setup, any ICT entity except
the substation entities can impose a Sybil attack and oblige other entities to forward
data to an entity that does not exist in that location resulting in increased packet
drop. In a node compromise attack, the attacker physically gets hold of the node
and reprograms it to launch attacks like wormhole attack and packet eavesdropping
attack [5]. These attacks can also be launched by an outsider’s node deployed in
the network. The Sinkhole attack can be launched by attackers with high-quality
equipment, and they can inform the ICT entities about a high-quality entity in their
sensing zone, so that all the other entities select it to forward the data to the CC. This
node gathers all the data the CCs do not receive any information about the system
state from the network.
In order to provide security to the ICT infrastructure for the smart grid, the
following assumptions are made in the proposed work:
1. The substation equipments like the server and the gateway are trustworthy, and
it is impossible for an attacker to compromise those components. The servers
authenticate themselves to the sensors or PDCs to avoid flooding attack [5].
2. Each entity of the ICT network is provided with a global key GBK which an
attacker cannot get hold of even if the entity is compromised.
3. A unique set of elliptic curves is stored in the memory of each ICT entity for the
purpose of ECC and ECDH protocols. Also, in order to achieve those mecha-
nisms, it is assumed that any pair of entities in the network agrees upon all the
domain parameters of the elliptic curves stored in them.
4. The two CC servers generate two separate one-way hash chains using SHA-1
hash function [6]. The starting key for each chain is randomly selected by each
of the servers. All the ICT entities are provided with same hash functions and
the last keys of the two chains so that they can authenticate the control messages
from the control center servers and avoid a Sinkhole attack.
5. Each ICT entity in the network is aware of its own ID, location information,
region ID and ID of the servers.
6. All the PDCs can communicate with other PDCs in the neighboring regions.
4 Secure and Energy-Efficient Routing Scheme of SERMT
The goal of the ICT network for a smart grid is to securely transmit the sensed data
from the sensors to the control centers and help in remote monitoring of the power
grid. In order to achieve this with the help of a WSN-based ICT network, a secure
and energy-efficient routing technique should be followed. In this section, the secure
routing scheme of SERMT is divided into four modules.
4.1 Module 1: Initial Trust Value Estimation of the ICT

Entities
In the first module of SERMT, the main control center server (Server1 ) sends
a number of test messages to the sensor nodes and PDC in its own region
to calculate their trust value (TVi ). The message format is Server1 → ∗ :
MSG||K x ||HMAC(GBK; MSG||Server1 ID||K x ) where * denotes a set of ICT enti-
ties, MSG is the test message, K x is a random key from the sequence of keys generated
by the hash function, and this key is never used again. A hash-based message authen-
tication code (HMAC) is generated over the MSG, server ID and the key using the
shared global key GBK and appended with the message, so that any non-legitimate
node which does not have the GBK and the server ID cannot separate the HMAC
from the original message and therefore cannot overhear the message from the server
to the other entities. The TVi is calculated by the server for each entity i using Eq. 1.
MSGdelivered
TVi = ∗ 100 (1)
MSGsent
Entities having a TVi of more than 40% are selected as trusted nodes, and any legit-
imate node having a lower TVi is added to the threat list of the server. Now, the node
X i with the highest TVi in that region is selected by the CC server to forward a request
message to the nodes closest to it in the adjacent regions. The format of the request
message is Server1 → X i : RQM||||K x ||HMAC(GBK; MSG||Server1 I D||K x )
where RQM is the request to calculate TVi , K x is a new random key from the
key chain, and HMAC is generated and appended with the message. First, the TVi of
the nodes of adjacent regions closest to X i are calculated by X i in the same manner.
If those nodes have a TVi greater than 40%, then the RQM is forwarded to them.
These nodes now calculate the TVi of the nodes in their respective regions and send a
list of blocked node IDs having low TVi to the CC server via X i . The message format
is ei → Server1 : Blocked_List||ei I D||HMAC(Blocked_List||ei I D). The Server1
updates the threat list using this data. This process is repeated till the threat nodes of
each region are identified by Server1 . Server1 shares the threat list with Server2 via
a shortest path consisting of nodes in the trusted list. After both the servers get the
list of trusted nodes in the network, the trust value of each node is shared with the
substation gateways by Server1 , via a shortest path to the substation consisting of
the most trusted nodes. This whole process is repeated at a regular time interval by
alternate CC servers. Each time the new trust valuesare assigned to the ICT entities,
both the servers compute a public–private key pair UServeri , vServeri and share their
public key with all the entities in the network by broadcasting.
4.2 Module 2: Data Sensing by the Smart Sensor Nodes

and Data Forwarding by Substation Gateways
The next phase of SERMT is the data sensing phase. In this phase, the smart
sensors collect data from the power entities and forward that to the substation
gateways. The substation gateway is assumed to be trusted. The CC gateways
forward the data directly to the connected servers, but they still need to send
data to the other CC. The gateways work differently for sending MU-based and
PMU-based sensor data to the CCs. In order to send the MU-based sensor data,
the gateways send a forward request message to each jth node (N j ) of that
region at one-hop distance from it. The format of the request is GWi → N j :
forw_ RQM||GWi I D||HMAC(GBK; forw_ RQM||GWi ID). Each N j receiving
the forw_RQM from the gateway sends back an acknowledgement
(ACK) appended
with information like their remaining
battery power BP j and its connectivity with
othernodes in the same region C j to the
sender in the format: N j → G i :
ACKBP j C j ||HMAC GBK; ACKBP j C j . In case a node receives forw_RQM
from multiple gateways in that region, it selects the one closest to it. The substation
gateways calculate a data forwarding score DF j for all the nodes sending back ACK
to it using Eq. 2.
DF j = BP j ∗ T V j ∗ C j (2)
Now, each gateway GWi generates a key pair having a public and a private key
UGWi , vGWi using ECC and forward the public key UGWi to the jth sensor node
having the highest DF j calculated by itself using the message format: GWi →
N j : UGWi ||GWi I D||HMAC(GBK; UGWi ||GWi I D). The Nj receiving the public
key from the GWi also generates a private–public key pair v N j , U N j using ECC
and sends back the public key and its ID in the similar format to GWi . GWi now
computes a point in the elliptic curve (xk , yk ) = vGWi .U N j and N j computes a point
(xk , yk ) = v N j .UGWi where xk , the x coordinate of the computed point becomes their
shared secret. This is ECDH key exchange scheme where the shared secret calculated
by both parties is same—vGWi .U N j = vGWi .v N j .G = v N j .vGWi .G = v N j .UGWi ,
where G is the generator of the elliptic curve. This shared secret xk is the temporary
key for exchanging encrypted data between the GWi and N j . GWi now encrypts the
aggregated MU data from the buses of that substation using RC5 symmetric cypher
with xk and forwards that to N j by using NH-MAC in the format: GWi → N j :
EMD||GWi ID||HMAC(GBK; HMAC(xk ; EMD||GWi ID)). The gateways keep on
forwarding the encrypted MU data to the same node until the servers initiate a new
round of module 1.
At the same time, the PMUs also sense data from the power entities and send that
to the corresponding substation gateway. Each PMU containing substation gateway
now forwards forw_RQM to the EHSNs of that region within its range. In the similar
way, the ES j s receiving the forw_RQM sends back ACK to the gateway with only the
connectivity information. Now, the gateway recalculates the TV j for each ES j like
module 1 and selects the node with the highest current TV j . The same key exchange
scheme is established with the most trusted ES j as it was done with the N j s and the
RC5 encrypted PMU data is forwarded to E S j in the similar format with NH-MAC.
GWi s recalculate the TV j of the ES j s at regular intervals to repeat this process of
key and data exchange and do not wait for the servers to estimate TV j of the ES j s.
However, if a server reinitiates module 1, all the substation gateways wait till the
process is complete and they are updated with the new trust values of all the nodes
in their region. The CC gateways also use the same process to send data to the other
CC.
4.3 Module 3: Data Forwarding to the Control Centers
In this module, two separate routing techniques are followed by the N j s and ES j s
to forward data to the CCs. N j s receiving data from the GWi s first decrypt them
using the same temporary key xk . Then, a join request
(join_RQM) is broadcasted
in the format: N j → ∗ : join_ RQM N j I D HMAC(GBK; join_ RQM||N j ID)
to all other N s in that region by that N j . Malicious nodes present in the
network receiving the join_RQM cannot decode the HMAC as they do not have
the GBK and thereby they cannot join any legitimate data-carrying N j . Each
N j receiving ACK from other N s provides a cluster ID to each of the N s in

the format RegionID||Timestamp_for_ current_T Vi ||List_ of_ SubstationIDs. The
server assigned timestamp value for current TVi is same for all the nodes in a
common region, and the list of substation IDs denotes the data carried from which of
the substations in that cluster. Therefore, all the nodes getting the common cluster ID
are considered as nodes of a cluster. Multiple clusters can be formed in each region
of the smart grid.
After the cluster formation in every region,
a cluster head (CH) is selected for
each cluster using a candidate value CV j calculated using Eq. 3. by each node j in
the cluster and shared with all other nodes in the cluster.
CV j = BP j ∗ TV j ∗ Cn j (3)
In Eq. 3., BP j is the remaining battery power of the node, TV j is the current
trust value of the node, and Cn j is the connectivity of that node with the nodes of
other adjacent regions. The Cn j of a node is determined based on the number of
nodes in other regions that are within the sensing radius of the node j. Now, the CH
generates a public–private key pair like GWi and share the public key with each of the
nodes in the cluster. The cluster member nodes (CMNs) also generate the key pair,
send their public key to the CH and compute their own xk using ECDH. Each CMN
now forwards its data to the CH by encrypting the data with the xk using RC5 and
generating NH-MAC just like each GWi did. The CHs after receiving the data from
the CMNs decrypt them with the xk for that CMN. Now each of the CHs aggregates
the
dataobtained from the member nodes and encrypts that with the public key
UServer1 of the main CC server using ECC-based public key cryptography. In case,
the main CC fails, the network is informed about that and all data is forwarded to
the backup CC.
Now, a CH of the main CC region directly forwards the encrypted-aggregated data
to the main CC gateway. Other CHs use a modified version of Dijkstra’s shortest path
algorithm to send
the aggregated data to the main CC gateway. In this algorithm, the
weight values Wi j of the link between two nodes i and j are determined by Eq. 4.
Wi j = Di j /(BP j ∗ TV j ) (4)
In Eq. 4., Di j stands for the distance between the sender i and the receiver j,
B P j is the remaining battery, and TV j is the trust value of the node j. The same
route is followed by the nodes for sending data to the main CC until new nodes
are selected by GWi s for data forwarding. The main CC gateway generates the
HMAC over the received message to check its authenticity, and if a match is found,
the data is forwarded to Server1 . Server1 periodically sends backup to the Server2
via the shortest path from Server1 to Server2 , determined using the same modified
Dijkstra’s algorithm.
In order to forward the PMU-based sensor data to the CCs, each data-carrying ES j
node checks if the PDC of that region is adjacent to it. Each ES j is updated with the
location information and the current TV j of other ES j s and the PDC in that region
by the GWi s of that region. If a PDC is adjacent to an ES j , then the ES j first decrypts
the data and generates a public–private key pair to share the public key with the PDC.
The PDC also does the same and share back its public key with the ES j to complete
the ECDH key exchange. The node ES j now forward the encrypted NH-MAC-ed
data to it. ES j s which are not near a PDC, select the shortest path consisting of the
most trusted ESs from itself to the PDC in that region to send their public key, and
the same route is followed by the PDC to send back its own public key to the ES j .
Now, the encrypted data is forwarded through the same path to the PDC. The PDC of
each region acts in the similar way as a CH. It decrypts the data from each ES j using
the unique xk for that node and aggregates the whole data. A shortest path consisting
of PDCs only from each PDC of the network to both the control center gateways is
calculated only once and is never repeated until a PDC of a region is replaced or a
PDC fails. In case, the TVi of a PDC drops below the threshold, the Server1 adds it
to the blocked list and selects the most trusted ES j node to perform the function of
a PDC, until the actual PDC is replaced by a new one or its fault is diagnosed and
recovered by the operator. Now, the aggregated data is encrypted using the public
key of the Server1 by each PDC and forwarded to the main CC gateway. The same
process is repeated for Server2 . Both the CC gateways authenticate the received data
and forward it to their connected servers for evaluation.
The energy consumed by both the N nodes and ES nodes in transmitting a message
(E T ) is the summation of energy consumed by their power amplifier (E A ), and it is
a function of distance from which the message is sent, baseband DSP circuit (E B )
and front-end circuit (E F ). The energy consumed in receiving a message (E R ) is the
summation of E B , E F and the energy consumed by low-noise amplifier E L .
5 Simulation Results
In this section, a smart grid network of IEEE 118 bus system is considered. The
total network region is divided into eight regions, and the power grid is divided into
107 substations. Substation 61 is selected as the main CC, and it consists of three
buses—68, 69 and 116. Substation 16, consisting of buses—17 and 30, is selected
as the backup CC. The main CC is in region 1, and the backup CC is in region 2 of
the smart grid. In order to analyze the performance of SERMT in this network setup,
a total of 500 N type nodes, 300 ES type nodes and 8 PDCs (high-quality EHSNs
with longer communication range) are deployed in the network area and simulation
is performed using NS2.29 network simulator. The simulation results are compared
with the Lo-ADI [1] technique that incorporates a secure token exchange mechanism
among the WSN nodes. The simulation parameters are given in Table 1.
In Fig. 2, the average battery power consumed per hour by all types of sensor nodes
is shown for different attack intervals in both Lo-ADI and SERMT. It is observed
that the average battery power consumption of SERMT is a little higher than that of
Lo-ADI, and this battery power consumption does not decrease if the attack interval
is 1 s as in Lo-ADI due to connection revoke and packet drops, rather its battery
Table 1 Parameter list for

Parameter Description
simulation
Operating system Fedora
Simulator NS2.29
Compiler TCL
Rechargeable battery for EHSNs NiMH
Battery capacity of EHSNs 2000 (mAh)
Initial battery power for all nodes 150 (mAh)
Fig. 2 Avg. BP consumed

versus attack interval
power consumption increases if the attack interval is less as SERMT immediately

adds the malicious node into its threat list and resends the packet to some other
trusted node. Moreover, due to the presence of ESs, the average residual energy in
the network will always be more than that of Lo-ADI. In Fig. 3, it is proved that even
though SERMT consumes more power than that of Lo-ADI because of using several
power-consuming security mechanism computations and key exchange protocols, it
wins in terms of offering security to the smart grid. Figure 3 shows percentage of
packets dropped in presence of different number of malicious nodes in the network,
and it is observed that in SERMT, the packet drop percentage is much lesser than
that of Lo-ADI even when there are 35 malicious nodes in the network.
Fig. 3 Number of malicious

nodes versus packet drop
percentage
6 Conclusion and Future Works
The region-based remote monitoring adopted by SERMT helps in easy identification

of a failure in the power grid or an attack in the communication network of the smart
grid. Security of sensor data is of utmost importance in the smart grid as any alteration
of the data can lead to wrong decision by the operator, and this might badly harm
the smart grid. Data privacy is obtained in SERMT by the encryption/decryption
mechanisms, data integrity is verified by the hashed message authentication, and
authentication of data is done by the one-way hash chain; therefore, all the data
security measures are addressed by SERMT. One drawback of SERMT is that it
cannot deliver the sensor data from the sensors to the control centers in a very short
time due to the security measures it follows and this can be explored in future research
how the time delay be reduced in a secure communication technique for smart grid.
Designing a threat model with attacks on MUs, PMUs, gateways or servers and
analyzing the effect of cyber-attacks on power grid can be another direction of future
work.
References
1. G.S. Dhunna, I. Al-Anbagi, A low power WSNs attack detection and isolation mechanism for
critical smart grid applications. IEEE Sens. J. 19(13) (2019)
2. K. Basu, M. Padhee, S. Roy, A. Pal, A. Sen, M. Rhodes, B. Keel, Health monitoring of crit-
ical power system equipments using identifying codes, in International Conference on Critical
Information Infrastructures Security (CRITIS 2018), Kaunas (2018), pp. 29–41
3. W. Zhu, J.V. Milanović, Cyber-physical system failure analysis based on Complex Network
theory, in IEEE EUROCON 2017—17th International Conference on Smart Technologies, Ohrid
(2017), pp. 571–575
4. S. Roy, H. Chandrasekharan, A. Pal, A. Sen, A new model to analyze power and communica-
tion system intra-and-inter dependencies, in IEEE Conference on Technology for Sustainability
(SUSTECH 2020) (2020), Santa Ana
5. C. Karlof, D. Wagner, Secure routing in wireless sensor networks: attacks and countermeasures.
Ad Hoc Netw. Elsevier J. 1, 293–315 (2003)
6. S. Roy, Secure cluster based routing scheme (SCBRS) for Wireless Sensor Networks, in 3rd
International Symposium On Security in Computing and Communications (SSCC), Kochi (2015),
pp. 368–380
7. A. Pal, C. Mishra, A.K.S. Vullikanti, S.S. Ravi, General optimal substation coverage algorithm
for phasor measurement unit placement in practical systems. IET Gener. Transm. Distrib. 11(2),
347–353 (2017)
8. E.Y. Song, G.J. FitzPatrick, K.B. Lee, Smart sensors and standard-based interoperability in smart
grids. IEEE Sens. J. 17(23) (2017)
Balanced Accuracy of Collaborative
Recommender System
Akanksha Bansal Chopra and Veer Sain Dixit
Abstract Recommender systems (RSs) have emerged as a vital tool to handle the
overloaded and huge information very effectively, caused by exceptional growth of
means on Web. Collaborative recommender system (CRS) is one of the types of RSs
that are being used tremendously by majority of e-commerce industry in present
scenario. CRS help the users to opt for an item of their own interest easily from
largely available options on Internet. CRS use the ratings given by similar users to
make predictions to the interested users. However, these ratings are not always true
and authentic. These un-authentic ratings are the biased ratings that lead to various
attacks, such as Push and Nuke Attacks (P-N Attacks). In this work, P-N Attacks
are, first, injected on available data and then the effect of these P-N Attacks are
explored on the performance of CRS by using balanced accuracy (BAC) and other
state-of-the-art measure metrics.
Keywords Recommender systems · Collaborative recommender systems · Push

Attacks · Nuke Attacks · Mean Absolute Error · Balanced accuracy
1 Introduction
RSs are primarily being used by various e-commerce websites. Amazon,1 Youtube,2
Spotify,3 Netfilx4 are some of the famous commercial websites that use RS approach.
1 http://www.amazon.com.
2 http://www.youtube.com.
3 http://www.spotify.com.
4 http://www.netflix.com.
A. B. Chopra (B)
Department of Computer Science, SPM College, University of Delhi, New Delhi, India
e-mail: akankshabansal.asm@gmail.com
V. S. Dixit
Department of Computer Science, ARSD College, University of Delhi, New Delhi, India
e-mail: veersaindixit@rediffmail.com
342 A. B. Chopra and V. S. Dixit
RSs are decision support systems that make the task easy and convenient to the users
for selecting an item from a wide range of available items by filtering the information
on Web [1]. There are two main approaches of RSs—Collaborative recommender
system (CRS) [2] and content-based recommender system (CBRS) [3]. In CRS,
ratings for an item are collected on the rating scale from users. A user-item matrix is
then created for the collected ratings [4, 5] and similar users are identified by applying
any available similarity coefficient. Pearson correlation coefficient (PCC) is one the
popular similarity coefficients that can be used for finding similar users. CBRS work
on contextual information based on item’s features and user profile both. Hybrid
recommender system (HRS) is another approach that works on collective principle
of CRS and CBRS approaches [6].
Data is most vulnerable part in present era. So we need to secure it from different
attacks. When users are rating an item, they are expected not to disclose their private
information. The private information available on network is helpful for malicious
users. Various attacks can be injected by these malicious users resulting in inappro-
priate recommendation list. The attacks, thus degrade the performance and accu-
racy of RS. Shilling attack and Sybil attack are two common types of attacks that
lead to breach in privacy of recommender system [7]. In Sybil attack, multiple fake
identities are created by malicious user and false recommendations are injected
resulting in inappropriate recommendation list. Shilling attack is a type of attack
where misleading recommendations are injected and recommendation list is altered
[8–10]. Push Attacks and Nuke Attacks (P-N Attacks) are two types of shilling
attacks. In Push Attacks (P-Attacks), higher ratings (HRs) are injected by malicious
users to increase the recommendations for an item and alter the actual and genuine
recommender list. On contrary, in Nuke Attacks (N-Attacks), lower ratings (LRs) are
injected by malicious users to decrease the recommendations of an item. In present
commercial industry, P-N Attacks are injected to either escalate the reputation of
own firm (by injecting P-Attacks) amongst its users or to destroy the reputation of
rival firm (by injecting N-Attacks).
In conventional RSs, user-item ratings are collected to generate recommendations.
No emphasis was given to identify any biased ratings (HRs, LRs), if injected to the
CRS. As in CRS, ratings highly contribute in prediction of items, it is therefore neces-
sary to identify and classify the biased ratings. Actual and authentic ratings increase
the system’s efficiency, whereas biased ratings (HRs, LRs) affect the system’s accu-
racy adversely and lead to false recommendations to its users. Once the user-item
ratings are entered, it is expected the system should predict true recommendations
for an item.
The problem here is defined as to analyse the performance of CRS, when P-N
Attacks are present. P-N Attacks alter the ratings of items by injecting HRs and LRs,
to the system which leads in false recommendations of items to the users. The effect
of P-N Attacks on performance of CRS is analysed by evaluating various measure
metrics, such as precision, recall, F-measure, accuracy and balanced accuracy, at
different threshold values.
This Tpaper is structured in five sections. In Sect. 2, the related works done by
various researchers are discussed. Section 3 explains various measure metrics that are
Balanced Accuracy of Collaborative Recommender System 343
used in paper. The datasets and experimental evaluation results along with validation
of evaluations are shown in Sect. 4. Finally, Sect. 5 discusses the conclusion and
outlook for future work.
2 Related Works
Authors of paper [11] reviewed various applications of recommender systems

(RSs) in e-commerce, e-shopping, etc., and related recommendation techniques
are discussed. They also examined dimensions such as collaborative filtering and
BizSeeker that are used to develop a recommender system. Paper [12] proposed an
algorithm named as, TasteMiner that identifies partial users’ interest for efficient use
of neighbourhood space. In another paper, [13] used Pareto dominance method for
pre-filtering of less promising users from k-neighbour selection process keeping on
promising users for recommendations. Challenges such as scarcity of data, scala-
bility, shilling attacks, grey sheep, etc., with their solutions are discussed by other
authors in their research work [14].
In research paper [15], authors have surveyed the privacy issues related to recom-
mender systems and have proposed algorithms to improve performance of recom-
mender system. The authors emphasized to provide enhanced solution to privacy
issues. In another paper, authors have used trust-based collaborative filtering algo-
rithm to recommend tourism-related features such as hotel, restaurant, etc., at one
single portal accurately [16]. Security risks in Web applications have been analysed
by the authors of research work [17]. The authors have proposed design patterns for
vulnerabilities. The cost and efforts associated with privacy are also discussed in the
paper. The authors of paper [18] have discussed privacy and the risks involved in
sharing private information while making recommendations in their work. In another
paper [19], the authors have identified threats to information systems and suggested
solutions through policy decisions and awareness programs. How the privacy in
recommender system affects user psychology is discussed by the authors in their
paper [20]. The authors of paper [21] presented a review about ethics and security in
their paper. They have proposed security mechanism in compliance to best ethical
practises. A model for attack strategy considering constraints on attack attributes
and attack order has been presented to provide error–tolerant graph and techniques
to measure similarities between attack strategies [22]. Logical attack graphs are
generated that represent logical dependencies and configuration information among
attacks [23]. The attack paths have been identified by the authors and a recommender
system is proposed that could classify cyber attacks in future [1, 24].
Research in paper [25] proposed a new similarity model for an improved recom-
mender system. The authors considered local information of users and global user
behaviour to develop the model. In another paper [26], authors studied the effect on
performance of recommender system in their research using Mean Absolute Error
metric. In this research, balanced accuracy metric is used to support and extend the
state of art research. Authors of paper [27] have discussed various algorithms—user
based, and item based, to improvise the recommender system’s performance and
predictive accuracy of recommender system is focussed. F-measure is used to eval-
uate the quality of recommender system. Authors have used cross-domain approach
to improve performance of recommender system [28]. The authors have focused on
to provide highly accurate recommender list and to protect user’s private data while
making recommendations in their research [29]. In paper [30], the authors have
thrown light on limitations of presently being used recommender systems. They
have also discussed to improve recommendations and make recommender system
more widely available. In paper [31], authors have reviewed more than 80 existing
approaches available for academic literature recommendations. The authors evalu-
ated and discussed about various shortcomings and weaknesses of these approaches.
The study shows and concludes that there exists no agreement to compare and
evaluate existing academic literature recommender approaches. Authors of paper
[32] have proposed a CF approach using Entropy and Manhattan distance similarity
measures for an improved recommender system.
3 Measure Metrics Used
3.1 User-Item Matrix Construction
Let us construct a user-item matrix model for collaborative recommender system

(CRS). Let us assume that ‘uth’ user gives rating ‘r’ in the rating scale for ‘ith’ item.
Then,
ru,i = U . I
where
ru,i = rating for item ‘i’ given by user ‘u’,
U = {u 1 , u 2 , . . . , u n } holds ‘n’ number of users,
and I = {i 1 , i 2 , . . . , i N } holds ‘N’ number of items.

Next, R be defined as the accumulated user–item ratings, such that

ru,i = I ; R = ru,i = I T
Finally, the recommendations for Ith item are generated by finding average values
of nearest neighbours. Mathematically, degree of similarity between user-item ratings
can be calculated by using Pearson’s correlation coefficient (PCC).
PCC
n N N N
n n
N a =b=1 i=1 ra,i rb,i − a =b=1 i=1 ra,i a =b=1 i=1 rb,i
= 2 2
n N n N n N n N
N a =b=1
2
i=1 ra,i − a =b=1 i=1 ra,i N a =b=1
2
i=1 rb,t − a =b=1 i=1 rb,i
(1)
where ra,i is ‘ath’ user rating for ‘ith’ item and rb,i is ‘bth’ user rating for ‘ith’ item
3.2 Mean Absolute Error
Accuracy is a measure used to find how diligently recommendation knots with user’s
likings. It is a key metric that helps to improve recommendation eminence. Mean
absolute error (μ E ) is one of the metric that can be used to compare recommendation
scores with true user rating. μ E is calculated as an average of expected and actual
rating values. Mathematically,
1
n
μE = |ei − ai | (2)
n i=1
where ‘ei ’ is expected rating and ‘ai ’ is actual rating. The lesser value of μ E implies
higher performance.
3.3 Mean Accumulated Error
In our study, we have used Book-crossings (BX) dataset. There are 1.1 million ratings
in BX dataset. Since, it is hard to work with so many single rating values, therefore,
to handle this huge data, we find mean accumulated ratings (μ A ), defined as
1
n
μA = ri (3)
n i=1
where ‘r i ’ is total number of ratings given for a particular book ISBN.
3.4 Precision and Recall
In recommendation systems, it is desired to recommend only top n items to the

user. To achieve this, precision at threshold n (P@n) and recall at threshold n (R@n)
metrics are computed by considering only topmost results and performing evaluation
at a given cut-off rank. In notion, they may be defined as
tp tp
P@n = ; R@n = (4)
tp + f p t p + fn
3.5 F-measure
Precision and recall are sometimes used harmonically together in F measure. The
significance of using F measure is that harmonic mean in comparison to average
mean, penalises extreme values. Thus, F-measure at threshold n (F@n) is defined as
P@n ∗ R@n
F@n = 2 ∗ (5)
P@n + R@n
3.6 Accuracy and Balanced Accuracy
Accuracy (ACC) is a metric used to evaluate classification models. Accuracy is

defined as proportion of number of true recommendations to total number of
recommendations. Mathematically, accuracy is computed by formula
τ p + τn
ACC = (6)
τ p + τn + f p + f n
where
τ p is true positive, means an interested item and is recommended by user,
τn is true negative, means an un-interested item and is not recommended by user,
f n is false negative, means an interested item but not recommended by the user,
f p is false positive, means an un-interested item but recommended by the user.
ACC may not yield better results when datasets are imbalanced. In such situation,
balanced accuracy (BAC) is used. BAC measures the overall performance of the
system using the formula
TR P + TR N
BAC = (7)
2
where ‘TRp ’ is True Positive Rate and ‘TRn ’ is True Negate Rate and are defined
mathematically as
τp τn
TR P = and TR N = . (8)
τ p + fn τn + f p
Following algorithm is developed and implemented for the work.

4 Dataset and Experiment
To conduct the experiment, a dataset, namely Book-crossings (BX) is taken [33]. The
dataset holds 270,000 books and 90,000 users. There are 1.1 million ratings for the
available books in the BX dataset. The dataset has explicit and implicit ratings. The
explicit ratings are given on a scale of ‘1–10’, where rating ‘1’ is the lowest rating
and rating ‘10’ is the highest rating. Implicit ratings are expressed by ‘0’. There are
three tables in BX dataset, namely BX-Users, BX-Books and BX-Book-Ratings that
store the Users, ISBN values of Books and Book Rating information, respectively.
4.1 Evaluation Results
The experiment is performed in two steps. In first step, actual book ratings are
considered. Then, Push Attacks (P-Attacks) are randomly injected to the dataset.
The effect of P-Attacks is observed at various proportions and performances of
CRS after injecting the attacks is compared with the performance of original CRS.
Performance is evaluated using measures: Precision, recall, F-measure, accuracy
and balanced accuracy. Similarly, Nuke Attacks (N-Attacks) are arbitrarily injected
at different proportions in the dataset and performance pattern of CRSs is observed.
The algorithms for P-N Attacks are repeatedly applied to different proportions
of datasets and results are evaluated. It is imperative to observe that in our experi-
ment, users’ numbers are not significant. The main emphasis is on book ratings only.
The average (medium) book rating is taken as ‘5’ while rating ‘10’ (highest) and
rating ‘1’ (lowest) is considered as malicious. In the experiment, the mean accu-
mulated ratings (μ A ) are calculated by summing up the ratings for same ISBN in
original dataset. Next, P-N Attacks, at different proportions are arbitrarily injected
with ratings ‘10’ and ‘1’, respectively. Table 1 shows evaluative results for mean
absolute error. Corresponding to results, Fig. 1 shows the pattern of deviation of
ratings after being attacked from actual rating value.
Various performance measures are evaluated at a threshold value n. The dataset
is injected with P-N Attacks to generate biased recommendations. Injection of P-N
Attacks also made the dataset unbalanced. So to analyse the performance of CRS,
Table 1 Results for actual book ratings and μ E for injected P-N Attacks
Attack Actual 0.1% 0.2% 0.3% 0.4% 0.5% 0.6% 0.7% 0.8% 0.9% 1.0%
proportion
μ A (P-Attacks) 4.77 4.72 4.73 4.74 4.75 4.75 4.75 4.78 4.79 4.80 4.81
μ E (P-Attacks) – 0.05 0.04 0.03 0.02 0.02 0.02 0.01 0.02 0.03 0.04
μ A (N-Attacks) 4.77 4.70 4.71 4.72 4.73 4.73 4.69 4.68 4.68 4.68 4.67
μ E (N-Attacks) – 0.07 0.06 0.05 0.04 0.04 0.08 0.09 0.09 0.09 0.10
Fig. 1 Mean absolute error Mean Absolute Error wrt P-NAttacks

pattern for P-N Attacks
0.15 Push AƩacks
0.1
E
0.05 Nuke AƩacks
0
Attack Proportions
Table 2 Results of performance after injecting P-Attacks

Measures P-Attacks proportions
0.1–0.6% 0.7% 0.8% 0.9% 1.0%
P@n 0.25 0.50 0.50 0.50 0.50
R@n 0.25 0.14 0.13 0.11 0.10
F@n 0.25 0.22 0.20 0.18 0.17
ACC@n 0.45 0.36 0.27 0.18 0.09
BAC 0.41 0.45 0.40 0.31 0.05
Table 3 Results of performance after injecting N-Attacks

Measures N-Attacks proportions
0.1–0.6% 0.7% 0.8% 0.9% 1.0%
P@n 0.50 0.50 0.50 0.50 0.50
R@n 0.20 0.17 0.14 0.13 0.10
F@n 0.29 0.25 0.22 0.20 0.17
ACC@n 0.55 0.45 0.36 0.27 0.09
BAC 0.52 0.48 0.45 0.40 0.05
balanced accuracy (BAC) is calculated. Tables 2 and 3, show the results for the
performance of CRS when P-N Attacks are present.
4.2 Validation
The purpose of validation is to ensure the quality of experiment performed for given
dataset and data-subsets. To confirm the correctness of results, the experiment is
repeated for different proportions—1/2th, 1/4th, 1/40th and 1/64th rating values of
dataset. At different proportions of dataset, mean absolute error (E ) is calculated
for P-N Attacks. The original dataset is divided into data-subsets. This approach is
helpful for comparative study of results. The evaluative results for actual book ratings
and E for injected P-Attacks at various proportions are shown in Table 4. To validate
350
Table 4 Results for actual book ratings and E for injected P-N Attacks
Attack proportions Actual 0.1% 0.2% 0.3% 0.4% 0.5% 0.6% 0.7% 0.8% 0.9% 1%
1/2th MAR(P-Attacks) 4.99 5.00 5.02 5.04 5.05 5.07 5.08 5.10 5.11 5.13 5.14
MAE(P-Attacks) – 0.02 0.03 0.05 0.06 0.08 0.09 0.11 0.12 0.14 0.16
MAR(N-Attacks) 4.97 4.97 4.97 4.97 4.97 4.97 4.98 4.98 4.98 4.98 4.98
MAE(N-Attacks) – 0.00 0.00 0.00 0.01 0.01 0.01 0.01 0.01 0.01 0.02
1/4th MAR(P-Attacks) 4.84 4.86 4.87 4.89 4.90 4.92 4.93 4.95 4.96 4.98 4.99
MAE(P-Attacks) – 0.01 0.03 0.04 0.06 0.07 0.09 0.10 0.12 0.13 0.15
MAR(N-Attacks) 4.84 4.85 4.85 4.85 4.85 4.85 4.85 4.86 4.86 4.86 4.86
MAE(N-Attacks) – 0.00 0.00 0.00 0.01 0.01 0.01 0.02 0.02 0.02 0.02
1/40th MAR(P-Attacks) 4.92 4.94 4.94 4.96 4.98 4.98 5.00 5.02 5.03 5.05 5.08
MAE(P-Attacks) – 0.02 0.02 0.04 0.06 0.06 0.08 0.10 0.11 0.13 0.16
MAR(N-Attacks) 4.49 4.49 4.49 4.49 4.50 4.50 4.50 4.50 4.50 4.50 4.50
MAE(N-Attacks) – 0.00 0.00 0.00 0.01 0.01 0.01 0.01 0.01 0.01 0.01
1/64th MAR(P-Attacks) 4.49 4.50 4.51 4.52 4.52 4.54 4.55 4.56 4.57 4.58 4.61
MAE(P-Attacks) – 0.01 0.02 0.04 0.04 0.05 0.06 0.07 0.08 0.10 0.12
MAR(N-Attacks) 4.42 4.42 4.41 4.42 4.41 4.43 4.43 4.44 4.43 4.44 4.43
MAE(N-Attacks) – 0.00 0.01 0.00 0.01 0.01 0.01 0.02 0.01 0.02 0.01
A. B. Chopra and V. S. Dixit
Table 5 Results of performance after injecting P-Attacks

Datasets Measures P-Attacks proportions
0.1–0.6% 0.7% 0.8% 0.9% 1.0%
1/2th P@n 0.50 0.50 0.50 0.50 0.50
R@n 0.25 0.14 0.33 0.50 1.00
F@n 0.33 0.22 0.40 0.50 0.67
ACC@n 0.64 0.36 0.73 0.82 0.91
BAC 0.55 0.45 0.60 0.69 0.95
1/4th P@n 0.50 0.50 0.50 0.50 0.50
R@n 0.25 0.14 0.33 0.50 1.00
F@n 0.33 0.22 0.40 0.50 0.67
ACC@n 0.64 0.36 0.73 0.82 0.91
BAC 0.55 0.45 0.60 0.69 0.95
1/40th P@n 0.50 0.50 0.50 0.50 0.50
R@n 0.25 0.14 0.33 0.50 1.00
F@n 0.40 0.22 0.40 0.50 0.67
ACC@n 0.73 0.36 0.73 0.82 0.91
BAC 0.60 0.45 0.60 0.69 0.95
1/64th P@n 0.33 0.50 0.50 0.50 0.50
R@n 0.33 0.14 0.33 0.50 1.00
F@n 0.33 0.22 0.40 0.50 0.67
ACC@n 0.64 0.36 0.73 0.82 0.91
BAC 0.54 0.45 0.6 0.69 0.95
performance and accuracy of these datasets, evaluations are presented in Tables 5 and
6, respectively. Table 5 shows the evaluation results of performance after insertion
of P-Attacks at different proportions and results for N-Attacks are shown in Table 6.
The results in Tables 5 and 6 are calculated by using various measures—Precision,
recall, accuracy and balanced accuracy, at threshold value n.
The results in Table 4 show that the balanced accuracy (BAC) of the CRS may
vary from 50 to 90% approximately, depending upon the size of dataset and size
of injected attacks. The experiment also validates that performance and accuracy of
any recommender system have direct relation with the size of attack (P-Attacks or
N-Attacks) injected to it. The more injected P-N Attacks lead to more biasness in the
recommendations predicted for an item. Pictorially, the interpretation for Table 4 is
depicted through Fig. 2 for P-Attacks and through Fig. 3 for N-Attacks, respectively.
The figures show variations in performances of CRS at different proportions due to
push and nuke attacks injected.
The results for various performance measures at different proportions of dataset
at a given threshold value n, when P-N Attacks are injected are given in Tables 5
and 6, respectively. The datasets at 0.1–0.6% are considered as small sampled and
Table 6 Results of performance after injecting N-Attacks

Datasets Measures N-Attacks proportions
0.1–0.6% 0.7% 0.8% 0.9% 1.0%
1/2th P@n 0.83 0.83 0.83 0.83 0.83
R@n 1.00 0.50 0.50 0.50 0.50
F@n 0.91 0.63 0.63 0.63 0.63
ACC@n 0.91 0.45 0.45 0.45 0.45
BAC 0.92 0.25 0.25 0.25 0.25
1/4th P@n 0.86 0.80 0.80 0.80 0.80
R@n 0.40 0.40 0.40 0.40 0.40
F@n 0.53 0.53 0.53 0.53 0.53
ACC@n 0.36 0.36 0.36 0.36 0.36
BAC 0.20 0.20 0.20 0.20 0.20
1/40th P@n 0.75 0.88 0.88 0.88 0.88
R@n 1.00 0.70 0.70 0.70 0.70
F@n 0.86 0.78 0.78 0.78 0.78
ACC@n 0.91 0.64 0.64 0.64 0.64
BAC 0.94 0.35 0.35 0.35 0.35
1/64th P@n 0.83 0.83 0.83 0.83 0.83
R@n 1.00 0.50 0.50 0.50 0.50
F@n 0.91 0.63 0.63 0.63 0.63
ACC@n 0.91 0.45 0.45 0.45 0.45
BAC 0.92 0.25 0.25 0.25 0.25
Fig. 2 Mean absolute error Mean Absolute Error for P-Attacks

pattern for P-Attacks
0.2 1/2th Dataset
0.15
1/4th Dataset
E 0.1
1/40th Dataset
0.05
0 1/64th Dataset
Attack Proportions
collective evaluations are performed. Greater than 0.6% proportions are considered as
large samples and evaluations are performed separately for each dataset proportion.
Corresponding to Table 5, Fig. 4 depicts the pattern of performance of CRS for
injected P-Attacks at different proportions of datasets. The figure shows the variation
in performance of CRS using various mathematical measures at different proportions
Fig. 3 Mean absolute error Mean Absolute Error for N-Attacks

pattern for N-Attacks
0.025 1/2th Dataset
0.02
0.015 1/4th Dataset
E
0.01
1/40th Dataset
0.005
0 1/64th Dataset
Attack Proportions
Performance Analysis for P-Attacks at different dataset proportions

1.2
1
0.8
Attack
0.6 0.1% - 0.6%
Proportions
0.4 0.70%
0.2
0.80%
0
0.90%
Performance
bACC
R@n
bACC
R@n
F@n
F@n
ACC@n
P@n
ACC@n
P@n
1.00%
Datasets 1/2th 1/4th 1/40th 1/64th
Fig. 4 Performance analysis of CRS injected with P-Attacks
of datasets with different sizes of P-Attacks injected to these datasets. Similarly,

corresponding to Table 6, pattern of performance of CRS for injected N-Attacks at
different proportions of datasets are shown in Fig. 5. The figure shows the variation
in performance of CRS using various mathematical measures at different proportions
of datasets with different sizes of N-Attacks injected to these datasets.
Performance Analysis for N-Attacks at different dataset proportions

1.2
1
0.8
Attack
0.6 0.1% - 0.6%
Proportions 0.4
0.2 0.70%
0
0.80%
Performance
R@n
R@n
R@n
bACC
R@n
P@n
F@n
bACC
ACC@n
P@n
F@n
ACC@n
bACC
P@n
F@n
ACC@n
P@n
F@n
bACC
ACC@n
0.90%
1.00%
Datasets 1/2th 1/4th 1/40th 1/64th
Fig. 5 Performance analysis of CRS injected with N-Attacks

5 Conclusion
A recommender system requires being efficient and accurate so as to generate correct

recommendations of items for users. In our research, a CRS is constructed by
collecting user-item ratings a matrix model is constructed. P-N Attacks are then
injected to the original dataset and various mathematical measure such as—Mean
absolute error, precision, recall, accuracy at threshold value n are evaluated to analyse
the performance of CRS. The results show, the P-Attacks escalates the accuracy of
CRS where N-Attacks degrades CRS’s accuracy. It is to note that, when P-N Attacks
are injected to the original dataset, the dataset became biased. Balanced accuracy is,
therefore, used to evaluate the effect of performance on RSs. The study also shows
the effect of performance on CRS at various attack proportions. While conducting
the experiment, it is concluded that at least 1/64th of dataset values are required to
reach at results. This means that the experiment may not yield good results on very
small datasets.
In this paper, only P-N Attacks from non-authentic users are studied and no
acquaintance about background of users is recognized. Our future work will address
this critical issue along with contextual information given by users while rating and
will try to identify and classify P-N Attacks injected to dataset available on Web.
Our study will also try to classify P-N Attacks in group recommender system by
enhancing multi-expert system [34].
References
1. N. Polatidis, E. Pimenidis, M. Pavlidis, H. Mouratidis, Recommender systems meeting secu-

rity: from product recommendation to cyber-attack prediction, in International Conference on
Engineering Applications of Neural Networks, Athens, Greece, 25–27 Aug 2017
2. R. Meena, K.K. Bharadwaj, A genetic algorithm approach for group recommender system
based on partial rankings. J. Intell. Syst. 29(1), 653–663 (2018)
3. J. Parra-Arnau, D. Rebollo-Monedero, J. Forné, A privacy-protecting architecture for collabora-
tive filtering via forgery and suppression of ratings, Data Privacy Management and Autonomous
Spontaneous Security (Springer, Berlin, 2011), pp. 42–57
4. Y. Koren, R. Bell, C. Volinsky, Matrix factorization techniques for recommender systems.
Computer 8, 30–37 (2009)
5. I. Christensen, S. Schiaffino, Matrix factorization in social group recommender systems, in 2013
12th Mexican International Conference on Artificial Intelligence (IEEE, 2013), pp. 10–16
6. H. Mehta, S.K. Bhatia, P. Bedi, V.S. Dixit, Collaborative personalized web recommender system
using entropy based similarity measure. IJCSI 8(6), 3 (2011)
7. J. O’Donovan, B. Smyth, Is trust robust? An analysis of trust-based recommendation, in IUI
’06: Proceedings of the 11th International Conference on Intelligent User Interfaces (ACM
Press, New York, NY, USA, 2006), pp. 101–108
8. C. Dellarocas,. Immunizing online reputation reporting systems against unfair ratings and
discriminatory behavior, in ACM Conference on Electronic Commerce (2000), pp. 150–157
9. E. Friedman, P. Resnick, The social cost of cheap pseudonyms. J. Econ. Manage. Strat. 10(2),
173–199 (2001)
10. S.D. Kamvar, M.T. Schlosser, H. Garcia-Molina, The Eigentrust algorithm for reputation
management in P2P networks, in WWW ’03: Proceedings of the 12th International Conference
on World Wide Web (ACM Press, New York, NY, USA, 2003), pp. 640–651
11. J. Lu, D. Wu, M. Mao, W. Wang, G. Zhang, Recommender system application developments:
a survey. Decis. Support Syst. 74, 12–32 (2015)
12. B. Shams, S. Haratizadeh, TasteMiner: Mining partial tastes for neighbor-based collaborative
filtering. J. Intell. Inf. Syst. 48(1), 165–189 (2017)
13. F. Ortega, J.L. Sánchez, J. Bobadilla, A. Gutiérrez, Improving collaborative filtering-base
recommender systems results using Pareto dominance. Inf. Sci. 239, 50–61 (2013)
14. X. Su, T.M. Khoshgoftaar, A survey of collaborative filtering techniques. Adv. Artif. Intell.
2009(Section 3), 1–19 (2009)
15. J. Wang, Q. Tang, Recommender systems and their security concerns (2015)
16. P. Bedi, S.K. Agarwal, Richa, Trust and reputation-based multi-agent recommender system.
Int. J. Comput. Sci. Eng. 16(4), 350–362 (2018)
17. A. Dalai, S. Jena, Evaluation of web application security risks and secure design patterns, in
International Conference on Communication, Computing & Security, ICCCS 2011, Odisha,
India, 12–14 Feb 2011, pp. 565–568
18. S.K. Tony, F.D. Lam, J. Riedl, Do you trust your recommendations? An exploration of secu-
rity and privacy issues in recommender systems, in ETRICS, ed. by G. Muller, LNCS 3995
(Springer, Berlin, 2006), pp. 14–29
19. S. Hettiarachchi, S. Wickramasinghe, Study to identify threats to information systems in orga-
nizations and possible countermeasures through policy decisions and awareness programs to
ensure the information security (2016)
20. B. Zhang, N. Wang, H. Jin,. Privacy concerns in online recommender systems: influences of
control and user data input, in 10th Symposium On Usable Privacy and Security (SOUPS 2014)
(2014), pp. 159–173
21. B. Bashir, A. Khalique, a review on security versus ethics. Int. J. Comput. Appl. 151(11), 13–17
(2016)
22. P. Ning, D. Xu, Learning attack strategies from intrusion alerts, in Proceedings of the 10th
ACM Conference on Computer and communication security—CCS ’03 (2003), p. 200
23. X. Ou, W.F. Boyer, M.A. McQueen, A scalable approach to attack graph generation, in 9th
ACM Conference on Computer and Communications Security (2006), pp. 336–345
24. N. Polatidis, M. Pavlidis, H. Mouratidis, Cyber-attack path discovery in a dynamic supply
chain maritime risk management system. Comput. Stand. Interfaces 56, 74–82 (2018)
25. H. Liu, Z. Hu, A. Mian, H. Tian, X. Zhu, A new user similarity model to improve the accuracy
of collaborative filtering. Knowl. Based Syst. 56, 156–166 (2014)
26. A.B. Chopra, V.S. Dixit, Performance analysis of collaborative recommender system:
a heuristic approach, in International Conference on Information, Communication and
Computing Technology (Springer, Singapore, 2019), pp. 178–193
27. S.B. Shinde, M.A. Potey, Research paper recommender system evaluation using coverage. Int.
Res. J. Eng. Technol. 3(6), 1678–1683 (2016)
28. Richa, P. Bedi, Parallel context-aware multi-agent tourism recommender system. Int. J. Comput.
Sci. Eng. 20(4), 536–549 (2019)
29. G. Li, Z. Cai, G. Yin, Z. He, M. Siddula, Differentially private recommendation system based
on community detection in social network applications. Secur. Commun. Netw. 2018 (2018)
30. G. Adomavicius, A. Tuzhilin, Toward the next generation of recommender systems: a survey
of the state-of-the-art and possible extensions. IEEE Trans. Knowl. Data Eng. 17(6), 734–749
(2005)
31. J. Beel, B. Gipp, S. Langer, C. Breitinger, Paper recommender systems: a literature survey. Int.
J. Digit. Libr. 17(4), 305–338 (2016)
32. W. Wang, G. Zhang, J. Lu, Collaborative filtering with entropy-driven user similarity in
recommender systems. Int. J. Intell. Syst. 30, 854–870 (2015)
33. Dataset: http://www2.informatik.uni-freiburg.de/~cziegler/BX/

34. R. Meena, S. Minz, Group recommender systems—an evolutionary approach based on multi-
expert system for consensus. J. Intell. Syst. 1 (2018). (ahead-of-print)
Using Transfer Learning for Detecting
Drug Mentions in Tweets
Laiba Mehnaz and Rajni Jindal
Abstract Drug mention detection in tweets is defined as a binary classification

task that aims to distinguish tweets that mention a drug from those which do not.
In our work, we aim to use transfer learning for automatic classification of tweets
mentioning a drug. To our knowledge, we are the first to use transfer learning in
drug detection using the UPennHLP Twitter Pregnancy Corpus. We find this corpus
challenging as the positive class(i.e., mentioning a drug) is only 0.2% of the total
number of tweets, leading to a highly imbalanced dataset. We carry out several exper-
iments using pre-trained language models such as BERT, BioBERT, ClinicalBERT,
SciBERT, BioMed-RoBERTa, ELECTRA, and ERNIE 2.0. We further analyze the
performance of each of the pre-trained models based on the training data distribu-
tion, training data size, and different pre-training objectives. According to our results,
BioMed-RoBERTa achieves the highest F1 score of 0.85 for the positive class.
Keywords Social media · Transfer learning · Drug name detection ·

Pharmacovigilance
1 Introduction
Recent research [1] reveals that Twitter is being increasingly used by adults to post
tweets regarding their personal and current medical conditions. The report also
states that adults using twitter also get influenced by online trends or discussions
related to medication and health. Hence, Twitter acts as an excellent resource for
researchers to use it for various medical applications. Detecting drug or medication
mentions from tweets helps in multiple tasks and fields such as pharmacovigilance,
syndromic surveillance, and monitoring drug abuse. Pharmacovigilance [2] activi-
L. Mehnaz (B) · R. Jindal

Delhi Technological University, New Delhi 110042, India
e-mail: laibamehnazbt2k16@dtu.ac.in
R. Jindal
e-mail: rajnijindal@dce.ac.in
358 L. Mehnaz and R. Jindal
ties range from detecting, assessing, understanding, and preventing adverse effects
of any drug-related problem. Adverse drug reactions produce devastatingly undesir-
able effects on the body after the drug intake. In 1999, for the treatment of type 2
diabetes mellitus, a drug called Rosiglitazone [3] was introduced, which was later
withdrawn due to concerns of a potential risk of increased fatal myocardial infarction,
also known as a heart attack. And since tweets pose a big challenge of containing
misspellings and informal words for clinical terms, building robust systems that can
detect drug mentions from tweets can further help in detecting adverse drug reaction
from tweets as well. We define the task of drug detection as a binary classification
task, distinguishing tweets that mention a drug from those that do not. The definition
of a drug is as per the definition provided by the FDA. The dataset that we use for
drug detection is called the UPennHLP Twitter Pregnancy Corpus [4]. It is a recently
released dataset for drug detection, released during the shared task of the Fifth Social
Media Mining for Health Applications 2020. We are specifically interested in this
corpus as it reflects the naturally occurring imbalance in tweets between both the
classes. For example, “Shoulda took some more NyQuil.” represents a tweet men-
tioning a drug, and “I see all these dads with their kids and it makes me sad for my
baby :(” represents a tweet that does not mention a drug.
The only work done on this corpus, to our knowledge is done by [4], where
they use Kusuri [4], a model which is an ensemble of deep neural networks. In our
approach, we use transfer learning by using pre-trained language models and fine-
tuning them on our dataset. We analyze the performance of these pre-trained models
on a dataset that is highly imbalanced and also is in the medical domain. We analyze
the effect of the size as well as the domain of the dataset used while pre-training
these models. We also analyze the effect of different pre-training objectives. We
carry out experiments with BERT [5], BioBERT [6], ClinicalBERT [7, 8], SciBERT
[9], BioMed-RoBERTa [10], ELECTRA [11] and ERNIE 2.0 [12].
2 Dataset
To create the dataset, [4] collected 112 timelines of women on twitter that had declared
their pregnancy and tweeted throughout their pregnancy. The authors manually anno-
tated the tweets mentioning a drug and tweets that did not. This dataset is highly
imbalanced between the two classes, with only 0.2% of the tweets mentioning a
drug name. We particularly chose this dataset as we wanted to analyze the perfor-
mance of transfer learning on the class imbalance. As the task is still going on and the
test set has not been released, we downloaded the training dataset and the validation
dataset as available. We use the validation dataset as our test set, as we have the
ground truth labels, and we can evaluate the performance of our pre-trained mod-
els. The original training dataset provided to us from the organizers of the SMM4H
2020 Task 1 contains 55419 tweets. Out of the total number of tweets, only 146
tweets mention a drug, and the rest 55,273 tweets do not mention any. Due to this
class imbalance in the dataset, we oversampled the positive tweets(i.e., mention a
Using Transfer Learning for Detecting Drug Mentions in Tweets 359
Table 1 Breakdown of tweets in UPennHLP Twitter pregnancy corpus, after the oversampling and
train-dev split on the training dataset
Dataset Mentioning a drug Not mentioning a drug
Training 49,803 49,688
Development 5470 5585
Test 35 13,818
drug) by simply copying them. After oversampling, our training dataset consisted of
110,546 tweets in total. We randomly split 90% of the dataset into the training set
and the rest 10% as the development set. We use the validation dataset provided by
the organizers as the test set. The test set consists of a total of 13,853 tweets, of which
only 35 belong to the positive class(i.e., mentioning a drug), and the rest 13,818 do
not. Table 1 shows the breakdown of the tweets in all the datasets.
3 Method
Transfer learning is an approach in which large neural networks are pre-trained

on data-rich tasks and then are fine-tuned on various downstream tasks. Transfer
learning is supposed to “transfer” the common and general-purpose abilities to these
downstream tasks. This approach has been responsible for achieving state-of-the-art
in most of the NLP benchmarks [5]. Transfer learning has not been used on our
dataset as of now, and we wish to explore the performance of the same. We use
the following list of pre-trained language models: BERT, BioBERT, ClinicalBERT,
SciBERT, BioMed-Roberta, ERNIE 2.0, and ELECTRA.
4 Experiments
In this section, we describe the various pre-trained models used during our experi-
ment.
BERT is a pre-trained language model used for various downstream tasks in
NLP after fine-tuning it on the specific task. It uses the Transformers [13] archi-
tecture. We use the BERT-base uncased model, implemented in PyTorch [14] via
the Transformers library [15] using the same hyperparameters as the original model
[5]. We use BioBERT, ClinicalBERT, SciBERT, and BioMed-RoBERTa through the
Transformers library. BioBERT is a domain-specific language model that further
trains BERT with a corpus that is specific to the medical domain. We use BioBERT-
base v1.0 which is trained on Wikipedia, BooksCorpus [16], PubMed abstracts and
PMC full-text articles. ClinicalBERT is also a domain-specific pre-trained language
model, which further trains BioBERT on the MIMIC-III V1.4 [17] dataset, which
is a collection of approximately 2 million clinical notes. SciBERT also uses the
same architecture as used by BERT. It trains BERT on scientific publications in var-
ious domains. The training data consists of a random sample of 1.14M papers from
Semantic Scholar [18]. Reference [9] have released two kinds of models. One, where
they train BERT further on the scientific publications corpus and one, where they
train BERT from scratch on the scientific publications corpus with its separate vocab-
ulary, called scivocab. We use the second model, SciBERT-scivocab-uncased, which
is trained from scratch. BioMed-Roberta trains RoBERTa [19] further on 2.68M
full-text papers from S2ORC [20]. ELECTRA is a pre-trained language model that
aims to make the pre-training objective of BERT computationally more efficient. In
BERT, only 15% of the input tokens are masked. And hence, BERT ends up learn-
ing only from 15% of the input tokens by predicting the word for that particular
index. Whereas, given the same input tokens, ELECTRA learns from every token,
by distinguishing original tokens from synthetically generated tokens. It is trained
as a discriminator as compared to BERT, which is trained as a generator. We use the
implementation of ELECTRA provided by the authors [11]. We have used two mod-
els, ELECTRA-small and ELECTRA-base to understand the impact of the model
size and the number of parameters on the performance on our task. ERNIE 2.0 is
another pre-trained model which essentially differs from the rest of the models used
so far. It aims to capture lexical, syntactic, and semantic information as well from the
training data, instead of just capturing knowledge about the co-occurrence of words
and sentences. It uses continual pre-training, where multiple pre-training tasks are
incrementally constructed, and the pre-trained model learns via continual multi-task
learning. We use ENRIE 2.0’s implementation provided by the authors [12]. For
all the models mentioned above, we have used the hyperparameters used in their
respective original papers.
We have chosen to use the F1 score for the positive class (i.e., tweets that men-
tion medications) to evaluate the performance of the classifier. The results on the
UPennHLP Twitter Pregnancy Corpus are given in Table 2. As can be seen, BioMed-
Roberta performs the best, and ELECTRA-small performs the worst. BioMed-
RoBERTa’s success can be attributed to its improved training strategies over BERT
itself. RoBERTa improves on several pre-training strategies, from training the model
for a longer time over additional data to removing the next sentence prediction
pre-training objective and dynamically changing the masking pattern applied to the
training data. As well as, further pre-training RoBERTa on domain-specific data,
which in this case is a corpus of scientific papers. Another interesting point to note
here is that the scientific papers used to pre-train the model consist of 18% of papers
from the computer science domain and 82% from the broad medical domain. It shows
the benefit of training on a dataset that is of a similar domain if the source data and
Table 2 Results: F1 score for the “positive class” (i.e., mentioning a drug)
Pre-trained model F1 score for the positive class
BERT 0.78
BioBERT 0.8
ClinicalBERT 0.8
SciBERT 0.83
BioMed RoBERTa 0.85
ERNIE 2.0 0.83
ELECTRA small 0.73
ELECTRA base 0.78
the target data happen to be from different domains and do not happen to share the
vocabulary. On the other hand, ELECTRA-small has the worst performance of all,
which we speculate to be because of the difference in the domain of the source data
and the task data. We also notice that ELECTRA-small contains much lesser param-
eters compared to the BERT-base and all the other pre-trained models in our list, and
decreasing the number of parameters has hurt the performance drastically. This list of
experiments is not at all comprehensive, but within these experiments as well, we can
see that large compute models do translate to better performance. We see BERT-base
and ELECTRA-base, both achieve the exact score, not benefiting from ELECTRA’s
replaced token detection pre-training objective. We again speculate that this could
be because our dataset is in the medical domain and maybe this could be the reason
that ELECTRA-base did not benefit from its replaced token detection pre-training
objective. Maybe if compared on a target domain that is very similar to the training
data used while pre-training both BERT and ELECTRA, then can we hope to see
a difference in the performance due to the difference in choice of the pre-training
objective.
We see a consistent increase in the F1 score of the pre-trained models as soon as
we train them further on a similar domain as the fine-tuning task. This proves a very
strong point that immediately applying the knowledge learned from a large general
corpus does not benefit as much as, first updating the weights for the specific task-
related domain and then fine-tuning for our task. It also raises an interesting question
of whether we can have universal pre-trained models that can perform competitively
on a variety of domains. BioBERT, ClinicalBERT, SciBERT, and BioMed-RoBERTa,
all consistently improve their performance over BERT, while sharing the same archi-
tecture and size. SciBERT has a higher F1 score than both BioBERT and Clinical-
BERT. This increase could be because of two reasons. According to the increase in
performance, it looks like the multi-domain scientific corpus suits the need of our
fine-tuning task better than PMC full-text articles, PubMed abstracts, and MIMIC-III
dataset, on which BioBERT and ClinicalBERT are trained on respectively. It could
also be because SciBERT-scivocab-uncased has been trained from scratch and has a
different vocabulary from that of BERT’s, built from the Semantic Scholar corpus.
SciBERT’s vocabulary has the same size as that of BERT-base’s vocabulary and is
considerably disjoint. The overlap between both the vocabularies is just 42% show-
ing the difference in the distribution of words in the scientific domain and the general
domain. And also, that the distribution of words in the training corpus in SciBERT
is more similar to our task data than compared to.
The most interesting result is that of ERNIE 2.0. This model scores the same
result as SciBERT without being trained on data which is even close to the domain
of scientific papers. Without being trained on any medical data, it still performs better
than BERT in understanding our task data, which highlights the fact that ERNIE 2.0
is truly capturing other aspects of language such as lexical, syntactic, and semantic
information and not just co-occurrence of words and sentences. This makes it easy for
ERNIE 2.0 to classify drug mentioning tweets even when there is a class imbalance
present in the dataset. ERNIE 2.0 learns representations from a variety of tasks
and is truly able to capture these aspects of information simultaneously, effectively,
and efficiently without forgetting the previously learned knowledge by leveraging
continual multi-task learning. ERNIE 2.0’s additional data collected from Reddit
and the usage of Discovery data [21] while pre-training, gives it an almost equal
boost in performance as compared to training BERT on a similar domain before
fine-tuning on the particular task. Providing us with information, that to increase the
performance and understanding of a model toward a particular domain, pre-training
with huge unlabeled data in the similar domain is not the only solution. This pre-
trained model gives us a ray of hope on having a universal pre-trained model that
works competitively on various domains and has a general “understanding”.
6 Conclusion
In this paper, we have analyzed the performance of several pre-trained language

models on the task of detecting drug mentions in tweets, the dataset for which presents
a massive class imbalance and ambiguity. Our work is the first work that uses transfer
learning on the UPennHLP Twitter Pregnancy Corpus for detecting drug mentions
in tweets, showing that transfer learning works well even when there is a huge
class imbalance present in the dataset. It also works well with the informal structure
of tweets. Through our analysis, we show that BioMed- RoBERTa performs the
best out of all the models we have used. We illustrate the power of contextualized
embeddings, further training on domain-specific unlabeled data, and the effect of
different pre-training tasks in these models. We also find that reducing the size of the
model and the parameters drastically reduce the performance as well. As future work,
we aim to explore more pre-trained language models that use various pre-training
objectives, and also look into incorporating information from knowledge graphs for
drug detection as well.
References
1. K. O’Connor, P. Pimpalkhute, A. Nikfarjam, R. Ginn, K.L. Smith, G. Gonzalez, Pharma-

covigilance on twitter? Mining tweets for adverse drug reactions, in AMIA annual symposium
proceedings, pp. 924–933. Published 2014 Nov 14
2. R. Leaman, L. Wojtulewicz, R. Sullivan, A. Skariah, J. Yang, G. Gonzalez, Towards internet-
age pharmacovigilance: extracting adverse drug reactions from user posts to health-related
social networks, in Proc. BioNLP (2010), pp. 117–125
3. C. Day., C.J. Bailey, Rosiglitazone, in xPharm: the comprehensive pharmacology reference
(2007), pp. 1–4. https://doi.org/10.1016/B978-0-12-801238-3.97237-4
4. D. Weissenbacher, A. Sarker, A. Klein, K. O’Connor, A. Magge, G. Gonzalez-Hernandez, Deep
neural networks ensemble for detecting medication mentions in tweets. J. Am. Med. Inform.
Assoc. 26(12), 1618–1626 (2019). https://doi.org/10.1093/jamia/ocz156
5. J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional trans-
formers for language understanding (2019). arXiv:1810.04805
6. J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C.H. So, J. Kang, BioBERT: a pre-trained biomedical
language representation model for biomedical text mining. Bioinformatics 36(4), 1234–1240
(2020)
7. E. Alsentzer, J.R. Murphy, W. Boag, W. Weng, D. Jin, T. Naumann, M.B. McDermott, Publicly
available clinical BERT embeddings (2019). arXiv:1904.03323
8. K. Huang, J. Altosaar, R. Ranganath, ClinicalBERT: modeling clinical notes and predicting
hospital readmission (2019). arXiv:1904.05342
9. I. Beltagy, A. Cohan, K. Lo, SciBERT: pretrained contextualized embeddings for scientific text
(2019). arXiv:1903.10676
10. S. Gururangan, A. Marasović, S. Swayamdipta, K,Lo, I. Beltagy, D. Downey, N.A. Smith, Don’t
stop pretraining: adapt language models to domains and tasks (2020). arXiv:2004.10964
11. K. Clark, M. Luong, Q.V. Le, C.D. Manning, ELECTRA: pre-training text encoders as dis-
criminators rather than generators (2020). arXiv:2003.10555
12. Y. Sun, S. Wang, Y. Li, S. Feng, H. Tian, H. Wu, H. Wang, ERNIE 2.0: A Continual Pre-training
Framework for Language Understanding (2019). arXiv:1907.12412
13. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, I. Polosukhin,
Attention is All you Need. arXiv:1706.03762
14. A. Paszke, S Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N.
Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani,
S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, S.: PyTorch: an imperative style,
high-performance deep learning library, in NeurIPS, ed. by H. M. Wallach, H. Larochelle, A.
Beygelzimer, F. d’Alché Buc, E. B. Fox & R. Garnett, pp. 8024–8035 (2019)
15. T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf,
M. Funtowicz, J. Brew, HuggingFace’s Transformers: State-of-the-art Natural Language Pro-
cessing (2019). arXiv:1910.03771
16. Y. Zhu, R. Kiros, R.S. Zemel, R. Salakhutdinov, R. Urtasun, A. Torralba, S. Fidler, Aligning
books and movies: towards story-like visual explanations by watching movies and reading
books, in 2015 IEEE International Conference on Computer Vision (ICCV), pp. 19-27 (2015)
17. A.E. Johnson, T.J. Pollard, L. Shen, L.W. Lehman, M. Feng, M. Ghassemi, B. Moody, P.
Szolovits, L.A. Celi, R.G. Mark, MIMIC-III, a freely accessible critical care database. Sci
Data 3, 160035 (2016). https://doi.org/10.1038/sdata.2016.35
18. W. Ammar, D. Groeneveld, C. Bhagavatula, I. Beltagy, M. Crawford, D. Downey, J. Dunkel-
berger, A. Elgohary, S. Feldman, V.A. Ha, R.M. Kinney, S. Kohlmeier, K. Lo, T.C. Murray, H..
Ooi, M.E. Peters, J.L. Power, S. Skjonsberg, L.L. Wang, C. Wilhelm, Z. Yuan, M.V. Zuylen,
O. Etzioni, Construction of the literature graph in semantic scholar (NAACL-HLT, 2018)
19. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoy-
anov, RoBERTa: A Robustly Optimized BERT Pretraining Approach (2019). arXiv:1907.11692
20. K. Lo, L.L. Wang, M.E. Neumann, R.M. Kinney, D.S. Weld, S2ORC: The Semantic Scholar
Open Research Corpus. arXiv: Computation and Language (2020)
21. D. Sileo, T.V. Cruys, C. Pradel, P Muller, Mining Discourse Markers for Unsupervised Sentence
Representation Learning. NAACL-HLT (2019)
A Comprehensive and Critical Analysis
of Cross-Domain Federated Identity
Management Deployments
Tejaswini Apte and Jatinderkumar R. Saini
Abstract Sharing the identity in cross-domain application deployment requires

a controlled environment to ensure privacy of the user or request. The paper presents
various protocols for Federated Identity Management, available in literature for iden-
tity sharing in cross-domain application deployment. Also, the implementation and
limitations of each protocol is discussed. Considering cloud as an application of cross-
domain deployment, the study covers the entity communication channel along with
identity visibility in cloud. During the deployment, the communication channel for
identity sharing can be designed with multiple protocols. The challenge is to protect
privacy of identity while maintaining the compatibility among multiple protocols.
Keywords Federated Identity Management (FIM) · SAML · OAuth · OpenID

connect · Service Provider (SP) · Identity Provider (IdP) · OpenStack
1 Introduction
In cross-domain application deployment where identities are getting shared with

required attributes among various applications; assembling and disassembling the
critical information for authentication and authorization in controlled manner is a
need of the hour. Transmission of the sensitive identity information over the insecure
network has proven itself a challenging risk.
Literature has discussed the concept of Centralized and Decentralized Identity
Management to control the exposure of identity among various applications [1, 2].
Identity Management (IM) provider service maintains the identities sharing mech-
anism across various service domains and distributed environments. Authentication
T. Apte (B) · J. R. Saini

Symbiosis Institute of Computer Studies and Research, Symbiosis International Deemed
University, Pune, India
e-mail: apte.tejaswini@gmail.com
J. R. Saini
e-mail: saini_expert@yahoo.com
366 T. Apte and J. R. Saini
engine of IM provider verifies the identity and then provides respective authoriza-
tion for accessing the various services. Policy driven access mechanisms are used to
control the access to a given resource [3]. However, sharing the identity attributes
for cross-domain application deployment poses the risk of privacy.
Federated Identity Management (FIM) minimizes the risk by controlled attribute
sharing. It uses the required policies to share the relevant attributes across the
network channels. Thus, with same attribute sharing user can access secured infor-
mation from various service domains. These domains confirm the users’ identities
by trusted policies to confirm the authorization scope. Hence, the need of sharing
of various attributes to access different applications over the network is drastically
reduced; resulted in improved usability and security. Some of the challenges related
to processes, technology and privacy in FIM are mentioned in the literature. Further,
the “trust framework” design with the required approaches is proposed to mitigate the
risk while sharing the identities across cross-domain application deployment [4–6].
Various policies and protocols are available in the literature to share the iden-
tities in cross-domain deployment. Section 2 briefly outlines the same. Section 3
exemplifies the comparative assessment and entities (User, Service Provider, Identity
Provider, EndPoint) communication required for various FIM protocols such as the
SAML, OAuth and OpenID and its implementation in cloud. Further the conclusion
is presented in Sect. 4.
2 Literature Review
Literature reveals, various Identity and Access Management (IAM) policies are
implemented across cross-domain deployment of an application. Role Based Access
Control (RBAC) policy is a most acceptable service solution to provide secu-
rity feature in cross-domain infrastructure. By using specific role hierarchies and
constraints the respective security policies can be expressed in a wider range based
on the function assigned to the user. These RBAC policies promote the authorization
of various users using Separation of Duties. The RBAC policy model defines basic
elements such as users, roles, permissions, operations, objects and their relation type.
The objective of RBAC policy model is to strictly define the boundary of each role
and their hierarchy with respect to CRUD operations. The assignment from the user
and permissions to respective roles can be from one to one or many to many [7].
Xenserver provides six predefined roles; in which the permissions can be changed
based on requirements. However, no new roles can be added or removed; which
results in limited scalability in a distributed architecture environment [8]. Litera-
ture also reveals the extension of authorization model which can be extended to
have multiple entities according to the tenancy, for example single tenancy covers a
tuple of (Subject, Privilege and Object); which can be extended to multi-tenancy by
creating a tuple of (Issuer, Subject, Privilege, Object). Further, it can also be extended
to support multiple objects. However, this model is not suitable for the requirement
of orchestrating the infrastructure; as the dynamic extension by model affects the
A Comprehensive and Critical Analysis of Cross-Domain … 367
performance negatively [9]. The distribution of role based access control over various
Internet technologies is also introduced to overcome the limitation of dynamic exten-
sion, but this approach does not cover support for various policy models of different
components in cross-domain application deployment [10]. To provide the security
over distributed infrastructure, the protection mechanism for securely transferring
the attribute is proposed in CloudSeal [11]. It uses various symmetric key algorithms
for efficient attribute transformation including the revocation broadcast with the limi-
tation of group membership control for revocation library. The key policy along with
attributes based encryption provides more scalability and accountability to the user
with respect to access privileges for revocation [12]. On the other hand, the latency
of an attribute transfer request is a major concern for privacy. The centralized access
control policies are also discussed to implement access control to minimize latency
[13].
AcaaS is based on providing the “Access Control as a Service.” It delegates
the responsibility of access and policy management to Service Provider (SP); these
policies are then converted to enforceable policies to individual cloud provider. The
enforcement of policies and its level is determined by the user. Further, on behalf
of the user; SP accepts these policies and enforcement for Identity Provider (IdP)
to verify user credentials. However, it doesn’t reflect flexible delegation mechanism
and revocation requirements [14].
To support flexible delegation mechanism, Federated Identity Management (FIM)
provides access to cross-domain application deployment by Single Sign On (SSO).
It shares the attributes of identity with the service being accessed after a successful
login. Attributes are shared securely and do not contain any private or sensitive
information. It is architected by two entity roles viz Service Provider (SP) and Identity
Provider (IdP). Service Provider transmits authentication request from user to IdP.
Identity Provider shoulders validation of user credentials [15].
The strong collaboration and trust is required by FIM to validate credentials across
organizational boundaries. These boundaries encapsulate the trust and relationship
among multiple IdPs. In order to have seamless transactions there exist number
of FIM standard protocols viz. Security Assertion Markup Language (SAML),
Open Authentication (OAuth), and OpenID Connect (OIDC). The communication
in SAML is architected by XML to exchange authentication and authorization data
across organizational boundaries. In SAML, the IdP authenticates users and generates
the SAML assertion, which is then transmitted to browser along with the authoriza-
tion. This SAML assertion later is presented to the organizational boundary to support
identity sharing across cross-domain deployment. However, it is not optimized for
the large data transactions via multiple API calls [16]. Also, SAML centrify signature
which are not signed by Certification Authority are prone to cloning [17].
OAuth defines a protocol to enable user to grant authentication to the respective
services [Relying Parties (RP)] to access other authorized services [Identity Provider
(IdP)] with a required access token. The token contains the authorization scope and
timeline to access the required resource for the user’s accounts. In OAuth architecture,
the access tokens are prone to leak from RP through direct communication with IdP.
OAuth enables the ease for sharing the data by providing the transparent environment
along with a decision about validity of token to the user. Due to the said ease, the
phishing attacks are more common in authentication via OAuth [18, 19]. OpenID
connect is lightweight protocol as compared to SAML and OAuth. The objective of
OpenID connect is to provide a defined interface for user authentication along with
RP registration, signing, encryption of messages, and logout [20, 21].
Though, Federated Identity Management (FIM) is in use by many cross-domain
services; but it is greatly limited by users’ ability to interact with multiple IdPs
per session. In addition, federated identity has introduced many security and privacy
requirements as it promotes the sharing of credentials over loosely coupled networks.
FIM governs the Single Sign On (SSO) across cross-domain network. SSO regulates
sharing of identity information as per the domain requirement. However, sharing
identity information across the domains is a greater challenge to manage privacy.
Attribute aggregation in FIM implemented by David W. Chadwick and George
Inman, provides the ease to link all IdP accounts. The accessibility of the respective
service can use such accounts to automatically authorize the required service [22].
In this view, we have explored cross-domain application, i.e., cloud (OpenStack) to
investigate the implementation of FIM protocols. In OpenStack, every component is
authenticated and authorized by keystone (Identity Service). Keystone promotes role
based access policies. These policies are then used to generate token for authoriza-
tion. The authorization information contains user credentials and respective project
information. The authentication information is transmitted via SAML or OpenID
connect protocols to support FIM in OpenStack [23]. Federated Identity Manage-
ment (FIM) ease the user to access the resources for cross-domain deployment; with
no additional logins. The said ease also introduces the risk of unauthorized access to
every resource.
In order to mitigate the risk the FIM protocols are explored to share only required
attributes for cross-domain deployment. However, during the transmission of creden-
tials the privacy preservation among multiple protocols is a significant limitation.
The said protocols are required to be compatible and interoperable in the cross-
domain communication for identity attribute sharing with no privacy conflicts. There
exists a need to architect a protocol; which can preserve privacy across cross-domain
deployment by providing the compatible layer for existing protocols. To maintain
the privacy, the entity information transmission across the domain is a significant
challenge; the subsequent Sect. 3 presents the entity communication in FIM proto-
cols SAML, OAuth and OpenID connect and their adoption in OpenStack cloud
deployment.
3 Methodology
Entity communication in FIM protocols has major impact on privacy of identity

across the domains. We have investigated various research papers available in liter-
ature to examine and evaluate entity communication pathway for identity sharing.
The subsequent paragraph explains the same.
Various entities or roles are responsible to build communication channel in SAML,

OAuth and OpenID connect, which in turn drives privacy of identity attribute sharing.
SAML promotes the communication from user to Service Provider (SP) to access
required resource; the said request also discovers the Identity Provider (IdP). Service
Provider then diverts the user request to Single Sign On (SSO) process for validation
of user identity with IdP. Once the user identity is validated, an assertion is generated
to access the targeted resource by the user. In addition, the Service Provider is then
responsible for granting the resource access to the user; this grant exists till it is
explicitly removed. OAuth enables users to share resources among multiple SP, which
includes the IdP as well. As a result, OAuth is not suitable to implement Federate
Identity Management (FIM). In order to support FIM, OpenID is developed to support
the layer of authentication to OAuth. User requests are forwarded to OpenID endpoint
for authentication to OpenStack by successfully validating the user identity with IdP.
Identity Provider then redirects OpenID auth response to the user to access required
resource [3, 18, 23, 24].
During the communication with the endpoints, user authentication data is exposed
to various vulnerabilities introduced by communication channel. Table 1, briefly
assesses the privacy preservation in these protocols along with the implementation
in OpenStack deployment.
As depicted in Table 1, to support Federated Identity Management, the protocols
SAML, OAuth and OpenID connect implements RBAC policies with XML and JWT.
Also, multiple entities are coordinating with each other to share identity details. In
order to address privacy preservation in cross-domain applications deployment, these
protocols must process the identity information with maximum compatibility while
honoring the privacy of said protocols. These protocols are inherited in OpenStack
deployment as external plug-in to support FIM.
4 Conclusion
We have assessed the concept of identity management and the respective policies
and protocols to implement cross-domain applications. Firstly, it describes avail-
able Identity Management policies including Federated Identity Management and
its protocols. Then, the respective protocols are assessed and compared with respect
to privacy and implementation in OpenStack. During our assessment we have discov-
ered that SAML shares the identity attributes with no consent from user and user
authentication remains valid even if explicitly removed in some cases. Hence, it
opens up larger threat for sharing the identity attributes in cross-domain deploy-
ment; whereas in OAuth and OpenID user identities are visible to third parties after
required user consent and also revoked immediately after removing the access to the
user for specified resources. Further, in cross-domain deployment these protocols are
working together; where identity leaks are frequently recurring. In this view, there is
a necessity to devise a protocol to implement compatibility with no privacy conflicts
in cross-domain deployment.
Table 1 Assessment of SAML, OAuth, OpenID connect to deploy Cross Domain Applications
S. No. Reference Parameter SAML OAuth OpenID Connect
1 Maler and RBAC Token XML, (User XML, JWT, JWT, (User
Reed [24] Generation Identity (Authorization Identity is
Naik and format and encoded in (client_id, encoded in token)
Jenkins [21] Parameters token) client_secret
Kyungoh encoded in
and Hunjin token)
[20]
2 Naik and Entities Identity Relying Party, Identity Provider,
Jenkins [21] Provider, Identity Service Provider,
Sun and Service Provider, User User, Endpoints
Beznosov Provider, User for
[18] communication
3 Maler and Privacy and Tokens are Token are Cryptographically
Reed [24] Sharing of user Encrypted and Encrypted Signed token; The
Naik and attributes user attributes In OAuth user user attributes are
Jenkins [21] are share with attributes are visible to third
other domain visible to third parties with User
applications parties with Consent.
without user User Consent. Authentication
consent; and Authentication revoked
user revoked immediately, once
authentication immediately, explicitly removed
remains valid once explicitly
even if removed
explicitly
removed in
some cases
4 Maler and Federated Promoted Less promoted Promoted
Reed [24] identity Support
Naik and
Jenkins [21]
Sun and
Beznosov
[18]
Brian and
Rhee [19]
5 OpenStack Implementation Deployed Deployed Deployed as light
[23] in OpenStack SAML as OAuth as weight protocol
Naik and Deployment Integrated plug-in in plug-in on top of
Jenkins [21] module with Identity OAuth 2.0 in
Brian and WebSSO and Service of OpenStack. It is
Rhee [19] regsite. It is OpenStack. It more often used
Sun and not optimized is vulnerable to due to its light
Beznosov for the large phishing weight nature
[18] data attacks
transactions
via multiple
API call
References
1. A.A. Pandey, J.R. Saini, Online Identity Management techniques: identification and analysis
of flaws and standard methods. Int. J. Innov. Res. Dev. 1(3), 06–21 (2012)
2. A.A. Pandey, J.R. Saini, An analysis of online identity management techniques. Int. J. Res.
Comput. Appl. Manage. 1(3), 53–55 (2012)
3. Y. Yang, X. Chen, G. Wang, L. Cao, An identity and access management architecture in cloud,
in Seventh International Symposium on Computational Intelligence and Design (2014)
4. A.A. Pandey, J.R. Saini, An investigation of challenges to online federated identity management
systems. Int. J. Eng. Innov. Res. 1(2), 50–54 (2012)
5. A.A. Pandey, J.R. Saini, Identity management in e-governance. Int. J. Emerg. Trends Technol.
Comput. Sci. 2(5), 51–56 (2013)
6. A.A. Pandey, J.R. Saini, Development of privacy protecting identity management system for
E-Government in India. Int. J. Adv. Netw. Appl. 94–100 (2014)
7. D.F. Ferraiolo, R. Sandhu, S. Gavrila, D. Richard Kuhn, R. Chandramouli, Proposed NIST
standard for role-based access control. ACM Trans. Inf. Syst. Secur. 4(3), 224–274 (2001)
8. Citrix, Modifying Default Role based Access Control Permissions for XenServer, Citrix (14
Apr 2016). https://support.citrix.com/article/CTX126442. Accessed 19 Apr 2020
9. J.M. Alcaraz Calero, N. Edwards, J. Kirschnick, L. Wilcock, M. Wray, Toward a multi-tenancy
authorization system for cloud services. IEEE Secur. Priv. 48–55 (2010)
10. L. Hu, S. Ying, X. Jia, K. Zhao, Towards an approach of semantic access control for cloud
computing. Cloud Computing, CloudCom (2009), pp. 145–156
11. H. Xiong, X. Zhang, D. Yao, X. Wu, Y. Wen, Towards end-to-end secure content storage
and delivery with public cloud, in Proceedings of the second ACM conference on Data and
Application Security and Privacy (2012), pp. 257–266
12. S. Yu, C. Wang, K. Ren, W. Lou, Achieving secure, scalable, and fine-grained data access
control in cloud computing, in Proceedings IEEE INFOCOM (2010)
13. V. EcheverrÍa, L.M. Liebrock, D. Shin, Permission management system: permission as a
service in cloud computing, in IEEE 34th Annual Computer Software and Applications
Conference Workshops (2010)
14. R. Wu, X. Zhangy, G.-J. Ahn, H. Sharifi, H. Xieyz, AcaaS: access control as a service for IAAS
cloud, in International Conference on Social Computing (2013)
15. S. Koussa, Comparing the Top 3 federated identity providers: OpenID, OAuth, SAML, software
secured (November 2, 2018). https://www.softwaresecured.com/federated-identities-openid-
vs-saml-vs-oauth/. Accessed 18 Apr 2020
16. PingIdentity, An introduction to identity federation and the SAML standard, PingIdentity.
https://www.pingidentity.com/en/lp/saml-101.html. Accessed 18 Apr 2020
17. Swisskyrepo, SAML Injection, github (June 9 2019). https://github.com/swisskyrepo/Payloa
dsAllTheThings/tree/master/SAML%20Injection. Accessed 18 Apr 2020
18. S.-T. Sun, K. Beznosov, The devil is in the (implementation) details: an empirical analysis
of OAuth SSO systems, in Proceedings of the 2012 ACM conference on Computer and
Communications Security (2012), pp. 378–390
19. O.M. Brian, K.-H. Rhee, A secure social networking site based on OAuth implementation. J
Korea Multimedia Soc. 19(2), 308–315 (2016)
20. H.-K. Oh, S.-H. Jin, The security limitations of SSO in OpenID, in 10th International
Conference on Advanced Communication Technology (2008)
21. N. Naik, P. Jenkins, Securing digital identities in the cloud by selecting an opposite federated
identity management from saml OAuth and OpenID Connect, in 11th International Conference
on Research Challenges in Information Science (RCIS) (2017)
22. D.W. Chadwick, G. Inman, University of Kent has proposed. Attribute Aggregation In
Federated Identity Management. Computer 42(5) (2009)
23. OpenStack, 2019-07-22, OAuth1 1.0a, OpenStack. https://docs.openstack.org/keystone/latest/

admin/oauth1.html/. Accessed 19 Apr 2020
24. E. Maler, D. Reed, The venn of identity options and issues in federated identity management.
IEEE Secur. Priv. 6(2) (2008)
Concentric Morphology Model
in Detection of Masses
from Mammogram: A Study
Madhavi Pingili, M. Shankar Lingam, and E. G. Rajan
Abstract Breast cancer is a victor among the most comprehensively saw malig-
nancies among ladies around the world. Clump territory from mammogram helps
in right on time affirmation for breast tumor. A PC-aided detection structure which
will see also perceive the subverting lumps in person’s breast in an unmistakable and
fiscally able way is required. It is the fantasy and reason for experts to have a CAD
framework with most important affectability and low fake positives per picture (FPI).
Affirmation of clump from person’s breast is badly arranged considering its plenteous
geomorphological attributes and questionable edges. Clump territory execution can
be overhauled by utilizing productive preprocessing instrument and geomorpholog-
ical features of the clump area. As a clump makes, it enrages the breast parenchyma
and widens by stirring up different coaxial layers. Geomorphological assessment
of these coaxial layers is a foundation in clump conspicuous verification figuring.
Particular CAD structures utilizing coaxial morphology show exist as a bit of the
creation. In this article, an endeavor has made to plot a piece of the current CAD
frameworks which utilize coaxial morphology appear for precisely on time and exact
zone of clumps.
General Terms Biomedical · Automated system
Keywords Breast cancer · Mammography · PC-aided detection · Clump

detection · Coaxial morphology model
M. Pingili (B)
CMR Engineering College, Hyderabad, India
e-mail: madhavipingili@gmail.com
Research Scholar, University of Mysore, Mysore, India
M. Shankar Lingam
University of Mysore, Mysore, India
e-mail: shankumacharla@gmail.com
NIRDPR, Hyderabad, India
E. G. Rajan
RG-ISIT, MGNIRSA, Hyderabad, India
374 M. Pingili et al.
1 Introduction
Breast cancer becomes the most imperative contamination among ladies both in
making and made nations. Global flourishing connection encounters projects that
for all intents and purposes 360,000 individuals flop pitiably a year and basically
900,000 newly emerged cases are spoken to dependably. In the event that recognized
at a starting time, it is conceivable to fix breast mischief and passing rate can be
decreased considering. Mammogram assessment is the most extensively seen and
gifted technique that is explored for perceiving clumps.
Mammography is the course toward utilizing less thick abundancy X-Rays to
see human being’s breast and is utilized as a symptomatic gadget. Radiotherapist
appraisals mammogram and recognizes lumps. In prior days, radiotherapists used to
look over the mammograms to curb blunders in recognizing verification. Despite the
way that twofold investigating was done, human slips up were incredibly typical and
it is expensive. So as to vanquish it PC-aided detection (CAD) and PC-aided diagnosis
(CADx) structure had presented. PC users arrangement perceived the greater part and
solicitation of clumps into kind and risky clumps were finished by CADx framework.
PC bolstered structure framework was conveyed as a manual for the radiotherapist and
it helps the radiotherapist in obvious affirmation of vibrant assortments from the norm
and hinders the measure of missed injuries. Eventually, CAD and CADx structures
are treated as a total framework which will perceive and depict clumps. It is finished
for precisely on time and distinct separating verification of breast ailment yet now
there isn’t to such an extent as a solitary CAD and CADx framework with no sham
positives and counterfeit negatives. Therefore, heaps of examines are continuing
around there to have a free of charge framework.
Clump affirmation from mammographic expects an essential part in breast ailment
disclosure. Accurate conspicuous confirmation of clump makes it conceivable to
perceive breast advancement at a starting time and it likewise maintains a strategic
distance from a patient without peril experiencing pointless clinical shows. Clump
affirmation is finished by isolating two perspectives of a human breast: the cranio-
caudal (CC) see, which is the totally observe, and the mediolateral slanted (MLO)
see which will be side-view captured at an edge. Exemplary of CC view also MLO
see are appeared in Fig. 1 [1].
Here are two fundamental zones in a mammograph, breast area and non-breast
domain. The breast area comprises pectoral muscle, breast tissues, or clump, and
the non-breast locale comprises dull foundation and foundation things. A piece of
fundamental indications of breast contamination that radiotherapists examine for
bunches of downsized scale calcifications, clumps, and compositional turns. A clump
may be depicted as an expanse incorporating injury perceived in various projections
[2]. Clumps may delineate through their structure and edge properties. Hardenings are
little stores of calcium that show up as pitiful marvelous marks on the mammograph.
They will depict through their sort and dispersing features. Building turns will be
the impacts passed on by the augmentation of mayhem and heartbreaking spots in
the mammograph, whereas the checking is finished. Territory of clumps is most
Concentric Morphology Model in Detection of Masses … 375
Fig. 1 Two sights of breast: the craniocaudal (CC) sight, mediodorsal oblique (MLO) sight
bothersome than every other improvement responses considering the way that their
segments can be progressively clouded or like basic breast parenchyma [2].
The motivation behind this article is to give an outline of the clump territory
figurings and also a piece of the current computer-aided design structures. This is in
addition needed to make the idea of more researchers to the monstrosity of manifold
coaxial layer quantifies in clump affirmation estimations.
Whatever is left of the article is filtered through which takes after. By Sect. 2,
the generic outline of clump exposure calculations is depicted. Every phase in PC-
aided detection frameworks and the estimations utilized to evaluating the execution
of clump distinctive verification figurings are in addition examined. By Sect. 3, a
piece of the current PC sustained detection frameworks are considered. Vitality of
clump explicit trademark like manifold coaxial layer models is in like way outlined
here. Zone 4 closes the article.
2 Mass Identification Algorithm
CAD framework helps the radiotherapist in finding clumps on the mammographs.

The broad extent of specialists may taking a shot at the affirmation of clumps in
mammograms. Clumps happen fit as a fiddle and structure. Hypothesized clumps
have more probability of danger. A theorized clump is portrayed through marks
oozing through the edges of the clump. As the dangerous clumps are not conjectured,
revelation of non-guessed clumps will likewise fundamental. The basic advances
required in the divulgence PC sustained detection and finding (CADx) of diagnostic
irregularities is appeared in Fig. 2 [1].
Fig. 2 Phases in Input Mammogram

identification (CAD) and
analysis (CADx) of CAD
mammographic aberrations
Stage One
Stage Two
Output: Lesions Detected (Marks or ROIs)
CADx
Output: Likelihood of Malignancy or
Management Recommendation (1)
Preprocessing
Feature ExtracƟon
Feature SelecƟon
ClassificaƟon
(2)
All around any clump affirmation calculation comprises two phases. In phase
one, the truth is to recognize questionable injuries at a inflated affectability. This
phase is called as zone phase. In phase two, the truth is to diminish the measure
of fake positives rather diminishing the affectability significantly. Game plan of the
questionable district into clump or common tissue is accomplished in phase two.
So this phase is known as depiction or breast rate. All of these phases comprise
of pretreated, consolidate uprooting, highlight choice and plan phases. In certain
approaches, a fragment of the techniques may fuse major procedures or be jumped
without a doubt.
Preprocessing step includes racket clearing, improvement and division stages.
Unsettling influence clearing and improvement are utilized for overhauling the
picture quality. The motivation behind the division experience in mammographic
picture appraisal is to center zones of intrigue (ROIs) containing all breast assort-
ments from the standard from the regular breast tissue. Another motivation behind
the division is to find the questionable injury hopefuls through the zone of intrigue.
In the part uprooting experience of the clump disclosure estimation, the segments
are figured from the qualities of the zone of intrigue. A fragment of segments discon-
nected from the zones of energy for the diagnostic picture are not tremendous while
watched lonely, yet rather in mix in with different parts, they can be essential for
depiction. The best game plan of fragments for getting out fake positives and for
depicting sore sorts as perfect or bargaining are picked in the segment confirmation
phase. In the depiction phase, the questionable regions are delegated clump or ordi-
nary tissue on the explanation of picked segments. The affirmation structure yield is
the larger part. These clumps or locales of intrigue (ROI) are depicted as duty to the
affirmation structure (CADx) [3].
Counts utilized as a bit of conspicuous evidence time of clump divulgence calcula-
tion are either pixel based or district dependent. In the pixel dependent methods, parts
are disconnected for every pixel and named regular or questionable. In the region
dependent procedure, ROIs are divided, and starting now and into the foreseeable
future, portions are expelled from every territory, they are thusly used to or breast
rate the area as questionable or unquestionable.
Affectability and false positives per image (FPI) are the conditions utilized for
enumerating the execution of clump distinctive verification tallies. Affectability is
the complete sum of confirmed positive etchings among the all out of number of
wounds and counterfeit positives per picture (FPI) is the measure of fake positive
inscriptions among absolute number of pictures. An ensured positive check is a stamp
accomplished by the PC sustained detection framework that relates to the district of
a sore. A fake positive stamp is a check made by the PC sustained detection structure
that does not relate to the zone of a physical issue. A plot of affectability versus FPI
is known as a free-response receiver operating characteristic (FROC) plot and this is
commonly explored to account the execution of the territory calculation. A feasible
CAD framework will have most unmistakable affectability and least FPI.
Moved Database for Screening Mammography (DDSM) and Mammographic
Image Analyses Society (MIAS) database are the imperative of all things considered
utilized databases of mammogram pictures by specialists with a definitive goal of
assessment, affirmation and finish of breast hurt.
Any clump zone tally ought to is capable of see all clumps in a mammograph
without erroneously picking an average tissue as clump. By and by, the basic trade-
off among affectability and FPI is an endeavor in order to decrease FPI will curb
affectability. So the main thought must be taken while endeavoring to diminish FPI.
FPI decay must be rehearsed rather decreasing affectability. Neither of the current
structure is accomplished 100% unmistakable verification exactness.
Different business CAD frameworks are accessible. Despite the way that they
have accomplished most vital accuracy in perceiving scaled down scale calcifica-
tions, clump affirmation precision must be pushed ahead. Two of the business PC
maintained exposure and affirmation structures are: R2 technology copy organizer
with a declared clump affirmation execution of 85.7% with 1.42 FPI and Brainy
System Software, Inc. (ISSI) with an uncovered clump ID execution of 87.4% with
3.32 FPI [1]. Most unprecedented precision can be gotten by picking fitting means
all of the times of clump ID calculation. A large portion of the current checks has
clump unmistakable evidence exactness of 90% at 1–10 FPI. So yet there is space
for updating the region pace of clumps in PC sustained detection.
3 Computer-Aided Detection of Clumps

from Mammographs
There is sweeping arrangement on the progress and examination of PC-aided detec-

tion frameworks in diagnosis. Various plans have been made for mammographic
clump ID. Unquestionable approaches have been utilized as a bit of these plans.
These approaches utilized obvious mix of picture managing structures for under-
standing the two times of clump conspicuous evidence tally. None of these systems
have winning concerning accomplishing a 100% exactness. Bigger part of these
methods was concentrating on the picture segments of the mammograph as opposed
to seeing attributes of clumps in mammogram, and these structures were not very
convincing in accomplishing most preposterous region rate.
Early affirmation plans utilized clear overhauling or sifting frameworks. Later,
some unusual strategies were made. Brake and Karssmeijer [4] projected clump
divulgence plot utilizing single and multiscale styles; Mudigonda et al. [5] broke
down the utilization of a thickness cutting philosophy to fragment the section of
benefits (ROIs). These techniques have compassions nearly 90% with 1–10 coun-
terfeit positives for each picture (FPI). A piece of the late assessments pondered
vital features of clumps for affirmation. Timp and Karssmeijer [6] dissected the
break vagaries between two mammographic pictures in include space. They will be
profitable in discovering little injuries and compositional turns.
3.1 Custom of MCL Norms in Clump Recognition
Movement of a clump bothers the breast parenchyma and distributed by making

coaxial layers. Clumps have an emphasized central locale by few unique dimmer
coaxial layers. So edge and geomorphological parts are for all intents and purposes
every so often utilized for clump attestation. Contemplating this Eltonsy et al. [7]
examined the geomorphological portrayal of the layers of clump for stirring up a
motorized arrangement for ID of clumps. This procedure centers around the disclo-
sure time of clump affirmation figuring through preview mammograph to pick ques-
tionable breast areas that will comprise dangerous clumps. These areas are the open
doors for the ensuing phase (gathering) of the PC-aided identification framework [8].
By this store up from the beginning in the wake of beating, the breast locale into
50 areas with various force expands, a framework lead is related with make another
picture with decreased amount of pellets. By then, these grainy locales are dissected
utilizing the preparing segment to perceive the measure of coaxial layers. At long last,
coaxial get-together with at most significant amount of coaxial layers and with wide
range of speculation are picked as the last questionable zones. Speculation essential
is associated with the grounds that it is clinically settled that frightful clumps will
when all is said in done packaging hypotheses [7].
This technique has been tried utilizing mammographs from Digital Database for
Screening Mammography (DDSM). Craniocaudally perspectives of the mammo-
graphs have been explored as a bit of this assessment. This procedure when executed
utilizing 42 biopsy indicated clumps as a bit of aggregate (21 harmful clumps and
the rest of the 21 big-hearted clumps) itemized 85.7% affectability with a normal of
0.53 fake positives for each picture [7]. After progressively far-reaching examina-
tion of the structure on a more prominent game plan of mammographs, the watched
execution was 92% affectability by 3.26 FPI [9].
Tourassi et al. [9] did a development of the exertion completed by Eltonsy et al.
[7]. They projected a sham positive diminishment system utilizing a phony neural
structure that cementings highlight and information-based appraisal of questionable
mammographic locales. This game plan will go about as the second time of the
CAD structure (game plan). The framework plays out the going with strides: (1)
Preprocessing and division of questionable areas, (2) feature dependent and learning-
dependent assessment of questionable region and (3) classification for false positive
diminishing.
The preprocessing segment utilized the game plan made by Eltonsy et al. [7] in
order to see and parcel the questionable region. Highlight evaluation examines both
the directional and textural qualities of the questionable zones. Highlight assessment
module gets a hero among the most outstanding depictions of dangerous clumps.
Comparing to include appraisal, an information relied unit is utilized for purifying
the sham positive diminishment handle. This unit utilizes an orientation library where
mammographic issues with acknowledged certainty are dealt with, and essential data
is utilized as the resemblance model among a solicitation case and recorded case.
Information relied assessment passes on a choice record that trials the comparative
correspondence among the solicitation case and the chronicled case. The sections
of highlight assessment stage and the choice record from learning relied appraisal
stage are joined into an extraordinary assurance utilizing a back affecting created
neural structure. This assessment utilized mammographs from Digital Database for
Screening Mammography (DDSM). Course of action is spoken to have accomplished
87.4% affectability with 1.8 sham positives for each picture [9].
Eltonsy et al. [10] projected an other coaxial layers (MCL) dependent include
to perceive clumps in mammographs. The projected region plan is a pick manu-
facture estimation that depends concerning a geomorphological prototype of breast
danger progression. Geomorphological evaluation of the coaxial layer prototype is
the foundation of MCL affirmation figuring. The estimation includes three stages:
Firstly, the breast areas are pretreated by division and pulverization philosophies. By
at that point, the questionable central regions are perceived utilizing learning-based
thinking. At long last, two remarkable measures are related to takeout counterfeit
positives. In the zone sort out following to limiting the central areas with the question-
able morphology, least segment standard is utilized to perform beginning evacuation
of questionable districts. By then manifold coaxial layer prototypes are related for
picking clumps. False positive decreasing is drilled by the assessment of relative
repeat and least segment standard.
They utilized mammogram pictures from Digital Database for Screening

Mammography (DDSM) database for their tests. They picked 270 CC perspectives
of mammographic cases with biopsy displayed dangerous clumps. The slightest bit
of the cases was utilized for preparing and the other half to test. The execution nitty
gritty by the producers is 92% affectability to frightful clumps at 5.4 FPI [10].
Gao et al. [11] presented another strategy which consolidates coaxial morphology
prototype (MCL) with geomorphological part analysis (MCA). This framework
vanquishes the setbacks in MCL rules. Exploration of MCA for pretreating has broad-
ened the conspicuous verification exactness. Here, MCA is utilized for declining
the picture into fragments smooth part and surface section. Promoted preparing is
accomplished on piece-wise soft part. By then clumps are perceived utilizing coaxial
layer models. Gigantic phases in this figuring are: pretreating, geomorphological part
abstraction and regulate relied disclosure.
In this strategy, mammograph pictures are pretreated utilizing MCA. MCA is
a decay system considering sad depiction of information. This technique depends
upon the uncertainty that every pennant is a straight blend of a few nuclear signs
of progressively clear commencement. For each nuclear standard lead to be isolated
there exists a jargon that empowers its headway utilizing a desolate depiction. In
like way, it is ordinary that the different word references are wasteful in tending
to exchange practices in the blend. This word reference close by an appropriate
pursuit calculation sifting for the sparsest depiction prompts the ideal fragment.
Over complete word references are utilized as a bit of this strategy. Undecimated
wavelet transforms and local DCT will be the best selection of changes for word
reference depictions as it possibly withdraws surface part and segmented smooth
section. Piece-wise soft parts are then restricted to various power layers utilizing
different force limits. The entirety of the areas in these free layers is utilized for
extra preparing. By then, geomorphological parts are ousted through these sheets to
pick the questionable sections. Geomorphological fragments utilized as a bit of this
strategy join healthiness, eccentricity, degree and differentiation. Supervise-based
affirmation stage is utilized for extra launch of superfluous districts. For lessening
sham, positives least division perspective and appraisal of relative repeat are utilized
[11].
This course of action was tried utilizing 100 merciful and 50 bargaining risks
from the Digital Database for Screening Mammography database. It declared the
affectability of such course of action is 99% in sabotaging, 88% in liberal and 95.3%
in a wide extent of risks with 2.7, 3.1 and 2.83 FPI solely [11].
A synopsis of the execution of techniques portrayed here are making review is
given in Table 1. It is ludicrous to cause a connection among these specific figuring’s
as they too had not organized and tried identical data records. This might be viewed
by the chart that the utilization of various data records with a tantamount figuring
offsets undeniable outputs.
A CAD will have most outrageous execution if the affectability of the structure
is most noteworthy and FPI is least. PC supported structure system chooses the
disclosure accuracy of breast development area. Therefore, the prerequisite for a
PC-aided identification system through most outrageous affectability and least or
Table 1 Synopsis of clump

Author No. of images Sensitivity (%) FPI
identification algorithms
exploring MCL norms Eltonsy et al., 2004 42 85.7 0.53
Eltonsy et al., 2004 150 92 3.26
Tourassi et al., 2005 150 87.4 1.8
Eltonsy et al., 2005 270 92.1 5.4
Xinbo Gao et al., 2010 150 95.3 2.83
no FPI is maximum. Advent of MCL standards in clump acknowledgment contrive

has enhanced the precision of clump recognizable proof. Amid the current scheme’s
revelation plot exploring geomorphological module scrutiny and coaxial layer show
have most extraordinary affectability by exchanging off FPI. It must be seen that this
arrangement has achieved 99% affectability for compromising clumps with 2.7 FPI.
Therefore, decline in FPI can enhance the precision [12].
Primary affirmation of breast advancement is essential to diminish the passing

frequency. With a specific extreme target to make it conceivable, a robotized terri-
tory structure is required. Different clump disclosure plans have been made now
a PC-aided identification framework with 100% precision notwithstanding every-
thing stays as a specialist’s fantasy. In this article, an attempt is performed to see
a clump affirmation plot which has spurred affectability and declined fake positive
rate. Through the making layout, it is originated that the greater part of the present
PC-aided identification structures has affectability about 90% with 1–10 counterfeit
positives for each picture. Presentation of MCL models has surged the exactness
of clump disclosure tallies. The clump affirmation plan utilizing a mix of geomor-
phological bit assessment (MCA) and diverse coaxial layer rules (MCL) have most
exceptional affectability (99% with dangerous clumps with 2.7 FPI). Therefore, this
clump territory figuring has unbelievable forthcoming update prospects.
Decay of counterfeit positives will update the execution considering. The utiliza-
tion of clump zone explicit qualities has an incredible part in unquestionable proof
and zone of clumps. Perceiving proof of fitting segment will curb the false positives. It
is common that the advancement of novel sections, for example, Gaussian allotment
qualities of clump areas will lessen false positives as power dispersal of clump locale
look like 2D prognosis of Gaussian surface. Utilization of restored word references
for MCA can propel redesign the affirmation happens as intended.
References
1. M.P. Sampat, M.K. Markey, A.C. Bovik, Computer-Aided Detection and Diagnosis in
Mammography (Elsvier Academic Press, 2005)
2. B. Senthil Kumar, G. Umamaheswari, A review on pc Aided Detection and diagnosis-towards
the detection of breast cancer. Eur. J. Sci. Res. (2011)
3. J. Tang, R.M. Rangayyan, J. Xu, I. El-Naqa, Y. Yang, Computer-aided detection and diagnosis
of breast cancer with mammography: recent advances. IEEE Trans. Inf. Technol. Biomed.
(2009)
4. G.M. te Brake, N. Karssemeijer, Single and multiscale detection of clumps in digital
mammograms. IEEE Trans. Med. Imag. (1999)
5. N.R. Mudigonda, R.M. Rangayyan, J.E. Leo Desautels, Detection of breast clumpses in
mammograms by density slicing and texture flow-field analysis. IEEE Trans. Med. Imag.
(2001)
6. S. Timp, N. Karssemeijer, Interval change analysis to improve pc aided detection in
mammography. Med. Image Anal. (2006)
7. N. Eltonsy, H. Erin Rickard, G. Tourassi, A. Elmaghraby, Geomorphological coaxial layer
analysis for detection of suspicious clumps in screening mammograms. IEEE Eng. Med. Biol.
Soc. (2004)
8. J.L. Starck, M. Elad, D. Donoho, Redundant multiscale transforms and their application for
geomorphological component analysis. Adv. Imag. Electron Phys. (2004)
9. G.D. Tourassi, N.H. Eltonsy, J.H. Graham, C.E. Floyd, A.S. Elmaghraby,. Feature and knowl-
edge based analysis for reduction of false positives in the computerized detection of clumps in
screening mammography. IEEE Eng. Med. Biol. (2005)
10. N.H. Eltonsy, G.D. Tourassi, A.S. Elmaghraby, A coaxial morphology model for the detection
of clumps in mammography. IEEE Trans. Med. Imag. (2007)
11. X. Gao, Y. Wang, X. Li, D. Tao, On combining geomorphological component analysis and
concentric morphology model for mammographic clump detection. IEEE Trans. Biomed. Eng.
(2010)
12. Z.A. Sasinas Alias Haritha, R. Shreeja, Use of coaxial morphology model in detection of
clumps from mammograms: a review, in Conference on Advances in Computational Techniques
(CACT) 2011 Proceedings published in International Journal of Computer Applications®
(IJCA)
A Survey on DevOps Techniques Used
in Cloud-Based IOT Mashups
M. Ganeshan and P. Vigneshwaran
Abstract Development life cycle involves writing scripts, testing, bug finding,
recoding, documentation, and deployment, and this activity consumes more time.
DevOps looks at automating the entire process. DevOps is a total software and
template delivery system that gives importance to the interaction between opera-
tions and development fellows. Choosing the right technologies and crafting an IoT
infrastructure need a sophisticated comprehensive methodology which can be diffi-
cult. DevOps consists of a set of people-oriented, technology-oriented, organizational
methods, and tools that face challenges one after the other so that a smooth, easy, and
flexible system is obtained. There are many STS practices involved in the develop-
ment and crafting cloud-based IOT Mashups. We provide few theoretical points of
the major IoT threats mentioned in the latest papers and a brief of research activity
undertaken in DevOps world to overcome these demerits. This paper briefs the corre-
lation between DevOps technical practices and the tools and engineering that enable
uninterrupted integration and transmission, test mechanization, and quick implemen-
M. Ganeshan (B)
Department of Computer Science and Engineering, Jain Deemed to be University, Bengaluru,
India
e-mail: link2ganeshan@gmail.com
P. Vigneshwaran
Department of Computer Science and Engineering, Faculty of Engineering and Technology, Jain
University, Bengaluru, India
e-mail: vigneshwaran05@gmail.com
384 M. Ganeshan and P. Vigneshwaran
tation. This paper also gives insight on the tools that unite to give a “tool chain” of
technology and services to excel the modular tasks of the template deployment and
schedule of activities.
Keywords IOT · DevOps · Mashups · Cloud · Technical automation practices
1 Basic Concepts
We present a theoretical structure of basic principles of DevOps framework. The goal

of our structured plan of study of the papers is to gain an understanding of principles
of DevOps, based on the literature, to aid the analysis and understanding of DevOps
challenges.
1.1 Scholarly Reviewed Literature
The survey questions mentioned in Table 1 are answered and helpful for completion
of the survey paper. After selection of papers, the papers are divided into seven
groups, such as methodologies, result report, influence, fundamentals, challenges,
knowledge, and implementation as shown in Fig. 1. The divisions are explained in
the following list:
• Methodologies: Gives tools, methods, templates and plans to improve DevOps.
These give ground rules for choosing implementation automation [8], metrics
values plays important role in deciding the nature of deployment [9], how to aid
communication among devices across the world using sophisticated network for
cloud based technology [10], and how to reconstruct the old infrastructure to latest
supportive platform reducing faults at initial stage and make the system reliable
for IoT devices connected on cloud network [11].
Table 1 Survey research questions

Survey question Source
SQ1: What does DevOps culture in IoT Mashups mean? References [1, 2]
SQ2: What are the problems that lead to the need for implementation of References [3–6]
DevOps technology?
SQ3: What are the major advantages of utilizing DevOps culture in IoT References [1, 2]
mashups?
SQ4: What are the probable major problems/limitations of using DevOps References [3, 4, 6]
technology/tools?
SQ5 What is continuous integration/continuous deployment? References [3, 7]
SQ6 What problems are encountered when implementing DevOps? References [5, 7]
A Survey on DevOps Techniques Used in Cloud-Based IOT Mashups 385
Fig. 1 Review literature

Qty:15
categories
3
5 Quantity
11
17 Category
19
24
26 Qty:8
27 2 Qty:7 Qty:7
30 6 25 4 Qty:6
32 7 39 14 10
38 12 40 16 15 Qty:4
46 18 41 21 28 1 Qty:3
47 20 42 22 29 8 13
48 33 43 36 31 9 34
50 34 49 37 35 23 45
Method Result. Influen Fundame Knowled Implem
Challenges
oligies report ce ntals ge entaion
• Result report: Points toward the end result of companies implementing DevOps
culture in the deployment of APIs between private and public cloud components.
• Influence: Deals with how DevOps culture improves the efficiency in the area
of APIs development between the private architecture and cloud-based network
architecture [12], there is importance of designing security criteria before the
implementation of crafted template [13], or the impact of DevOps culture in the
APIs development in specific cases of real-time computation of data research [14].
• Fundamentals: Covers the basic theory needed for cultivating DevOps culture
in continuous delivery and data integration using the concept of the deployment
pipeline [15].
• Challenges: Explains the threats present in implementation of DevOps culture.
Few sources deal with specific threats like DevOps used in old domains and new
domains supporting the new technologies [16], most of the devices connected in
IoT are embedded systems in nature with multiple computation nature [17], or
communication APIs between different IoT mashup devices with heterogeneous
characteristics [18].
• Knowledge: Deals with the problems faced while training people in DevOps.
Papers in this sub-division introduce teaching styles [19], as wel l as explain
the key attributes, abilities and exposure needed for DevOps technical engineers,
leads and experts [20, 21].
• Implementation: This sub-division deals with the different prototypes for DevOps
culture tools for deployment and scripting the templates [3] and shows how a
company identifies and shortlists a set of standards tools to implement and gauge
respective tools in DevOps activity [22].
2 Role of DevOps in IOT Mashup
DevOps, with its focus on reducing delivery cycle times, can increase the speed at
which IoT heterogenous devices are available in the technology market throughout
the globe. Google states that the number of IoT heterogenous devices will exceed
non-IoT by 2025. The technologies and principles of IoT will have a major impact on
organizations with regard to planning business, management of risk, and design of
architecture and network. Any error occurring in any module regarding security and
data integration of IoT Mashup architecture can affect the whole DevOps process.
3 Implementing DevOps Architecture Style
The course in the needs of DevOps architecture is the utilization of connected

portable microservices with suitable APIs for the micro services engineering along
with DevOps methodologies [23]. An appropriate tool to orchestrate virtualization
mentioned in Table 2 can be decided by the DevOps team to achieve a continuous
integration and a recovery system.
Table 2 DevOps tools to orchestrate virtualization technology

Tools Description
CloudSlang CloudSlang is an open-source tool, we can use or customize ready-made
YAML-based workflow
Dockers Docker, and its open-source its being widely adopted for real-world data
backs
OpenVZ Open-source container-based virtualization for Linux.
Kubernetes Kubernetes is a free package for container-composition system used in
automation of APIs and monitoring of code
Containership Containership simplify the build/test/deploy pipelines in DevOps
Packer Packer is user friendly used in automation of machine image creation of
different types
Linux Containers Image building tool for LXC/LXD, Support for a lot of distributions and
architectures
Nomads Nomad is an open-source clustered scheduling engine. It can run different
applications of micro-service, batch, containerized and non-containerized
applications
Table 3 Survey on IoT Mashups security

Year Reference Contributions
2016 [24] A brief discussion of vulnerabilities faced by the edge side layer of IoT
2017 [25] Survey on using end points for IoT devices on cloud computing planform
2017 [26] Discussion on similarities and differences between IoT orchestration using
cloud computing and fog computing
2017 [27] A brief discussion on most relevant limitations of IoT devices
2017 [28] Discussion on security key factors to be improved on targeted endpoints in
network-based IoT services for heterogeneous devices
2017 [29] Security issues related to the IoT middle ware
2018 [30] Security mechanism for IoT security like SDN and NFB
2019 [31] Trust management techniques for Internet of Things
3.1 DevOps Tools for Template/Scripts/Code Management
Script/code management DevOps system is in need of tools mostly intended to

improve relationship among developers of different domains. DevOps tools are
most required for today’s world of bigdata to deploy continuous integration/delivery
system. While using traditional approach in deployment and code management,
the delivery concepts cannot match the data handling for present world data/code
processing requirement. Even developer and operational team cannot manually
conduct source code/data management. Therefore, automation scripts are needed
to obtain workflow information that can affect the activities carried out in the work
frame. For any IoT Mashup instance, a code error can create a bug which may result
workflow disturbance and data leak that affects system security [8] and Table 3 lists
some of the IoT Mashups Security issues faced in the years mentioned against them.
3.2 Thingworx Versus AWS in IoT Platform
While comparing IoT platforms, choosing a better platform is not an issue since you
do not have to worry about choosing among options. You can pick Thingworx [32]
for the industry expertise and then use AWS for your cloud platform as an integral
solution. Thingworx is a third-party development tool for implementing mechanical
drives in industry platform for high-level IoT architecture.
AWS IoT Core can automate devices in report generation and limitless logs main-
tained using S3 services [33]. Management and maintenance are cost effective, with
less waiting time. The communication can be upscaled to as many heterogenous
devices as required. AWS IoT Core is compatible with protocols like HTTP, MQTT
and WebSocket’s [34]. Information security is provided by a protocol TLS [35].
4 Conclusion
DevOps culture in deployment and orchestration automation tools for IoT should
be sophisticated to provide information exchange through well-designed APIs to
confirm lesser downtime in information transfer between IoT mashup heterogenous
devices. Edge location in cloud platform must be well defined by appropriate proto-
cols for specific DevOps tools. In this survey, we have explored the different tech-
nologies on DevOps culture to automate the deployment, orchestration, monitoring
and recovery approaches for IoT Mashup devices on cloud platform by studying a
systematic mapping approach. The references were categorized as shown in Table 4
for conceptual study. Table 5 gives the list of cloud IoT features.
Many challenges were realized in building a data mashup service between IoT
smart gadgets. But, the number of electronic gadgets and mechanical drives are
increasing rapidly and are inter-dependent. Wearable technology in coming days is
ready with big data lake for analysis. Every human being has increased his count
of personal electronic smart devices that are connected to network for computation.
All the smart devices require data computation and the process for analyzing the
data needs DevOps culture. This rapid growth of technology and IoT devices needs
DevOps culture in deployment and Ochestration to ensure friendly Web services.
Technically, well-defined research is required in DevOps culture to provide the
services with reliable performance. Cloud-based services and tools for IoT Mashups
is helpful for conducting high-end workflow. AWS IoT Core provides end-to-end
services and embeds DevOps culture in deployment, monitoring, and recovery to
increase workflow reliability and efficiency by applying more focus on the DevOps
tool-based automated APIs and devices configuration from cross platform.
Table 4 Study of references

Reference Category Concepts Goal
References [36, 37] Source code Automate the versioning on Human
management DevOps tools collaboration
Reference [38] Improve the DevOps culture Continuous
and methods in scripts/code delivery and
management recovery using
tool for code
management
References [18] Overcoming the barriers Decision support
between different system (DSS) is
Tool platform implemented for
specific
technologies to
obtain desired
results
References [1, 20, 36, Continuous Develop APIs with can handle Continuous
39] integration continuous information flow delivery;
between domains
References [1, 12, 19, DevOps can be used in testing Optimize the
40, 41] environment method to secure
devices in IoT
domain, using
cross-layered
network
Reference [42] Deployment Real time monitoring/recovery Continuous
automation delivery
References [37] Implement continuous delivery Reliability
and integration of system
References [43–45] Infrastructure as code To improve APIs
Virtualization, containerization connectivity for
cloud services, automation smooth
information flow
References [1, 40] Monitoring and DevOps tools can improve the Reliability
References [3, 10, 26, logging scalability of instances. Provide To migrate
27, 35] reliable work platform with existing
automatic metric value and infrastructure to
alerting new domain on
cloud platform
Reference [24] IoT as a network Three layers in technological Bridge between
of networks innovation—cloud-pipe-device cloud and IoT
Table 5 Survey on cloud IoT features

IoT cloud provider Network protocols Data store Push not Geo BLOB Trigger
Amazon REST/MQTT ✓ ✓ ✓ ✓ ✓
Azure MQTT/AMQP/REST ✓ ✓ ✓ ✓ ✓
Google REST ✓ ✓ ✓ ✓ ✓
Cloud-foundry REST ✓ ✖ n/a ✓ ✖
Kinvey REST ✓ ✓ n/a ✓ ✓
Heroku MQTT ✓ ✖ n/a ✓ ✖
Parse REST ✓ ✓ ✓ ✓ ✓
Bluemix MQTT ✓ ✓ ✓ ✖ ✓
Dream-factory REST ✓ ✓ n/a ✓ ✓
References
1. D. Cukier, DevOps patterns to scale web applications using cloud services, in Proceedings of
the 2013 Companion Publication for Conference on Systems, Programming, & Applications:
Software for Humanity (SPLASH ’13) (ACM, 2013), pp. 143–152. (Code: A35)
2. A. Khattab, H.S. Hamza, S.M. Khattab, Enhancing user experience in IoT mashup using
semantic technology, in 2018 IEEE Global Conference on Internet of Things (GCIoT),
Alexandria, Egypt (2018), pp. 1–6. https://doi.org/10.1109/gciot.2018.8620164
3. K. Nybom, J. Smeds, I. Porres, On the impact of mixing responsibilities between Devs and Ops,
in International Conference on Agile Software Development (XP 2016) (Springer International
Publishing, 2016), pp. 131–143. (Code: S18)
4. J. Roche, Adopting DevOps practices in quality assurance. Commun. ACM 56(11), 38–43
(2013). (Code: A74)
5. F.M.A. Erich, C. Amrit, M. Daneva, A qualitative study of DevOps usage in practice. J. Softw.
Evol. Proc. 29(6), e1885 (2017). (Published online in Wiley InterScience, www.interscience.
wiley.com, https://doi.org/10.1002/smr)
6. K. Beck, C. Andres, Extreme Programming Explained: Embrace Change (Addison-Wesley
Professional)
7. A. Rahman, Characteristics of defective infrastructure as code scripts in DevOps, in Proceed-
ings of the 40th International Conference on Software Engineering (ICSE ’18) (ACM, 2018),
pp. 476–479. (Code: A806)
8. J. Wettinger, V. Andrikopoulos, F. Leymann, Automated capturing and systematic usage of
DevOps knowledge for cloud applications, in 2015 IEEE International Conference on Cloud
Engineering (IEEE, 2015), pp. 60–65. (Code: A61)
9. N. Forsgren, M. Kersten, DevOps metrics. Commun. ACM 61(4), 44–48 (2018). (Code: B21)
10. K. Magoutis, C. Papoulas, A. Papaioannou, F. Karniavoura, D.-G. Akestoridis, N. Parotsidis, M.
Korozi, A. Leonidis, S. Ntoa, C. Stephanidis, Design and implementation of a social networking
platform for cloud deployment specialists. J. Internet Serv. Appl. 6(1) (2015). (Code: S3)
11. N. Basiri, N. Behnam, R. de Rooij, L. Hochstein, L. Kosewski, J. Reynolds, C. Rosenthal, Chaos
engineering. IEEE Softw. 33(3), 35–41 (2016). (Code: A76); J. Clerk Maxwell, A Treatise on
Electricity and Magnetism, 3rd edn., vol. 2 (Clarendon, Oxford, 1892), pp. 68–73
12. M. Shahin, M.A. Babar, L. Zhu, The intersection of continuous deployment and architecting
process: practitioners’ perspectives, in Proceedings of the 10th ACM/IEEE International
Symposium on Empirical Software Engineering and Measurement (ESEM ’16) (ACM, 2016),
pp. 44:1–44:10. (Code: A64)
13. L. Zhu, D. Xu, A. B. Tran, X. Xu, L. Bass, I. Weber, S. Dwarakanathan, Achieving reliable
high-frequency releases in cloud environments. IEEE Softw. 32(2), 73–80 (2015). (Code: B12)
14. M. de Bayser, L.G. Azevedo, R. Cerqueira, ResearchOps: the case for DevOps in scientific
applications, in 2015 IFIP/IEEE International Symposium on Integrated Network Management
(IM) (2015), pp. 1398–1404. (Code: I40)
15. J. Humble, J. Molesky, Why enterprises must adopt DevOps to enable continuous delivery.
Cutter IT J. 24(8), 6 (2011). (Code: B5)
16. T. Laukkarinen, K. Kuusinen, T. Mikkonen, DevOps in regulated software development: case
medical devices, in Proceedings of the 39th International Conference on Software Engineering:
New Ideas and Emerging Results Track (ICSE-NIER ’17) (IEEE Press, 2017), pp. 15–18. (Code:
A26)
17. L.E. Lwakatare, T. Karvonen, T. Sauvola, P. Kuvaja, H.H. Olsson, J. Bosch, M. Oivo, Towards
DevOps in the embedded systems domain: why is it so hard? in 49th Hawaii International
Conference on System Sciences (HICSS) (IEEE, 2016), pp. 5437–5446. (Code: A42)
18. E. Diel, S. Marczak, D.S. Cruzes, Communication challenges and strategies in distributed
DevOps, in 11th IEEE International Conference on Global Software Engineering (ICGSE)
(2016), pp. 24–28. (Code: I19)
19. H.B. Christensen, Teaching DevOps and cloud computing using a cognitive apprenticeship
and story-telling approach, in Proceedings of the 2016 ACM Conference on Innovation and
Technology in Computer Science Education (ITiCSE ’16) (ACM, 2016), pp. 174–179. (Code:
A47)
20. S. Chung, S. Bang, Identifying knowledge, skills, and abilities (KSA) for DevOps-aware server-
side web application with the grounded theory. J. Comput. Sci. Coll. 32(1), 110–116 (2016).
(Code: A16)
21. W. Hussain, T. Clear, S. MacDonell, Emerging trends for global DevOps: A New Zealand
Perspective, in Proceedings of the 12th International Conference on Global Software
Engineering (ICGSE ’17) (IEEE Press, 2017), pp. 21–30. (Code: A25)
22. B. Snyder, B. Curtis, Using analytics to guide improvement during an agile/DevOps
transformation. IEEE Softw. 35(1), 78–83 (2018). (Code: I7)
23. A. Sill, Cloud standards and the spectrum of development. IEEE Cloud Comput. 1(3), 15–19
(2014). (Code: I67)
24. A. Mosenia, N.K. Jha, A comprehensive study of security of internet-of-things. IEEE Trans.
Emerg. Topics Comput. 5(4), 586–602 (2017)
25. W. Yu, F. Liang, X. He, W.G. Hatcher, C. Lu, J. Lin, X. Yang, A survey on the edge computing
for the internet of things. IEEE Access 6, 6900–6919 (2018)
26. J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, W. Zhao, A survey on internet of things: archi-
tecture, enabling technologies, security and privacy, and applications. IEEE Internet Things J.
4(5), 1125–1142 (2017)
27. Y. Yang, L. Wu, G. Yin, L. Li, H. Zhao, A survey on security and privacy issues in internet-of-
things. IEEE Internet Things J. 4(5), 1250–1258 (2017)
28. L. Chen, S. Thombre, K. Järvinen, E.S. Lohan, A. Alén-Savikko, H. Leppäkoski, M.Z.H.
Bhuiyan, S. Bu-Pasha, G.N. Ferrara, S. Honkala, J. Lindqvist, L. Ruotsalainen, P. Korpisaari,
H. Kuusniemi, Robustness, security and privacy in location-based services for future IoT: a
survey. IEEE Access 5, 8956–8977 (2017)
29. A.H. Ngu, M. Gutierrez, V. Metsis, S. Nepal, Q.Z. Sheng, IoT middleware: a survey on issues
and enabling technologies. IEEE Internet Things J. 4(1), 1–20 (2017)
30. I. Farris, T. Taleb, Y. Khettab, J. Song, A survey on emerging SDN and NFV security
mechanisms for IoT systems. IEEE Commun. Surv. Tutorials 21(1), 812–837 (2018)
31. U. Din, M. Guizani, B.-S. Kim, S. Hassan, M.K. Khan, Trust management techniques for the
internet of things: a survey. IEEE Access 7, 29763–29787 (2019)
32. L. Bass, I. Weber, L. Zhu, DevOps: A Software Architect’s Perspective (Addison-Wesley
Professional, 2015)
33. B. Beyer, C. Jones, J. Petoff, N.R. Murphy, Site Reliability Engineering: How Google Runs
Production Systems (O’Reilly Media, 2016)
34. J. Bonér, D. Farley, R. Kuhn, M. Thompson, The Reactive Manifesto (2014). https://www.rea
ctivemanifesto.org/. Accessed Aug 2018
35. S. Neely, S. Stolt, Continuous delivery? Easy! Just change everything (well, maybe it is not
that easy), in 2013 Agile Conference (2013), pp. 121–128. (Code: B23)
36. X. Bai, M. Li, D. Pei, S. Li, D. Ye, Continuous delivery of personalized assessment and feedback
in agile software engineering projects, in Proceedings of the 40th International Conference on
Software Engineering: Software Engineering Education and Training (ICSE-SEET ’18) (2018),
pp. 58–67. (Code: I818)
37. L. Bass, The software architect and DevOps. IEEE Softw. 35(1), 8–10 (2018). (Code: I33)
38. K. Brown, B. Woolf, Implementation patterns for microservices architectures, in Proceedings
of the 23rd Conference on Pattern Languages of Programs (PLoP ’16), Article 7 (The Hillside
Group, 2016), pp. 7:1–7:35 (2016). (Code: A104)
39. L. Chen, Continuous delivery: huge benefits, but challenges too. IEEE Softw. 32(2), 50–54
(2015). (Code: B15)
40. P. Debois, Devops: a software revolution in the making. Cutter IT J. 24(8), 3–5 (2011). (Code:
B4)
41. C. Ebert, G. Gallardo, J. Hernantes, N. Serrano, DevOps. IEEE Softw. 33(3), 94–100 (2016).
(Code: A54)
42. A. Balalaie, A. Heydarnoori, P. Jamshidi, Microservices architecture enables DevOps:
migration to a cloud-native architecture. IEEE Softw. 33(3), 42–52 (2016). (Code: A2)
43. M. Kersten, A cambrian explosion of DevOps tools. IEEE Softw. 35(2), 14–17 (2018). (Code:
I808)
44. N. Forsgren, V.A. Brown, G. Kim, N. Kersten, J. Humble, 2016 State of DevOps Report (2016).
https://puppet.com/resources/whitepaper/2016-state-of-devops-report
45. H. Barua, The role of configuration management in a containerized world (2015). https://www.
infoq.com/news/2015/12/containers-vs-config-mgmt
46. M. Callanan, A. Spillane, DevOps: making it easy to do the right thing. IEEE Softw. 33(3),
53–59 (2016). (Code: A67)
47. G.G. Claps, R.B. Svensson, A. Aurum, On the journey to continuous deployment: technical
and social challenges along the way. Inf. Softw. Technol. 57, 21–31 (2015). (Code: B13)
48. R. de Feijter, S. Overbeek, R. van Vliet, E. Jagroep, S. Brinkkemper, DevOps compe-
tences and maturity for software producing organizations, in Enterprise, Business-Process
and Information Systems Modeling (Springer, 2018), pp. 244–259. (Code: S805)
49. A. Dyck, R. Penners, H. Lichter, Towards definitions for release engineering and DevOps,
in 2015 IEEE/ACM 3rd International Workshop on Release Engineering (2015), p. 3. (Code:
B26)
50. D.G. Feitelson, E. Frachtenberg, K.L. Beck, Development and deployment at Facebook. IEEE
Internet Comput. 17(4), 8–17 (2013). (Code: B7)
51. J. Gray, A conversation with Werner Vogels. ACM Queue 4(4), 14–22 (2006). (Code: B3)
52. J. Humble, Continuous delivery sounds great, but will it work here? Queue 15(6), 57–76 (2017).
(Code: B22)
53. M.G. Jaatun, Software security activities that support incident management in Secure DevOps,
in Proceedings of the 13th International Conference on Availability, Reliability and Security
(ARES 2018) (ACM, 2018), pp. 8:1–8:6. (Code: A803)
54. M.G. Jaatun, D.S. Cruzes, J. Luna, DevOps for better software security in the cloud, in Proceed-
ings of the 12th International Conference on Availability, Reliability and Security (ARES ’17)
(ACM, 2017), Article 69, pp. 69:1–69:6. (Code: A85)
55. H. Kang, M. Le, S. Tao, Container and microservice driven design for cloud infrastructure
DevOps, in 2016 IEEE International Conference on Cloud Engineering (IC2E) (2016), pp. 202–
211. (Code: I58)
56. M. Leppanen, S. Makinen, M. Pagels, V. Eloranta, J. Itkonen, M.V. Mantyla, T. Mannisto,
The highways and country roads to continuous deployment. IEEE Softw. 32(2), 64–72 (2015).
(Code: B14)
57. Z. Li, Q. Lu, L. Zhu, X. Xu, Y. Liu, W. Zhang, An empirical study of cloud API issues. IEEE
Cloud Comput. 5(2), 58–72 (2018). (Code: I802)
58. H.H. Olsson, H. Alahyari, J. Bosch, Climbing the “Stairway to Heaven”—a multiple-case study
exploring barriers in the transition from agile development towards continuous deployment of
software, in 38th Euromicro Conference on Software Engineering and Advanced Applications
(2012), pp. 392–399. (Code: B17)
59. C. Pang, A. Hindle, Continuous maintenance, in 2016 IEEE International Conference on
Software Maintenance and Evolution (ICSME) (2016), pp. 458–462. (Code: I55)
60. R. Punjabi, R. Bajaj, User stories to user reality: a DevOps approach for the cloud, in 2016
IEEE International Conference on Recent Trends in Electronics, Information Communication
Technology (RTEICT) (2016), pp. 658–662. (Code: I17)
61. M. Rajkumar, A.K. Pole, V.S. Adige, P. Mahanta, DevOps culture and its impact on
cloud delivery and software development, in 2016 International Conference on Advances in
Computing, Communication, Automation (ICACCA) (2016), pp. 1–6. (Code: I48)
62. S. Van Rossem, W. Tavernier, D. Colle, M. Pickavet, P. Demeester, Introducing development
features for virtualized network services. IEEE Commun. Mag. 56(8), 184–192 (2018). (Code:
I77)
63. R. Siqueira, D. Camarinha, M. Wen, P. Meirelles, F. Kon, Continuous delivery: building trust
in a large-scale, complex government organization. IEEE Softw. 35(2), 38–43 (2018). (Code:
B29)
64. E. Woods, Operational: the forgotten architectural view. IEEE Softw. 33(3), 20–23 (2016).
(Code: A82)
65. H. Yasar, K. Kontostathis, Where to integrate security practices on DevOps platform. Int. J.
Secure Softw. Eng. (IJSSE) 7(4), 39–50 (2016). (Code: A58)
66. H. Yasar, K. Kontostathis, Secure DevOps process and implementation, in 2016 IEEE Cyber-
security Development (SecDev), Boston, MA (2016), p. 166. https://doi.org/10.1109/secdev.
2016.048
67. xMatters Atlassian DevOps Maturity Survey Report 2017 (2017). https://www.xmatters.com/
press-release/xmatters-atlassian-2017-devops-maturity-survey-report/
68. How Netflix Thinks of DevOps (2018). https://www.youtube.com/watch?v=UTKIT6STSVM
69. H. Beal, Where are you on the DevOps Maturity Scale Webcast (2015). https://www.youtube.
com/watch?v=a50ArHzVRqk
70. R. Brigham, DevOps at Amazon: A Look at Our Tools and Processes (2015). At AWS re:
Invent 2015, https://www.youtube.com/watch?v=esEFaY0FDKc. Accessed June 2018
71. D. Brown, Our DevOps journey—Microsoft’s internal transformation story, in DevOneConf
2018 (2018). https://www.youtube.com/watch?v=cbFzojQOjyA. Accessed July 2018
72. D. Budgen, P. Brereton, Performing systematic literature reviews in software engineering, in
Proceedings of the 28th International Conference on Software Engineering (ICSE ’06) (ACM,
2006), pp. 1051–1052
73. N. Ceresani, The Periodic Table of DevOps Tools v.2 Is Here (2016). https://blog.xebialabs.
com/2016/06/14/periodic-table-devops-tools-v-2/. Accessed Apr 2018
Digital Assistant with Augmented Reality
Vijay Verma, Aditya Sharma, Ketan Kumar Jain, and Sushant Adlakha
Abstract From CRTs to virtual displays, the perception of the digital world we know
is continually evolving. Technologies like computer vision and object recognition can
be used in conjunction with augmented reality to create an interactive and enhanced
user experience of the real world. People are familiar with digital assistants like
Siri on iPhone, Google Assistant on android, Cortana on Windows, or Alexa and
Google Home in households. As the digital assistant market increases, so does their
functionalities and capabilities. They perform a wide variety of tasks ranging from
answering the most basic questions, to playing a song, telling jokes, setting alarms,
reading audiobooks, and even take control of IoT devices in a smart home system.
With this paper, we aim at incorporating the same in an environment where the user
can see and interact with the agent, which increases user satisfaction and enhances
the overall experience.
Keywords Augmented reality · Virtual assistants · Object tracking ·

Human–computer interaction · 3D modeling · Computer vision
1 Introduction
Augmented reality (AR) is a technique that involves overlaying computer graphics

on real-world objects [1]. Unlike virtual reality (VR), which creates an artificial
environment, augmented reality uses the existing environment and integrates digital
V. Verma (B) · A. Sharma · K. K. Jain · S. Adlakha

National Institute of Technology, Kurukshetra, Kurukshetra, Haryana, India
e-mail: vermavijay@nitkkr.ac.in
A. Sharma
e-mail: aditya_11610192@nitkkr.ac.in
K. K. Jain
e-mail: ketan_11610193@nitkkr.ac.in
S. Adlakha
e-mail: sushant_11620015@nitkkr.ac.in
396 V. Verma et al.
information with the user’s environment in real time. It tries to detect a marker, which
is a distinguishable landmark in the environment over which the digital graphics are
rendered. However, the rendered graphics exist in virtual space only and not presented
holographically. A closely related term is mixed reality (MR), which, in addition to
the above features, allows the user to physically interact with the rendered virtual
image.
AR can be used in various domains for a variety of purposes. It is used as an assis-
tive system for performing human tasks, thereby enhancing the ease and proficiency
of the functions. In surgery [2], AR can be used to identify the patient’s affected
anatomy, which gives a much deeper insight in comparison to the two-dimensional
scan of a body part. AR can also change the perspective in the military vehicle and
aviation training [3], where the cost of failure is high, or the equipment is only avail-
able at a select few sites. In product manufacturing, most of the working part remains
obscured beneath layers, investigating which might cause permanent damage to the
product. Such procedures are critical, expensive, and rarely can be performed in a
controlled environment, but can be conveniently taught using AR, where the user
can learn in a more intuitive environment and the ability to interact, which would
otherwise not be possible. Further, remote teaching and distance learning of the
same complex procedures can also become possible without the establishment of
infrastructure. The same technique can be brought into the education system [4]
to explain complex natural, geographical, or biological phenomena with ease. An
average person learns better by observing and listening to something than by merely
reading something, and this specific property of the human mind can be exploited
to accelerate learning. AR can be leveraged to provide real-time information to the
tourists about the various places by using GPS data and the buildings as a reference
point [5]. When used in the field of entertainment, AR can provide boundless oppor-
tunities to amuse one in a real-time 3D environment setting of games [6], movies, and
other forms of media. The concept can also be utilized for virtual tryout of dresses
and accessories [7] to allow users to get an insight whether the product is suitable
for them before making the transaction.
To make AR possible, video cameras are used to grab the visual input and recog-
nize where the graphic overlay is to be rendered. A processing system, possibly a
computer, then translates the read image and projects the virtual objects onto the
user’s screen, which can be a wearable component. Additional hardware compo-
nents such as audio systems, infrared sensors for tracking the location of real-world
objects, and biosensors might be used to get more information about the user and
their environment.
Wearable systems come equipped with various electronic sensors which provides
a user with greater ease of use. If such a model is backed by an artificially intelligent
core, it will serve as the foundation of intelligent virtual assistants, which can perform
tasks specific to the user in a one-to-one relationship. Such an agent may learn to
perform better with time, just like a human personal assistant would behave.
For a digital assistant to be effective, the primary requirements [8] are simplicity,
flexibility, and ease of interaction. Hence, the voice-based IO interface serves best,
as it requires the least cognitive effort from the user’s end. In this paper, we aim to
Digital Assistant with Augmented Reality 397
build a virtual assistant which not only is capable of understanding and executing
user’s voice commands but also provides a visual embodiment of self, making its
actions more affable to the user [9, 10].
2 Related Works
The tech giants like Google, Microsoft, Apple, and Facebook have taken a keen
interest in enhancing user experience in their current application using AR and MR
(mixed reality). In the academic contrast, several researchers have devoted their time
to explore viable applications and techniques as well. Saquib et al. [11] presented an
approach that assists beginners to generate real-time augmented video presentations
with a variety of hand gestures, where the users first have to map the gestures with the
graphical objects via a direct manipulation interface, which are later interpreted by
the system in real time to generate the required graphics. Reuter et al. [12] proposed
to integrate AR using the Microsoft HoloLens [13] into Unified Modeling Language
(UML) teaching as 3D views for easier comprehension of the relationships between
different models as compared to the conventional teaching methods. Prakash et al.
[14] suggested a new technique for dealing with the problem of improper illumina-
tion in augmented reality. It primarily observes those objects that tend to reflect light,
which is further used for real-time approximation of lightning conditions. This could
enable AR developers to work with liquid and shiny metals, etc. For ensuring low
computation cost, it is run in parallel as a separate thread. Hartholt et al. [15] intro-
duced an approach to boost morale of persons affected with autism spectrum disorder
by presenting a virtual humanoid to help them prepare for a meeting or an interview.
The rendered agent is capable of tracking eye gaze movement and flicker, response
delay and duration, and provides feedback to the user by normalizing their score on
a scale of ten. Kyto et al. [16] proposed to combine head and eye movement together
for a better range of selection of deliberate actions. Deshmukh et al. [17] aimed to
reinvent the education system by bringing out a low-cost high-fidelity application to
the masses by making use of Vuforia SDK with Unity to render a three-dimensional
layout of a pre-scanned object, which can then be studied in-depth to understand its
function and operation. Gary et al. [18] proposed an approach to create custom levels
and content within a given video game by exploring the concept and design of AR
game editors using Microsoft HoloLens for building up the AR environment. This
allows the users to edit and play community-made maps designed in the setting of a
home environment. Schmeil et al. [19] combined the idea of integrating the human
interface with a digital location tracking system by rendering a virtual humanoid
to guide and navigate the user through the city. Xu [20] introduced a digital image
processing algorithm for detection of the marker, which is 11 * 11 two-dimensional
array of black and white color pattern, where the image is captured via cell phone
camera parallelly taking into account rotation of the marker.
398 V. Verma et al.
3 Proposed Approach
With our work, we aim to build a digital personal assistant with AR, operable using
a voice input–output interface. The assistant would be capable of real-time interac-
tion, understanding and executing user commands, read and identify text, process
information, and fetch results based upon the user query. Thus, a user may be able to
extract digital assistance in a variety of forms utterly different from each other from
a single computing machine.
The notion of a digital assistant using AR starts from the recognition of the plane
on which the digital assistant would be projected. The use of computer vision allows
distinguishing between the varieties of objects captured in the image frame based
on their color. The object that we herein take into consideration is referred to as a
marker. Preprocessing of the image frame using computer vision libraries allows us
to detect edges, contours, and the number of sides present in the contour selected,
which help us to obtain the coordinates and the orientation of the marker in the image
frame. This shall be used to render the model of the virtual assistant on the plane
normal to the plane of orientation of the marker. The camera used for recognition
will capture some distortions (both radial and tangential) due to non-conformance
of the plane of marker and axis of camera focus; hence, normalizing, i.e., removing
distortions from the image shall also be taken into account.
The entire framework is designed in Python. FreeGLUT, an open-source alter-
native to the OpenGL Utility Toolkit, is utilized for creating and managing appli-
cation window frames with the inherent ability to render 2D and 3D graphics. For
image processing, OpenCV is used for the purpose of detecting the marker pattern.
When coordinates of the recognized marker are obtained, a 3D model designed using
Blender is rendered at those specific coordinates. Features build are highly modular
for ease of upgradability.
4 Experimentation
To overlay an AR glyph or object on a real-world object, a pattern distinguishable

from the environment, viz. marker, is to be identified from the image frame. This can
be accomplished using image processing libraries like OpenCV, Scikit-Image, etc.
After a few trials, it was found that the camera to be used also needs to be calibrated
if better marker designs are to be used. But since the axial plane of the camera lens
is not always in coherence with the plane of the marker, which is prone to movement
as well, radial and tangential distortions tend to occur. Radial distortions occur when
the curvature of line or shape changes, in the sense that straight line appears curved.
As our webcam is static in the position, i.e., not parallel to imaging plane or marker
plane. Hence, some part of the image which is closer to the webcam appears larger and
vice versa, giving a perspective effect, which is referred to as tangential distortion.
Camera calibration [21, 22] returns intrinsic parameters like focal length and the
optical center of the camera lens, along with extrinsic parameters like translational
and rotational vectors. The camera is calibrated against a grid of black and white
squares, as shown in Fig. 1. These parameters can deeply affect how the image is
recognized and the spatial position of an object in the image on the screen.
The marker that we consider is a two-dimensional array, which consists of 5 × 5
elements, and each element is either black or white. As shown in Fig. 2, the feature
of the marker is fixed with a black horizontal and vertical guide bar (5 elements long)
at the edges. For each white cell, we select a value 1, and for each black cell, we
select a value 0. This marker data is stored in a file for identifying whether the correct
marker is recorded from the image or not.
To capture the image frames continuously from the webcam, it is run inside a
parallel thread, which gives us the current frame, whenever prompted. Then, the
current frame is checked for having a specific marker.
For removing the distortion, the camera’s distortion coefficients and other details
of the camera were retrieved from camera calibration data processed in the previous
Fig. 1 Calibration of the camera
Fig. 2 Markers used for rendering assistants using a black and white grid
400 V. Verma et al.
step. For checking the marker pattern in the image, the contour (a curve connecting all
the continuous points with similar color or intensity along the periphery) is found.
It is ensured that the selected contour has four sides by using the approxPolyDP
function of OpenCV. If the contour has four sides, then the endpoints of the contour
are calculated (i.e., coordinates of all four vertices). Then, the contour is divided into
nine equal parts comprising of three rows and three columns. From the center point
of each of those nine parts, it is determined whether the pixel value lies in black or
white color’s threshold range. The black color threshold is set at 100, while that of
the white color is set at 155. If the pixel value satisfies the criteria, then an entry is
made in a 3 × 3 array, representing nine parts of the pattern under consideration.
Then, the 3 × 3 array is matched with the actual marker pattern. If a perfect match
occurs, then the recognized pattern is confirmed as a marker, and then translational
and rotational vectors are returned.
The model view for the virtual assistant, which is to be rendered on the user’s
screen, is designed in Blender 3D. Blender is a free and open-source 3D computer
graphics software toolset used for creating 3D-printed models, interactive 3D appli-
cations, animated clips, motion graphics, and computer games. First, the skeletal
framework for the model is designed. Care was taken to model the moving parts for
animation. For facial animations, sets of separate action frames were designed to
call up the desired effect. Two similar robot-like models were generated, and texture
mapping was generated to give them a unique feel.
The designed models were then exported as wavefront objects, which provide
data regarding the vertices, vertex textures and vertex normals, and how the faces
are to be formed. This is a standard file format and is acceptable by most rendering
libraries. The designed models were then tested for compatibility with OpenGL [23]
in a PyGame host window, to check for proper 3D orientation. OpenGL or Open
Graphics Library is a cross-language, cross-platform API for rendering 2D and 3D
vector graphics, with inherent support for assembling of vertex points, rasterizing,
and shading. OpenGL makes use of the GL display lists generated from the data of the
wavefront file to create set of coordinates and their corresponding texture mappings,
which are called to generate a frame. The window and event handlers are provided
by GLUT. This has a non-returning function calling a display function over and over
again, which takes the frame as input and redraws the subsequent frame in the next
processing cycle.
Once the virtual model is imported into OpenGL, it is rendered at the specific
coordinates whenever a marker pattern is recognized by the webcam. Being vector
in form, the rendered objects retain their shape and are automatically scaled up or
down with change in position of marker without loss in quality of the model.
The virtual assistant rendered on screen is able to interact with the user via voice
commands (Fig. 3). This is implemented via text to speech and speech recognition
libraries. The user may select one of the available features, and the control is trans-
ferred to the respective feature module for execution. To provide ease of feature
upgradability, the followed approach is highly modular. For any new feature to be
added, only the feature handler file is to be modified.
Fig. 3 Assistant images

shown
The interface loaded at startup is a chatbot that can answer to most user queries,
but is not able to perform any computation. For specific operations, the user can
request a particular feature to be loaded. Some of the basic features implemented
are real-time news and weather updates, optical character recognition and language
translation, deliver presentations, play audiobooks, stream music, and videos.
5 Results
This paper presents an introductory approach to a contemporary problem which

is being researched and developed by major tech organizations of the world. The
projections hold well in ordinary lighting conditions.
The features being modular provide easy upgradability and maintenance with
ample scope for the addition of new features. Each feature just performs only one
task efficiently that is expected from it. The initial interaction of virtual assistant
happens with the user by the medium of a chatbot which is capable of answering basic
user queries that does not require any mathematical computation. Features added to
the assistant like streaming audio and video, optical character recognition, phase
translation are a step toward increasing its viability in the day-to-day requirements
of an average user. Some of such features is depicted in Figs. 4, 5, and 6.
6 Conclusion
This paper represents a hybrid combination of augmented reality applications along

with a personal digital assistant, and it provides the user with ease of interaction
and interface of an augmented reality application and a plethora of features of the
personal digital assistant. At present, our system is capable of answering the user’s
402 V. Verma et al.
Fig. 4 Translating native language to foreign language from voice input
Fig. 5 Streaming of a sample video from Internet. The video overlay is presented over a third
marker
Fig. 6 Detecting text from

the image in the camera
frame
queries via the implementation of chatbot and providing a user an option to stream
audios and videos and may even help them to read out or digitize the text, translate
phrases in foreign languages and many more.
In order to make the assistant understand a wider range of commands, we plan
on applying artificial intelligence to the module. We expect that in future iterations
we could provide the user the option to customize the looks of his digital assistant
along with making it compatible for mobile platform as well.
References
1. K. Maxwellm Augmented Reality—definition of Augmented Reality, http://www.macmillan

dictionary.com/buzzword/entries/augmented-reality.html
2. J.W. Meulstee, J. Nijsink, R. Schreurs, L.M. Verhamme, T. Xi, H.H.K. Delye, W.A. Borstlap,
T.J.J. Maal, Toward holographic-guided surgery. Surg. Innov. (2019). https://doi.org/10.1177/
1553350618799552
3. L. Fan, J. Chen, Y. Miao, J. Ren, Y. Wang, Multi-vehicle cooperative military training simulation
system based on augmented reality, in Adjunct Proceedings of the 2019 IEEE International
Symposium on Mixed and Augmented Reality, ISMAR-Adjunct 2019 (2019)
4. C.S.C. Dalim, H. Kolivand, H. Kadhim, M.S. Sunar, M. Billinghurst, Factors influencing the
acceptance of augmented reality in education: a review of the literature (2017)
5. D.D. Han, T. Jung, Dublin AR: implementing augmented reality (AR) in tourism. Inf. Commun.
Technol. Tour. 2014 (2013). https://doi.org/10.1007/978-3-319-03973-2
6. J. Hamari, A. Malik, J. Koski, A. Johri, Uses and gratifications of pokémon go: why do people
play mobile location-based augmented reality games? Int. J. Hum. Comput. Interact. (2019).
https://doi.org/10.1080/10447318.2018.1497115
7. M. Yuan, I.R. Khan, F. Farbiz, S. Yao, A. Niswar, M.H. Foo, A mixed reality virtual clothes
try-on system. IEEE Trans. Multimedia (2013). https://doi.org/10.1109/TMM.2013.2280560
8. G. Chollet, J. Boudy, A. Esposito, G. Pelosi, Mines-t, I.: Building the next generation of personal
digital assistants, in International Conference on Advanced Technologies for Signal and Image
Processing, vol. 1 (2014), pp. 458–463
9. S. Haesler, K. Kim, G. Bruder, G. Welch, Seeing is believing: improving the perceived trust
in visually embodied Alexa in augmented reality, in 2018 IEEE International Symposium on
Mixed and Augmented Reality Adjunct (ISMAR-Adjunct) (2018), pp. 204–205. https://doi.org/
10.1109/ISMAR-Adjunct.2018.00067
10. K. Kim, L. Boelling, S. Haesler, J. Bailenson, G. Bruder, G.F. Welch, Does a digital assistant
need a body? The influence of visual embodiment and social behavior on the perception of intel-
ligent virtual agents in AR, in 2018 IEEE International Symposium on Mixed and Augmented
Reality (ISMAR) (2019), pp. 105–114. https://doi.org/10.1109/ISMAR.2018.00039
11. N. Saquib, R.H. Kazi, L.Y. Wei, W. Li, Interactive body-driven graphics for augmented video
performance, in Conference on Human Factors in Computing Systems—Proceedings (2019)
12. R. Reuter, F. Hauser, D. Muckelbauer, T. Stark, E. Antoni, J. Mottok, C. Wolff, Using augmented
reality in software engineering education? First insights to a comparative study of 2D and
AR UML modeling, in Proceedings of the 52nd Hawaii International Conference on System
Sciences, vol. 6 (2019), pp. 7798–7807. https://doi.org/10.24251/hicss.2019.938
13. Y. Liu, H. Dong, L. Zhang, A. El Saddik, Technical evaluation of HoloLens for multimedia: a
first look. IEEE Multimedia (2018). https://doi.org/10.1109/MMUL.2018.2873473
14. S. Prakash, A. Bahremand, L.D. Nguyen, R. LiKamWa, GleaM: an illumination estimation
framework for real-time photorealistic augmented reality on mobile devices, in MobiSys 2019—
Proceedings of the 17th Annual International Conference on Mobile Systems, Applications,
and Services (2019), pp. 142–154. https://doi.org/10.1145/3307334.3326098
15. A. Hartholt, S. Mozgai, E. Fast, M. Liewer, A. Reilly, W. Whitcup, A.S. Rizzo, Virtual humans
in augmented reality: a first step towards real-world embedded virtual roleplayers, in HAI
2019—Proceedings of the 7th International Conference on Human-Agent Interaction (2019),
pp. 205–207. https://doi.org/10.1145/3349537.3352766
16. M. Kytö, B. Ens, T. Piumsomboon, G.A. Lee, M. Billinghurst, Pinpointing: Precise head- and
eye-based target selection for augmented reality, in Proceedings of the 2018 CHI Conference on
Human Factors in Computing Systems, Apr 2018. https://doi.org/10.1145/3173574.3173655
17. S.S. Deshmukh, C.M. Joshi, R.S. Patel, Y.B. Gurav, 3D object tracking and manipulation in
augmented reality. Int. Res. J. Eng. Technol. 5, 287–289 (2018)
18. G. Ng, J.G. Shin, A. Plopski, C. Sandor, D. Saakes, Situated game level editing in augmented
reality, in Proceedings of the Twelfth International Conference on Tangible, Embedded, and
Embodied Interaction, Jan 2018, pp. 409–418. https://doi.org/10.1145/3173225.3173230
404 V. Verma et al.
19. A. Schmeil, W. Broll, MARAA mobile augmented reality-based virtual assistant, in Proceed-
ings—2007 IEEE Virtual Reality Conference (2007), pp. 267–270. https://doi.org/10.1109/VR.
2007.352497
20. Q. Xu, Visual code marker detection, pp. 2–4 (2006)
21. A. Wu, H. Xiao, F. Zeng, A camera calibration method based on OpenCV, in ACM International
Conference Proceeding Series (2019), pp. 320–324. https://doi.org/10.1145/3378065.3378127
22. Docs.opencv.org: OpenCV: Camera Calibration, https://docs.opencv.org/master/dc/dbb/tut
orial_py_calibration.html
23. E. Angel, D. Shreiner, Introduction to modern OpenGL programming, in ACM SIGGRAPH
2012 Courses, SIGGRAPH’12 (2012)
An Experimental Approach Toward Type
2 Diabetes Diagnosis Using Cultural
Algorithm
Ratna Patil, Sharvari Tamane, and Kanishk Patil
Abstract The performance of cultural algorithm for diabetes prediction has been
studied systematically and result has been presented in this paper. The performance
of the proposed method is compared with prevailing approaches to measure the
efficacy of this method. The study shows that cultural algorithm can be employed for
determining the best features for building the type 2 diabetes prediction model. In this
experimental study, cultural algorithm is used for searching the best subset of features
to enhance the efficiency of the predictive model. The selected subset of features was
assessed using artificial neural network to check the quality. Proposed method was
tested on diabetes dataset collected from the local hospitals and compared with other
metaheuristic approaches.
Keywords Artificial neural network · Classification · Cultural algorithm · Feature

selection · Knowledge source · Type 2 diabetes
1 Introduction
One of the challenges faced by biomedical scientist is to analyze the data which
consists of smaller number of samples with multiple features out of which many are
irrelevant. Efficient processing of high-dimensional data has increased the research
opportunities and improved health care services. Recent research work shows that
extracting the most important features from the high dimensional dataset and using it
for classification has improved the accuracy of predictive models. Reduced number
R. Patil (B)
Noida Institute of Engineering and Technology, Greater Noida, India
e-mail: ratna.nitin.patil@gmail.com
S. Tamane
Jawaharlal Nehru Engineering College, Aurangabad, India
e-mail: sharvaree73@yahoo.com
K. Patil
J J Hospital, Mumbai, India
e-mail: kanishk.nitin.patil@gmail.com
406 R. Patil et al.
of features makes the learning model more efficient and easier to interpret. Selec-
tion of features is broadly classified into three methods which are filter, wrapper,
and ensemble. Filter approach for selecting features uses statistical tests to find out
important features for the classification model. Wrapper approach works together
with the classification algorithm and the performance of the predictive model is
taken into consideration to choose best feature subset. To obtain the most important
feature subset which gives optimal classification performance is a NP Hard problem.
To search the best subset in a search space is an exhaustive search process. Alter-
native to this exhaustive search will be to employ metaheuristic approaches [1, 2].
These algorithms are application independent and produce optimal solution itera-
tively thereby improving the quality of subsets of features. For tuning the parameters
of classifiers and selection of optimized feature subset, metaheuristics-based hybrid
feature selection classifiers based on swarm intelligence [3], ant colony, genetic
algorithm, cultural and gravitational search algorithm may be developed.
The study has been presented in the following manner. Section 2 describes the
cultural algorithm and correlated basic concepts. Proposed methodology and its flow
diagram are presented in Sect. 3. Experimental results and its analysis have been put
forth in Sect. 4 and conclusion drawn in Sect. 5.
2 Cultural Algorithm
The cultural algorithm has been presented by Reynolds in 1994. The cultural algo-
rithm (CA) is induced by the growth of the human culture and its effect on indi-
viduals as well as forthcoming generations. Process of cultural evolution in human
society has invigorating effect on this algorithm. Information is stored in the common
area and shared by population which leads to the evolution process in the human
society. The shared ideas, knowledge by population are referred as a culture. This
shared knowledge directs the behavior and actions of the society. The algorithm
has three components as shown in Fig. 1. These are (1) population (2) Belief space
(3) Communication protocol for interaction between population and belief space.
Figure 1 demonstrates the framework of CA algorithm [4]. The population space
has autonomous solution agents. The belief space has global knowledge repository.
The acceptance function and influence function are used for interaction between
the population and belief space. Belief space encompasses four knowledge sources,
namely normative, situational, history, and topographical.
1. Normative knowledge source: Agents during mutation use normative knowledge
source. This knowledge source stores the features and the possible values that
the feature can take from the training dataset. Minimum and maximum values
of numeric attributes are stored by normative knowledge source. Normative
knowledge is updated by (1) and (2).
An Experimental Approach Toward Type 2 Diabetes Diagnosis … 407
Fig. 1 Components of cultural algorithm

f k (t) f k (t) < fl (t)
fl (t + 1) = (1)
fl (t) otherwise

f k (t) f k (t) > f u (t)
f u (t + 1) = (2)
f u (t) otherwise
where lower limit and upper limit of the fitness value at the tth iteration are
denoted by fl (t) and f u (t), respectively. Position of kth cultural individual is
given by yk (t) and f k (t) is the fitness value of yk (t).
2. Situational knowledge source (SKS): Situational knowledge source contains the
best pattern found through the evolutionary development process. It denotes a
front-runner for the other entities to follow. In this system, for the crossover
operation, instead of arbitrarily selecting individual, agents use the leader. It
pushes the offspring nearer to the best point found. Situational knowledge is
updated by (3).

S(t) f S (t) < f Yg (t)
S(t + 1) = (3)
Yg (t) otherwise
where, at the time t, the best cultural individual is given by Yg (t). f S (t)) and
f Yg (t) are the fitness value of S(t) and Yg (t), respectively.
3. Topographical knowledge source: A map of the fitness landscape of the problem
is produced throughout the evolutionary process. It contains a set of cells which
store the best individual found in each cell. It produces an ordered list of the best
b cells as per the fitness value of the best individual on each of them. In this, the
influence function attempts to push the offspring to any one of the b cells in the
ordered list.
408 R. Patil et al.
Fig. 2 Genetic transmission versus cultural transmission
4. History knowledge source: To search patterns in the environmental changes,

history knowledge source is used. The best individual position found preceding
to each environmental change is recorded in the form of a list. W is the size of
the list mentioned.
Human culture covers behaviors, artifacts, and ideas that is cultured and communi-
cated between individuals and changes over time period. This course of transmission
and change is reminiscent of Darwin’s principle of parentage with variation through
natural selection [1]. Figure 2 shows the genetic transmission versus cultural trans-
mission which transpires in a short span. It has been learned that cultural transmission
is more complex than genetic transmission. An offspring can get social individuali-
ties not only from its parents as well as from nonparental adults and peers. Hence,
the occurrence of a cultural characteristic in the population is relevant beyond just
the likelihood that an individual’s parents had that peculiarity. In the literature, it
has been proved that in most of the cases, the more common a cultural trait is in the
population, the more likely it is for an individual to have the chance to get it through
social learning [5]. However, the dimensions of the population may also have its
impact on the ongoing transmission, and thus existence, of a cultural trait.
The pseudocode for cultural algorithm is given below:
1. Start
2. Set t = 0.
3. Initialization of population space population(t).
4. Initialization of belief space belief(t).
5. Repeat the following until termination condition is found.
a. Variations of individuals in the population space population (t + 1) ←
population (t)
b. Evaluate population, population(t)
c. Adjust (Belief space (t), Accept(population(t)))
d. Influence(population(t), belief(t))
e. t = t + 1.
It is hard and expensive to obtain the optimal solution using conventional methods,
specifically for high value of number of features in a dataset. Cultural algorithm is a
metaheuristic approach which may be employed to accelerate the search for the best
feature’s subset.
The proposed hybrid approach is used to improve the learning and convergence
of the search techniques. The algorithm is functioning at two sages. First stage is
a population or cultural stage. Binary form is used for representing chromosomes
(one implies that the particular feature is selected and zero denotes not selected). In
the first stage, individual candidate solutions are transmitting their characteristic into
objective function in the problem domain. In the second stage, the knowledge or belief
space records information learnt by generations. This information is made available
to the current generation. Figure 3 depicts the flow diagram of proposed approach.
The hybrid approach uses cultural algorithm for feature selection in combination
with the artificial neural network [6, 7]. The fitness function used for the population
space is based on the MSE computed by developing artificial neural network (ANN).
The objective function is developed using ANN due to its ability to learn and model
nonlinear complex relations between inputs and outputs. Artificial neural network
can deduce unseen relations on unknown data and makes the model more generalized.
4 Analysis and Experimental Results
As a part of this study, the real-time dataset for developing a model for early diagnosis
of type 2 diabetes was collected from local hospitals. The total samples collected are
1133. This dataset has 32 attributes. The dataset was verified from physician. Missing
values are handled and outliers are detected in the pre-processing stage. The class
variable has three values. They are diabetic, prediabetic, and healthy. The param-
eter values giving good results were decided and used for the developed predictor
model. It is a challenge to search the best predictor features out of 32 attributes. In
this study, cultural evolutionary algorithm is used for searching the problem space
to find all the prospective feature subsets and artificial neural network is employed
for evaluation of the performance of each subset. The proposed approach was imple-
mented in MATLAB and configured to enhance the accuracy. Features selected using
proposed approach are “HDL” “No. of times Pregnant” “LDL” “Family History”
“Blurred Vision” “PPG” “BMI” “VLDL” “Fatigue” “FPG”. Multi-class confusion
matrix output obtained from the developed model is summarized in Table 1.
To assess the model’s performance accuracy, error, precision, sensitivity, and
specificity are computed as per the formulas. Predictor model’s overall parameters
and corresponding values are shown in Table 2.
410 R. Patil et al.
Fig. 3 Flow diagram of proposed CA algorithm

Table 1 Summary of multi-class confusion matrix output

Labels True positive False positive False negative True negative
Actual class 1 (Healthy) 416 56 18 643
Actual class 2 (Prediabetic) 226 36 81 790
Actual class 3 (Diabetic) 358 41 34 700
Table 2 Overall values

S. No. Parameter Value
1 Accuracy 0.8826
2 Error 0.1174
3 Sensitivity 0.8693
4 Specificity 0.9403
5 Precision 0.8804
6 False positive rate 0.0597
7 F1 score 0.8726
8 Matthews correlation coefficient 0.8169
9 Kappa 0.7359
To compute the various performance measures given in Table 2, following terms

are used.
• True positive (TP): is predicted to be positive with actual class positive.
• False negative (FN): is predicted negative with actual class positive.
• True negative (TN): is predicted to be negative with actual class negative.
• False positive (FP): is predicted positive with actual class negative.
Accuracy of the developed model is computed by Eq. 4.
(TP + TN)
Accuracy = (4)
(TP + FP + FN + FP)
Accuracy obtained is 88.26% and depicted in the form of confusion matrix is

shown in Fig. 5.
Error is a misclassification and computed by Eq. 5.
(FP + FN)
Error = (5)
(TP + FP + FN + FP)
Sensitivity is the ratio of the true positive predictions to the total number of actual
positives. The best sensitivity is one. It is also called as recall and computed by Eq. 6.
TP
Sensitivity = (6)
(TP + FN)
412 R. Patil et al.
Specificity is the ratio of true negatives to the total number of

Actual negatives. The best value of specificity is one. Equation 7 is used to compute
specificity.
TN
Specificity = (7)
(TN + FP)
Precision is the ratio of true positives to the total predicted positives computed by
Eq. 8.
TP
Precision = (8)
(TP + FP)
False positive ratio (FPR) is the ratio of the number of false positive predictions
to the total number of negatives. Zero is the best FPR value. FPR is given by Eq. 9.
FP
FPR = (9)
(FP + TN)
Harmonic mean of recall and precision is called F1 score. Equation 10 computes

F1 score.
2 ∗ (precision ∗ recall)
F1Score = (10)
(precision + recall)
Matthews correlation coefficient (MCC) is used to measure the quality of the

classifier. It is computed by Eq. 11. Positive one represents a perfect predictive
model and negative one represents inverse prediction.
(TP ∗ TN) − (FP ∗ FN)

MCC = √ (11)
(TP + FP) ∗ (FP + TN) ∗ (TP + FN) ∗ (FN + TN)
The comparison of observed accuracy with an expected accuracy is done by

Cohen’s Kappa coefficient and its value is computed by Eq. 12.
(observed accuracy − expected accuracy)

Kappa value = (12)
1 − expected accuracy
The best cost at each iteration is plotted in the form of graph shown in Fig. 4.
The X-axis of the graph refers the iteration number and the Y-axis gives the best cost
obtained in the respective iteration.
A confusion matrix describes performance of the developed classifier in tabular
form and is shown in Fig. 5.
An epoch is used to indicate the number of passes through the complete raining
set. The best validation performance obtained is 0.20712 at epoch 32. Figure 6 shows
the validation performance at training, validation, and testing phases.
Fig. 4 Best cost versus number of iterations
Fig. 5 Confusion matrix

414 R. Patil et al.
Fig. 6 Validation performance
Receiver operating characteristics (ROC) curve is plotted with true positive rate
on Y-axis against false positive rate on X-axis. The ROC curve for proposed approach
is displayed at Fig. 7. The test is more efficient when the graph is closer to the top left
corner. This curve tells us how much the model is skilled in distinguishing among
classes. Therefore, this predictive model has shown a better result.
5 Conclusion
In this experimental study, cultural algorithm in combination with artificial neural

network is developed for optimization of feature selection problem in a diabetes
dataset. The accuracy obtained for the developed prediction model for type 2 diabetes
is 88.3%. The developed model will detect the prediabetic stage. Sedentary lifestyle,
obesity, and consuming junk food are known to play a huge role in the pathophysi-
ology of type II diabetes mellitus. Detection of the prediabetic stage will therefore
serve as a wake-up call for individuals to initiate lifestyle modifications as suggested
by their physician and prevent further progression of the disease. Lifestyle modi-
fications include daily exercise and limiting intake of food high in saturated fat,
trans fatty acids. In the future, the social knowledge derived from the individuals in
the cultural algorithm may be applied to solve the computer security problems in
numerous applications.
Fig. 7 ROC curve
References
1. O. Oloruntoba, G. Cosma, A Modified Cultural Algorithm for Feature Selection of Biomedical

Dat (Springer, Cham, 2019)
2. R.N. Patil, S.C. Tamane, Upgrading the performance of KNN and Naïve Bayes in diabetes
detection with genetic algorithm for feature selection. Int. J. Sci. Res. Comput. Sci. Eng. Inf.
Technol. 3(1), 2456–3307 (2018)
3. X. Wang, W. Hao, Q. Li, An adaptive cultural algorithm with improved quantum-behaved particle
swarm optimization for sonar image detection. Sci. Rep. (2017)
4. H.C. Kuo, C.H. Lin, Cultural evolution algorithm for global optimizations and its. J. Appl. Res.
Technol. 11, 510–522 (2013)
5. B. Zaeri, Z. Kobti, Improving robustness in social fabric-based cultural algorithms with two
new approaches in population and belief spaces, in Proceedings of the Thirtieth International
Florida Artificial Intelligence Research Society Conference (2017)
6. P. Ratna, S.C. Tamane, PSO-ANN-based computer-aided diagnosis and classification of diabetes,
in Smart Trends in Computing and Communications, Proceedings of SmartCom 2019 (2019)
7. X. Xue, M. Yao, R. Cheng, A novel selection operator of cultural algorithm. Adv. Intell. Soft
Comput. 123, 71–77 (2017)
Predicting Students Performance
Through Behavior and Computational
Thinking in Programming
Vinayak Hegde, H. N. Meghana, R. Spandana, and M. S. Pallavi
Abstract The advancement and achievement of the students are known as primary
success metrics for many universities around the world. Hence, universities invest
heavily in resources to reduce the performance gap between academically good
performers and low-performing students. The high rate of failure in computer
programming courses can attribute several factors. This paper explores the educa-
tional history of students, their fields of study, and their approaches to learning
applied to programming courses. Considering students’ attendance, assignment
marks, program practice at home, willingness to work, interaction in class predicts
student’s behavior. By considering students’ lab test marks, willingness to apply
a new algorithm predicts student’s computational thinking. The data are collected
from undergraduate students by conducting lab tests and surveys among teachers.
Finally, considering student behavioral and computational thinking parameters, and
by using simple linear regression, the attributes which are supported and which are
not supported for predicting student performance can be finalized.
Keywords Computational thinking · Student behavior · Predicting student

performance · Education computing · Educational data mining
V. Hegde (B) · H. N. Meghana · R. Spandana · M. S. Pallavi

Department of Computer Science, Amrita School of Arts and Sciences, Amrita Vishwa
Vidyapeetham Mysuru Campus, Mysuru, Karnataka, India
e-mail: vinayakhegde92@gmail.com
H. N. Meghana
e-mail: meghanahn13@gmail.com
R. Spandana
e-mail: Spandana47741@gmail.com
M. S. Pallavi
e-mail: palls.ms@gmail.com
418 V. Hegde et al.
1 Introduction
This paper works on predicting student performance in programming through

students’ behavior and computational thinking. Most of the students in computer
programming show poor performance in their academics. Student background plays
a major role like age, gender, attendance level at the University, programs they are
pursuing and pursued at the high school, and also, it depends on the number of times
they have taken a particular programming course and some of the students will be
from the non-computer background, and it will be very difficult for them to learn
programming language. So, we should take opinions on the difficulties in learning
programming, and also, they will be facing problems in writing, reading as well as
designing a simple programming code.
To overcome the problem, in this study, we have followed the protocol as Vo
Ngoc Hoi research paper [1]. We try to predict student computational thinking and
their behavior by giving tests on programming language which include algorithms
and executable programs and taking feedback from the teachers on students’ regu-
larity, assignments, internal marks, and interaction in class. Computational thinking
attributes like lab test marks, willingness to work, and behavioral attributes like
attendance, assignment marks, program practice at home, willingness to work, and
interaction in class are considered for both computational thinking attributes and
behavioral attributes, and converting them to a common ‘n’ form is shown in the
Table 3. Then apply the correlation matrix for Table 3 data. After getting the result of
the correlation matrix from Tables 4 and 5, apply principal component analysis (PCA)
to get eigenvector values. After applying the PCA, screen plot, active variables, the
active component graph will be displayed. For the eigenvalues from Tables 6 and 7,
simple linear regression will be applied to find out which all attributes are supported
and which are not. The positive values which we get by applying simple linear
regression will be considered as supported, and negative values which we get will
be considered as not supported.
2 Literature Review
It presents the thought of foreseeing scholarly execution dependent on computational

thinking. The exhibition of 900 and eighty-two understudies in forty segments were
evaluated over the two-year time frame. These outcomes have likewise inferred that
the appraisal of computational thinking can be utilized as an early intercession appa-
ratus to improve the understudies’ maintenance, movement, and graduation rates
in STEM-related orders [2]. Computational thinking is viewed as a general skill,
right now further expound on what computational thinking is and present instances
of what should be instructed and how. First, we position computational thinking in
paper’s work with LOGO and point out difficulties in characterizing computational
thinking and examine the center and fringe parts of a definition [3]. The explanation
Predicting Students Performance Through Behavior … 419
between computational intuition assignments in a single side, understudy’s focused

on aptitudes, and the sorts of issues they expect to fathom, we led a little scope
pilot contextual analysis with two gatherings of understudies for the basic and center
grades. We found a divergence between the sorts of abilities evaluated, the effort-
lessness of the undertakings, and the age gatherings, which makes it hard to show
up to some steady end [4]. Numerous variables influence this scholarly presenta-
tion, and therefore, it is important to establish a prescient knowledge digging model
for the exhibition of understudies to discern the contrast between high students and
moderate student understudy. The rough information was preprocessed to top off
missing qualities, turning qualities into another and important characteristic/variable
determination in one single structure [5]. The proof of quality for the new content
mining plans owed exceptional physical, articulation, and commotion. However, the
information mining works with viably, the update finds designs having issues on
the open research particularly in the content mining as space. Thus, the information
preparing apparatus, for the most part, works in the forecast and characterization of
information [6]. Right now, mark forecast models have been created utilizing insti-
tutional inside databases and outside open information sources. The consequences
of observational investigation for college understudies’ first-year point expectation
show that forecast models dependent on the institutional interior and outside infor-
mation sources furnish better execution with increasingly precise models contrasted
with the models dependent on just institutional inner information sources [7].
As a consequence of the rapid expansion of knowledge in instructive context,
instructive knowledge mining resulted in the development of techniques to investigate
the interesting kinds of information emanating from instructive settings and the use
of such strategies to obtain understudies more readily and the settings in which they
learn. [8]. Worldwide the design of schools has a relatively high whittling down the
pace. Approximately 35% of the main year understudies in different design programs
do not usually make it into the next year. The purpose behind this test is to identify
the elements that fill in as great indications of whether an understudy is going to
drop out or flop the programme [9]. Numerical models have been used to anticipate
understudies’ exhibitions in the expectation of illuminating the help group to mediate
at the beginning time of the in danger understudy at the college. Right now, utilized a
mix of institutional, scholarly, segment, mental and financial variables to anticipate
understudies exhibitions utilizing a multilayered neural system (NN) to characterize
understudies’ degrees into either a decent or essential degree class [10]. The extent
of this paper is to the calculation to foresee the evaluation of understudies to give
convenient and proper admonition to understudy the individuals who are in danger.
A study cum exploratory technique was adopted in the present analysis to construct
a database, and it was created from both an important and an optional source. This
work will assist the instructive foundations with identifying the understudies who
are in danger and to and give better extra preparing to the powerless understudies
[11].
420 V. Hegde et al.
Attendance
Behavior
Interaction in Predicting
Class
Assignment
Computational
Marks
Thinking
Willingness
toWork
Program Willingness to
Practice in Apply New Lab Test
Home Algorithm
Fig. 1 Research model for predicting students performance through behavior and computational
thinking in programming
3 Methodology
3.1 Research Model and Hypotheses
This study sought to predict the student’s computational thinking and behavior in
class and laboratory. Figure 1 depicts the research model with relevant attributes like
attendance, interaction in class, assignment marks, willingness to work, program
practice at home are the parameters considered for predicting student behavior. Lab
test and willingness to apply the new algorithm are the parameters considered for
predicting student computational thinking. Behavior and computational thinking are
the two paths for the prediction of student overall performance.
3.2 Attendance
Attendance is the concept of students, to know whether the students are present in
class or not. It plays a major role in predicting student’s behavior. The educational
success of a student is based on student attendance [12]. Participation inspiration
is a great extent affected by fulfillment with the activity circumstance and different
internal and outer external pressure to attend [13].
H1: Attendance has a positive significant effect on behavior.
3.3 Interaction in Class
Interaction is the concept of knowing student attentiveness in class, by clarifying

doubts with the teachers, answering questions. The instructor mainly used positive
engagement to demonstrate unity and to maintain a friendly relationship with the
students, records to provide fair and unambiguous guidance, negative good manners
to reduce the students’ pressure [14].
H2: Interaction in class as a positive impact on behavioral significance.
3.4 Assignment Marks
Assignments are given to make students better on particular topics. In training,

it shows that self-assessment is conceivable and is a possibly important guide in
educating and learning [15].
H3: Assignment marks have a positive significant effect on behavior.
3.5 Willingness to Work
Students who are motivated in certain works show more interest in doing those
works. Not all students show a willingness to work as the subject interest will be
less. Students should show eagerness is to learn the subject in the classroom. This
examination utilizes a subjective technique, especially the contextual analysis with
the utilization of perception, survey, and meeting to gather the information [16].
H4: Willingness to work has a positive significant effect on behavior.
3.6 Program Practice at Home
Learning the things at college cannot be remembered by students, but practicing it

at home can. The more practice they do, the quality of the knowledge increases.
Practicing programs at home to study the program carefully so that they know a lot
about it and they can try to apply new algorithms.
H5: Program practice at home has a positive significant effect on behavior.
422 V. Hegde et al.
3.7 Lab Test
The laboratory test is a concept to assess students’ practical knowledge in a specific

course. Conducting lab tests for students makes it easier to predict their computational
skills. The technical research tools as part of cooperative online learning helps to
improve the development of logical competencies [17]. A laboratory environment
motivates students to understand and conduct well in examinations. The approach
must be obtained positively by students at the college [18].
H6: Lab test has a positive significant effect on computational thinking.
3.8 Willingness to Apply the New Algorithm
It is a concept of creative thinking and applying own logic. People with more compu-
tational skills will try to find out new things rather than getting stuck to the existing
ones. Great software engineers are essential than ordinary ones in terms of showing
a willingness in software creation [19].
H7: Applying a new algorithm has a positive significant effect on computa-
tional thinking.
4 Materials
A survey was conducted among teachers to collect students’ behavioral data, and
a test was conducted among students to collect computational data. A survey had
attributes like attendance, interaction in class, assignment marks, and willingness to
work, program practice at home. The test consisted of basic programming questions,
where students were supposed to write the proper logic and expected output. A total
of 60 students’ data were collected.
5 Results
By specifying the student’s name, the third person can know the private data of a
student. In order to maintain privacy for student’s data, specific code is given for
each student so that we can predict their data privately, without revealing their name
for the third person (Tables 1 and 2).
The Table consists of a student’s code followed by seven parameters. V1, V2,
V3, V4, and V5 are the behavioral attributes, the data of this attributes are collected
by conducting survey. V6 and V7 are the computational thinking attributes, the data
are collected by conducting a test. Teachers were expected to offer ratings out of 5,
Table 1 Replacing each student’s name with a specific code

Name Code
Student 1 A
Student 2 B
Student 3 C
Student 4 D
Student 5 E
Table 2 Displays the parameters for predicting student’s behavior

Code V1 V2 V3 V4 V5 V6 V7
A 76 Good 7 Average Average 10 Yes
B 89 Very Good 8 Average Poor 12 Yes
C 94 Excellent >8 Excellent Excellent 14 Yes
D 77 Good 6 Average Poor 9 No
E 73 Good 7 Average Poor 10 Yes
where 5 was regarded as excellent, 4 as very good, 3 as good, 2 as poor, and 1 is

average. By taking all of the above criteria into consideration, teachers were meant
to offer an overall ranking for each student out of 5.
V1—Attendance, V2—Interaction of students in class and laboratory, V3—
Assignment marks, V4—Willingness to work, V5—Program practice at home,
V6—Lab test, V7—Willingness to apply new algorithm.
Table shows the model measurement performance. Nearly, seven parameters were
considered. We need to convert all the attributes to the common ‘n’ form. Here, in
Table 3, behavioral attributes are converted into three entries, and computational
thinking attributes are converted into two entries. The threshold values are taken as
per the higher education requirement.
A matrix of correlation is a table showing the coefficients of correlation between
collections of variables. To get the correlation matrix, we have considered Table 3
data. Table 4 is the correlation matrix for behavioral attributes. Table 5 is the
correlation matrix for computational thinking attributes.
In principal component analysis (PCA), we can pick either the matrix of covari-
ance or the matrix of correlations for components. Here, we have already chosen the
correlation matrix in Tables 4 and 5. By applying PCA to the correlation matrix, we
will get eigenvector values as shown in Tables 6, 7 and 8.
424 V. Hegde et al.
Table 3 Displays measurement process performance

Constructs Items Loading
Attendance <70 8
<80 39
>80 13
Interaction in class Poor 18
Average 36
Excellent 6
Assignment marks <4 3
4–8 54
>8 3
Willingness to work Poor 15
Average 41
Excellent 4
Program practice at home Poor 19
Average 34
Excellent 7
Lab test <8 18
<15 42
Willingness to apply new algorithm Yes 44
No 16
Table 4 Correlation matrix (behavioral data)

Attendance Interaction in Assignment Willingness to Program
class marks work practice at
home
Attendance 1
Interaction in 0.885022355 1
class
Assignment 0.988654046 0.944911183 1
marks
Willingness 0.902843815 0.999208234 0.957185973 1
to work
Program 0.819466004 0.992064533 0.89625816 0.986276801 1
practice at
home
Table 5 Computational data

Lab test Willingness to work
Lab test 1
Willingness to work −1 1
Table 6 Applying PCA behavioral data

Eigenvectors
F1 F2 F3 F4
Attendance −0.489 0.390 −0.409 −0.169
Interaction in class 0.416 0.436 0.656 −0.141
Assignment marks 0.392 0.631 −0.440 −0.258
Willingness to work 0.494 −0.036 −0.369 0.705
Program practice at home 0.435 −0.507 −0.272 −0.623
Table 7 Computational data

Eigenvectors
F1
Lab test 0.707
Willingness to apply the new algorithm −0.707
Table 8 Calculating simple regression

Constructs Items Loading Sum Average
Student attendance <70 8 3.8025 10.2513
<80 39 20.4619
>80 13 6.4895
Interaction in class Poor 18 9.1765 10.2513
Average 36 18.8497
Excellent 6 2.7277
Assignment marks <4 3 1.1155 10.2513
4–8 54 28.5229
>8 3 1.1155
Willingness to work Poor 15 7.5658 10.2513
Average 41 21.5367
Excellent 4 1.6529
Program practice at home Poor 19 9.7139 10.2513
Average 34 17.7749
Excellent 7 3.2651
Lab test <8 18 9.1765 15.6253
<15 42 22.0741
Willingness to apply new algorithm Yes 44 23.1489 15.6253
No 16 8.1017
426 V. Hegde et al.
Fig. 2 Simple line segment plot that shows the fraction of total variance in the data explained.
Scree plot for eigenvalues
6 PCA Graphs for Behavioral Data
This table shows the calculation part. Data in column items and loading are from
Table 3. Column sum is the data calculated by using simple standard regression for
the data from Tables 3, 6 and 7. Column average is the final regression data which
we got by taking an average of the column sum values (Figs. 2 and 3).
Table 9 shows the outcome, where we have all direct parameters. By applying a
simple regression algorithm, negative values that we get will be considered as not
supported, and positive values will be considered as supported. To calculate simple
regression, we will consider Tables 5 and 7 data along with the loading values from
Table 3. By calculating simple regression, we have got all positive values. Hence, all
parameters are supported.
7 Discussion
This paper focuses on selecting the proper attributes for predicting student’s perfor-
mance in programming through behavior in class and laboratory and their computa-
tional thinking. For predicting student’s behavior, a survey was conducted for teachers
to collect student’s data. The survey consisted of five parameters V1, V2, V3, V4,
Fig. 3 Positively correlated variables merge. Negatively correlated variables are placed at opposite
sides of the origin of the plot (opposed quadrants). The distance between the variables and the origin
in the factor map tests the consistency of the variables. On the factor diagram, variables that are
distant from the origin are well defined
Table 9 Specifying supported and nonsupported parameters

Hypothesis Regression path Effect type Simple regression Remarks
H1 Attendance Direct 10.2513 Supported
H2 Interaction in class Direct 10.2513 Supported
H3 Assignment marks Direct 10.2513 Supported
H4 Willingness to work Direct 10.2513 Supported
H5 Program practice at home Direct 10.2513 Supported
H6 Lab test Direct 15.6253 Supported
H7 Willingness to apply the new Direct 15.6253 Supported
algorithm
428 V. Hegde et al.
and V5 as displayed in Table 2. By conducting the survey, we could collect 60 data.

To predict students computational thinking, a test was conducted for 15 marks. The
test had basic C and C++ questions. V6, V7 from Table 2 is the test data. Five sample
data of both surveys and tests are shown in Table 2. Now, we need to convert all
the parameters to common ‘n’ form. The first five parameters are from the survey,
and the remaining two are from the test. Converting behavioral parameters by 3
entries and Computational thinking parameters by 2 entries as shown in Table 3.
Then, using Table 3 data, we will calculate the correlation matrix for behavioral data
and computational data as shown in Tables 4 and 5. Considering data from Tables 4
and 5, apply principal analysis of components (PCA) to obtain eigenvector values
as shown in Tables 6 and 7. Using eigenvector values from Tables 6 and 7 along
with the loading values from Table 3, let us calculate simple standard regression.
Table 9 is the outcome of this paper, where we have got simple regression values as
all positive and no negative value. This paperwork is to know which all parameters
are truly supported and which are not supported. To find out this, we have chosen
a simple regression analysis. This is a statistical method for quantifying the rela-
tionship between one independent variable based on the results and one dependent
variable. As independent and dependent variables, we have the following parameters.
Attendance—Independent variable, Interaction in class—Independent variable,
Assignment marks—Dependent variable, Willingness to work—Independent vari-
able, Program practice at home—Independent variable, Lab test—Dependent
variable, Willingness to apply new algorithm—Dependent variable.
8 Future Work
The attributes like attendance, interaction in class, assignment marks, willingness to

work, program practice at home, lab test, willingness to apply the new algorithms
are influential attributes to decide students’ performance. The data collected from
the above-mentioned attribute can be used for predicting the academic performance
of the student.
9 Conclusion
All parameters considered shall be accepted among the constructs tested in the
modified simple regression model. All the parameters considered for prediction are
strongly supported to predict student success by their computational thinking and
behavior in class and laboratory. The key aim of this paper is to identify the low-and
high-performance students. This attribute selection helps to predict the academic
performance of the students by teachers to recognize low-performing students and
train them to match the level of high performing students. In doing so, the orga-
nization gets 100% result, placement opportunities, and a good name in society as
well.
References
1. V. N. Hoi, Understanding higher education learners’ acceptance and use of mobile devices for
language learning: a Rasch-based path modeling approach. Comput. Educ. 146, 103761 (2020)
2. R.J. Haddad, Y. Kalaani, Can computational thinking predict academic performance? in 2015
IEEE Integrated STEM Education Conference (IEEE, 2015), pp. 225–229. (March)
3. J. Voogt, P. Fisser, J. Good, P. Mishra, A. Yadav, Computational thinking in compulsory
education: towards an agenda for research and practice. Educ. Inf. Technol. 20(4), 715–728
(2015)
4. T. Djambong, V. Freiman, Task-based assessment of students’ computational thinking skills
developed through visual programming or tangible coding environments, in International
Association for Development of the Information Society (2016)
5. B.K. Bhardwaj, S. Pal, Data mining: a prediction for performance improvement using
classification (2012). arXiv preprint arXiv:1201.3418
6. U. Harilatha, N. Sudhakaryadav, Predicting educational performance of a student failure and
dropout by using data mining techniques. Int. J. Comput. Sci. Inf. Technol. (IJCSIT) 5(6) (2014)
7. F. Sarker, T. Tiropanis, H.C. Davis, Students’ performance prediction by using institutional
internal and external open data sources (2017)
8. F. Berhanu, A. Abera, Students’ performance prediction based on their academic record. Int.
J. Comput. Appl. 131(5), 0975–8887 (2015)
9. K.F. Li, D. Rusk, F. Song, Predicting student academic performance. In 2013 Seventh Interna-
tional Conference on Complex, Intelligent, and Software Intensive Systems (IEEE, 2013, July),
pp. 27–33
10. S. Al-Sudani, R. Palaniappan, Predicting students’ final degree classification using an extended
profile. Educ. Inf. Technol. 24(4), 2357–2369 (2019)
11. A.A. Saa, Educational data mining & students’ performance prediction. Int. J. Adv. Comput.
Sci. Appl. 7(5), 212–220 (2016)
12. B. Bartanen,. Principal quality and student attendance. Educ. Resear. 0013189X19898702
(2020)
13. R.M. Steers, S.R. Rhodes, Major influences on employee attendance: a process model. J. Appl.
Psychol. 63(4), 391 (1978)
14. D. Rahayuningsih, M. Saleh, S.W. Fitriati, The realization of politeness strategies in EFL
teacher-students classroom interaction. Engl. Educ. J. 85–93 (2020)
15. C. Adams, K. King, Towards a framework for student self-assessment. Innov. Educ. Training
Int. 32(4), 336–343 (1995)
16. Z. Latifah, D. Sudana, F. N. Yusuf, Investigating the factors influencing students’ willingness
to communicate in english. ISLLAC: J. Intensive Stud. Lang. Lit. Art Cult. 4(1), 91–100 (2020)
17. E.X. Tan, Improving logical thinking skills through jigsaw-based cooperative learning
approach, in Preparing 21st Century Teachers for Teach Less, Learn More (TLLM) Pedagogies
(IGI Global, 2020), pp. 162–181
18. O. Vojinovic, V. Simic, I. Milentijevic, V. Ciric, Tiered assignments in lab programming
sessions: exploring (2020)
19. P.L. Li, A.J. Ko, A. Begel, What distinguishes great software engineers? Empirical Softw. Eng.
25(1), 322–352 (2020)
Deployment of Ambulance
with Maximum Area Coverage by Using
Particle Swarm Optimization Algorithm
Amandeep Kaur
Abstract Efficiency of any service depends on several factors. The accuracy of these
factors represents the actual performance of that service. In this paper, a service of
such kind is implemented to evaluate an optimum value of a specific factor. Basically,
in this paper, ten ambulances deployment is studied to get maximum value of area
where these ambulances are used to serve. Technique used for this purpose is particle
swarm optimization algorithm. To analyze this algorithm, MATLAB software is
used and it has analyzed that there is an efficient increase in area coverage after
implementing this algorithm.
Keywords Efficiency · Coverage · Particle swarm optimization algorithm ·

Ambulance deployment
1 Introduction
With the passage of time, advancement in technology serves us with great efficiency
of equipments. One of the most important concerns these days is the facility of
medical treatment in an emergency. This is to say, if any accident occurs, then the
only thing that should be up to that place is ambulance or hospital. It is not possible
to provide so many hospitals in a specific area. To overcome this drawback, certain
kind of emergency vehicles need to be deployed. The basic work of these vehicles is
to approach the accident location as soon as possible and to give the first aid to the
injured ones. The early approach of vehicles is also needed in several other services
such as delivery companies, courier service as well as hire taxi and many more. The
basic methodology of this technique is very much similar to a communication in
which several nodes are working to send message to other location in short time with
less efforts.
As earlier said, it is difficult to build so many hospitals in one area but the deploy-
ment of emergency vehicles is a suitable approach to establish a perfect protocol with
A. Kaur (B)
Baba Banda Singh Bahadur Engineering College, Fatehgarh Sahib, India
e-mail: amandeep.kaur@bbsbec.ac.in
432 A. Kaur
Fig. 1 Ambulance network
PSO techniques and algorithms. A number of experiments have done in the field of
communication to get an optimum approach of sending messages. One example of
it is as shown in the below diagram. The main objective of this paper is to study and
understand the importance of swarm optimization algorithm in the deployment of
emergency vehicles (Fig. 1).
2 Particle Swarm Optimization Algorithm: An Overview
An optimum solution of any adversity depends upon the techniques used to curb that
problem. Till now, enormous algorithms have used to obtain the best solutions of
optimization problems. Few of the techniques out of these are such as nontraditional
optimization technique used for military purposes. Several others are as follows:
• Genetic algorithms (GA)
• Ant colony optimization
• Fuzzy optimization
Deployment of Ambulance with Maximum Area Coverage … 433
• Neural networks.
In 1995, an algorithm was introduced in the category of these algorithms named
as PSO algorithm (particle swarm optimization algorithm). Mathematical operators
used in it became the root cause of its popularity. It is able to resolve the problems
related to continuous problems, discrete, integer variable type and nonlinear.
2.1 PSO Algorithm Parameters
Not every time but in certain case, the performance of the swarm optimization algo-
rithm is affected by its parameters. The rate of the effect of these parameters basically
depends upon the sort of problems. The basic parameters are as follow:
• Swarm size
• Iteration numbers
• Velocity component.
Swarm size
Number of particles is the most important part of initialization of this algorithm. The
number of particles is also known as swarm size. The benefit of having large swarm
size is that it provides a large space to search for better solutions. On the other side,
the only drawback of large search space is that it consumes a lot of execution time.
Most of the experiments studied in past time had used swarm size of [20, 60].
Iteration Numbers
Second important thing to consider during execution of it is the number of iterations.
These iterations make the results less and more optimum. It is preferred that number
of iterations should be average because less number of these give premature results
while large number of iterations make the experiment more complex.
Velocity Component
It is the last most important factor to keep in mind. The position of particles changes
as per the value of this factor. There are also some sub-parameters of velocity such
as shown in the following Table 1.
Table 1 Velocity parameters

Parameters Description
Weight inertia Information about previous movements.
Cognitive component Performance measurement
434 A. Kaur
2.2 Steps of Optimization
The idea of this technique is basically related to the behavior shown by the animals.
In this optimization technique, swarm keyword is used to represent the population
and particle is used to identify the member of population.
1. Initial population as well as the directions of movement of it is selected randomly
at the initial stage.
2. Every particle starts moving to the directions given by the initial stage with in
the search space. While moving, it is necessary for each particle to remember
the best locations of its neighbors.
3. After this, the positions and velocity components are selected their values
dynamically. Best locations are used to calculate these dynamic values.
4. All the above steps are performed again and again those are called iterations.
5. At the termination stage, all particles get their better positions.
6. All the above five steps are repeated till the swarm is not located on best position.
3 Literature Review
Ref. No. Algorithms Conclusions

[1] Hybrid differential evolution algorithm Proposed hybrid algorithm performance
is much better than the simple PSO
[2] Virtual force-based co-evolutionary It is a rapid, effective and robust technique
particle swarm optimization algorithm
[3] Multi-objective PSO algorithm Performance improves with an increase in
number of population.
[4] Multi-objective particle swarm Distance-based energy conservation
algorithm
[5] Multi-objective PSO technique efficient packet delivery, reduction in
energy consumption
[6] Voronoi diagram Large network implementation
[7] Particle swarm optimization (PSO) Drastically reduction of energy
consumption
[8] PSO-based algorithms Maximizes the area coverage, less energy
consumption
4 Proposed Work
4.1 Problem Statement
Efficiency of any service depends on several factors. The accuracy of these factors
represents the actual performance of that service. In this paper, a service of such kind
is implemented to evaluate an optimum value of a specific factor. Basically, in this
paper, ten ambulances deployment is studied to get maximum value of area where
these ambulances are used to serve. Technique used for this purpose is particle swarm
optimization algorithm. To analyze this algorithm, MATLAB software is used and it
has analyzed that there is an efficient increase in area coverage after implementing
this algorithm.
4.2 Objective of Study
The main objective of this work is to implement the PSO algorithm by using ten
particles at one time for optimizing the efficient coverage of ambulance in a specific
area.
4.3 Methodology
The idea of this technique is basically related to the behavior shown by the animals.
In this optimization technique, swarm keyword is used to represent the population
and particle is used to identify the member of population. Initial population as well
as the directions of movement of it is selected randomly at the initial stage. Every
particle starts moving to the directions given by the initial stage with in the search
space. While moving, it is necessary for each particle to remember the best locations
of its neighbors. After this, the positions and velocity components are selected their
values dynamically. Best locations are used to calculate these dynamic values. All
the above steps are performed again and again those are called iterations. At the
termination stage, all particles get their better positions. All the above five steps are
repeated till the swarm is not located on best position.
4.4 Flow of Research Work
As described in the above steps, the flow of working of hybrid algorithm is as follows.
The following flow chart explains the steps of methodology in the form of flow
(Fig. 2).
436 A. Kaur
Fig. 2 Flow of algorithm
5 Result Analysis
In the first step, input is given for ten particles. The locations and velocity components
are selected randomly. The following table as well as the diagram represents the total
Table 2 Parameters of
Parameter description Value of parameters
evolutionary algorithms
Swarm size 10
c1 1
c2 1
r1 rand value between 0 to 1
r2 rand value between 0 to 1
Range 3
Table 3 Area coverage for

Description Initial area Proposed algorithm
ten ambulances
coverage area coverage
Total coverage of ten 545.70 545.70
ambulances
Actual coverage of 387.50 545.70
ten ambulances
and actual coverage of area. Both these values are then compared with initial coverage
that is conducted without proposed algorithm (Tables 2 and 3).
From the above tables, it is clear that the random selection of positions of ambu-
lances covers less area. The implementation of proposed algorithm gives the best
locations of ambulances those give the maximum coverage as compared to the
initial deployment. The following graph represents the results of proposed algorithm
(Fig. 3).
6 Conclusion and Future Scope
The result analysis shows that the PSO algorithm covers area that is mostly equal
to actual area. In this paper, this algorithm provides an efficient placement of ambu-
lances. In the end, same dimensions of area are covered by less number of ambulances
in very significant manner. In each case, the coverage increases as the hybrid tech-
nique is applied. This algorithm has proved a better placement strategy for maximum
area coverage.
In future, this same algorithm will be implemented with certain variations in
parameters.
438 A. Kaur
Fig. 3 Representation of
coverage of ten ambulances
References
1. A. Bagherinia, I. Maleki, M.M. Tabrizi, S.R. Khaze, A new approach for area coverage in wsn.
IJMNCT 3(6), 61–75 (2013)
2. M.J. Jie, W. Sheng, W. Xue, An improved CPSO (co-evolutionary particle swarm optimization)
for WSN with dynamic deployment. Sensors 7(3), 354–370 (2007)
3. P.M. Pradhan, V. Baghel, M. Bernard, Energy efficient WSN using multi-objective particle
swarm optimization (MPSO), in IEEE International Advance Computing Conference, Patiala,
India (2009)
4. A. Vahe, D. Mehdi, H. Samaneh, M.N. Ghazanfari, PSO Based Node Placement Optimization
for Wireless Sensor Networks. IEEE, Islamic Azad University Tehran, Iran, ISBN: 978-1-4244-
8605-2 (2011)
5. N. Devarajan, K. Sheela, S. Rani, Multiobjective deployement in wireless sensor networks.
International Journal of Engineering Science and Technology (IJEST) 4(4), 1262–1266 (2011)
6. A.A. Ammar, A.A. Bt, M. YH. Alias, W. Mohemmed, A WSN’s coverage optimization algorithm
based on PSO and Voronoi diagram, in International Conference on Networking, Sensing and
Control, Japan, 26–29 March 2009
7. E. Natalizio, F. Guerriero, V. Loscrí, Particle swarm optimization schemes based on consensus
for wireless sensor networks, in 15th ACM International Conference on Modeling, Analysis and
Simulation of WMSN, Paphos, Cyprus
8. A. Aziz, N. Azlina, Wireless sensor networks coverage-energy algorithms based on particle
swarm optimization. Emirates Journal for Engineering Research 18(2), 41–52 (2009)
Detection of DDOS Attacks Using
Machine Learning Techniques:
A Hybrid Approach
Datla Anurag Varma, Ravi Ashish, V. Venkata Sai Sandeep, B. Venkatesh ,

and R. Kannadasan
Abstract The essential advantage of cloud computing is that it flexibly scales to

fulfill various needs and it provides the sufficient environment that scales up and
downsizes quickly as indicated by the interest, so it needs incredible security from
DDoS attack to handle vacation impacts of DDoS attacks. Circulate DoS assaults
fall on the classification of basic assaults that bargain the accessibility of the system.
These assaults have gotten refined and keep on developing at a quick pace so to
identify and to take down these assaults have had become a difficult undertaking. In
this project, we have designed a hybrid algorithm which consists of a combination of
several machine learning techniques to train a model which can be used to detect and
classify the type of DDoS attack with greater accuracy than that of each individual
machine learning technique used in the hybrid model.
Keywords Machine learning · DDoS attacks · Cloud services · Blending

technique · Hybird model · Cyber-crime control · Flooding assaults · Dynamic
assignment
1 Introduction
Cloud computing gives its clients to get to the different cloud benefits and lets them
to get to the storage information and the computational assets with low information
overhead. Cloud computing is quickly developing and has pulled in the clients toward
the cloud environment. It offers the clients to redistribute the information assets from
the remote areas when they need them. Cloud computing model has three compelling
service models and arrangement models [1]. Figure 1 represents the services of the
cloud, and the primary qualities provided to cloud clients are: on request service to
its clients, wide system get to, asset pooling, quick versatility and estimated service
[2].
D. A. Varma · R. Ashish · V. Venkata Sai Sandeep · B. Venkatesh (B) · R. Kannadasan

SCOPE, Vellore Institute of Technology, Vellore, Tamilnadu 632014, India
e-mail: Venkatesh.cse88@gmail.com
440 D. A. Varma et al.
Fig. 1 Cloud services
In a DDoS assault, a few exclusively organized PCs are utilized to target net-
work assets and servers which lead to overloading of messages, distorted bundles,
and demands for the association which brings about the refusal of the service to the
authentic client [3]. DDoS assaults on the cloud are highly complicated and require
strategies that are made to discover vulnerabilities and flooding assaults. The DDoS
assaults are fundamentally characterized or partitioned based on network conven-
tion and application layer. The transport and network level assaults are performed
utilizing the ICMP, TCP, and UDP conventions [4]. The HTTP DDoS assaults are
performed on high- and low-rates situations; every one of them significantly affects
the victim in question. The high-rate assault floods the casualty with countless solic-
itations via requests, while the situation in the low rate is that the casualty is given by
moderate and traded-off requests which consequently ends in the weary of assets [5].
The DDoS assaults performed on the nature of cloud computing are most principally
application layer conventions which conveys a few requests following the correspon-
dence convention which are then difficult to recognize in the system network layer
because of their pattern which coordinates the corresponding requests, in this man-
ner making the customary defense frameworks inapplicable [6]. The overloading
of DDoS assaults on the cloud is of different classifications, for example, request
flooding assaults and session assaults, slow reaction assault, and uneven assault.
All these flooding assaults create traffic which looks like that of an authentic client
which gets harder for the objective to recognize assault and real traffic accordingly
hindering the services for a client [7]. Cloud computing is broadly spread stage for
all services. Cloud platform is basic for the framework to complete the information
security issues. Security of the framework is performed by the firewall. The frame-
work has its impact on inside usage yet not on the exterior data [5]. DDOS assault is
inclining on web in the upcoming premise. Certain strategies must be completed for
keeping the framework from the assault. Dynamic assignment must be embraced,
yet the cloud runs in an uncertain manner [8].
Detection of DDOS Attacks Using Machine Learning Techniques: A Hybrid Approach 441
Fig. 2 Typical DDoS attack
In distributed DoS ambush, there are at any rate 100 individual assaulters attacking
the single system. Due to the higher number of assaulters, DDoS ambush is more
convincing and undermining than a typical DoS attack. The assaulters have order over
ace zombies, which, successively, have authority over slave zombies, as appeared in
the Fig. 2 [9].
Further in this paper, Sect. 2 explains about the related work, Sect. 3 explains about
the methods used , Sect. 4 explains about the experimental results and implementation
part, and finally in Sect. 5, we conclude the work.
2 Related Work
To identify and counter, denial-of-service assaults have become a difficult undertak-

ing. These assaults have gotten refined and keep on developing at a quick pace. They
fall on the classification of basic assaults that bargain the accessibility of the system.
Security upgrade and enhancement is acknowledged by the following stages:
(i) confirmation stage in which Enhanced Rivest Shamir Adleman (ERSA)-based
computerized signature is used for client verification, (ii) assessment stage misuses
fuzzy inference system (FIS) for guaranteeing required security and reliability level,
and (iii) encryption stage includes with ERSA calculation for encrypting information
to be put away in cloud.
The current existing framework utilizes various ways to deal with counter DDOS
assaults, for example, CAPTCHA puzzle. This offers a fundamental methodology for
assault alleviation, yet the framework flops because of its unproductiveness appeared
by recent concentrates and studies on controlling the DoS assaults [10].
CAPTCHAs endeavor to recognize bots from humans and are regularly used to
keep away from such assaults. Nonetheless, CAPTCHAs themselves ingest sources
and habitually are conquered. Kill-Bots diminishes CAPTCHA overhead by driv-
ing customer verification into the kernel. Notwithstanding, Kill-Bots requires core
adjustments, which can be infeasible.
Other models are as per the following: A computerized digital signature for net-
work stream examination utilizing meta-heuristic strategies was intended to know
the anomalous traffic which at that point demonstrated the better exactness or pre-
cision in the discovery of a DDoS assault; however, it failed to distinguish typical
DDoS attacks [11].
Another strategy called Seven which depends on the adaptive and selective ver-
ification which is for the most part used to counter system network layer DDoS
assaults. The method takes a shot at the idea of the notion of a state despite the fact
that the DDoS assaults on the application layer do not have a notion of state. This
component naturally falls flat against HTTP POST flooding assault on the grounds
that the tremendous measure of reflectors is utilized to send payloads [12].
3 Proposed System
To overcome all these shortcomings, machine learning approach is preferred. In

proposed system, we have taken one dataset which has 23 features including source
address, destination address, packet ID, from node, to node, packet type, packet
size, packet rate, etc., classifying the values based on the features including source
and destination address using machine learning techniques which includes SVM,
KNN, and naïve Bayes in order to detect the DDoS attack and calculating the model
accuracies of the same. In order to consider worst-case possible, we have used the
following characteristics: sigmoid kernel for SVC, 5 n_neighbors for KNN, and
default for Naïve Bayes. Figure 3 shows the model classification diagram.
After splitting the data into training and test data in order to obtain new data, the
new training data is further divided into training set validation set, and it is represented
as shown in Fig. 4. This technique is called blending ensembling technique.
Blending follows a very similar methodology as stacking however utilizes just a
validation set from the train set to make predictions. As it were, dissimilar to stacking,
the predictions are made on the validation set and test set as shown in Fig. 5.
Figure 6 represents the proposed system architecture. These predictions are made
by three classification models which includes support vector machine (SVM),
K-nearest neighbors (KNN) and Gaussian naïve Bayes model with the same char-
acteristics which are mentioned before. Each model provides two prediction sets,
namely validation prediction set and test prediction set. Hence, each validation pre-
diction set from the three models are concatenated into validation input. Likewise,
each test prediction set from the three models are concatenated into test input.
Fig. 3 Model classification diagram
Fig. 4 Blending ensemble

technique
Fig. 5 Predictions are made

on the validation set and the
test set
Finally, the newly obtained data, i.e., validation input, will be trained with ran-
dom forest technique to create the hybrid model. This model is used to make final
predictions on the test input to get the final predicted output which in turn compared
with the actual test data, and the accuracy is been calculated.
Fig. 6 Proposed system architecture
4 Results and Discussions
Therefore, the hybrid model is used for making analysis of the DDoS attack on the
cloud computing environment and thereby calculating the accuracy.
As shown in Fig. 7, Series 1 indicates the first series testing of each model and
calculating their respective accuracies. As we can see that the hybrid model provides
better accuracy than that of each individual model. Series 2 and Series 3 indicate the
second and third series testing of each model, respectively. For every test series, we
can get better accuracy for the hybrid model when compared to that of other models.
Fig. 7 Model accuracy comparison
5 Conclusion
Utilizing this hybrid model will give more precision prediction level of DDoS assault.
Likewise, it is increasingly viable when contrasted with different models. This con-
cept is utilized to anticipate the early predictions of DDoS by utilizing machine
learning strategies. The effect of denial of services will be distinguished by uti-
lizing this hybrid algorithm, and the defect can be controlled by taking necessary
precautions such as informing the cyber-team about the attack probability of that
particular site, removing credential cookies, neglecting to upload any sensitive piece
of information in order to create a caution regarding the security of that particular
site.
6 Future Works
The future implementation will be focused on implementing 100% accuracy of the

hybrid model and make a real-time system analysis to curb the denial of services.
Also, the detection and analysis of DDoS attacks on several websites have been very
difficult due to technological advancements in dark web, so we will be designing an
add-on tool which can not only detect the attack but also alert the user regarding the
attack in real time.
References
1. N. Subramanian, A. Jeyaraj, Recent security challenges in cloud computing. Comput. Electr.

Eng. 71, 28–42 (2018)
2. P. Medeira, J. Grover, M. Khorjiya, A survey on detecting application layer ddos using big data
technologies. J. Emerg. Technol. Innov. Res. (JETIR) (2019)
3. A. Lohachab, B. Karambir, Critical analysis of ddos-an emerging security threat over IoT
networks. J. Commun. Inf. Netw. 3(3), 57–78 (2018)
4. I. Sreeram, V.P.K. Vuppala, Http flood attack detection in application layer using machine
learning metrics and bio inspired bat algorithm. Appl. Comput. Inform. 15(1), 59–66 (2019)
5. S. Toklu, M. Şimşek, Two-layer approach for mixed high-rate and low-rate distributed denial
of service (ddos) attack detection and filtering. Arab. J. Sci. Eng. 43(12), 7923–7931 (2018)
6. A. Praseed, P.S. Thilagam, Ddos attacks at the application layer: challenges and research
perspectives for safeguarding web applications. IEEE Commun. Surv. Tutor. 21(1), 661–685
(2018)
7. R. Swami, M. Dave, V. Ranga, Software-defined networking-based ddos defense mechanisms.
ACM Comput. Surv. (CSUR) 52(2), 1–36 (2019)
8. K. Narayanaswamy, S. Malmskog, A. Sambamoorthy, Systems and methods of per-document
encryption of enterprise information stored on a cloud computing service (ccs) (October 30
2018). US Patent 10,114,966
9. A.R. Wani, Q. Rana, U. Saxena, N. Pandey, Analysis and detection of ddos attacks on cloud
computing environment using machine learning techniques, in 2019 Amity International Con-
ference on Artificial Intelligence (AICAI) (IEEE, 2019), pp. 870–875
10. K.M. Prasad, A.R.M. Reddy, K.V. Rao, Ensemble classifiers with drift detection (ECDD) in
traffic flow streams to detect ddos attacks. Wirel. Pers. Commun. 99(4), 1639–1659 (2018)
11. I. Makhdoom, M. Abolhasan, J. Lipman, R.P. Liu, W. Ni, Anatomy of threats to the internet
of things. IEEE Commun. Surv. Tutor. 21(2), 1636–1675 (2018)
12. S.J. Marck, R.C. Smith, System and method for mitigating distributed denial of service attacks
in a cloud environment (February 22 2018). US Patent App. 15/241,920
Revisiting Algorithms for Resource
Discovery in Unstructured Peer to Peer
Networks
Abhijith Jayan, Subi George Samuel, and P. Jayakumar
Abstract Peer to peer is an overlay network of interconnected nodes with high

scalability, reliability, efficiency, and low setup cost. A fundamental problem in p2p
networks is resource discovery. Several graph-based resource discovery algorithms in
unstructured p2p networks are existent. The performance analysis of these algorithms
were conducted under constrained settings like fixed hop count, fixed maximum
number of neighbours, etc. In this work, we revisit these algorithms on random
networks. Such analysis can give better insights and hence more realistic results,
because real-world networks are often random.
Keywords Unstructured p2p networks · Random network · Resource discovery
1 Introduction
Peer-to-peer (p2p) networks are distributed overlay networks having no centralized

authority or hierarchical organization. Every node in a p2p network is both a server
and a client at any point of time. The structure of a network refers to the way nodes
are connected, i.e. topology of the overlay. P2p networks are either structured or
unstructured. Structured p2p networks organize its peers to form a specific predefined
topology, but an unstructured p2p network does not form a specific structure or
topology. Some examples of structured p2p networks are Chord [1], CAN [2], etc.,
whereas Bit Torrent, Gnutella [3], etc., are examples of unstructured networks.
A. Jayan (B) · S. G. Samuel · P. Jayakumar

Department of Computer Science and Applications,
Amrita Vishwa Vidyapeetham, Amritapuri, India
e-mail: abhijithjn3@gmail.com
S. G. Samuel
e-mail: subigsamuel15@gmail.com
P. Jayakumar
e-mail: jayakumarpai@gmail.com
448 A. Jayan et al.
Unlike traditional client–server networks, p2p networks have many useful features
like high scaling, better accessibility, reliability and robust connections, graceful
management of failure or withdrawal of nodes, etc. References [4, 5] for a review.
Another advantage of p2p networks is that it is easy to implement the network and is
less expensive to maintain it, when compared to traditional client–server networks.
However, the resources in a p2p network are also distributed. Consequently, a new
operation called resource discovery in p2p network becomes relevant. Keeping the
large size of p2p networks in mind, it is evident that finding any resources in p2p
network is a challenging problem. In structured networks various results have been
existent [6]. In this article, we review the following resource discovery algorithms
in unstructured p2p networks. (1) flooding, (2) random walk, (3) random walk with
neighbours table and two hybrid algorithms, (4) flooding with Random Walk, and (5)
flooding with random walk with neighbours table [7].
The performance of these algorithms were evaluated using a network simulator
in [7]. The simulation assumed certain predefined values for routing parameters
like maximum number of hops, maximum number of neighbours of a node, etc. A
theoretical upper bound for the hop count is given by the diameter of the network.
It appears to be more logical to fix the TTL (time to live flag) at the diameter of the
network rather than fixing a predefined number. Predefined parameters potentially
can degrade the degree of randomness in the network. So we use random networks
like G(n, p) and G(n, m) networks [8]. The purpose of the project is to conduct a
theoretical study of the parameters like average hop count, number of failures, etc.,
on random graphs. Random graph models are popularly used for analysing the p2p
algorithms [3].
In Sect. 2, we give a review of the five algorithms. Then, in Sect. 3, the experimental
setting we use is explained and results are discussed, and Sect. 4 concludes the article.
2 Resource Discovery Algorithms
As mentioned earlier, the resources in a p2p network are also stored in multiple
peers. The problem of finding a peer having the required resource is called resource
discovery problem. This is similar to the routing problem in traditional client–server
network. Two different approaches are used to design algorithms. One is algorithms
based on graph model, and the other is probabilistic algorithms [9] for routing. Our
study is based on the graph theoretic algorithms.
In the graph theoretic approach, the overlay network is modelled as a graph where
a vertex represents a node and an edge represents a logical connection between two
nodes. It must be noted that the overlay network is not aligned with the underlying
physical network. See Fig. 1.
Graph-based routing algorithms are designed by using various structural proper-
ties of graph like connectivity, degree, locality of neighbours, clusters, etc. However,
in unstructured networks, such choices except connectivity of nodes may become
highly unpredictable. The five graph-based algorithms [7] that we study perform a
Revisiting Algorithms for Resource Discovery in Unstructured … 449
Fig. 1 Underlay and overlay

network
TTL-bounded search in the connected graph that models the p2p network. For exam-
ple, in flooding, the query is forwarded to the neighbours and then to the neighbours
of neighbours, etc., until the destination node is reached or the TTL is expired, which
is akin to bounded search. An overview the five algorithms is given below.
Flooding is a simple and one of the first resource discovery techniques designed
for unstructured peer-to-peer networks. In flooding, a peer initiates a search query and
forwards it to all neighbours. If none of the receiving peer has the required resource,
the query is forwarded to the next level of neighbours. This process is continued
until the node with required resource (destination henceforth) is reached or the TTL
is expired. On each forwarding, the TTL value is reduced by one, thus ensuring the
progress. If a query fails to reach the destination within a predefined TTL bound,
it will be discarded. Also, to prevent duplication, each query is assigned a unique
number, which will be accepted only once by a peer, and any duplicates received
thereafter will be discarded.
In random walk (RW), a peer initiates a search by sending a query to one of its
randomly chosen neighbour. If it is the destination, then the forwarding process is
aborted; otherwise, the query is forwarded to a randomly chosen neighbour of the
current node. Random walk also is bound with TTL. The query forwarding stops
when TTL is expired or if the destination is reached.
The algorithm random walk with neighbour table (RWNT) uses the routing infor-
mation in the neighbouring node. Each node maintains a table consisting the informa-
tion of its neighbours. At first, the source checks the routing table of its neighbours.
If destination is in neighbouring table, then query is forwarded to that peer; other-
wise, a random neighbour is chosen and query is forwarded to that neighbour. The
query will be forwarded in the same manner until destination is reached or the TTL
450 A. Jayan et al.
is expired. This algorithm requires additional messages to fetch routing information

in neighbours table.
The hybrid algorithm flooding with random walk (FRW) sends the query to all the
neighbouring nodes (flooding). If destination is not in neighbours, then the searching
is done using random walk algorithm and the query is forwarded one of its random
neighbour. Random walk algorithm will be done until destination is reached or the
ttl value is expired.
Flooding with random walk with neighbouring table (FRWNT) is another hybrid
algorithm. Similar to FRW, this algorithm also has two phases of operation. In the
first phase, the source node that initiates a query and sends it to all neighbours, as in
flooding. If none of these neighbours has the required resource, all of them perform
a lookup in a table that stores resource information of their neighbours. Note that
every peer in the network stores such a table and is periodically updated. If this table
lookup finds a destination, then the query will be forwarded to that peer. Otherwise,
a search is initiated from the current peer, using the random walk technique. This
constitutes the second phase of the algorithm.
3 Experimental Setup and Results
The algorithms are implemented in Python, and the NetworkX Package is used to
generate random graphs. The random graphs G(n, m) and G(n, p) are generated for
various values of n, m, and p, and the corresponding performances are recorded.
The parameters observed are hop count, failure rate, and message load. For a given
graph, the whole process is repeated for various source–destination pairs and the
corresponding values are recorded and averaged.
Although, a theoretical upper bound for the TTL is the diameter of the graph,
it would be too stringent an upper bound for unstructured networks. So, we relax
it to twice the diameter of the graph. Also, we evaluate the performance of these
algorithms in random graph for various edge probabilities ranging from 0.3 (sparse
graph) to 0.9 (dense graphs). The two random graphs G(n, m) and G(n, p) have
shown similar results. So, we consolidate them and the results are presented in next
section.
Average Hop: Whenever a query passes from one peer to another, the hop increases
by one. In other words, hop counts the number of edges in the path from source
to destination. Average hop is obtained by averaging the hops across all source
destination pairs for a given graph. Unlike the previous work, we count edges in the
successful path only. The average hop is plotted against the probability of edges. It
has been observed that flooding outperforms the other algorithms because it finds
the shortest path from source to the destination, and consequently, the hop count is
minimized.
The other algorithms shows almost similar pattern and near values. The results are
depicted in Fig. 2.
Revisiting Algorithms for Resource Discovery in Unstructured … 451
Fig. 2 Average hop versus probability of edges
Fig. 3 Failure rate versus probability of edges
Failure Rate: Following the previous work, failure rate is defined as the number
of times an algorithm fails to reach the destination before the TTL expires. Since the
TTL is fixed at twice the diameter of the network, a failure in this case means that
the packet has traversed at least twice the required distance through the network,
but failed to reach the destination. It has been observed that flooding has the lowest
failure rate compared to other algorithms because in flooding all possible paths are
explored, which ensures that no valid path is missed.
Also, it has been observed that the failure rate decreases when probability of
edges increases. This is obvious because as the probability of edges increases the
connectivity of the graph increases. The results are depicted in Fig. 3.
452 A. Jayan et al.
Messages Load: This parameter is used to capture the number of message packets
circulated in the network whilst a search query reaches destination from source.
This parameter indirectly measures the congestion that may occur in the network. It
has been observed that RWNT and RW have the lowest message load because both
of them follow a search path that is linear. Flooding has the maximum because it
sends messages to all neighbours (maximum branching). The message load of other
algorithm lies between that of RW and flooding.
4 Conclusion
Routing algorithms for unstructured p2p networks is a developing area with a lot of
research potential. The algorithms we analysed show that hop count and failure rate
is less for flooding, but the message load is prohibitively high. Hence, it can lead
to high bandwidth consumption and congestion in the network. By incorporating
suitable mechanisms to limit this message load explosion, it is possible to develop a
message load-controlled flooding algorithm. The two hybrid algorithms give an idea
in this direction, which may be explored further.
References
1. F. Dabek et al., Building peer-to-peer systems with chord, a distributed lookup service, in Pro-
ceedings Eighth Workshop on Hot Topics in Operating Systems (Elmau, Germany, 2001), pp.
81–86. https://doi.org/10.1109/HOTOS.2001.990065
2. S. Ratnasamy, P. Francis, M. Handley, R. Karp, S. Shenker, A scalable content-addressable
network, in Proceedings of the 2001 Conference on Applications, Technologies, Architectures,
and Protocols for Computer Communications (2001), pp. 161–172
3. M. Faloutsos, P. Faloutsos, C. Faloutsos, On power-law relationships of the internet topology.
ACM SIGCOMM Comput. Commun. Rev. 29(4), 251–62 (1999)
4. E.K. Lua, J. Crowcroft, M. Pias, R. Sharma, S. Lim, A survey and comparison of peer-to-peer
overlay network schemes. IEEE Commun. Surv. Tutor. 7(1–4), 72–93 (2005)
5. N.J. Navimipour, F.S. Milani, A comprehensive study of the resource discovery techniques in
Peer-to-Peer networks. Peer-to-Peer Netw. Appl. 8(3), 474–92 (2015)
6. T.K. Prasad, P. Jayakumar, G.P. Sajeev, A K-clique based clustering protocol for resource dis-
covery in P2P network, in 2018 International Conference on Advances in Computing, Commu-
nications and Informatics (ICACCI) (IEEE, 2018), pp. 2475–2480
7. L. Bashmal, A. Almulifi, H. Kurdi, Hybrid resource discovery algorithms for unstructured peer-
to-peer networks. Procedia Comput. Sci. 109, 289–96 (2017)
8. B. Bollobás, B. Béla, Random graphs (Cambridge University Press, Cambridge, 2001)
9. D. Tsoumakos, N. Roussopoulos, Adaptive probabilistic search for peer-to-peer networks, in
Proceedings Third International Conference on Peer-to-Peer Computing (P2P2003) (Linkoping,
Sweden, 2003), pp. 102–109. https://doi.org/10.1109/PTP.2003.1231509
802.11 Frame-level Network IDS for
Public Wireless Networks
Anish Sujanani and Shashidhar Pai
Abstract The growth in capability of consumer electronics, intelligent systems,

and wireless networking technology has brought about an unprecedented scale in
communication by networked devices. With scale of capability comes information
security issues that present themselves through all layers of software and network
stacks. In this paper, we propose a lightweight network intrusion detection system
for link-physical layer devices operating the 802.11 suite of protocols. Though many
attempts have been made at detecting abnormal behavior through protocol modifi-
cation, firmware augmentation, and machine learning, a novel approach has been
proposed which is suitable for non-compute-intensive embedded devices in a dis-
tributed network, synonymous to applications of smart-city wide area networks and
relatively smaller wireless local area networks.
Keywords Network security · Intrusion detection · Wireless networking · IoT
1 Introduction
The evolution of wireless networking and mobile computing has brought about a
massive leap in application capability, and system potential, however, has also given
rise to a slew of security concerns which need to be addressed to reduce risk and
provide sustainable and safe communication systems. The 802.11 protocol [1] is a
family of protocols governed by the IEEE that dictates communication of ‘frames’
between network access points and hosts at the link and physical layer (L2/L1) of
the OSI networking model. The 802.11 family specifies data plane, control plane,
and management plane network operations over the radio frequency spectrum, with
different variants of the protocol being implemented at different frequency bands and
A. Sujanani (B) · S. Pai

Bangalore, Karnataka, India
e-mail: ansujanani@gmail.com
S. Pai
e-mail: shashidharpai95@gmail.com
454 A. Sujanani and S. Pai
radio channels. Security concerns present in this layer are quite unique compared
to those found in upper layers in the sense that they are largely bound to physical
media and electronic interference vectors rather than purely virtualized, software
vulnerabilities. Though the majority of network-based intrusion detection systems
(NIDS) are based on software-defined upper layers of the OSI model, it is crucial to
secure the link/signal layer as this serves as the transmission medium in the com-
munication pipeline and moreover may be relatively more feasible to implement in
constrained embedded environments where custom network stacks and rigorous soft-
ware are costly. Considering the progression of a digitised society, IoT environments
and ‘smart-city’ concepts, the implementation of massive public WAN for civilians
is not unforeseen.
Prior machine learning-based approaches have involved recognizing 802.11 network-
ing anomalies through autoencoder-based ladder networks that combine supervised
and unsupervised learning techniques [2] and fuzzy networks [3]. Support vector
machines (SVM) for classification with particle swarm optimized clustering of fea-
tures for efficient feature extraction have been implemented [4]. A comparative study
on Bayesian models, decision trees, SVM, and rule learning has been proposed [5].
Attempts have been made involving the use of discrete event systems and parame-
ters such as the retry bit, sequence number, and association responses for evil twin
detection [6]. Denial-of-Service (DoS) attack detection has been proposed through
MAC-layer firmware modification and Markov chaining for threshold estimation [7].
DoS attack detection has been attempted through maintaining a valid device MAC
address table as well as limiting the number of frames forwarded per device [8]. Fea-
tures such as the rolling mean of de-authentication packets have been correlated with
network throughput to indicate DoS attacks [9]. DoS mitigation has been proposed
by the integration of a message integrity code into association, authentication, dis-
association, and de-authentication frames [10]. Electromagnetic signal monitoring
has been proposed for anomaly detection [11].
Though many approaches suggest reasonably high accuracy in terms of detection
and mitigation, they are costly to implement in terms of computing capability and
system complexity, especially so in a mobile-agent environment. Approaches have
involved modification to firmware on devices, modification of the protocol to add
encryption tokens which may be hardware dependent and compute-intensive statis-
tical modeling and learning networks. We propose a lightweight solution, apt for
small-scale compute devices that inspects management frames of the 802.11 proto-
col that approach a heuristic intrusion detection system via rule generalization and
thresholding. In order to maximize detection capability and accuracy while still main-
taining a resourcefulness in terms of compute and storage capability, we consider
the following frames:
1. Beacon Frames: Such frames are periodically broadcasted by 802.11 compatible
access points to advertise their presence and services including authentication
mechanisms, quality of service terms, and hardware capability details.
2. Authentication Frames: A device choosing to join the network managed by
an access point will have to authenticate to the network before being allowed
802.11 Frame-level Network IDS for Public Wireless Networks 455
to associate and transfer data. There are a variety of authentication mechanisms

available; however, we consider the WPA2-PSK protocol which offers a four-way
handshake providing a device-specific session key, encrypted with the network
authentication key.
3. De-authentication Frames: De-authentication takes place through mutual flush-
ing of session parameters on both the access point and the client. The client will
then need to re-authenticate with the WPA2-PSK key before associating with the
network during the next session.
4. Association Frames: These frames are sent to synchronize information across
parties and aid in management plane operations. Once authenticated to an access
point, devices will automatically re-connect on availability and will not have to
re-authenticate unless forced to de-authenticate.
5. Disassociation Frames: Temporary disconnects in sessions are implemented
through the transmission of these packets. As opposed to de-authentication, ses-
sion caches are maintained on both devices and automatic re-connection is allowed
without having to re-authenticate, based on device capability.
2 Capturing and Parsing 802.11 Management Frames
Network Interface Cards (NICs) are typically configured to operate in two modes—
managed mode and monitor mode. The choice of switching between the modes is
regulated by the operating system interacting with the card. The more commonly
available managed mode implies that the client has to be associated with a wireless
access point (AP) through the NIC before sending or receiving data. From the client’s
perspective, communication in this mode is limited to packets to and from the asso-
ciated access point only. In contrast, monitor mode allows a client to independently
interact with traffic passing through the physical medium, i.e., radio signal. Hence,
monitor mode is invoked by privileged operating system processes for access point
discovery before managed mode is activated for device association.
Since the 802.11 protocol transfers frames at the link layer of the OSI model, seg-
ments within the packets that belong to upper layers will be encrypted with session
keys based on the authentication mechanism. In order to maintain encryption and to
improve performance, we discard control and data plane packets and rely purely on
management frames. A NIC that supports monitor mode has been attached to our
machine and has been configured to listen for traffic exclusively from a white list
of MAC (physical) addresses belonging to our test network. The test environment
depicted in Fig. 1 comprises of an 802.11b supported wireless access point providing
service on the 2.4 GHz frequency band, multiple client personal machines, mobile
devices, a Raspberry Pi to emulate embedded electronics and the detection machine.
Each of these devices have automated and manual processes performing AP opera-
tions, video streaming, and Internet usage to generate traffic.
The IDS is Python based and uses raw sockets on the NIC’s virtual interface. Details
of specific management frames and fields parsed are shown in Table 1.
Fig. 1 Network architecture overview
Table 1 Management frames and fields Parsed

Beacon Authentication De-authentication Association Disassociation
Frame subtype Yes, 0 × 80 Yes, 0 × 0b Yes, 0 × 0c Yes, 0 × 00 Yes, 0 × 0a
Frame flags Yes Yes Yes Yes Yes
Transmitter Yes Yes Yes Yes Yes
address
Receiver address Yes Yes Yes Yes Yes
BSSID Yes Yes Yes Yes Yes
Sequence number Yes No No No No
Timestamp Yes No No No No
Interval Yes No No No No
Tag (0) length Yes No No No No
Tag (0) Yes No No No No
value—SSID
Capture Yes Yes Yes Yes Yes
timestamp
Parse timestamp Yes Yes Yes Yes Yes
The proposed software architecture depicted in Fig. 2 involves independent threads

for capturing, parsing, and analyzing packets through thread-safe data structures.
This allows for manipulation and tuning of components to improve parsing perfor-
mance and reduce alerting latency. Elasticity is enforced by the fact that the parsing
thread duplicates each parsed packet into alert-specific queues, which can be scaled
horizontally for extending attack vector detection. This architecture has the added
benefit of independent instrumentation and metric generation for each alert, provid-
ing specific telemetry and potential for enhancement.
Fig. 2 Software architecture overview
3 Attack Vector Specification
3.1 Evil Twin Access Points
An Evil Twin Access Point is characterized by an access point deployed by a bad actor
with the same SSID as a known-good access point. Though most modern devices do
not automatically connect to these access points due to their fingerprinting via Basic
Service Set Identifiers (BSSID—Physical address), the user may voluntarily choose
to connect to it due to the similarity in Service Set Identifiers (SSID—Network
name). The rogue access point effectively serves as a man in the middle and will be
able to intercept, modify, and re-route packets while servicing requests as a known-
good access point would, giving the user a false sense of security. Instrumentation
of the protocol on the bad actor’s machine can also lead to a client unintentionally
providing legitimate-network credentials by authenticating to the evil twin. As part of
a standalone deployment or in a distributed routing network, we detect this vector by
looking for beacon frames advertising white-listed SSIDs having different BSSIDs
than registered.
3.2 Forced De-authentication and WPA-Handshake Brute

Forcing
De-authentication on an 802.11 network involves a mutual disconnect and flushing

of session information between both the access point and the client. Connecting back
to the network involves authenticating to the network first, followed by associating
with a session. Such mechanisms are subject to abuse by bad actors and typically
lead up to credential harvesting and denial-of-service attacks. The process is char-

acterized by a bad actor transmitting (flooding) multiple 802.11 de-authentication
frames to an access point and client device with transmitter addresses spoofed to rep-
resent legitimate communication. This stream of packets misleads the access point
into believing that the client did in fact, request for de-authentication and proceeds
to terminate the session and flush authentication caches. Upon reconnecting to the
network, the authentication handshake involves creating a unique device-specific
session key encrypted with the access point authentication key for secure communi-
cation. A bad actor now listening for the handshake is able to intercept the encrypted
session key and brute-force it offline, avoiding detection mechanisms by the network.
Discovering the encryption key reveals the session key for that client, rendering the
entire communication stream between the client and the host compromised. This
primary session key may also provide access to join the network, thereby increas-
ing the risk of security events for all connected clients. We detect such attacks by
applying a threshold to the number of de-authentication frames sent to the access
point over a period of time. These values can be tuned to detect different tools and
have been implemented through a device-specific hash table and a sliding window
to look for multiple thresholding conditions across parameters such as number of
packets sent, total time elapsed since packet flood, and delay between packets for a
packet-in-stream counter. For our tests, we have included a tuple of 64 packets in a
5 second stream with a minimum 60 milliseconds in between them. Multiple tuples
may be configured to detect transmission patterns of different tool sets that may be
used by bad actors.
3.3 Denial of Service (DoS)/Distributed Denial of Service

(DDoS)
Denial of Service refers to the disruption or unavailability of a service to a legitimate

client, usually due to system outage by malicious activity. DoS on the 802.11 layer
is initiated by a wide variety of vectors including signal jamming, signal interfer-
ence, and evil twins among many others. We consider a vector that prevents control
and data plane sessions through the continuous transmission of disassociation and
de-authentication frames to the access point and a legitimate client. Such attacks can
either target a single device or an entire network by transmitting spoofed frames. We
categorize this attack by looking for sustained transmission of de-authentication and
disassociation frames sent to the access point, regardless of the source address to
detect targeted or generalized DoS. Thresholds for the alert were defined empirically
at 128 frames and use a sliding window mechanism and a hash table of each authen-
ticated device to maintain context up to the previous 5 seconds. Multiple counters
and time-windows may be defined to detect attacks triggered by different tool sets.
3.4 Unauthorized Device Traffic
Attackers may be able to bypass blacklist checks by access points for their device
parameters by spoofing information in 802.11 frames to gain and maintain access
on the network. In order to detect unauthorized device access, we look for frames
transmitted by blacklisted physical (MAC) addresses. Looking at management plane
frames provides good detection capability as these are needed before data transfer
can take place. Considering a distributed network of wireless routers, the process
of updating and synchronizing information on all nodes across geography becomes
quite a complex challenge. The system system allows for access points to look for
blacklisted devices trying to interact with the network within the signal range of
the transceiver. This mechanism provides an additional measure of security through
redundancy and compensates for issues such as latency and failure of synchronized
updates as well as access point failure in a distributed wide area network.
4 Results
From Table 2 and Fig. 3, we observe that despite a sharp increase in the number
of devices on the network and the accompanying jump in network traffic, parse
latency and alert latency increase relatively marginally. Sensitivity of the agent in
terms of being able to sense packets as well as the proposed software architecture for
parsing and running heuristics on packets is maintained, irrespective of the increase
in devices, traffic, and percentage of malicious packets.
5 Conclusion
With significant advances in wireless networking, mobile computing, and societal

digitisation, the period of unified massive scale wide area networks is nearing. The
magnitude of growth warrants the need for efficient, lightweight, and scalable security
systems to sustain and progress development. The above system proposes a novel
approach to an 802.11 frame-level IDS for wireless networks that is compute-efficient
and is apt for distributed topology. Future work includes extension of heuristics for
detection of different attacks and a secure policy broadcast channel for access points
to communicate and update the state of their environments. The applicability of the
same to massive scale public WANs for smart cities, WAN for organizations, and
relatively smaller domestic WLAN networks advances the state of network security
by detecting vectors that are not present in software-defined upper layers of associated
centralized OSI model networks.
460
Table 2 Parse latency and alert latency for test-sets

Malicious Parse latency (s) Alert latency (s)
packets
Evil twin De-auth DoS Unauth Evil twin De-auth DoS Unauth
device device
1 Device—10,000 Packets 1964 0.000630 0.000322 0.001365 0.001881 0.001205 0.131371 0.038018 0.051167
2 Devices—20,000 Packets 1589 0.001355 0.000489 0.002395 0.002980 0.027018 0.130987 0.237640 0.064309
4 Devices—40,000 Packets 3719 0.002812 0.000544 0.002560 0.002458 0.052014 0.131101 0.176611 0.068994
6 Devices—60,000 Packets 1561 0.002549 0.000657 0.012494 0.004091 0.077435 0.131287 0.255493 0.088564
A. Sujanani and S. Pai
(a) Parse Latency (b) Alert Latency
Fig. 3 Variation of performance for different test-sets
References
1. Ieee standard for information technology–telecommunications and information exchange

between systems local and metropolitan area networks—specific requirements—part 11: Wire-
less lan medium access control (mac) and physical layer (phy) specifications. IEEE Std 802.11-
2016 (Revision of IEEE Std 802.11-2012) (2016), pp. 1–3534
2. J. Ran, Y. Ji, B. Tang, A semi-supervised learning approach to IEEE 802.11 network anomaly
detection, in 2019 IEEE 89th Vehicular Technology Conference (VTC2019-Spring) (2019), pp.
1–5
3. D.W.F.L. Vilela, A.D.P. Lotufo, C.R. Santos, Fuzzy artmap neural network ids evaluation
applied for real IEEE 802.11w data base, in 2018 International Joint Conference on Neural
Networks (IJCNN) (2018), pp. 1–7
4. M. Usha, P. Kavitha, Anomaly based intrusion detection for 802.11 networks with optimal
features using svm classifier. Wirel. Netw. 23(8), 2431–2446 (2017)
5. M. Agarwal, S. Biswas, S. Nandi, Detection of de-authentication dos attacks in wi-fi networks:
a machine learning approach, in 2015 IEEE International Conference on Systems, Man, and
Cybernetics (2015), pp. 246–251
6. N.S. Selvarathinam, A.K. Dhar, S. Biswas, Evil twin attack detection using discrete event
systems in IEEE 802.11 wi-fi networks, in 2019 27th Mediterranean Conference on Control
and Automation (MED) (2019), pp. 316–321
7. J. Soryal, T. Saadawi, IEEE 802.11 dos attack detection and mitigation utilizing cross layer
design. Ad Hoc Netw. 14, 71–83 (2014)
8. A. Gupta, M. Garg, Dos attacks on IEEE 802.11 wireless networks and its proposed solutions.
Available at SSRN: https://ssrn.com/abstract=1645757
9. M. Agarwal, S. Biswas, S. Nandi, Detection of de-authentication denial of service attack in
802.11 networks, in 2013 Annual IEEE India Conference (INDICON) (2013), pp. 1–6
10. R. Singh, T. Parval Sharma, Detecting and reducing the denial of service attacks in wlans, in
2011 World Congress on Information and Communication Technologies (2011), pp. 968–973
11. J. Villain, V. Deniau, A. Fleury, E.P. Simon, C. Gransart, R. Kousri, Em monitoring and clas-
sification of iemi and protocol-based attacks on ieee 802.11n communication networks. IEEE
Transactions on Electromagnetic Compatibility 61(6), 1771–1781 (2019)
Machine Learning for Malicious URL
Detection
Gold Wejinya and Sajal Bhatia
Abstract In recent years, Web-based attacks have become one of the most common
threats. Threat actors tend to use malicious URLs to intentionally deceive users and
launch attacks. Several approaches such as blacklisting have been implemented to
detect malicious URLs. These unreliable approaches were also accompanied with
strenuous task of maintaining an up-to-date blacklist URL database. To detect mali-
cious URLs, machine learning techniques have been explored in recent years. This
method analyzes different features of a URL and trains a prediction model on an
already existing dataset of both malicious and benign URLs. This paper proposes
a MuD (Malicious URL Detection) model which utilizes three supervised machine
learning classifiers—support vector machine, logistic regression and Naive Bayes—
to effectively and accurately detect malicious URLs. The preliminary results indicate
that Naïve Bayes algorithm produced best results.
Keywords Malicious URL · Machine learning · Support vector machine ·

Logistic regression · Naïve Bayes
1 Introduction
Studies have shown that at least 47% of the world’s population is online at any given
time [1]. Most users especially people using home computers are a lot more vulnera-
ble to external attacks and threats because they have very little knowledge of some of
the attacks that can be launched against them and how to be protected [2]. These users
cannot differentiate between a legitimate and malicious URL because most malicious
Web sites have similar content as the legitimate ones. Threat actors take advantage
of this opportunity by creating malicious Web contents and lure unsuspecting users
G. Wejinya · S. Bhatia (B)

Sacred Heart University, Fairfield, CT 06826, USA
e-mail: bhatias@sacredheart.edu
G. Wejinya
e-mail: wejinyag@mail.sacredheart.edu
464 G. Wejinya and S. Bhatia
to these URLs. Identification of these malicious URLs and creating awareness to

users is one way to reduce or prevent these phishing and malware attacks. Different
researchers have proposed various methods to identify malicious URLs. Machine
learning algorithms have come to be one of the most effective techniques for detect-
ing malicious URLs. Research has been carried out to compare different machine
learning algorithms for their accuracy in detecting malicious URLs.
Our proposed MuD(Malicious URL Detection) model describes a process of
detecting malicious URLs using three different machine learning algorithms and
evaluates which algorithm works best. It explores different algorithms and imple-
ments a working prototype that classifies URLs as benign or malicious based on
certain features of a URL that were seen to be relevant in the detection of mali-
cious URLs. The previous work done to detect malicious URLs using blacklisting or
heuristics is found to lack reliability as they lack capabilities to detect new malicious
URLs.
The rest of the article is organized as follow: Sect. 2 gives an overview of perti-
nent work in malicious URL detection. Section 3 gives a detailed description of the
proposed MuD (Malicious URL Detection) model. Section 4 presents experimental
results and compares the performance of different classifiers. Section 5 summarizes
the paper and presents directions for future research work in this area.
2 Background and Related Work
Over the past few years, there have been studies on the detection of malicious URLs
using machine learning techniques in order to deter attacks originating from such
URLs. This MuD model proposed in this paper makes an effort to stop similar attacks
by effectively and efficiently detecting malicious URLs. This paper enumerates more
on the work done by Vanhoenshoven et al. [3] where malicious URL detection is
mapped as a binary classification problem on a publicly available dataset consisting
of 2.4 million URLs and 3.2 million features. Their experimental results demonstrate
that most classification methods achieve acceptable prediction rates without requiring
advanced feature selection techniques or domain expertise. Their obtained results
showed that random forest and multi-layer perceptron attained the highest accuracy.
Singh and Goyal [4] proposed malicious URL detection by focusing on the
attribute selection. Their analysis was based on the fact that machine learning tech-
niques and selection of attributes are more important than any other aspect. They con-
sidered twenty-five attributes that can help find malicious Web sites. These attributes
were analyzed in regards to resources (computational) needed for extraction, pro-
cessing and accurately classifying the predicted malicious URLs. Based on their
analysis, the top five attributes for detecting malicious URLs were identified to be
geographical location, URL properties, HTTPS-enabled Web site, domain name and
DNS WHOIS information.
Manjeri et al. [5] focused on the feature selection technique after the URL dataset
has been fed into the prediction model. The authors also addressed the issues of class
Machine Learning for Malicious URL Detection 465
imbalance in several classifiers and proposed incorporating a reduced number of

features, based on their importance, required for classification. They also proposed
the use of rule mining algorithms such as Apriori and FPGrowth to generate IF-
THEN rules which helped in establishing relationship among these features [5]. They
concluded that when comparing class imbalance against random forest, decision tree,
logistic regression, k-NN and SVM and random forest had 96% accuracy. Cui et
al. [6] emphasized on using keyword matching to detect malicious URLs. They used
statistical analysis for gradient learning and sigmoidal threshold level for feature
extraction in their proposed techniques. Naïve Bayes, SVM classifiers and decision
trees were used for validation and efficiency calculations of their proposed approach
which has a good detection performance and an accuracy rate of 98.7%.
3 Proposed Malicious URL Detection (MUD) Model
The proposed model is an approach for malicious URL detection based on the features
of already existing malicious URLs using machine learning. With any URL, 15
features are selected and categorized as either lexical-, host-based or content-based.
The classifiers used in this project are trained using the dataset [7], and subsequently,
URLs fed into the model are classified as being either malicious or benign. In order to
properly build and train the proposed MuD model (Fig. 1) to effectively and efficiently
detect malicious URLs, the following steps were taken—(a) data gathering; (b) data
preprocessing; (c) classification; (d) training and testing; (e) evaluation.
Fig. 1 MuD (Malicious URL Detection) model for categorizing URLs

Fig. 2 URL components [1]
3.1 Data Gathering
As we will be working with URLs, it is important to clearly understand what a URL

is and what it is comprised of (Fig. 2). A universal resource locator (URL) is the
address of a Web page located in the World Wide Web (WWW), and it is displayed
in the Web browser’s address bar. A URL usually contains a protocol. Some common
protocols that are used today includes Hypertext Transfer Protocol (HTTP) which
helps with site navigation, Hypertext Transfer Protocol Secure (HTTPS) which is
just like HTTP but has a layer of security, file transfer protocol (FTP) which allows
for the exchange or transfer of files between systems and Domain Name Server
(DNS). Cyberattacks are usually performed using these compromised URLs that are
malicious [8]. The first step is to gather URLs, which can be done by visiting some
Web sites that offer malicious URLs links and also finding good URLs.
3.2 Data Preprocessing
This step is used to transform raw data to a clean dataset. It is an important part
in machine learning because the quality of the data used immensely affects the
capacity of the proposed model to learn. In this step, the data is cleaned, null values
in the dataset are replaced or removed, and data is rescaled. Data preprocessing is
important as it makes it possible for the dataset to be used on different machine
learning algorithms. One of the important steps of data preprocessing is feature
extraction.
3.2.1 Feature Extraction:
There are different types of features that can be extracted and used in classify-
ing URLs. Features can be lexical, which are URL length, number of characters
and special characters (“//” , “.”, “@”) [2]. Host-based features are retrieved from
host properties of the URLs which give access to the location of the host, the IP
address and information about the registration of the domain. “Due to the difficulty
of obtaining new IPs, these features are very important to detect malicious URLs” [2].
Content-based features are features obtained from downloading the URLs content.
It is important in prediction as it helps find malicious code that can be embedded
into the HTML by threat actors. Some other features that can help in the detection
are the page rank which has to do with how many times a URL is visited.
3.2.2 Dataset Used
An existing dataset from UCI machine learning repository was used which has
already been preprocessed with 11,055 URLs both malicious and benign and 30
features extracted [7]. The dataset was cleaned to select specific features that we
need and came down with about 15 features. We chose some features from [5] which
explained different features of URL that can help determine if a URL is malicious
or not and three of the top five features from [4] which can be used to predict malign
URLs based on computational resource requirement and classification accuracy. The
following features were used:
1. Having an IP address: checks if the URL has an associated IP address
2. URL length: checks for the number of characters. Threat actors at times use this
feature to hide the malicious parts of the URL
3. Having @ symbol: This makes the Web browser to ignore all the content pre-
ceding the ‘@’ symbol and the real address subsequently follows the symbol
4. Having double slash ‘//’: The existence of ‘//’ within the URL path implies that
the user will be redirected to another site. The ‘//’ in URLs is usually located
on the sixth position of HTTP and position for HTTPS.
5. Link in tags: Legitimate Web sites offer link in tags that offer metadata about
the HTML document
6. Having a sub-domain: Sub-domain is a division of a domain that helps to orga-
nize an existing Web site into a separate site, and also, it helps organize and
navigate to different sections of a main Web site
7. Age of Domain: This information can be extracted from WHOIS database.
Majority of the phishing Web sites have a relatively small age of domain.
8. DNS record: It provides the registration details of the Web resource. The pres-
ence or absence of this information in the DNS WHOIS information has been
found to be linked to maliciousness
9. HTTPS token: This makes HTTPS-enabled Web sites less likely to host mali-
cious content.
10. Page-Rank: This feature aims to measure the relative importance of a Web page
on the Internet. In the dataset used for this paper, it is observed that around 95%
of phishing Web pages have no Page-Rank
11. Google Index: It examines if a Web site is indexed by Google or not. Phishing
Web pages, which are merely accessible for a short period, are often no found
to be indexed by Google
12. Iframe: Iframe HTML tag is well known in being used to download malicious
JavaScript exploits [5]
13. Redirect: When a URL is redirected, it is often linked to malicious behavior.
The feature checks whether the URL is redirected or not.
14. Pop-up window: JavaScript Window Open() pop-ups are often used for adver-
tisements and injecting exploits.
15. Favicon: This is a graphic associated with a specific Web page. If the image is
found to be loaded from a different domain than the one shown in the address
bar, the Web page is likely to contain malicious content.
3.3 Classification
Machine learning is a branch of science which enables computers to learn, adapt,

extrapolate patterns and communicate with each other without explicitly being pro-
grammed to do so and minimal human intervention. Machine learning approaches
use a set of URLs to train a model and based on the statistical properties, learn to
classify the URL as malicious or good. Classification models also known as classi-
fiers usually offer a range of solutions when faced with classification problems. Some
popular classifiers are decision trees, support vector machine, Bayesian networks,
random forests and k-nearest neighbors. This paper uses three classifiers viz; support
vector machine (SVM), logistic regression and Bayesian network (Naïve Bayes).
Support Vector Machine (SVM): is one of the powerful supervised learning
techniques, and it can be used for classification and regression problems. It is usually
used for two-group classification problems. SVM is a non-probabilistic binary linear
classifier and begins with a set of training examples, with each example labeled to
belong to one of the two categories. SVM separates the examples into the categories
dividing it by finding a hyperplane that has the largest distance to the closest data
point of any class which tends to minimize the generalization error of the SVM
classifier.
Naïve Bayes: is a classification algorithm and it is Naïve because it assumes that
every feature of the given variable is independent of each other [4]. It is a probabilistic
classifier. This model is used in the prediction of classification of a new instance by
calculating the probability of each instance in the dataset.
Logistic Regression: This classification algorithm usually gives a properly cal-
ibrated probability and can be used in categorizing whether a URL is malicious or
not, since it only supports binary classification problems. Its goal is to get the best
model which explains the relationship between the outcome and a set of independent
variables.
3.4 Training and Testing
In this step, the dataset is divided into a training and testing set. The training set
is usually 70% of the dataset and the rest is used for testing. The classification
model learns from the training set. The testing set should not include any part of the
training set, and it is used for performance evaluation of the model. When the model
is properly trained, it will be able to perform a task on new data.
3.5 Evaluation
The next step is to evaluate the performance of our classifiers. To be able to correctly
evaluate our models, we make use of confusion matrix as shows in [9] and use F1
score, accuracy, precision and recall as the evaluation metrics
F1 Score: It is a function of precision and recall, calculated using the average of
precision and recall.
Precision × Recall
F1 Score = 2 × (1)
Precision + Recall
Accuracy: This is defined as the overall success rate of the URL prediction tech-
nique.
TP + TN
Accuracy = (2)
TP + TN + FP + FN
Precision: This is the ratio of the positive predictions of URLs that are correctly
classified.
TP
Precision = (3)
TP + FP
Recall: Can be seen as out of all the positive classes (URLs), how much was
actually correctly predicted
TP
Recall = (4)
TP + FN
TP: number of true positives, actual malicious URLs classified correctly

TN: number of true negatives, actual benign URLs classified correctly
FP: number of false positives (error 1), benign URLs classified as malicious
FN: number of false negatives (error 2) malicious URLs classified as benign.
The proposed MuD model is tested on three machine learning classifiers, and the
accuracy, recall and precision of each classifier are recorded. Analysis of the per-
formance of each machine learning classifier used is done based on the precision,
recall, F1-score and accuracy when applied to the UCI machine learning repository
dataset [7]. Classifiers are tasked with the ability to analyze different patterns and be
able to identify the differences between new and the existing patterns.
Confusion matrix (Table 1) is used to evaluate the performance of the three classi-
fiers (Naive Bayes, support vector machine and logistic regression) in detecting and
categorizing URLs as malicious and benign. The highest accuracy score of 100% was
observed for the Naive Bayes, SVM had an accuracy of 98% and logistic regression
gave an accuracy of 96% as shown in Table 2.
Low recall rate as compared to the precision and accuracy as in the case of logis-
tic regression usually indicates that the model will most likely classify bad URLs as
good, and this results more malicious URLs being undetected. However, when preci-
sion values are low, it indicates that the classifier will classify good URLs as bad. The
graphical representation in Fig. 3 shows the confusion matrix which uses accuracy,
F1 score, precision and recall for each of the classifiers to evaluate their performance.
With the features selected from the dataset [7], we observed that the classifiers used
achieved high classification precision and accuracy of above 90%. In classifying
URLs, SVM ran slower as compared to Naive Bayes and logistic regression. Using
lexical-, host- and content-based features of URLs played an important role in the
performance our classifiers. Testing the prediction model with logistic regression
having an accuracy of 96%, it was tested with several malicious and benign URLs
from Phishtank to know how well it classifies URLs.
Table 1 Confusion Matrix for Performance Evaluation of MuD Model

Predicted outcome
Positive Negative
Actual value Positive True positive False negative
Negative False positive True negative
Table 2 Evaluation of classifiers in MuD Model

Classifier Accuracy (%) F1 Score (%) Precision (%) Recall (%)
Support vector 98 9 98 95
machine
Naïve Bayes 100 100 100 100
Logistic 96 93 97 90
regression
This paper addresses the widespread cybersecurity concern where threat actors
bypass security defenses and use URLs to launch various forms of malicious attacks
on unsuspecting individuals. In order to prevent such attacks, the paper proposes the
use of machine learning algorithms to detect malicious URLs. The proposed MuD
(Malicious URL Detection) model is trained using an existing dataset which contains
11,055 URLs, each with 15 unique features, and is applied to three different machine
learning classifiers—support vector machine, logistic regression and Naïve Bayes.
After training and testing the algorithms, it is observed that Naïve Bayes classifier
recorded the highest accuracy.
As part of the future work in this direction, the authors plan to investigate the
performance of the proposed model by assigning distinct weights to different feature
components of the URL. The proposed model is also planned to be deployed online
by integrating it as a Web browser plug-in capable of warning users of potential
malicious URLs in real time. URLs clicked or typed will be checked based on its
features to determine if it is malicious or not. If it is malicious or suspected to be
malicious, there will be a pop-up informing the user of the potential threat and it
will be temporarily blocked except the user chooses to still navigate to the URL.
The future work also includes evaluating the proposed model against more recent
and diverse datasets along with using additional classifiers such as decision trees and
random forest.
References
1. S. Halder, S. Ozdemir, Hands-On Machine Learning for Cybersecurity (Packt, 2018)

2. M. Ferreira, Malicious URL detection using machine learning algorithms, in Digital Privacy
and Security Conference (2019), p. 114
3. F. Vanhoenshoven, G. Nápoles, R. Falcon, K. Vanhoof, M. Köppen, Detecting malicious URLs
using machine learning techniques, in 2016 IEEE Symposium Series on Computational Intelli-
gence (SSCI) (IEEE, 2016), pp. 1–8
4. A. Singh, N. Goyal, A comparison of machine learning attributes for detecting malicious web-
sites, in 2019 11th International Conference on Communication Systems & Networks (COM-
SNETS) (IEEE, 2019), pp. 352–358
5. A.S. Manjeri, R. Kaushik, M. Ajay, P.C. Nair, A machine learning approach for detecting mali-
cious websites using URL features, in 2019 3rd International conference on Electronics, Com-
munication and Aerospace Technology (ICECA) (IEEE, 2019), pp. 555–561
6. B. Cui, S. He, X. Yao, P. Shi, Malicious URL detection with feature extraction based on machine
learning. Int. J. High Perform. Comput. Netw. 12(2), 166–178 (2018)
7. A. Asuncion, D. Newman, UCI Machine Learning Repository (2007)
8. D. Sahoo, C. Liu, S.C. Hoi, Malicious URL Detection Using Machine Learning: A Survey. arXiv
preprint arXiv:1701.07179 (2017)
9. V.M. Patro, M.R. Patra, Augmenting weighted average with confusion matrix to enhance clas-
sification accuracy. Trans. Mach. Learn. Artif. Intell. 2(4), 77–91 (2014)
Implementation of OSPFv3 in IPv4
and IPv6 for Establishing a Wide Area
Network
Fatema Zaman Sinthia, Afsana Nasir, Bijan Paul,

Mohammad Rifat Ahmmad Rashid, and Md Nasim Adnan
Abstract Open shortest path (OSPFv3) is a latest routing protocol which is made
for IPv4 and IPV6 address. We know OSPFv2 routing protocol which is made for
IPv4. Currently, we are turning in IPv6. So OSPFv2 will not work on IPv6. For IPv6,
OSPFv3 protocol discovered which has taken very good attention in the recent years.
As OSPFv3 is new protocol, it is still not efficient in every communication scenario,
but the current world needs very fast network system. So for establishing a wide
area network where multiple areas work together, it is very essential to solve this
issues that will make a very efficient communication system. This paper presents
efficient communication system for wide area network and after that OSPFv3 will
apply on IPv4 and IPv6. From this paper, we also get an idea regarding what will be
the scenario when same protocol applied on different addressing (IPv4 and IPv6).
Keywords Simulation · Virtual-link · Redistribution · Broadcast address
F. Z. Sinthia (B) · A. Nasir · B. Paul · M. R. A. Rashid

University of Liberal Arts Bangladesh, Dhaka, Bangladesh
e-mail: fatema.sinthia.cse@ulab.edu.bd
A. Nasir
e-mail: afsana.nasir.cse@ulab.edu.bd
B. Paul
e-mail: bijan.paul@ulab.edu.bd
M. R. A. Rashid
e-mail: rifat.ahmed@ulab.edu.bd
M. N. Adnan
Jashore University of Science and Technology, Jashore, Bangladesh
e-mail: nasim.adnan@just.edu.bd
474 F. Z. Sinthia et al.
1 Introduction
‘Wide area network system’ is a wireless sensor network where multiple of areas
connecting throughout various kind protocol where they all can work simultaneously.
To control over in a large traffic network area, OSPFv3 networking protocol plays a
significant platform. We can define a large area by making areas into subareas such as
area 0, area 1, area 2 up to area n and can established an easy platform for controlling.
Controlling over to the entire area one specific area needs to define as backbone area
and other will be the entire subcontinent. Other area must connect directly with this
backbone area. But the problem is if any subarea does not connect directly with
backbone area, then the whole wide area network will not work. That means, one
area cannot communicate with other area without being connected directly with the
backbone area.
Beforehand there are many work done by this relevant protocol yet none can reach
the significant efficient result of the entire platform. Several of paper focusing on
this protocol working efficiency performance; merely this paper is able to catch the
implementation of IPV4 and IPV6 platforms. Even not only in the combinations of ip
platform but this paper also able to catch the real-life simulation of multiple working
protocol environment to the entire full zone. Mainly to maintain and catch efficient
result of a wide area where several of protocols performing simultaneously was our
main motivation behind that paper.
2 Related Works
In previously one of paper published, they work with OSPFv3 protocol which is
applied in IPv6 and measured the performance [1, 4, 5, 6, 9]. In these research arti-
cles, they are illustrating new algorithms for network implementation to establishing
minimal link capacity that is required to route in case of node failure without over-
load. The uniqueness of this paper is present when other protocol come and add to
the OSPFv3 in wide area network then how we add these two protocols and what
will be the environment. This paper also provides the solution how subarea relates
to backbone area. So, to be continuing this task implement in ipv4 and in next part
is in ipv6 platform.
3 Routing Protocol
There are many of protocol we know. For this specific solution, I have chosen OSPFv3
protocol. I describe little bit of EIGRP. In here, there in no role of EIGRP but in wide
area network I also add EIGRP because of got to know how two protocols added in
wide area network.
Implementation of OSPFv3 in IPv4 and IPv6 for Establishing … 475
3.1 OSPFv3
One of the most popular link-state routing protocol is open shortest path first version
3 (OSPFv3) which is based on unicast address families (AFs) of IPv6 and IPv4.
The working principle of routing protocol OSPfv3 is it uses shortest paths based on
the destination tables which is developed by each IP packet from source to destina-
tion. For any updates and acknowledgements issue, OSPFv3 is followed by multi-
cast traffic. The addresses FF02::5 and FF02::6 are applied for OSPF routers and
OSPF-designated routers and 224.0.0.5 and 224.0.0.6 are alternative new addresses,
respectively [6, 7].
3.2 EIGRP
Enhanced Interior Gateway Routing Protocol (EIGRP) is classless, distance vector

protocol that uses the concept of an autonomous system to describe a set of contiguous
routes that run the same routing protocol and share routing information also includes
the subnet mask in its route updates. It is also sometimes referred to as hybrid routing
protocol or advanced distance vector protocol. Instead, it sends traditional distance
vector updates that include information about networks plus the cost of reaching
them from the perspective of the advertising route.
4 Scenarios of IPv4 and IPv6
4.1 IPv4
In 1983, IPv4 was first integrated in the ARPANET. IPv4 is the fourth stage of
the internet protocol platform. The IPv4 makes use of a 32-bit address intention
allocating to store 2ˆ32 addresses which is allowing more than 4 billion addresses.
Features of IPv4: Addressing: 32 bit addresses restrictions the space to 4,294,967,296
(232) addresses. Loopback Address: In IPv4: 127.0.0.1. Identify Network Portion:
To identify Assign Default Route: To assign Default Route in IPv4 network we use:
0.0.0.0. There are three different categories of addressing modes in IPv4: (i) Unicast
Addressing Mode; (ii) Broadcast Addressing Mode; (iii) Multicast Addressing Mode.
4.2 IPv6
The latest version of computer networks internet protocol is internet protocol version
6(IPv6). The feature of IPv6 is that it makes very strong platform especially inter-
networking in packet-switched architecture. Like for multiple IP networks uses end-
to-end datagram transmission it works very simultaneously. The security of IPv6
is ensured with the help of integrated Internet Protocol Security (IPsec). The cryp-
tographic keys for encryption and authentication are supported by IPv6. The main
advantage of IPv6 is it uses large range address range which is 128 bits approxi-
mately 3.4 × 1038 addresses whereas in IPv4 is 32 bits addressing. IPv6 is able to
make wide network. Internet protocol version 6(IPv6) also gives identification and
location system of computer networks. IPV6 also uses broadcast multicast traffic
instead [2, 5].
Features of IPv6: (a) Addressing: 128 bits; (b) Loopback Address: In IPv6:::1/128;
(c) Identify Network Portion; (d) To identify network in IPv6 we use: Prefix;
Types of network in IPv6: Multicast: Multicast is point to multipoint and unicast
address. In unicast addressing, there are three address: (i) Global Unicast Address:
Routable or Public Address; (ii) Unique Local Address: No routable or Private
Address; (iii) Link Local Address: New Concept (Auto generate).
5 States of Entire Environments
Lots of simulation tools found like Netsim [8], Packet Treasure, NS2, etc. All those
shows various way of simulation platform. The reason behind that we chose GNS3
because of it provides the real-life environment. This provides a place where we can
test more than 20 different network vendors and also can analysis hope-by-hop with
net path. Here, every router works in real environment mood [9].
5.1 Network Scenario
Defining the full coverage zone in a single platform a demonstration is given by

Figs. 1 and 2. Setting up the full zone we established in two different ip platform via
figure those can be easy to assemble visualization.
Fig. 1 Scenario of IPv4
Fig. 2 Scenario of IPv6
5.2 Describe Network Scenario for IPV4 APPLIED iN

OSPFV3
There is a problem in OSPFv3; backbone area (area 0) is the main area. So, subarea
must connect with main area. But if subarea does not connect with main area directly,
if one subarea connects with another sub area, then the wide network cannot have
established. And if the other environment will add with the OSPF, what the total
environment will be we can get. So, we can see our scenario that EIGRP (here
another protocol adds with OSPF) and area 1 is directly connected with area 0. But
area 2 and area 3 do not connect directly with area 0. So, via virtual-link, we can
build this wide area network system.
Virtual-link is logical linkup for those area who are not directly connected with
backbone area. Every router must have router id. We are going to describe the virtual-
linkup process. Here is the example that will be better for understanding. In router
1–9, every router has their own router id. Router 2, 4, 6, 8 are the area border router
(ABR) which is contain two areas’ network. So every ABR have to tell their own
router id in other ABR of each area. Here router 2 will tell router id to router 4 and
router 4 also will tell router id to router 2. It is vice versa for both (ABR) of each
area. So similarly, router 4 router id will tell router 6 and router 6 router id will tell
router 4. Then lastly router 6 router id will tell router 8 and router 8 router id will tell
router 6. In every (ABR), we must define area so that they can detect the path. In this
process, we can logically connect area 2 and area 3 with backbone area, while area
2 and area 3 were not connected with backbone area previously. Now they related to
backbone area. When we will understand they are connected?—when we will trace
router 9 to router 1, that time we will see every network in every area will be included
to the router 9. So, router 9 will be hold the summary of every network.
(i) Virtual-link: Router 4: area 1 virtual-link 1.1.1.7. Router 5: area 1 virtual-link
1.1.1.5 and area 2 virtual-link 1.1.1.9. Router 8: area 2 virtual-link 1.1.1.7.
(ii) Redistribution for Communication: When another area who is containing
different protocol and including with OSPF area, here we must do redis-
tribution. Area 1 is containing EIGRP (Enhanced Interior Gateway Routing
Protocol) and area 0, 1, 2, 3 is contain OSPF protocol. Here, router 2 is ABR
router who carrying two environments network. So, in this router, we set the
redistribution that EIGRP and OSPF both protocols can work and continue the
communication.
(iii) Why redistribution need: This is the process where we two protocols can
connect. So here without redistribution, we cannot connected OSPFv3 and
EIGRP.
5.3 Describe Network Scenario for IPv6 APPLIED IN

OSPFV3
Now, the process which we followed in OSPFv3 into IPv4, same will apply on IPv6.
But there is some difference we got when we applied. IPv6 do not use subnet mask
and wail card mask. So when we configure the area, we just use in prefix. IPv6 is so
advanced addressing system that we do not need the wail card mask and subnet mask.
We saw that the same environment is working in IPv6 properly (redistribution and
virtual-link). For ‘wide area network system,’ applying IPv6 and the whole system
communicate one area to other area like IPv4 works. But the passing of the packet is
not successful when they are traveling area to area. In IPv6, we applied redistribution
and virtual-link like IPv4 successfully. But some of packets loss showing in IPv6
which we will describe on next part.
Fig. 3 All network (left) and redistribution (right) of R2
6 Simulation of IPv4 and IPv6
This part is completely applied for real-life simulation platform (illustrated in

Fig. 3). This sector is chosen some of areas router performance for catching the
full summarization of the entire zone.
6.1 Simulation of IPv4
In this way, full protocol simulation is organized in IPv4 environment. For router 1:
Here, we can see the entire summary of every area from router 1 to router 9, each of
networks, subnets and connection. This is possible for virtual-link which is logically
connected by redistribution method. Getting the full summarization of the entire area
yet the main routing root protocol is set up by EIGRP via virtual-link. For router 2: In
this router, we did redistribution. So, router 2 carrying two protocol’s network such as
EIGRP and OSPFv3. Figure 3 showing all area network of entire zone and showing
all networking address of two protocols after applying redistribution method. For
router 9: Router 9 showing how much path passing if router 9 ping to router 1. This
is possible for virtual-link so that we can see every area’s network. Packet loss: Here,
we find successful ping rate of every packet.
7 Simulation of IPv6
Every experiment we did in IPv6 which we also did in IPv4. Like the reviewing of
the Summary of every area’s network, redistribution, path detection all those have
also experimented in IPv6 [8] . We were successful in all those process in IPv6. Now,
I will show the packet loss of IPv6.
Packet loss: In the same time, when we started our packet ping from IPv6 that
success rate is not 100%, some of packet is lost. 1, 2, 0, 2 packets delivered out of 5
packets in per ping.
Fig. 4 Showing the success rate IPV4 and IPV6
Security Issue: In this project, we like to describe the way where controlling of
one area to other area. IPv6 is more secure than IPv4. We can control router 1(EIGRP
Area) to router 6(OSPF) area. For example, stay from router 6 one can easily enter in
router 1 in various processes and one can also work in router 1 after done your work
one can exit from router 1, and back to the router 6 again. To do enable this term, we
had to provide security. So that none can easily entered in that router environment
and we can able to insure security.
8 Result Analysis
OSPFv3 successfully applied in IPv4 and IPv6. All of process of OSPFv3 is working
successfully in IPv4 and Ipv6. But IPv6 is losing packet more than Ipv4. From
simulation part, we get that IPv4 (Fig. 4).
Those comprising able to show the rate of success packet transforming. Packet
delivers 100% and Ipv6 packet deliver was not 100%. IPv6 always losing some
packets. The reason behind that IPv6 losing packets, currently we are using in IPv4.
IPv6 is not still implementing. But as recently as our devices recommended also are
IPv6 supported so here network is IPv4 and device is IPv6 supported. When this
two-addressing network we are connecting, there is some collusion occurred that
is why IPv6 losing some packets. Now, here we showed a graph representation of
losing packets in IPV4 and IPv6. Via graph, we can clearly be reviewing the changes
of comprising of IPv4 & IPv6. Once we will turn in IPv6, this losing rate of the
packet will come in 0 percent like today’s IPv4. But in IPv6 is securing rather than
IPv4 in case of ensuring security system.
9 Conclusion
Get reviewing the form this two platforms, we can ensure that within the flexibility
of IPv4 and IPv6 both are work in same way. Limitation of IPv6 is that sometime this
environment shows the packet loss scenario, the reason behind we are still not turn
into the IPV6 network. So, our devices still IPV4 supported. But the best advantages
of this one is we can ensure the security system of within this, we can travel easily one
area to other area. We can entire into any area by staying in any another. On another
way in case of IPv4 it always works in same way of the entire environment. In my
paper, I am focusing that issue that this protocol work in a wide area where the entire
area is defined with in many sub areas via platform of IPv4 and IPv6 and compare
of them is losing packets rates. We see IPv6 losing packets than IPv4. But IPv6 is
secure, without security we cannot establish wide area network and we will be turn
in IPv6 because of IPv4 is going to finish. So IPv6 is better for future networking
communication.
References
1. J.V. Jancy, S. Kumar, A.K. Sahoo, V. Kumar, Performance evaluation of OSPFv3 routing protocol
on IPv6 heterogeneous network. J. Innov. Sci. Eng. Res. 2(1). ISSN: 2456-8619. Available: http://
www.jiser.net/wp-content/uploads/2018/01/70.pdf
2. TechTarget. (2000–2019). IPv6 (Internet Protocol Version 6) [Online] Available: https://search
networking.techtarget.com/definition/IPv6-Internet-Protocol-Version-6
3. Science Engineering Medicine. The New NETSIM Simulation Program [Online]. Available:
https://trid.trb.org/view/348792
4. S.T. Chandel, S. Sharma, Performance evaluation of IPv4 and IPv6 routing protocols on wired,
wireless and hybrid networks. Int. J. Comput. Netw. Appl. (IJCNA) 3(3). Available: http://ijcna.
org/Manuscripts/Volume-3/Issue-3/Vol-3-issue-3-M-02.pdf
5. Z. Ashraf, A. Sohail, M. Yousaf, Performance analysis of network applications on IPv6 cloud
connected virtual machine. I.J. Comput. Netw. Inf. Secur. 12, 1–9 (2019). Available: http://www.
mecs-press.org/ijcnis/ijcnis-v11-n12/IJCNIS-V11-N12-1.pdf
6. R. Whitfield, S.Y. Zhu, A comparison of OSPFv3 and EIGRPv6 in a small IPv6 enterprise
network. Int. J. Adv. Comput. Sci. Appl. (IJACSA) (2015)
7. Paloaltonetworks. Configure ospfv3 [Online]. Available: https://docs.paloaltonetworks.com/
pan-os/8-1/pan-os-admin/networking/ospf/configure-ospfv3.html
8. T.R. Velieva, A.V. Korolkova, D.S. Kulyabov, Designing installations for verification of the
model of active queue management discipline RED in the GNS3, in 2014 6th International
Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT),
St. Petersburg, Russia (IEEE, 2014). Electronic ISBN: 978-1-4799-5291-5
9. A. Hinds, A. Atojoko, S.Y. Zhu, Evaluation of OSPF and EIGRP routing protocols for ipv6.
2(4). Available: https://www.researchgate.net/profile/Shao_Ying_Zhu/publication/272911281_
Evaluation_of_OSPF_and_EIGRP_Routing_Protocols_for_IPv6/links/5731b93408ae08415e
6b79d1.pdf
Obstacle Avoiding and Fire Extinguishing
Robot for Everyday Fire Security
Moumita Kabir, Fatima Noor Popy, Bijan Paul,

Mohammad Rifat Ahmmad Rashid, and Khan Raqib Mahmud
Abstract At present, in Bangladesh, almost all the advanced working places have
fire alarm systems. But the fire alarm systems only cover a certain area. And in
most cases, by the time it alerts people, the fire has already caused a devastating
situation and is out of control. Since fire accidents can be completely natural or man-
made; humans can’t be fully reliable to prevent it from happening. In this modern era,
humans are willing to rely on robots more than other humans. Thus, replacing humans
with automated systems (Robots) can provide a good solution to this problem. And
robotics technology is also becoming much more advanced than before. In this paper,
we will be discussing the system design and impact of a robot that can move around
by detecting obstacles and avoiding it for free path navigation and more importantly
to detect fire and extinguish it by spraying water, if there is any. This Robot can be
used to provide fire security in industrial places as well as in homes. One can charge
this automated system and use it every day for fire security purpose.
Keywords Flame detection · Fire extinguisher · Object detection · L293D IC ·

Connectivity · pinMode
M. Kabir (B) · F. N. Popy · B. Paul · M. R. A. Rashid · K. R. Mahmud

University of Liberal Arts, Bangladesh, Dhaka, Bangladesh
e-mail: moumita.kabir.cse@ulab.edu.bd
F. N. Popy
e-mail: fatima.noor.cse@ulab.edu.bd
B. Paul
e-mail: bijan.paul@ulab.edu.bd
M. R. A. Rashid
e-mail: rifat.ahmed@ulab.edu.bd
K. R. Mahmud
e-mail: raqib.mahmud@ulab.edu.bd
484 M. Kabir et al.
1 Introduction
Every year in Bangladesh, over 1000 people die in fire accidents. Local newspapers
reported that last year at least 1590 people died in fire accidents. Most recent fire
accidents are tragic. The National Human Rights Commission (NHRC) has deemed
the Chawkbazar fire incident (on February 20, 2019) as Bangladesh’s third-largest
fire disaster till date causing 71 deaths. In the Banani fire incident (on March, 2019)
25 people died in a matter of hours. So, that is our motivation to develop this robot
so that this advanced movable fire fighting robot can save peoples life.
This robotic system is powered by Arduino Uno, which is the brain of our Robot.
As shown in (Fig. 1), the robot is placed onto a chassis that can move around using
2 gear motors with wheels by an L293D IC. We have used an ultrasonic sensor to
detect objects [1] within 20 cm range (according to our code). It lights a LED as a
notifier if it finds an object within 20 cm. We have also added a flame sensor which
senses the presence and intensity of fire [2] and rings the buzzer as a notifier if a fire
is present. Whenever it rings the buzzer, the robot will stop and spray water from the
water tank placed on top of a servo motor using a 5 V water pump. The servo motor
will rotate 90° so that it can spray water. The robot structure moves around to avoid
running into obstacles and detect fire and extinguish it concurrently. Whenever the
robot detects fire and an object within 20 cm range, it stops lights the LED (to notify
about the obstacle), rings the buzzer (to notify about the fire) and spray water (to
extinguish the fire). In case of only an obstacle, the robot structure lights the LED
(to notify about the obstacle) and uses the motors to turn left by driving the motors
through the L293D IC. The whole structure is connected to specific Arduino pins.
Necessary coding is done in Arduino Genuino (software) platform and uploaded
into Arduino Uno (hardware) to perform the mentioned mechanism. This paper is
organized as follows in Sect. 2 describes related work, System design and impact and
user satisfaction is demonstrated in Sects. 3 and 4. Section 5 describes Conclusion
and Future work and at the end adds the references.
Fig. 1 Obstacle and fire extinguishing robot

Obstacle Avoiding and Fire Extinguishing Robot for Everyday … 485
2 Related Works
There are already so many researches on developing a robot which can replace human
effort in case of fire accidents. There exist many related works that can be integrated
to develop a robotic system.
• Kirti Bhagat, Sayalee Deshmukh, Shraddha Dhonde, Sneha Ghag and Project
Guide Prof. Vijaya Waghmare’s work on an obstacle avoidance robotic vehicle
using ultrasonic sensor [3].
• Abdulkadir Cakir and Nyan Farooq Ezzulddin’s work on Fire-Extinguishing
Robot design using Arduino [4].
• Sunil Mathew, Gaikwad Sushanth, KR Vishnu, V. Vishnu Nair, and G. Vinoth
Kumar’s work on Fabrication of Fire-Fighting Robot [5].
• Mohammed Azher Therib, Mohammad Abdulhamza, Zaid Sadiq Hameed and
Mustafa Abdulkadhim’s work on Design and Implementation of Car Sensing
System (Obstacle Avoiding Robot) [6].
But in our paper we develop a system by which a robot can not only avoid the
obstacle but also detect and extinguish fire at the same time.
3 System Design
3.1 Tool Specification
See Table 1.
Table 1 Equipment names

Equipment Quantity
and quantity
Arduino Uno R3 1
Breadboard 1
Flame sensor 1
Ultrasonic sonar sensor (HC-SR04) 1
L293D motor driver IC 1
Servo motor mini 1
Water Pump 5 V 1
2 Wheel driver robot platform chassis 1
Buzzer 1
LED 1
Jumper wires 3 sets
Battery 9v 1
486 M. Kabir et al.
3.2 Tool Description
Arduino Uno: It is a microcontroller board which has 14 digital input/output pins

that are used to connect with external electronic circuits [7].
Breadboard: It is a way of constructing electronics without having to use a
soldering iron [8].
Flame sensor: Its fire detection range is 60° and it has 4 pin module—VCC,
Ground, A0 (Analog output), D0 (Digital output) [9].
Ultrasonic sonar sensor: Practically it can measure 2–8 cm distance [10].
L293D Motor Driver IC: It has 16 pins and can control 2 DC motors at a time
[11].
Servo Motor Mini: It is used to rotate an object at specific angle or distance [12].
Water Pump 5 V: A 5 V water pump requires much power that other motors
[13].
Two Wheel Driver Robot platform chassis: The whole structure is placed on
this platform. It has 2 DC motors with two wheels which enable the robot structure
to be movable [14].
3.3 System Working Principle
All the mentioned tools are connected using jumper wires in a specific order to
perform required operations. Figure 2 shows the system flow of obstacle avoiding
Fig. 2 Flow of the system

design
Fig. 3 Connection diagram of the system design
and fire detecting mechanism of our Robot. When our system detects an obstacle or
fire through the sensors, the micro controller controls the movement of the motors
using the motor driver IC. When our Robot detects an obstacle (using ultrasonic
sensor) it uses the motors to turn right to avoid running into the obstacle and when
it detects fire (using flame sensor) it uses the motors to stop.
Figure 3 shows the connection diagram that is used to build the robot using the
specific tools. Connectivity of Ultrasonic Sensor HCSR04 to Arduino: (a) +VCC to
+5 V; (b) GND to GND; (c) TrigPin to 9; (d) EchoPin to 10.
Connectivity of Flame Sensor to Arduino: (a) +VCC to +5 V; (b) GND to GND;
(c) A0 to A0.
Connectivity of L293D IC to Arduino: (a) Pin4, Pin5, Pin12, Pin13 to GND; (b)
Pin8, Pin9, Pin16 to +5 V; (c) Pin2 to 2; (d) Pin7 to 3; (e) Pin10 to 4; (f) Pin15 to 5.
Connectivity of Water Pump to Arduino: (a) GND to GND; (b) VCC to +5 V.
Connectivity of Servo Motor to Arduino: (a) GND to GND; (b) VCC to +5 V;
(c) MyServo to 11.
3.4 Code Explanation
Necessary coding is done in Arduino Genuino (software) platform [15] and uploaded
into Arduino Uno (hardware) using the cable [16].
Step 1: Define variables to store the values coming from the sensors. Variables
for flame sensor: (a) define A0 as sensorPin, (b) take a Boolean variable (fire) and
initialize it as false to store the value coming from the sensor, (c) set 12 as output
pin for a buzzer and (d) take an integer variable (pos) for iteration. Variables for
ultrasonic sensor: (a) set 9 as trigPin to get output from the sensor, (b) set 10 as
echoPin to take input through the sensor and (c) set 7 as output pin for LED.
488 M. Kabir et al.
Variables for Motor Driver L293D IC: (a) define 2 as leftForward, (b) define 3 as
leftBackward, (c) define 4 as rightForward and (d) define 5 as rightBackward.
Variables to measure the distance in cm: (a) take two variables (duration and
distance) to measure distance in cm. Variable for the water pump: (a) define 6 as
pump. Variable for the Micro Servo: (a) take a variable (myservo).
Step 2: Put the Setup Code for pinMode in void setup(). Set pinMode for buzzer
to output. Set pinMode for ultrasonic sensor: (a) set pinMode for LED to output, (b)
set pinMode for trigPin to output and (c) set pinMode for echoPin to input.
Set pinMode for motor driver IC: (a) set pinMode for leftForward to output, (b)
set pinMode for leftBackward to output, (c) set pinMode for rightForward to output
and (d) set pinMode for rightBackward to output. Set pinMode for Servo Motor: a)
attach myservo with 11 and write 90 degree to rotate.
Step 3: Put Fire extinguishing code with condition to run when fire is detected
in void put_off_fire(): (a) set leftForward, rightForward to HIGH and leftBackward
and rightBackward to LOW to move forward, (b) use a for-loop to rotate the servo
motor, (c) set the pump to HIGH to spray water to extinguish the fire, (d) add delay
time, (e) set the pump to LOW and (f) set fire as false.
Step 4: Put the main code with conditions to run repeatedly in void loop().
Condition 1: (a) Check with a condition if there is an obstacle within 20 cm and
for the flame sensor the sensorPin is 1. If yes, then—(a) set the buzzer to LOW and
LED to HIGH and (b) set leftForward and rightBackward to LOW and rightForward
and leftBackward to HIGH to turn left.
Condition 2: Check with a condition if there is no obstacle within 20 cm and
for the flame sensor the sensorPin is 1. If yes, then—(a) set the buzzer and LED
to LOW and (b) set leftForward and rightForward to HIGH and leftBackward and
rightBackward to LOW to move forward.
Condition 3: Check with a condition if there is an obstacle within 20 cm and for
the flame sensor the sensorPin is 0. If yes, then—(a) set the buzzer and LED to HIGH,
(b) set leftForward and rightForward to HIGH and leftBackward and rightBackward
to LOW to move forward and (c) set the fire to true.
Condition 4: Check with a condition if there is an obstacle within 20 cm and for
the flame sensor the sensorPin is 0. If yes, then—(a) set the buzzer to HIGH and the
LED to LOW, (b) set leftForward and rightForward to HIGH and leftBackward and
rightBackward to LOW to move forward and (c) set the fire to true. Whenever fire is
set to true execute void put_off_fire().
4 Impact and User Satisfaction
This is an autonomous robot. This robot is used for avoiding obstacles, detecting fire
and extinguishing it. This robot can detect and avoid obstacles in its path and protect
the robot from any physical damages. The fire detection and extinguishment is the
primary requirement of this autonomous robot (Fig. 4).
Fig. 4 Robot extinguishing

fire
Table 2 Industry survey

Organization Survey people
DHAKA EPZ garments 300
Table 3 User satisfaction

Criteria Satisfaction level (%)
rate
Fire detection capability 70
Fire extinguishing capability 65
Our robot can be used in industries and also in homes where there is a chance
of fire affliction. It can cover a wide area to detect fire and extinguish it. It can be
used as a movable Surveillance System. It can also be used to assist people to cross
the roads safely with its obstacle avoiding feature. Once our robot is automated it
doesn’t require further human assistance. Our Robot can serve fire security service
for any place such as banks, shopping malls, universities etc. Through its capability
of detecting and extinguishing fire, this robot can save the lives of millions of people
around the world.
We have surveyed a garments company in Bangladesh. Table 2 shows the orga-
nization name and the number of people we have surveyed. Table 3 shows the user
satisfaction rate. We have considered two criteria’s for majoring user satisfaction:
(a) Fire Detection Capability rate and (b) Fire Extinguishing Capability rate for our
Robot.
By collecting this feedback we are working for the further development of the
system to make it usable for the welfare of human being.
5 Conclusion
This paper resembles the development components of the obstacle avoiding, fire
detecting and extinguishing robot. We have successfully assembled our Robot
following these components. The fire detection and extinguishment is the primary
490 M. Kabir et al.
requirement of this autonomous robot. However, we plan to add more features like
extinguishing fire with Carbon-di-oxide gas because it is more effective to use it to
extinguish fire rather than the water. We will also work on developing the structure
and design of the robotic system to give a more humanly shape. As future work, we
plan to add GPS system to our Robot to keep a track of the Robot. If we succeed to
add such modifications to this robot then it will be used effectively in the industrial
sectors. We also plan to design and launch this robot as a commercial product since it
would be budget friendly for many industrial sectors. Future modifications can make
this robot really useful for the wellbeing of mankind.
References
1. N. KGD, Features; Instructables (2017, December 27). Retrieved December 27, 2017, from
Instructables: https://www.instructables.com/member/Nayantha%20KGD/
2. Tech Ponder, Features; Instructables (2017, July 20). Retrieved July 20, 2017, from
Instructables Website: https://www.instructables.com/id/Flame-detection-using-Arduino-and-
flame-sensor/
3. K. Bhagat, S. Deshmukh, S. Dhonde, S. Ghag, Obstacle avoidance robot. Int. J. Sci. Eng.
Technol. Res. (IJSETR) 5(2), 439–442
4. A. Çakir, N. Farooq, E. Ezzulddin, Fire-extinguishing robot design by using arduino. IOSR J.
Comput. Eng. (IOSR-JCE), 113–118 (2016)
5. G.S. Sunil Mathew, Fabrication of fire fighting robot. Int. J. Innov. Sci. Res. 375–381 (2016,
April 2)
6. M.A. Mohammed Azher Therib, Design and Implementation of Car Sensing System (Obstacle
Avoiding Robot) (March 2016). Thesis of Bachelor. https://doi.org/10.13140/rg.2.2.15459.
66082
7. A. Aqeel, Introduction to Arduino Uno; The Engineering Projects (2018, June 21). Retrieved
June 21, 2018, from Technoprenuer: https://www.theengineeringprojects.com/2018/06/introd
uction-to-arduino-uno.html
8. Science Buddies, How to use a Breadboard; Tutorials (2019). Retrieved 2019, from Science
Buddies
9. REES52, Arduino Compatible IR Flame Sensor - SR017; REES52 (2018, May 9). Retrieved
April 21, 2019, from REES52: https://rees52.com/arduino-modules/99-flame-sensor-sr017
10. Components101. HC-SR04 Ultrasonic Sensor; Components101 (2017, September 16).
Retrieved September 16, 2017, from Components101: https://components101.com/ultrasonic-
sensor-working-pinout-datasheet
11. M. Aqib, L293D Motor Driver IC with Arduino; Electronics Hobbyist (2016). Retrieved 2016,
from Electronics Hobbyist
12. CircuitDigest, Servo Motor: Basics, Theory & Working Principles; Electronics (2015, August
1). Retrieved August 1, 2015, from CircuitDigest: https://circuitdigest.com/article/servo-
motor-basics
13. Arduino Project Hub (2019, July 9), Retrieved from DIY Aquarium Evapora-
tion Top Off System: https://create.arduino.cc/projecthub/atlasscientific/diy-aquarium-evapor
ation-top-off-system-3c9c6f?ref=tag&ref_id=pump&offset=0
14. Arduino Project Hub, Create: Arduino Project (2017, December 14). Retrieved from Obstacle
Avoiding Car: https://create.arduino.cc/projecthub/adam/obstacle-avoiding-car-a192d9?ref=
search&ref_id=obstacle%20avoiding%20car&offset=0
15. Arduino, (n.d.), Arduino. Retrieved 2020, from Software: Arduino: https://www.arduino.cc/en/
main/software
16. J. Nussey, (n.d.), Dummies. Retrieved 2020, from Computers: Arduino: Dummies: https://www.
dummies.com/computers/arduino/how-to-upload-a-sketch-to-an-arduino/
A Survey on Internet of Things (IoT):
Communication Model, Open
Challenges, Security and Privacy Issues
R. Venkadeshan and M. Jegatha
Abstract In recent years, the Internet of things is an emerging research topic, where
the Internet devices like computers, everyday objects, sensors, and actuators are
being linked through wired and wireless networks in the world and its computing
capabilities allowing the data exchanges and consumption with very minimal human
interaction. The IoT incorporates many multifarious end devices to provide open
access to digital data services over the Internet. In the meantime, security and privacy
are two inseparable areas of this technology. It has been observed that there are several
potential risks that exist in ensuring confidentiality, authentication, and integrity
among the IoT applications. In this paper, the authors begin with the brief introduction
of IoT protocol stack along with the security models supported by each layer. Later,
this survey analysis the security and privacy risk of the IoT model and its open
challenges. Finally, the authors discussed the approaches and a few recommendations
to ensure the fundamental securities of IoT applications.
Keywords IoT security · IoT protocol stack · Security and privacy risk · IoT
challenges · Security · Issues · IoT communication models · 6LoWPAN · RPL
1 Introduction
The Internet of things becomes the hottest topic in the research community by seam-
lessly integrating the devices into the virtual world at anytime, anywhere. In this
Internet world, there are billions of objects interconnecting with one another can
sense, share information, and communicate over the public or private network through
Internet Protocol (IP). The concept of interconnecting all objects in a day-to-day life
was equipped with identifiers by radio frequency identification (RFID) technique
R. Venkadeshan (B)
University of Technology and Applied Sciences - SHINAS, Shinas, Sultanate of Oman
e-mail: venkadengg@gmail.com
M. Jegatha
Chettinad College of Engineering and Technology, Karur, India
e-mail: vpragatha@gmail.com
494 R. Venkadeshan and M. Jegatha
Consumer
Securities Smart Home

& Cities
IoT
Health Care Applications Automobile
Energy &
Infra
Utilities
Agriculture
Fig. 1 IoT applications
[7]. These objects were establishing the communication and managed by sensors
without human interaction. With the help of Internet Protocol IPv6, the billions of
devices could obtain the communication address. As depicted in Fig. 1, the impact of
the Internet of things is spread over different industry sectors [4] to surfacing smart
framework in smart cities [9, 11], transportation, automobile industries, agriculture,
logistics, security strengthening, smart Web, medical care, smart home, and smart
lattice. Due to the pervasive nature of IoT, the primary concern of its application is to
address the security and privacy risk over the data that is being captured and stored
by the sensing devices.
At the time of IoT implementation, certain things need to be finalized such as who
controls the data, where the data has been stored, and how long the data has been
sensed.
In our further discussions, Sect. 2 focuses on the IoT communication protocol
stack [2] with different layers and its security requirements for deployment. The
authors also discuss the security and privacy issues that exist in the communications of
the IoT model in Sect. 3. In Sect. 4, the authors identify the different open challenges
and strategies for providing security for future work in the area. Finally, in Sect. 5, the
authors briefly address the guidance provided to ensure the IoT security to enabling
end-to-end Internet communications which involves sensing devices.
2 IoT Protocol Stack
Figure 2 illustrated the IoT layer model and its communication protocol stack, which
has been designed and is available by IEEE and IETF standardization bodies. The
main characteristics of IoT protocol stack are given below:
(1) Physical (PHY) and Medium Access Control (MAC) layers enable low-energy
communications with the support of IEEE 802.15.4 and there are set of rules
A Survey on Internet of Things (IoT): Communication Model … 495
Fig. 2 IoT communication Application Layer CoAP

protocols stack
Network/Routing Layer IPv6 , ROLL RPL
Adaptation Layer 6LoWPAN
MAC Layer IEEE 802.15.4

IEEE 802.15.4e
Physical Layer IEEE 802.15.4
implemented at lower layers of the stack which govern the communications and
that rest the base for higher layers of IoT protocol stack [5].
(2) The adaptation layer implements Internet Protocol IPv6 and Low-power Wire-
less Personal Area Networks (6LoWPAN) which uses 102 bytes at the higher
layers for the data transmission and IPv6 packets transmission over the IEEE
802.15.4 [1].
(3) In the routing/network layer, the Internet packets are routed over the 6LoWPAN
environs with the support of Low-power and Lossy Networks (RPL) Routing
protocol. RPL acted as a framework, rather than a routing protocol, to support
the requirements of IoT application requirements [6].
(4) In the application layer, the protocol called the Constrained Application Protocol
(CoAP) supports higher layer communications [2].
2.1 Security Requirements
The security requirements of different protocols in IoT protocol stack is to be

protecting in such a way that to offer assurances in the confidentiality, authenti-
cation, integrity, and non-repudiation of information transferred among the devices
[10].
2.2 Providing Security on Layers to Defend IoT
The convergence of millions of devices like sensors, actuators, and other smart objects
in an environment is stands for Internet of things (IoT) where the security challenges
[8] are huge. The primary task is to build a secure system that could provide all the
possible solution to threats over the IoT communications.
Application Layer
A Constrained Application Protocol (CoAP) supported by an application layer of
IoT protocol stack is being designed by Internet Engineering Task Force (IETF)
workgroup named Constrained RESTful Environments (CoRE) [3].
(a) NoSec—in this mode, CoAP messages have been transmitted without security
and it is in exercise provides no security for communications.
(b) PreSharedKey—this mode of security in sensing devices are preprogrammed
which allows the symmetric cryptographic keys to ensure secure communica-
tions among the connected devices.
(c) RawPublicKey—this mode of security will be employed by the sensing devices
that could be required public key infrastructure to enable public key-based
authentication.
(d) Certificates—this mode of security uses the public key-based authentications
mechanism X.509 certificate to provide security to the applications that entertain
the job of certificate validation for access.
Network Layer
The IETF uses route-over technique and Routing Protocol for Low-Power and Lossy
Networks (RPL) approach to handle the routing problems in mesh network as a
standard solution. The RPL protocol utilizes 128-bit keys of Advanced Encryption
Standard (AES)/counter with CBC-MAC (CCM) standards for supporting integrity
in MAC generation. The different security modes supported by RPL are:
(a) Unsecured—this is the default security mode of RPL, where there is no security
mechanism in routing control messages.
(b) Preinstalled—in this security mode, the participating devices should be
preconfigured with some symmetric keys to act as a router or a host.
(c) Authenticated—this mode is suitable for the devices acting as router. Initially,
the devices may have preconfigured key and security mode to establish the
connection with the network and then it obtains the dissimilar cryptographic
key from the central key authority to act as router.
Adaptation Layer
The adaptation layer utilizes the IPv6 Low-power Wireless Personal Area Network
(6LoWPAN) [1] to optimize the network payload using packet header compression
technique, and defining the mechanisms like address auto-configuration and neighbor
discovery. 6LoWPAN is a key technique to support the IoT communications over
the Internet and changes the perception of IPv6 that it is not suitable for constrained
low-energy wireless communication environments.
MAC and PHY Layer
In Physical (PHY) and Medium Access Control (MAC) layers, IEEE 802.15.4 stan-
dards [8] are employed to support the low-energy communications over the IoT
devices, as illustrated in Fig. 3. The wireless sensor network standards such as
ZigBee [12], ZigBee PRO, WirelessHART, and ISA 100.11a are the few adopted
by IEEE 802.15.4 standards. ZigBee standards targeting the home-based automation
and smart energy-based application profiles [12], whereas WirelessHART and ISA
100.11.a targeting industrial automation and control market applications. Achieving
Fig. 3 Security principles of

IoT
the reliability at PHY layer is possible with the help of various modulation tech-
niques such as Direct Sequence Spread Spectrum (DSS), Ultra-Wideband (UWD),
and Chirp Spread Spectrum (CSS). There are various security modes available at
IEEE 802.15.4 standards, which are well-known by their security promises and data
integrity level.
(a) Confidentiality—in IEEE 802.15.4 standard, defining security level is an
optional one, where the application may choose no security or other layers
security of the protocol stack. The applications that are requiring confidentiality
over the transmitted data may use Advanced Encryption Standard in Counter
(AES-CTR) security mode.
(b) Data Authenticity and Integrity—the security modes required by the appli-
cations supporting data authenticity and integrity are Cypher Block Chaining
(CBC) which provides Message Integrity Code (MIC) and Message Authenti-
cation Code (MAC) which affixed to the transmitted data.
3 Security and Privacy Risks in IoT
3.1 Security for IoT
In recent years, IoT communication devices are undergone various potential risks
in the security side which exploited the consumers into the harmful environment.
The security risks are (1) unauthentic data access and mishandling of personal data;
(2) emphasis security attacks on other participating devices; and (3) creating secu-
rity threats. Though, these security risks are existed in traditional communications
in Internet also, but in IoT, these risks get heightened. The different attack routes
are available to the attackers become accessible (“access anytime, anywhere”), and
globally connected (“access anyone”).
There are numerous attack threats that affect the IoT communications such as:
• Denial of service (DoS)—malicious attackers taken the control over the network,
e.g., jamming the channels.
• Physical damage—this threat is a subset of DoS where the attackers are lack of
technical knowledge but destroy the actual things and hinder the IoT services.
• Eavesdropping—this attack is passive in nature which targets different commu-
nication channels to extract the digital data from the flow.
• Node captured—Attackers targeting the data and also the node that are actual
physical location of data.
3.2 Privacy for IoT
The devices communicating over the IoT technologies are almost carrying the
personal data and confidential information which could not be transparent to the
outside world. There are numerous privacy inferences raised from the ubiquity of
IoT devices. Those devices should incorporate some features to ensure the risk-free
communications. There are several key places which require some advancement in
technology to deploy privacy:
• Lightweight cryptographic algorithms are needed to implement in constrained
environments such as RFID, sensors, smart cards, and healthcare devices.
• Techniques like homomorphic and searchable encryption of cryptographic are to
support privacy, authentication, and anonymity.
There are numerous privacy inferences raised from the ubiquity of IoT devices.
Those devices should incorporate some features to ensure the risk-free communica-
tions. The promising approaches facilitate the IoT in privacy-by-design principles to
manage their own data. For providing privacy to sensitive data, an advanced concept
like active bundles might be used, which is a container for sensitive data, metadata,
and a virtual machine (VM).
The features are:
• Privacy over context awareness
• Authenticating all participating devices
• Human-device trust relationship
• Fault-tolerance—robust environment
• Data access control—preserving data location
• Decentralized computing.
4 Challenges and Principles of IoT Security
The most challenging factor in IoT communication is to overcome the security risk
[4] that existing among the participating devices. IoT architectures are dealing with
different kinds of devices of billions of objects which can interact with each other
[8]. All these communications must be secured in such a way that the information
sharing and the services providing should be happened without any intervention and
disturbance. There are different attacks and threats targeting IoT entities through
various communication channels, denial of services, physical threats, and identity
productions [8]. The security principles as given in Fig. 3, to be designed to improve
the security of IoT devices and its communication with full sort of design, protocol,
and deployment.
4.1 Falsehoods in Internet
In IoT environment, many autonomous systems can offer false information in

substantial formats. IoT devices should always checks the transferred data from
the edge to guard autonomous false information from polluting the system.
4.2 System Isolation
The autonomous nature of devices in IoT communication should have an advantage

of isolation where the security countermeasures are never cut down even in the
absence of connectivity.
4.3 Defend Consistently
To protect the data in the communication pathway, the data encryption technique is
used. It is to consider carefully whether the encryption is being applied to the data
throughout the lifecycle uniformity and guarantee the data privacy.
4.4 System Reinforcing
Make sure that the IoT devices are not exposed to the viable features to minimize the
security attacks. In the IoT environment, all the ports and protocols which are unused
should be disabled, and the software installed unnecessary should be uninstalled.
4.5 Unpredictable Data
Devices in IoT are collecting massive of data through sensors and that may look
harmless at first, but later the system may reveal the sensitive information or complex
data to the outside world.
4.6 Targeting Weak Spot
Attackers in the pathways always look for the weakest devices and attempt to deed
it. So, it is to be ensuring that the security controls are maintaining across the inter-
faces in a system. The devices, like mobile, hidden interfaces in API’s, and resource
constrained sensors are must to enforce the security principles.
4.7 Accounting Scalability
It is to check the each design and security consideration of IoT into scalability.
Countermeasures for security must undergone different volumes of IoT devices.
5 Security Guidance
In Table 1, the authors attempted to addresses the different security challenges and
discussed some IoT security recommendations to meet out the challenges.
6 Conclusion
This paper begins brief introduction about the IoT concepts and its applications, and
the needs of having secure communication to support the promising technology. It
describes the IoT protocol stack and security protocol measures for implementing
security requirements which handles the different security risks associated with layers
of IoT protocol stack. And addresses the security and privacy challenges incorporated
with the IoT communication model, which are the most essential factors because,
without sufficient security measures, this giant technology will harm the environ-
ment and not useful to the human societies. Finally, the security principles with the
guidance to provide secure structure for IoT-based Internet communication.
Table 1 Security challenges in IoT and recommendations

Challenges IoT security recommendations
Insecure Web interface Ensure the Web interfaces to check the ability of
HTTPS to guard the transmitted information.
Build the firewalls for the Web applications to
protect vulnerable attacks from outside through
Web interfaces
Inadequate authentication and authorization Ensure for strong passwords-based access to
applications requires authentication
Absence of encryption standards in transport Recommended to use the accepted encryption
practices and to avoid using proprietary
protocols. Ensure the secure transfer of data by
configuring SSL/TLS up to date
Inadequate security configurability Ensure that the encryption standards like
AES-256 ad AES-128 are made available as
default setting for transport
Uncertain software and physical security Ensure the software files are signed, validated,
and updated regularly by the devices before
installing it
References
1. 6lowPAN IETF Working Group, IPv6 over Low power WPAN (6lowpan)—Charter (Online).
Available http://datatracker.ietf.org/wg/6lowpan/charter
2. C. Bormann, A. Castellani, Z. Shelby, CoAP: an application protocol for billions of tiny Internet
nodes. IEEE Internet Comput. 16(2), 62–67 (2012)
3. Q. Chen, G.-r. Quan, B. Yu, L. Yang, Research on security issues of the internet of things. Int.
J. Future Gener. Commun. Netw. 6(6), 1–10 (2013)
4. F. Wang, L. Hu, J. Zhou, K. Zhao, A survey from the perspective of evolutionary process in
the internet of things. Int. J. Distrib. Sens. Netw. (2015). (Hindawi Publishing Corporation)
5. IEEE Standard for Information Technology—Telecommunications and Information Exchange
Between Systems—Local and Metropolitan Area Networks—Specific Requirement Part 15.4:
Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-
Rate Wireless Personal Area Networks (WPANs), IEEE Std. 802.15.4a-2007 (Amendment to
IEEE Std. 802.15.4-2006), pp. 1, 203, (2007)
6. IEEE Standard for Local and Metropolitan Area Networks—Part 15.4: Low-Rate Wireless
Personal Area Networks (LR-WPANs) Amendment 1: MAC Sublayer, IEEE Std. 802.15.4e-
2012 (Amendment to IEEE Std. 802.15.4-2011), (2011) 1–225, (2012)
7. L. Li, J. Chen, System security solutions of RFID system of internet of things sensing layer. J.
Net Secur. Technol. Appl. (6), 34–36 (2011)
8. J. Liu, L. Yang, Application of internet of things in the community security management, in
Third International Conference on IEEE Computational Intelligence, Communication Systems
and Networks (2011), pp. 314–318
9. H. Ning, Z. Wang, Future internet of things architecture: like mankind neural system or social
organization framework? IEEE Commun. Lett. 15(4), 461–463 (2011)
10. R. Venkadeshan, M. Chandrasekar, Effective communication in P2P network by intro-
ducing GOSIP-PHE algorithms, Published in Wireless Personal Communication (Springer
US Publisher, 2016). ISSN: 0929-6212. https://doi.org/10.1007/s11277-015-2625-6
11. P. Vijai, P.B. Sivakumar, Design of IoT systems and analytics in the context of smart city
initiatives in India, in 2nd International Conference on Intelligent Computing, Communication
and Convergence, ICCC 2016, vol. 92 (2016), pp. 583–588
12. ZigBee Alliance, Available online www.zigbee.org
Violence Detection Through Surveillance
System
Abhishek Deshmukh, Kshitij Warang, Yash Pente, and Nilesh Marathe
Abstract This project aims to deliver a system which detects violence in the crowd.
The model does not need human intervention to detect violence as it is automated.
A dataset has been collected for detecting whether violence is taking place or not.
The dataset includes both the scenarios, the ones which contain violence and ones
which does not. Then, the model is trained to analyze whether the scenario contains
violence or not. To detect scenarios which contain violent instances, various deep
learning algorithms are applied on the dataset. CNN and LSTM-based architectures
are experimented separately and in combination on this dataset. The model can be
easily implemented in CCTV camera systems because it is lightweight.
Keywords Convolutional neural network (CNN) · Recurrent neural network

(RNN) · Long short-term memory (LSTM)
1 Introduction
Over the years, the instances of public violence have increased. The weak security
measures have made it easier for violence to breakout in crowd. Violent crowd
instances are difficult to detect because of lack of security cameras and software.
There are CCTV cameras installed in some part of the cities. Moreover, security
forces tend to deploy drones in open fields to detect disturbance in crowd during the
A. Deshmukh (B) · K. Warang · Y. Pente · N. Marathe

Ramrao Adik Institute of Technology, Nerul, India
e-mail: abhishekrd03@gmail.com
K. Warang
e-mail: kwarang7@gmail.com
Y. Pente
e-mail: penteyash@gmail.com
N. Marathe
e-mail: nilesh.marathe@rait.ac.in
504 A. Deshmukh et al.
time of public gatherings. However, these security measures are costly and prone to
human error. They also need constant human supervision.
This project aims to deliver a system which detects violence in the crowd. The
model does not need human intervention to detect violence as it is automated. A
dataset has been collected which includes both the scenarios, the ones which contain
violence and ones which does not; and our model is trained according to it. The
model can be easily installed in CCTV camera systems because it is lightweight.
2 Related Work
This section reviews various researches done in order to detect violence through
surveillance. We have reviewed various works and gathered information which could
help our system to result in more accurate results.
In [1], they have experimented with three models: CNN, LSTM and CNN +
LSTM. The dataset is collected from YouTube. The CNN model has two convolu-
tional layers and 32 filters on each layer. Pre-trained model is used to leverage transfer
learning. The LSTM model has two LSTM layers, two fully connected layers and a
Softmax layer at top. In CNN + LSTM model, a CNN layer was added in front of
the LSTM model. CNN extracts features from videos and LSTM acts as classifier.
It was found out that adding CNN layer in front of LSTM layer improves accuracy.
In [2], the 1D temporal convolutional model used has four layers: an input layer,
convolutional layer, fully connected layer and the output layer. Body sensors are used
to collect data. CNN is used to detect activities like cooking, bathing and sleeping.
The data set that is used is publicly available. Results are analyzed in comparison
with performance of Naïve Bayes, HMM, HSMM and CRF. Although CNN produced
less accurate results as compared to LSTM, training time required by CNN is much
lesser. The disadvantage is sensors are required to collect the data and only simple
activities could be detected.
In [3], they have used four layers of convolution to extract features from the
input sensor data like accelerometer and gyroscope. There are two layers of recur-
rent neural network (LSTM) to establish temporal dynamics between the features.
Proposed architecture also has Softmax layer at the end. The advantage is that the
proposed architecture produced better results than CNN. But the disadvantage is that
the proposed architecture required 20% more training time than baseline CNN.
3 Proposed Work
The design of our system is shown in Fig. 1. The proposed system obtains the video
footage from the surveillance camera to begin with the process of analyzing the
footage for violence. As a single frame of the video cannot be used to determine
if the video contains violence, the system needs to analyze a sequence of frames.
Violence Detection Through Surveillance System 505
Fig. 1 System design
The video footage is broken down into frames, with each frame extracted after an
interval of 0.2 seconds. In the system, a convolutional neural network (CNN) is
used in association with a recurrent neural network (RNN). CNN is used to extract
the features from the video frames and to break them down into activation maps
that will be fed to the RNN. To analyze the sequence of frames where each frame
adds to the context of the next frame, a RNN is used. The CNN and RNN used are
VGG-16 (Visual Geometry Group) and long short-term memory (LSTM) network,
respectively.
For determining if the video has violence or not, multiple frames need to be
analyzed together. Hence, 3 s of the video are analyzed at a time to determine violence
in the video, then the next 3 s and so on. Each second is broken down into 5 frames,
each frame taken at an interval of 0.2 s. Thus, for every second, 5 frames are obtained,
resulting in 15 frames for 3 s which are to be analyzed. The frames are extracted
from the video using OpenCV. The frames obtained from the video are then passed
through the CNN model, VGG16 (Fig. 2).
VGG-16 [4] is a CNN model which consists of 16 layers: 13 convolutional and 3
fully connected layers. There are 4 distinct actions performed in these layers. To train
the VGG16 model, transfer learning is used. Transfer learning ensures that with less
amount of dataset the accuracy is still high. Instead of classifying, the classifier layer
of VGG16 is removed and the output of the penultimate Fully connected layer is given
as input to the RNN. When a frame is passed through the VGG16 model, the model
performs convolution on it using filters which are pre-trained. It uses max-pooling
and activation function like ReLU to extract the features from the input image, which
in this case is the frame from the video. Thus, for 3 s of video, a group of 15 transfer
values is obtained. These transfer values are passed to the RNN architecture, which
is LSTM (Fig. 3).
Fig. 2 VGG-16 layers
Fig. 3 Data flow diagram
LSTM [5] are a special kind of RNN. They are composed of a memory cell and
three regulators which are called gates. Unlike standard neural networks, LSTM
also has feedback connections. This enables them to form long term dependencies
between multiple frames and determine if there is presence of violence in the video.
The group of transfer values will be given as an input to the LSTM. The LSTM will
then determine if the 3 s of videos analyzed contain violence or not. The LSTM will
classify the video as violent or non-violent as a final result. If the video is classified
as violent, a mail having the frame of the video is sent to the respective authorities.
In this way, the authorities are intimated that violence is taking place.
4 Requirement Analysis
In order for the system to work smoothly, the user’s machine should satisfy the
following software and hardware requirements.
4.1 Software Requirements
Python 3.7 or above. Python programming language is mainly used for the
data analysis and classification part of the project. Python is used for backend
development.
Python IDE. The IDE that will be used for writing Python code is IDLE. It
provides a user-friendly environment for development of Python-based applications.
TensorFlow. TensorFlow is an open-source symbolic math library for dataflow
and differentiable programming across a range of tasks, and machine learning
applications like neural networks.
MobileNet. MobileNet [6] is an architecture that uses depth-wise separable convo-
lutions. This results in reduced number of parameters and lightweight deep neural
networks.
OpenCV. OpenCV stands for ‘open-source computer vision’. It is a program-
ming library that focuses on real-time computer vision. It supports the deep learning
frameworks like TensorFlow and Torch/PyTorch.
4.2 Hardware Requirements
• Hard disk with minimum of 250 GB.

• Minimum 4 GB RAM.
• Windows 7 or above.
• Laptop with Intel i5 processor or above.
5 Machine Learning Algorithm
To differentiate between a violent and a non-violent situation, we use machine

learning algorithm. All quick and impulsive physical actions are not violent in nature.
So machine learning approach is used to differentiate between these two. We have
used supervised learning technique. In supervised learning, the model is trained using
labeled dataset. The dataset which is used to train the model consists of different
classes for different values. Once the model is trained, it can be used to predict the
class of the unlabeled data which is given as an input to the model. After training
the model with a particular algorithm, other algorithms can also be used to obtain
higher accuracy and to increase the model’s optimization.
6 Analysis
Dataset was split into 70% for training and 30% for testing randomly, four times.
Then, the inception model was trained on these four datasets separately. Hence,
the result was four trained models. Then, these four models were tested on their
corresponding testing datasets. Fifteen frames of a video are considered for testing.
The violence score of those 15 frames was analyzed. The threshold value is 0.5,
meaning that if violence is detected in five frames, then the video is classified as
violent. After testing, the confusion matrix was prepared and accuracy was calculated.
The confusion matrix is shown (Fig. 4).
7 Result
See Figs. 5, 6 and 7.
The system uses CNN along with LSTM for human activity recognition, specifically
to classify a fighting scene. Footage from the CCTV will be sent to the system where
Fig. 5 Testing of frame from video
Fig. 6 Violence detected in video

Fig. 7 Sent mail when violence is detected
the system will extract a frame at every 0.2 s. If violence is detected, it also sends
a mail to the respective authority along with frame of the video. Implementation of
this model aims to reduce impact of crime. Since transfer learning is used to train
VGG16 model, it ensures that accuracy is high even though the size of dataset is
comparatively low. Further, the model can be trained extensively to identify different
categories of crimes as well. The advancement that can be done in the system is that
the location and time of the crime can also be sent to the local authorities in the mail.
References
1. S.A. Sumon, M.T. Shahria, M.R. Goni, N. Hasan, A.M. Almarufuzzaman, R.M. Rahman, Violent
Crowd Flow Detection Using Deep Learning, in Intelligent Information and Database Systems,
ACIIDS 2019, vol. 11431, Lecture Notes in Computer Science, ed. by N. Nguyen, F. Gaol, T.P.
Hong, B. Trawiński (Springer, Cham, 2019). https://doi.org/10.1007/978-3-030-14799-0_53
2. D. Singh, E. Merdivan, S. Hanke, J. Kropf, M. Geist, A. Holzinger, Convolutional and Recur-
rent Neural Networks for Activity Recognition in Smart Environment, in Towards Integrative
Machine Learning and Knowledge Extraction, vol. 10344, Lecture Notes in Computer Science,
ed. by A. Holzinger, R. Goebel, M. Ferri, V. Palade (Springer, Cham, 2017). https://doi.org/10.
1007/978-3-319-69775-8_12
3. F. Ordóñez, D. Roggen, Deep convolutional and LSTM recurrent neural networks for multimodal
wearable activity recognition. Sensors 16, 115 (2016). https://doi.org/10.3390/s16010115
4. K. Simonyan, A. Zisserman, Very Deep Convolutional Networks for Large-Scale Image
Recognition (2014). arXiv 1409.1556
5. Sepp Hochreiter, Jürgen Schmidhuber, Long short-term memory. Neural Comput. 9, 1735–1780
(1997). https://doi.org/10.1162/neco.1997.9.8.1735
6. A. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H.

Adam, MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
(2017). arXiv, abs/1704.04861
Influence of Big Data Capabilities
in Knowledge Management—MSMEs
Ravi Shankar Jha and Priti Ranjan Sahoo
Abstract In recent time innovation of open source technologies, commodity hard-

ware, cloud computing, data-driven application, visualization techniques have paved
their way for revolutionary changes in business models and data technology space.
This has resulted in creating the enormous opportunities for human creativity to
produce novel insights. In the “Era of Information” where huge amount of complexity
and varied of data being generated every day from multiple sources or point of sales
making business requirements/decisions more composite and time intensive. In infor-
mation dominated Era, Big data (emerging frontier) plays an instrumental role in
offering, insight and prediction is at the heart of almost every scientific/business
discipline. The discovery of insights and value from data is the central focus of the
industry to create new products and services to meet the changing demands. In digital
driven economy, Big Data capabilities and offering have been relentlessly empow-
ering organizations knowledge management platforms to ease knowledge manage-
ment practices including acquisition, creation, sharing storage and transfer to meet
organizations near term and future goals and accelerate toward success path. More-
over, Big Data capabilities helped MSME firms to leap forward in customer-centric
view analyzing transactional data across myriad sources and forming new insight and
model not only retain existing customers but also attract new customers or opportuni-
ties by offering tailor-made products and services. Most of those data centric insights
and models translated into organizational knowledge and being captured, groomed
and refined using knowledge management platforms and practices. According to
Wikibon (Technology Company), Big Data market revenues are projected to increase
from $42B in 2018 to $103B in 2027 with a CAGR of approx. 10.48%. Therefore,
among all the trends, buzz words, noises and quest for answers, it would be inter-
esting to see how Big Data is going to shape and influence data-driven ecosystem,
business models and process and practices of existing and emerging enterprises for
better tomorrow.
R. S. Jha (B) · P. R. Sahoo

KIIT School of Management, KIIT University (Institution of Eminence), Bhubaneswar, India
e-mail: ravis.jha@gmail.com
P. R. Sahoo
e-mail: prsahoo@ksom.ac.in
514 R. S. Jha and P. R. Sahoo
Keywords Big data · Knowledge management · Disruptive technologies · Micro ·

Small and medium size enterprise · Innovation
1 Introduction
Knowledge plays an integral role not only in the survival and well-being of an orga-
nization, but also helping them to achieve sustainable, scale, continuous innovation
and competitive advantage in connecting global ecosystem [3]. Therefore, it is essen-
tial to have well defined and established know-how, processes and frameworks to
nurture, grow and manage organization knowledge as a precious asset. Successful and
enterprising organizations have been significantly depending on organization knowl-
edge as a critical success factor to attain sustainable competitive advantage [6, 13].
Knowledge management is gaining popularity among organizations over recent years
because effective and efficient implementation of knowledge management practices
exhibiting bouquet of benefits to increase social equity of organizations by improving
overall performance and quality of products and services. Appropriate implementa-
tion of knowledge management practices, tools, process and framework help organi-
zation to achieve a desired conducive learning ecosystem, improve resource produc-
tivity, increase turnaround, better asset utilization, improve job satisfaction, increase
turnover and many more. Therefore, knowledge management machinery has been
promoted as strategic asset thereby seeking much needed attention like any other
asset in an organization for survival and success in short and long run.
2 Overview of Big Data and Its Value Chain
Over the years, the amount of data has been growing at an exponential rate. These data
are created by billions of people across the world through their digital platforms, busi-
ness connecting society, digital cameras and millions of sensors, connected objects
and communications devices sending and receiving data over the Internet. This has
resulted in the creation of large sets of data that is nearly impossible to handle and
process using traditional business intelligence tools. This brings us to the concept of
Big Data.
Big Data can be defined as huge volume of data (both structured and unstructured)
which requires advance database management system to capture, store, manage and
analyze [16, 17]. An International Data Corporation (IDC) report [8] forecasts that
“worldwide revenue for big data and business analytics (BDA) solutions to reach
$260 billion in 2022,” with a compound annual growth rate (CAGR) of 11.9%.
The core essentials of Big Data can be described by “three V’s”: Volume, Variety
and Velocity. Volume renders the large volume of data which can range from Exabyte
to Zettabyte. Variety refers to the data that has a specific format such as spreadsheets,
XML and DBMS which can be directly handled with computing equipment and data
Influence of Big Data Capabilities in Knowledge … 515
Fig. 1 Big data value chain
that does not have a specific format, such as images, videos and audio files. Velocity
outlines the speed at which data are created, transmitted, processed and analyzed so
that useful insights are delivered to the users. With time, other V’s have been added
as the characteristics which are Veracity (quality or reliability of the data), Value
(assessment/verification of their data).
In course of extracting values, data traverse through several phases right from
creation to destruction. The Big Data Value Chain (Big data monetization throughout
Big Data Value Chain [9] as illustrated in Fig. 1, can be used to construct the high-level
activities that comprise an information system. The Big Data Value Chain identifies
the following key high-level activities:
2.1 Data Generation
As the first step, data generation talks about the data (both structured and unstruc-
tured) that is generated from various sources including social media, transactional
data, sensor data, public and enterprise data with varying quality and quantity.
2.2 Data Acquisition
The data generated by the data source is forwarded to the data storage by means of a
data acquisition framework that implements a predefined protocol. The infrastructure
required to support the acquisition of big data must be able to deal with very high
transaction volumes in shared or scattered environment.
2.3 Data Pre-processing
This step appears to be extreme importance for upcoming analysis and exposition
phases as data gets collected continuously in real time which can be completely
strident and redundant. In a nutshell, it is treated as data mining technique. It consists
of methods like Cleaning (involves handling of missing data, noisy data), Reduction
(using Data Cube Aggregation, Numerosity Reduction etc.), Transformation (using
Normalization, Attribute Selection etc.) and Integration (building a single view of
data distributed over different sites) [4].
2.4 Data Storage
The data will come to storage system which needs to be highly scalable in nature that
satisfies the need of applications (such as Google, Facebook) which requires quicker
access to the data. One such good example of Data Storage is Hadoop Distributed
File System (HDFS).
2.5 Data Analysis
With data analysis, we aim to create sensible insights by finding correlations and
patterns across the data processed. There are four different ways in which this phase
can be conducted: Descriptive Analysis (targets to offer insights what has happened),
Predictive Analysis (aim to forecast on what might happen), Prescriptive Analysis
(seeks to determine the best decisions or outcome among various selections) and
Diagnostic Analysis (looking at the causes which leads to the achievement of the act
by looking at past).
2.6 Data Visualization and Exposition
With meaning dashboards, this phase helps to have visualize representation (such as
maps, graph, charts, 2D or 3D models) of the analyzed data. In data visualization,
data is abstracted and summarized. Spatial variables such as position, size, and shape
represent key elements in the data. The system should perform a data reduction, trans-
form and project the original dataset on a dashboard. It varies from histograms and
pie charts to more complex representations like heat maps and fever charts, enabling
individuals to explore data sets to identify correlations or unexpected patterns. It
identify scopes that need attention or improvement, elucidate which factors influ-
ence customer behavior and help us to understand which products/service to place
where. The ability to make quicker decision increases at this stage. Tools like Plotly,
Infogram, Oracle Visual Analyzer, QlikSense etc. are present for performing data
visualization and exposition activity.
Big Data Ecosystem can be formed around an organization, unrestricted tech-
nology platforms, within or across sectors which are created and supported by
numerous actors, their continuous data generation along with their interactions.
One such example of the ecosystems existence is within ecommerce sectors where
vast amount of data travel across various entities ranging from simple to complex
information.
3 Constructs of Knowledge Management and Its Practices
Knowledge is lifeline for any organization to grow and survive in the era of rapid
changing technology, disruptive business landscape, customers’ preference, evolu-
tionary technology, external factors (globalization) and many more. Also, in current
state of economy is equally known as “knowledge economy” in which knowledge
prevails over labor or natural resources which mull as “Strategic asset.” In brief terms,
Knowledge is culmination of data, information, experience that enable individual or
organization to understand and make informed decision at different stages of a busi-
ness event [11]. Therefore, it is evident that organization needs to lay strong foun-
dation to have robust Knowledge Management System and Practices to retain tacit,
external and explicit knowledge in the organization learning ecosystem to leverage
benefits. Additionally, attitude of Individuals and groups, organization culture, values
and norms, conducive knowledge sharing ecosystem, reward system and leadership
support are essential precursors for a sound knowledge sharing ecosystem [1].
Over the years, the objective of knowledge management has been the ability to
merge information from multiple perspectives to provide the insights required for
effective decision making. Knowledge Management practice briefly touch upon four
pillars discovering and capturing new knowledge, Knowledge acquisition, Knowl-
edge storing & sharing and Knowledge application. Inadvertently, Micro, Small- and
medium-sized enterprises (MSMEs), rapidly adopting both KM practices and evolu-
tionary technology enabled Knowledge Management systems preferably summa-
tion of Artificial Intelligence plus Machine Leaning or Big Data or IoT to inno-
vate products and services to accelerate innovation, enhance business performance,
obtain sustainable advantage and remove business impediments [10]. Knowledge
is broadly classified into two categories, tacit knowledge (comes from knowledge
surveys, questionnaires, individual & group interviews, focus groups, network anal-
ysis and observation) and explicit knowledge (includes document management, intel-
ligence gathering, data mining, text mining etc.). Organizations are involving all
possible stakeholders (internal and external) in their continuous endeavor to create
knowledge through formal and informal learning channels and leverage SECI Model
[12]. Knowledge management further expands to series of events including acquisi-
tion, creation, sharing, storing and transfer. The knowledge management with these
Fig. 2 Core pillars of

organizations knowledge
events comes under the umbrella of knowledge management practice. Knowledge

management practices are broadly involved with three key pillars as illustrated in
Fig. 2.
“People” primarily touch upon mid and senior level leadership support to incul-
cate the culture of conceiving knowledge management as essential part of today’s
operation. “Process” touch upon the underlying mechanism to codify or translate
tacit and explicit knowledge and share across various stages of knowledge manage-
ment practices. Lastly, “Know-how” technology plays a role of an integral enabler
which accelerates the knowledge management activities by offering range of evolu-
tional technology tools not only to perform Knowledge Management activities but
also enhance Business User Experience.
4 Role of Big Data in Knowledge Management
Big data capabilities and offerings are extensively being implemented in most of
the industries and services to enhance the competitive capabilities, help knowledge
management practices and improve efficiency and effectiveness. Moreover, extrac-
tion of enormous data, perform in-depth analysis and derive valuable insights (infor-
mation plus knowledge) by using various statistical techniques & machine learning
and ends at value creation. Figure 3 depicts the relationship among various data
stages, Big Data technologies layer and Knowledge management practices.
Fig. 3 Big data in knowledge management
5 Overview of MSMEs Firms
The Micro, Small and Medium Enterprises (MSME) sector has been emerged as
a promising and demanding sector of the Indian economy over few decades. It
offers holistic growth in the social-economic development of the country by fostering
entrepreneurship and spawning employment opportunities across the sectors. Appar-
ently, MSMEs are fundamental pillars to large industries as augmented units spending
time and effort in exploring new and better ways of doing manufacturing and services
and contributes significantly in the inclusive industrial development of the country.
The micro, small and medium businesses (SMEs) in India include 63.4 million units
and contribute around 6.11% of the manufacturing GDP and 24.63% of the GDP
from service activities as well as 33.4% of India’s manufacturing output [2].
In accordance with the provision of Micro, Small & Medium Enterprises Devel-
opment (MSMED) Act, 2006 the Micro, Small and Medium Enterprises (MSME)
are revised [5] and classified as given in the Table 1.
Many industries have already made the large investments in Big Data and are
reaping the benefits of it. For instance, within banking and ecommerce investments
Table 1 Revised MSME classification in India

Composite criteria: investment and annual turnover
Classification Micro Small Medium
Manufacturing and Investment < Rs. 1cr. Investment < Rs. 10cr. Investment < Rs. 20cr.
Services and Turnover < Rs. and Turnover < Rs. and Turnover < Rs.
5cr. 50cr. 100cr.
are all about managing and give a boost to the customer experience and thus resulting
increase in revenue and market share. With this, micro, small and medium enterprises
(MSMEs) have started exploring various ways to harness raw data. Having said
that, abundant of data, SMEs are ready to derive value and make wise, data-driven
decision to grow their businesses [8]. One way is to develop partnerships with big
data technologies which is offered by IBM and Oracle. Many analytical tools (like
IBM’s Watson Analytics) are available to all organizations (small or big) on a series
of models, thus giving SMEs the freedom to experiment without entailing significant
cost and develop applications across the knowledge management practices offering
enriching experience to customers and partners.
6 Drivers for Adopting Big Data in MSMEs
Big data is one of the cornerstones empowering MSME to uniquely tackle the
increasing problem of data, growing competitiveness and knowledge deprivation.
Following are few key drivers that compels firms including MSME to embrace Big
Data capabilities as one of their core business strategies [7].
• Ubiquitous Competitiveness: With growing competition resulting due to glob-
alization, off-sourcing, global economy and many more, it is imperative for firms
to harness raw data using Big Data capabilities and offerings to create difference
and deliver unique value proposition to clients and stakeholders.
• Decision-Making Time: In the era of information, firms generate huge amount
of complex and variety of data include structured and unstructured from multiple
touchpoints including social media, transaction data, sensor feeds, Enterprise data
and public data results in elongated decision-making process incurs additional
time.
• Surging need of Tailor-made Products/Services: Customer’s demands are
changing at rapid pace because to globalization, proliferation of digital awareness,
change in buying behavior and other reasons compel firms to build customized
products and service. In order to do so, firms need insight on customer’s demo-
graphics, buying patterns, recent trends and many more other influencers to fulfill
the demands of Millennial and generation Z.
• Digital Narrative: In the era of hyper-competitive world where organizations
are moving from brick and mortar to digital age and leveraging data to derive
information and build knowledge. Therefore, it is essential for organization to have
effective and efficient digitally empowered systems for sharing and replication of
data across all point of sales or touchpoints to improve customers experience
• Changing Business Models: With surpassing rise of disruptive Insurtech start-
ups and emerging technologies like Big Data, artificial intelligence, blockchain
and drones entering the mix, the industry is being compelled to rethink its busi-
ness model. Acknowledging the urgency to adapt to the new environment and
reinvent themselves to become future-ready, companies need to prepare them-

selves, focusing not only on technology, regulations and business models but also
speed, efficiency and risk.
7 Benefits and Challenges of Big Data
Knowledge based on the work of Nonaka (1994) and Huber (1991) defined as
“Knowledge is a justified personal belief that increases an individual’s capacity to
take effective action” [7]. With the rapid growth and proliferation of data due to advent
of digital revolution, enterprise started adopting Big Data capabilities across business
transactions to collect all the required information which can be further processed,
consolidated, and explored for mining the hidden insights (information/knowledge)
in order to boost business performance and competitiveness. Additionally, provides
most valuable and garner real time business insights through extensive data analytics
which helps business to make timely and accurate decisions [14]. The key benefits
and challenges of Big Data are depicted in Table 2.
Table 2 Key benefits and challenges of big data

Key benefits Challenges
Enable firms to launch new products and Real time and batch processing of raw data to
services (cross/upsell) based on insight analysis and other stages of value chain
generated by data captured across the POS or
touchpoints
Create new avenues for innovation and Acquisition of data from diverse sources that
disruptions comes in various shapes and sizes and store in
existing data set
Help to attain sustainable competitive Establish sustainable infrastructure to support
advantage robust storage and processing platforms
Increase operational efficiency and offer Additional investment on train existing
agility in customer management workforce to use Big data-driven systems
Improve decision-making time and reduce Define Data governance, regulation and
expense protection framework to safeguard customer’s
confidentiality
Improve planning and forecasting, monitor Support from leadership, especially CDO
quality and provide better pricing (Chief Data officer) to secure funding and able
to navigate other organizational barriers
Transform business models for future needs Establish data-driven culture across various
functions of organization
Create strong bonds with the end customers Shortage of external success stories/business
based on the derived statistics and enhance cases for SMEs to model their business
customer value and equity implementation
8 Conclusion
Growth of MSME is of utmost importance for a developing and under-developed

economy. This growth is sustainable given knowledge, data and information flows
freely seamlessly across the organizations. Therefore, knowledge management plays
a pivotal role in creating, acquiring, storing, transferring, sharing and disseminating
know-how across the value chain and touch every stakeholder. To make this process
effective and efficient, technology is the only real catalyst or platform. In coming
times, it would be interesting to see how induced cognitive ability of machines help
in our research, how advanced NLP (natural language processing) help in training
the workforce anytime and anywhere, how Blockchain ensures to build the trust in
the system making transactions tamperproof, how Big Data empowers knowledge
management practices, how AR (augmented Reality) reduces the bottleneck in the
assembly line by enabling new workforce to deliver etc. There will be more addition
to the list as technology paving a part to our KM practice. Moreover, Big Data, IOT
and other disruptive technologies helping firms to create smarter environment and
relying more on data-driven ecosystem to improvise on decision making and delight
stakeholders at every level of the business value chain [10].
Organizations settings are getting positively influenced by implementing knowl-
edge management practice such as quality of products or services, innovation ability,
improve way of working, organization performance etc. There is no unified approach
or framework to solve the barriers of knowledge management practices as it varies
from context to context basis in globally connected and competitive world. The busi-
ness events or transactions are shifting from transaction centric view to customer-
centric view to enhance user experience across each touch points with the help
to data and information available at disposal with the help of Big Data, IOT, AI
and other emerging technologies. Classical examples are Amazon and Zappos lean
toward customer-centric view and spending lot of time in creating a culture around
the customer and their needs. When a customer buys a laptop, an automatic sugges-
tion comes up on the web/mobile, suggesting for products like mouse, cooling pad,
mouse pad etc. In a similar way, At Warby Parker, with integrated data and a unique
point of sale system, customers work one-on-one with associates to find the perfect
pair of glasses with flexibility of paying via tablet rather than repeating their pref-
erences, Kaiser Permanente uses innovative technology like virtual appointments
that allow providers to see more patients more conveniently and many more such
examples in which data oriented technologies act as catalyst to delight customer’s
experience and increase mindshare. However, on the hindsight these technologies
including Big data, IOT, Machine learning and cloud computing bring plethora of
challenges [15] such as data and privacy, scalability and visualization of data, data
storage and resource etc. In due course, organizations including MSME and large
organizations are relentlessly pondering upon creatively improving their products
and services, embracing innovation, creating learning culture to tap organization’s
knowledge (raw data exploits Big data capabilities resulting into insight or knowl-
edge) to delight customer and occupy their mindshare. Therefore, it is imperative for
organization to pursue knowledge management practice at holistic level and carve

out dedicated strategies to leap from resource-based organization to knowledge-
oriented organization supported by gamut of disruptive technologies including Big
Data technology stack for a vibrant and enterprising tomorrow.
References
1. G.W. Bock, Y.G. Kim, Breaking the myths of rewards: an exploratory study of attitudes about
knowledge sharing. Inf. Resour. Manage. J. (IRMJ) 15(2), 14–21 (2002). https://doi.org/10.
4018/irmj.2002040102
2. CII, The Indian MSME sector (2020). https://www.cii.in/Sectors.aspx?enc=prvePUj2bdMt
gTmvPwvisYH+5EnGjyGXO9hLECvTuNuXK6QP3tp4gPGuPr/xpT2f
3. P.F. Drucker, Knowledge-worker productivity: the biggest challenge. Calif. Manag. Rev. 41(2),
79–94 (1999)
4. A.Z. Faroukhi, I. El Alaoui, Y. Gahi, A. Amine, Big data monetization throughout Big data
value chain: a comprehensive review. J. Big Data 7(1) (2020). https://doi.org/10.1186/s40537-
019-0281-5
5. Government of India, Aatmanirbhar Bharath strategic reforms and growth initiatives Part-
1 : businesses including MSMEs. Ministry of Micro, Small & Medium Enterprises. https://
msme.gov.in/sites/default/files/Atmanirbhar Presentation Part1Business including MSMEs13-
5-2020_0.pdf
6. R.M. Grant, Towards a knowledge-based theory of MNE. Strateg. Manag. J. 17, 109–122
(1996)
7. S.H. Hsu, H.P. Shen, Knowledge management and its relationship with TQM. Total Qual.
Manage. Bus. Excellence 16(3), 351–361 (2005). https://doi.org/10.1080/14783360500054111
8. International Data Corp., IDC Predicts Strong Growth for Big Data and Business Analytics
(2018). Www.Destinationcrm.Com. https://www.destinationcrm.com/Articles/CRM-News/
CRM-Across-the-Wire/IDC-Predicts-Strong-Growth-for-Big-Data-and-Business-Analytics-
126895.aspx
9. M. Iqbal, S.H.A. Kazmi, A. Manzoor, A.R. Soomrani, S.H. Butt, K.A. Shaikh, A study of big
data for business growth in SMEs: opportunities & challenges, in 2018 International Confer-
ence on Computing, Mathematics and Engineering Technologies: Invent, Innovate and Inte-
grate for Socioeconomic Development, ICoMET 2018—Proceedings, 2018-January (2018),
pp. 1–7. https://doi.org/10.1109/ICOMET.2018.8346368
10. R.S. Jha, P.R. Sahoo, Internet of things (IOT)—enabler for connecting world, in ICT for
Competitive Strategies (CRC Press, 2020), (pp. 1–6)
11. S. Massa, S. Testa, A knowledge management approach to organizational competitive advan-
tage: evidence from the food sector. Eur. Manag. J. 27(2), 129–141 (2009). https://doi.org/10.
1016/j.emj.2008.06.005
12. I. Nonaka, The knowledge-creating company. Harvard Bus. Rev. 69(6), 96–104 (1991)
13. I. Nonaka, P. Byosiere, C.C. Borucki, N. Konno, Organizational knowledge creation theory:
a first comprehensive test. Int. Bus. Rev. 3(4), 337–351 (1994). https://doi.org/10.1016/0969-
5931(94)90027-2
14. C. O’Connor, S. Kelly, Facilitating knowledge management through filtered big data: SME
competitiveness in an agri-food sector. J. Knowl. Manage. 21(1), 156–179 (2017). https://doi.
org/10.1108/JKM-08-2016-0357
15. A.M.H. Obaid, P.K. Pattnaik, S.K. Pani, Towards on cloud of things: survey, challenges, open
research issues, and tools, in Proceedings of the International Conference on Inventive Commu-
nication and Computational Technologies, ICICCT 2018, Icicct, (2018), pp. 1458–1463. https://
doi.org/10.1109/ICICCT.2018.8473212
16. Z. Polkowski, M. Nycz, Big data applications in SMES. Sci. Bull. Econ. Sci. 15(3), 13–24
(2016)
17. C. Tan, M. Haji, Big data educational portal for small and medium sized enterprises (SMEs), in
ACM International Conference Proceeding Series (2017), pp. 11–15. https://doi.org/10.1145/
3175684.3175688
Logistically Supervised Aspect Category
Detection Using Data Co-occurrence
for Sentiment Analysis
Anice John and Reshma Sheik
Abstract Word of mouth (WoM) always had an impact on consumer decision

making. Before making any purchase decision, family and friends are asked for
advice and recommendations. With the advent of Web, wlectronic WoM (EWoM)
came into being. One of the important forms of EWoM is product and service reviews.
These reviews contain large amount of information, and so, reading these reviews one
at a time is a tedious and time-consuming task. Therefore, summarization of reviews
is desirable. This can be done by identifying the aspects in the reviews. Aspects are
the specific topic that is being referred. The proposed approach is a hybrid of three
methods for aspect category detection. The first method is based on co-occurrence
between the indicators and categories, and the second method is based on logistic
regression, used to predict the probability of a dependent variable based on one or
more features. The third one based on Wu-Palmer similarity measure. Experimental
results show a good performance of the proposed hybrid approach.
Keywords Aspects · Aspect category · Summarization · Co-occurrence · Logistic

regression · Wu-Palmer · Similarity measure
1 Introduction
With the advent of Web and its technologies, people can now express their opinion
through them. Decisions regarding the purchase of product/services or about a polit-
ical decision or trends in market can be analyzed through the opinions expressed by
the people via social media networking sites like Facebook, Twitter, etc. Sometimes
it is not enough to identify the only sentiments but need to explore more about it.
For example, consider the sale of a particular product. Here, the manufacturer not
A. John (B) · R. Sheik

Department of Computer Science and Engineering, TKM College of Engineering, Kollam, India
e-mail: anicejohn1@gmail.com
R. Sheik
e-mail: reshmasheik@tkmce.ac.in
526 A. John and R. Sheik
only needs to understand the sentiments of the customer but also needs to know why
their product sales has reduced or the decrease in popularity of the product. Here, the
aspect-level sentiment analysis plays its role as this can be understood by identifying
the aspects of the reviews.
Identifying the aspects can be helpful in various areas that is, in business for
marketing of products and services, for customers regarding their purchase decisions
and in political field for analyzing the reaction of people regarding the new political
policies etc. For better understanding of this topic, we need to get a clear idea of
aspects and its categories. Aspect-based sentiment analysis (ABSA) identifies the
aspects in addition to its sentiments. The two main steps involved in ABSA are
aspect extraction and finding the polarity of the aspects. The primary focus of this
work is to detect the aspects in the restaurant reviews for sentiment analysis.
Aspects are the specific topics being referred to and the category to which they
belong is called the aspect category. The concept of aspect and aspect category can
be explained by the following example:
“The pizza was too good!!!”
In the above sentence, the aspect is “pizza,” and the category to which it belongs is
“food.” All the aspects will not be given explicitly in the sentences. Some aspects will
be implied within the sentence that is these aspects will not be explicitly mentioned
in the sentence.
A supervised method consisting of three approaches is proposed. One of the
methods uses co-occurrence between lemma/dependencies and aspect categories [1],
and the second approach employs logistic regression which successfully classifies
the default category “anecdotes/miscellaneous.” The third method uses similarity
measure to detect the aspects of reviews with unknown words. This work is organized
in the following manner—Sect. 2 includes the related works, which gives a broader
view of the various methods in use for aspect detection. Section 3 incorporates the
methodology with which the proposed system is based on. Section 4 includes results
and evaluation, and conclusion of this work is given in Sect. 5.
2 Related Work
Aspects are the specific topic that is being referred in a sentence or a document. Many
methods have been proposed for aspect detection. Early works on aspect detection
focused on finding the explicit aspects of a sentence. Explicit aspects are those aspects
which are explicitly given in the sentences. A frequency-based method is proposed
in [2]. Since the opinion features/aspects are mostly nouns and noun phrases, they
are extracted using a POS Tagger. The frequent feature, that is the noun and noun
phrases, is then found by association rule mining. After this, the opinion words are
extracted. For this, the nearest adjective to the frequent feature is considered as the
opinion words for each of the review sentences. This method is good to find the
explicit aspects of the sentences. However, it fails to find the implicit aspects.
Logistically Supervised Aspect Category Detection … 527
Moghaddam and Ester [3] refined the frequency-based approach in their work by
introducing a filter based on patterns to identify and remove the non-aspect terms.
Similar to [2], the noun and noun phrases are extracted using POS tagger. The frequent
nouns are selected as aspects, but all the frequent nouns and noun phrases might not
be aspects. Next for each of the known aspects, the matching phrases are found. Their
POS is saved as pattern. Generalized sequential pattern (GSP) [4] is then applied to
find the frequent pattern. A filtering factor is proposed which filters out the non-
aspect patterns. Even though the method filters out the non-aspect terms, still it fails
to find the implicit aspects.
In [5], a multi-knowledge based approach is proposed. A keyword list which
contains an initial opinion seed word is used. Here, feature-opinion pair relations are
mined using a dependency grammar graph which is then used to find the relevant
opinion-feature pair in the test set. This work finds the explicit aspects, but only two
simple cases of implicit aspects are detected. Implicit aspects were found for those
sentences having specific mapping between feature words and opinion words and
for short sentences.
Implicit aspects were identified by a clustering method in [6]. Here, the implicit
aspects which were assumed to be the sentiment words were mapped with the explicit
aspects. The mapping is done by clustering the set of explicit aspects and sentiment
words iteratively, and in each iteration, the results of the previous set are used to
find the pair wise similarity of the set. After clustering, the strongest n links between
aspects and sentiment words forms the mapping. In this method, only adjectives were
considered as opinion words and fail to find the non-adjective opinions. Another
method proposed in [7] uses phrase dependency parsing to find the aspects. This
work proposes a phrase dependency parsing to find the features. The advantage of
this method is that it finds the multi-word aspects of the review sentences.
A two-stage co-occurrence-based association rule mining is employed in [8]. In
the first stage, all the opinion words and explicit features are identified. Then, a co-
occurrence matrix is created which stores the frequency of the opinion word and the
feature. Based on the co-occurrence matrix, association rules are mined. The second
stage is the application of the explicit rules to a given set of implicit sentences.
For this, the explicit aspects are clustered. When a new opinion word occurs in a
sentence, it is checked with the set of explicit rules. Among these matched rules, the
cluster consisting of the largest number of features is selected, and the corresponding
representative word of the cluster is considered as the implicitly mentioned aspect.
However, this method assumes that implicit aspects always relate to an explicit aspect.
A dependency-based approach is used in [9]. Here, a dependency parser is used
to find the syntactic relation which links an opinion word and targets or features.
This method only requires an initial opinion lexicon to start the double propagation
process. The basic idea of the double propagation is to extract opinion word and
targets using the seed opinion lexicon, and then, these extracted opinion words and
targets are used for further extraction. This is performed until no new opinion words
or targets are identified. However, this method only considers relationship between
opinion words and targets to find the features.
Association rule mining is used in [10]. Here, annotated explicit aspects are
not required, instead it uses double propagation algorithm of [9]. This method not
only focuses on the association between feature words and opinion words but also
considers the association between the feature words and the other notional words.
Garcia-Pablos et al. [11] proposed an approach which also uses double propagation
technique described in [9]. A co-occurrence-based approach is used in [12]. Here,
the co-occurrence-based approach records the frequency of the co-occurrence of the
lemmas and annotated categories. For each of the unprocessed sentences, a score is
calculated and whenever the score of a particular category is greater than its associ-
ated threshold, then that category is assigned as the aspect category for that particular
sentence. However, this method did not perform well as expected.
3 Methodology
The proposed method is a combination of three approaches namely: co-occurrence

approach [1], logistic regression (LR) approach, and Wu-Palmer approach. The
working principle behind these approaches is explained in the following sections.
3.1 Co-occurrence Approach
The co-occurrence approach utilizes co-occurrence frequency between

lemmas/dependencies and categories of a sentence. Each dependency is composed
of three components: governor word, dependent word, and relation type. These
grammatical dependencies are more specific than lemmas because they give the
relationship between the words of a sentence. For example, by knowing when a
word will be used as a subject or a modifier, it can affect aspect detection to a great
extent. All but the determinant relation from the Stanford parser [13] is used. The
idea of co-occurrence approach is inspired from the work of [1]. The basic steps of
this approach are as follows:
Find lemmas/dependencies: After pre-processing the dataset using POS tagger,
lemmatizer, and dependency parser of the Stanford Core NLP, the lemmas and the
dependency forms are obtained.
Calculate the weight matrix: The occurrence frequencies of lemma/dependencies
are found, and the insignificant ones are removed. Their co-occurrence frequency
with the categories are found and stored as co-occurrence matrix. For each entry in
this matrix, conditional probability is calculated and stored as weight in the weight
matrix.
co-occurrence frequency of lemma/dependency
weight =
occurrence frequency of lemma/dependency
Threshold estimation: An optimal threshold is selected using linear search, one for
lemma and three for the three types of dependency for each category.
Category detection: For each sentence in the test set, the maximum conditional
probability for lemmas and dependencies are calculated. If this maximum condi-
tional probability is greater that the threshold value of a category, then that particular
category is detected as the aspect category of that sentence.
Since it is not clear what actually belong to the miscellaneous category, it is
given when no other category is detected by the approach even when it might have
a separate category. So in order to rectify this, logistic regression is used.
3.2 Logistic Regression Approach
Logistic regression (LR) [14] is one of the most commonly used machine learning
algorithms inspired from the field of statistics. It is employed for classification prob-
lems and is based on probability concepts. It predicts the value of a numeric categor-
ical variable depending on its relationship with other predictor variables. LR uses a
sigmoid function also called as the logistic function to map the predicted values to
its corresponding probabilities.
Here, in this work, LR is used to predict the aspect categories of reviews. The
steps are as follows:
Create the dataset: Each of the reviews in the training set is considered as a sample.
The input to the LR is a set of features and its weight. Each sample will consist
of several features or independent variables. The features considered here are the
lemmas for each of the review sentences. The dependent variables will be the anno-
tated category set. The basic idea is that LR predicts the class labels of each of the
samples using the features.
Building the model: This step will create a model for LR to predict the aspect
categories. The features are the lemmas, and feature values are their occurrence
frequencies. Let us now consider the conditional probability that a feature belongs
to a particular category, and LR uses the logit function to obtain this. Weights are
assigned to the features of each of the sample review sentences. Next, these weights
should be adjusted in order to effectively predict the aspect categories. This method
of adjusting the weights is called learning.
Training the model: Here, in this step, the weights are adjusted such that maximum
probable outcomes are obtained. Logistic regression is performed using Sci-Kit learn.
The best fit is found using stochastic gradient descent (SGD) classifier. Fitting the
model basically implies learning the relationship between features and class labels.
At this stage, the model has finished learning and ready to predict the categories.
Make predictions: The final step is to ask the fitted model obtained in the previous
step to predict the aspect categories of the unknown sentences of the test set. Finally,
these predictions can be compared with the actual category of the sentences to
determine the evaluation scores.
3.3 Wu-Palmer Approach
Previously mentioned approach detects aspects using the words of the sentences from
the training set. As long as the words of test sentence are found in the training set, the
aspects can be detected. Now what happens if sentences with unknown words appear
in the test set, the above problem can be addressed by using Wu-Palmer similarly
measure.
Semantic similarity measures to what extent a word X is related to another word
Y. Detecting the semantic similarity is broadly used in applications related to natural
language processing. There are many similarity measures used for this purpose, and
one among them is called the Wu-Palmer similarity measure. Wu-Palmer similarity
is a scoring method which measures the relatedness of two synsets in the WordNet
[4, 15]. It calculates that how alike two words are with respect to one another. To
understand more about Wu-Palmer similarity measure, there is a need to know some
basic concepts of WordNet which is a lexical database for English language consisting
of group of synonyms and definitions. The set of synonyms of words in a WordNet
is called “synsets.”
To find the aspects of sentences with unknown lemmas, Wu-Palmer similarity is
employed. After finding the lemmas if there are one or more unknown lemmas in
the sentence, then their corresponding synset is found. Here, only nouns and verbs
of synset are considered since most of the aspects are nouns and adjectives. After
finding the synsets of the unknown lemmas, they are compared with the known
lemmas obtained in the previous approach. The synset of the unknown lemmas is
compared with both the known lemmas and their corresponding synset. The similarity
values will be in the range 0–1. The maximum similarity measured word is taken,
and its corresponding category is predicted as the aspect category for that particular
sentences.
3.4 Hybrid Approach
The proposed system is a hybrid approach comprising of the three methods, that
is, co-occurrence, LR, and Wu-Palmer approaches. The co-occurrence approach is
considered as the basic approach, but the anecdotes/miscellaneous category was not
handled efficiently. So, whenever the co-occurrence approach predicts the category
as anecdotes/miscellaneous, the LR approach is checked for its prediction, and its
predicted category is assigned as the aspect category for that sentence. Next, the
aspect category for the sentences with unknown lemmas is detected using the Wu-
Palmer approach. Figure 1 gives a basic idea of the proposed method.
Fig. 1 Proposed hybrid approach
4 Results
Evaluation is mainly done by analyzing the performance of the hybrid approach of

aspect detection. That is determining how well the fitted model performs in predicting
its correct category. The performance of the approach with only the lemmas and with
both the lemmas and dependencies can clearly be understood from the following
Tables 1 and 2. It is seen that using both the lemmas and grammatical dependencies
gave a much better result than using only the lemmas for aspect category detection.
The proposed method gave a fair performance in detecting aspects with an F1
score of 83%. The overall performance of the hybrid approach is shown in Fig. 2. It
can be clearly seen that the hybrid approach efficiently classifies the miscellaneous
Table 1 Evaluation scores of

Category Precision Recall F1
the method with only lemmas
indicators Service 44.00 100.00 61.11
Food 74.00 97.00 83.95
Miscellaneous 84.00 93.00 88.27
Price 0.00 0.00 0.00
Ambience 54.00 86.00 66.34
Table 2 Evaluation scores of

Category Precision Recall F1
the method with lemmas and
dependencies Service 61.00 100.00 75.78
Food 86.00 97.00 91.17
Miscellaneous 100.00 94.00 96.91
Price 23.00 80.00 35.73
Ambience 69.00 92.00 78.86
Fig. 2 Comparison of Lemmas Lemmas+Dependencies Hybrid approach

evaluation scores using only
lemmas, both lemmas and 77 91 83
93
dependencies and hybrid 68 76
approach 51 75
60
Precision
Recall
F1
category. Also, aspects for sentences with words previously not mentioned were also
detected to a certain extent using Wu-Palmer approach. This method can be further
improved since it does not perform as good as expected with very large reviews, but
with short reviews, it performs effectively. However, the combination of the three
approaches does give a fair evaluation in aspect detection.
5 Conclusion
In this work, a hybrid method comprising of three approaches has been proposed
that is useful for online review summarization by detecting the aspect categories of
sentences. The first approach uses co-occurrence between indicators and categories,
and the second approach uses logistic regression. The third approach uses Wu-Palmer
similarity measure with the usage of WordNet to find the aspects of sentences with
words not previously defined in the training set. This work shows how the combi-
nation of the above three methods improves the performance of aspect category
detection for sentiment analysis. Experimental results show that the proposed hybrid
approach is effective in predicting aspects to a good extent.
References
1. K. Schouten, O. Van Der Weijde, F. Frasincar, R. Dekker, Supervised and unsupervised aspect
category detection for sentiment analysis with co-occurrence data. IEEE Trans. Cybern. 48(4),
1263–1275 (2018)
2. M. Hu, B. Liu, Mining opinion features in customer reviews, in AAAI, vol. 4, no. 4, pp. 755–760
(2004)
3. S. Moghaddam, M. Ester, Opinion digger: an unsupervised opinion miner from unstructured
product reviews, in Proceedings of the 19th ACM International Conference on Information
and Knowledge Management, ACM, 2010, pp. 1825–1828
4. G.A. Miller, WordNet: a lexical database for English. Commun. ACM 38(11), 39–41 (1995)
5. L. Zhuang, F. Jing, X.Y. Zhu, November. Movie review mining and summarization, in Proceed-
ings of the 15th ACM international conference on Information and knowledge management,
ACM, 2006, pp. 43–50
6. Q. Su, X. Xu, H. Guo, Z. Guo, X. Wu, X. Zhang, B. Swen, Z. Su, Hidden sentiment association
in chinese web opinion mining, in Proceedings of the 17th international conference on World
Wide Web, ACM, 2008, pp. 959–968
7. Y. Wu, Q. Zhang, X. Huang, L. Wu, Phrase dependency parsing for opinion mining, in Proceed-
ings of the 2009 Conference on Empirical Methods in Natural Language Processing: vol. 3
(Association for Computational Linguistics, 2009) pp. 1533–1541
8. Z. Hai, K. Chang, J.J. Kim, Implicit feature identification via co-occurrence association
rule mining, in International Conference on Intelligent Text Processing and Computational
Linguistics (Springer, Berlin, Heidelberg, 2011), pp. 393–404
9. G. Qiu, B. Liu, J. Bu, C. Chen, Opinion word expansion and target extraction through double
propagation. Comput. Linguist. 37(1), 9–27 (2011)
10. Y. Zhang, W. Zhu, Extracting implicit features in online customer reviews for opinion mining,
in Proceedings of the 22nd International Conference on World Wide Web, ACM, 2013, pp. 103–
104
11. A.G. Pablos, M. Cuadros, G. Rigau, V3: unsupervised generation of domain aspect terms for
aspect based sentiment analysis, in Proceedings of the 8th International Workshop on Semantic
Evaluation (SemEval, 2014), pp. 833–837
12. K. Schouten, F. Frasincar, F. De Jong, Commit-p 1wp3: a co-occurrence based approach to
aspect-level sentiment analysis, in Proceedings of the 8th International Workshop on Semantic
Evaluation (SemEval, 2014), pp. 203–207
13. M. de Marneffe, C.D. Manning, Stanford Typed Dependencies Manual, Sept 2008
14. D.G. Kleinbaum, K. Dietz, M. Gail, M. Klein, M. Klein, Logistic Regression (Springer, New
York, 2002)
15. C. Fellbaum, WordNet, in Theory and applications of ontology: computer applications
(Springer, Dordrecht, 2010), pp. 231–243
Automated Dietary Monitoring System:
A Novel Framework
Samiul Mamud, Punyasha Chatterjee, Saubhik Bandyopadhyay,
and Suchandra Bhandari
Abstract Now, the world is struggling through an unexpected adverse situation due
to the COVID-19 pandemic. Governments are spending significant resources so that
people can get rid of the disease and develop and maintain their health. Healthy
lifestyle and disease preclusion are a socio-economic issue now. According to WHO
[1], diet is an important factor to gain immunity and protect people from severe
ailments. Therefore, it is required to stay acquainted with the equation between
a balanced diet and a healthy life. Unfortunately, in the developing countries like
India, there is no proper dietary assessment system specially for the rural people.
The popular traditional manual approaches are biased, subjective and inaccurate.
Our aim is to develop a low-cost, power-aware novel framework for an automated
dietary monitoring system using wearable sensors that can monitor eating habits,
food intake, nutritional content of the food and calorie intake along with a feedback
to the user. The system is potentially applicable in Indian scenario and effective to
analyse versatile Indian foods.
Keywords Food intake detection · Body sensor network · Food database ·

Nutrition database · Food-type detection and food volume detection · Eating
behaviour
S. Mamud (B) · P. Chatterjee · S. Bhandari

School of Mobile Computing and Communication, Jadavpur University, Kolkata, India
e-mail: samiulmamud1995@gmail.com
P. Chatterjee
e-mail: punyasha.chatterjee@gmail.com
S. Bhandari
e-mail: suchandra0609@gmail.com
S. Bandyopadhyay
Computer Science & Engineering Department, Amity University, Kolkata, India
e-mail: saubhikit@gmail.com
536 S. Mamud et al.
1 Introduction
Due to the present pandemic situation of the world for COVID-19 outbreak, the
perspective of life has been suddenly changed in a different way: now, health care
is the primary concern in every country, even sometimes ignoring the skyrocketing
expense, incurring the medical service cost. Governments are eager to spend suffi-
cient resources to save people and offer them a healthy lifestyle. Also, the people are
now very much conscious about their health and protective livelihood. According
to World Health Organization (WHO), people who eat a well-balanced diet tend to
be healthier with stronger immune systems and lower risk of chronic illnesses and
infectious diseases [1].
Hence, diet is an important lifestyle feature related to plummeting interminable
ailments. A nutritional diet intake can enhance the physical and mental health. Over-
diet and excessive weight gain act as an indoctrinate factor for diseases such as
heart attack, stroke, osteoarthritis, cancers, breathing problems, crumbling, amnesia
and psychological effects [2]. A person not receiving essential nutritive elements
due to under-diet can also be exposed to serious health problems such as anorexia
nervosa, pancreatic cancer, rheumatoid arthritis, sarcopenic obesity and organ failure
[3]. Therefore, it is required to stay acquainted with the equation between a balanced
diet and a healthy life.
Unfortunately, in the developing countries like India, there is no proper dietary
assessment system, especially for the rural people, who are almost unaware of their
food habits. Despite India’s 50% increase in GDP since 1991, more than one-third
of the world’s malnourished children live in India [4]. It is necessary to build a
user-friendly automated dietary monitoring approach that does not interfere with the
person’s routine activity rather than helping him/her to easily adapt to the system.
The popular dietary monitoring approaches are manual, using conventional self-
report method, self-report method using smartphones [5], food frequency question-
naire (FFQ) [6] etc., which are proven to be inaccurate and biased. Wearable sensors
can be a potential solution for a smart, automated dietary assessment system, since
the system is not dependent on people’s judgments and can provide accurate real-
time data for food intake. Another essential advantage for wearable sensors is that it
provides convenience for users, specially for patient, child and elderly.
Keeping in mind the Indian scenario, our aim is to develop a low-cost, power-
aware novel framework for an automated dietary monitoring system using wearable
sensors that can monitor eating habits, food intake, nutritional content of the food
and calorie intake, along with a feedback to the user. Our approach can be applied
to the people who often lack the literacy skills to provide accurate reports.
The rest of the paper is organized as follows: Section 2 discusses the related
works, Sect. 3 describes the proposed automated dietary monitoring system, Sect. 4
illustrates the conceptual workflow of the proposed system, Sect. 5 highlights the
merits of the scheme, and Sect. 6 concludes the paper.
Automated Dietary Monitoring System: A Novel Framework 537
2 Related Work
In this section, we provide a review of the existing works on wearable dietary moni-
toring system. We mainly discuss the acoustic, visual, inertial and fusion approaches,
as these are related to our proposed strategy.
In the acoustic approach, chewing [7–10] and swallowing [11] detection is done
using wearable microphones by capturing sound waves and then data are analysed.
However, these are effective for specific kinds of foods only. Visual approach captures
food’s image or video with the smartphone camera [12, 13] as well as image sensors.
Sun et al. [14] developed a system called eButton, where the camera is embedded
in a minuscule button. Shang et al. [15] used camera and laser beam to capture 3D
images and estimated food volume and food type. Inertial approach targets the hand
movement and the motion of the eating process by gyroscopes and accelerometers
[16]. These are not very efficient because they can detect false hand movements.
The fusion approach merges two or more wearable approaches to increase accu-
racy, utilization, simplifying and quickening data processing. Liu et al. [17] proposed
a wearable device where the camera is triggered by a microphone as soon as chewing
activity is detected. The system requires much storage for video streaming and
is complex to analyse the video data. Sen et al. [18] proposed an inertial–visual
technique, which comprises smartphone and smart-watch. Here, the smart-watch’s
camera is triggered by the hand gesture sensor. If it is not triggered at right time, proper
image capture will not be possible. Fontana et al. [19] proposed an automatic inges-
tion monitor (AIM) method that comprises a piezoelectric, accelerometer and propin-
quity sensor to distinguish jaw movements, body motion and hand-to-mouth motions
during ingestion. AIM lacks to provide real-time feedback to individuals about their
eating behaviour, the type and caloric value of food could not be determined, could
not perceive liquid intake.
Finally, due to the complex structure and texture of the Indian food, none of
the existing approaches are found to be perfectly suitable for the Indian scenario.
Also, the systems are costly and power-hungry as most of the approaches are using
smartphones and smart-watches to capture food intake information and no specific
power management scheme is used.
3 Proposed Automated Dietary Monitoring System
Our aim is to design a novel framework for a fusion-based automated dietary moni-
toring system using wearable sensors, applicable to Indian food dataset. Also, our
system will be low cost, comfortable for wearing and energy-aware, with a real-time
feedback to the user. The proposed system is shown in Fig. 1. It has three basic
components: body sensor network (BSN), remote server and Web application. In the
next subsections, we will describe each component in detail.
538 S. Mamud et al.
Fig. 1 Proposed automated dietary monitoring system
3.1 Body Sensor Network (BSN)
BSN is formed by sensors, placed on and around a human body, to observe several
physiological parameters for healthcare monitoring systems [20]. The main tech-
niques it covers are sensing events, data fusion and transmitting the data via wireless
communication to a base location for future analysis. In our model, the BSN consists
of an image sensor, audio sensor and an accelerometer sensor for capturing informa-
tion regarding food intake and the eating behaviour of a person. The image sensor,
clubbed with a processor, behaves as a single component and acts as the coordi-
nator node in the BSN. It can be fixed at chest or shirt collar. It captures images
during eating. The audio sensor is fixed at the ear and captures chewing/swallowing
sounds. The accelerometer sensor is worn as a wristband and captures hand move-
ments during eating. The audio and accelerometer sensors can send their data to the
coordinator node through intra-body communication using Wi-Fi. The sensors are
equipped with energy harvesting circuitry so that they can recharge their batteries
from ambient energy sources like body temperature, hand movement/vibration, light
energy, etc. From the coordinator node, data reach to the remote server via a gateway
node.
3.2 Remote Server
The data that we collect from the sensors are kept in the Food Database in the remote
server for future processing. A Nutrition Database, containing the nutritional values
of each food type, is also maintained in the server. By analysing these data, we
can calculate the nutrient information, overeating episodes, unhealthy food intake,
calorie earned, etc., of a person. Users can also give their feedbacks, which are stored
in the Feedback Database.
3.3 Web Application
To make the system interactive with the user or professional health workers, a Web
application (mobile app) can be integrated. Through this, the person concerned can
be aware of his eating habits, overeating episodes, diet recommendations, etc. On the
other hand, the medical/health professionals can view the data for their purpose. Users
can also provide their feedbacks, depending on which the system can be improvised.
However, in case of children/patient the mobile app feedback system is only used by
the health workers.
4 Conceptual Workflow of the System
The system takes food image, chewing/swallowing sound, hand movement as input
from the sensor data. At first, it identifies the food plate, which is assumed to be
placed on a table/floor in front of the user. Then, the specific types of food which
is placed on the plate are detected. During eating, the eating behaviour is observed.
After completion of eating episode, the consumed food volume/quantity is estimated,
and nutritional value is calculated. Lastly, a feedback is provided to the user. The
workflow of the whole system is given in the flow chart as shown in Fig. 2.
Individual steps of the flow diagram are described below.
Step 1 (Food Plate Identification): The camera sensor captures the initial image of
the food plate containing the food. The image is further analysed to identify the food
plate.
Step 2 (Food-Type Classification): The audio sensor captures different eating
episodes like chewing/swallowing throughout the eating process. Also, the camera
sensor captures initial food image, as mentioned in Step 1. By analysing these audio
and image data, food type can be identified and is stored in the Food Database.
Step 3 (Food Volume Estimation): The consumed food volume can be measured by
analysing the initial and leftover food, captured by the camera sensor. The volume
information is stored in Food Database.
540 S. Mamud et al.
Fig. 2 Workflow of
proposed system
Step 4 (Nutritional Value Calculation): After the estimation of the type and volume
of the consumed food, the data are consulted with the Nutrition Database to find out
the exact calorie value and nutritional components of the food intake.
Step 5 (Eating Behaviour Analysis): Eating behaviour is affected by habit. Some
participants exhibit night eating, evening/weekend overeating, bad eating gestures,
etc. With the help of the accelerometer sensor, mounted on the wrist, the proposed
system can provide helpful data. Integrating this with image and audio information,
the system can analyse the eating behaviour of a person.
5 Merits of the Scheme
Accuracy: The proposed BSN consists of an image sensor, audio sensor and an
accelerometer sensor for capturing information regarding food intake and activity of
a person during eating. Now, by using only the image sensor, it is possible to identify
the consumed food type and food volume, which are the most important data to be
analysed. But, to make the system more accurate, reliable and efficient with respect
to versatile Indian food types, we are taking the advantages of the other two types of
sensors. For example, it may be difficult to identify a bowl of milk or a bowl of curd
by an initial image, as both are white in colour and having same kind of textures.
But it becomes easily possible to classify these by tracking the chewing/swallowing
sound.
Cost-effectiveness: We have used wearable sensors for capturing food intake infor-
mation instead of using smartphones or smart-watches, as those are much costlier
gadgets, compared with the sensors. Also, the camera sensor is the compulsory one,
and the other two units, i.e. audio sensor and accelerometer sensor, can be avoided
to reduce the cost, with a little bit of compromising accuracy. Due to the cost-
effectiveness of the system, it can be used in mass to monitor a set of people (each
BSN for individual) during any communal eating (e.g. children taking midday meal
of school).
Energy-aware: The sensors are assumed to be equipped with energy harvesting
circuitry so that they can recharge their batteries from ambient energy sources like
body temperature, hand movement/vibration, light energy, etc. This will make the
system energy-efficient and sustainable.
Comfortable: Our proposed system is user-friendly and comfortable; elders as well
as children can wear, as it is lightweight and easy to manage. The audio sensor is
like an earphone; the image sensor is badge-like; the accelerometer sensor looks like
a wristband, and anyone can comfortably carry these. Also, except the image sensor
unit, the optional components can be removed to increase the ease of use.
6 Conclusion
The paper proposed a novel framework for a wearable sensor-based automatic dietary
monitoring system, specially emphasizing on the Indian food intake monitoring. The
system is low cost, accurate, power-aware, comfortable and scalable. In future, the
system is to be implemented practically, mostly in the rural areas of India, such as to
monitor the diet of children in schools, where midday meal is distributed, and several
feasibility studies must be carried out.
References
1. World Health Organization, Newsroom. https://www.who.int/news-room/campaigns/con

necting-the-world-to-combat-coronavirus/healthyathome/healthyathome—healthy-diet. Last
accessed 25 May 2020
2. M. De Onis, M. Blössner, E. Borghi, Global prevalence and trends of overweight and obesity
among preschool children. Am. J. Clin. Nutr. 92(5), 1257–12649 (2010)
542 S. Mamud et al.
3. J.V. White, P. Guenter, G. Jensen, A. Malone, M. Schofield, Academy malnutrition work

group, ASPEN malnutrition task force and ASPEN board of directors: consensus statement:
academy of nutrition and dietetics and american society for parenteral and enteral nutrition:
characteristics recommended for the identification and documentation of adult malnutrition
(undernutrition). J. Parenter. Enteral Nutr. 36(3), 275–283 (2012)
4. S. Singh, S. Srivastava, A.K. Upadhyay, Socio-economic inequality in malnutrition among
children in India: an analysis of 640 districts from national family health survey (2015–16).
Int. J. Equity Health 18(1), 203 (2019)
5. C.C. Tsai, G. Lee, F. Raab, G.J. Norman, T. Sohn, W.G. Griswold, K. Patrick, Usability and
feasibility of PmEB: a mobile phone application for monitoring real time caloric balance.
Mobile Netw. Appl. 12(2–3), 173–184 (2007)
6. R. Fallaize, H. Forster, A.L. Macready, M.C. Walsh, J.C. Mathers, L. Brennan, E.R. Gibney,
M.J. Gibney, J.A. Lovegrove, Online dietary intake estimation: reproducibility and validity
of the Food4Me food frequency questionnaire against a 4-day weighed food record. J. Med.
Internet Res. 16(8), e190 (2014)
7. K. Yatani, K.N. Truong, BodyScope: a wearable acoustic sensor for activity recognition, in
Proceedings of the 2012 ACM Conference on Ubiquitous Computing (2012), pp. 341–350
8. T. Olubanjo, E. Moore, M. Ghovanloo, Detecting food intake acoustic events in noisy recordings
using template matching, in 2016 IEEE-EMBS International Conference on Biomedical and
Health Informatics (BHI), IEEE (2016 Feb), pp. 388–391
9. S. Päßler, M. Wolff, W.J. Fischer, Food intake monitoring: an acoustical approach to automated
food intake activity detection and classification of consumed food. Phys. Meas. 33(6), 1073
(2012)
10. E.S. Sazonov, J.M. Fontana, A sensor system for automatic detection of food intake through
non-invasive monitoring of chewing. IEEE Sens. J. 12(5), 1340–1348 (2011)
11. P. Lopez-Meyer, S. Schuckers, O. Makeyev, E. Sazonov, Detection of periods of food
intake using support vector machines, in 2010 Annual International Conference of the IEEE
Engineering in Medicine and Biology, IEEE (2010, Sept), pp. 1004–1007
12. F. Zhu, M. Bosch, T. Schap, N. Khanna, D.S. Ebert, C.J. Boushey, E.J. Delp, Segmentation
assisted food classification for dietary assessment, in Computational Imaging IX, vol. 7873,
p. 78730B (2011).. International Society for Optics and Photonics
13. F. Zhu, M. Bosch, I. Woo, S. Kim, C.J. Boushey, D.S. Ebert, E.J. Delp, The use of mobile
devices in aiding dietary assessment and evaluation. IEEE J. Sel. Top. Signal Process 4(4),
756–766 (2010)
14. M. Sun, L.E. Burke, Z.H. Mao, Y. Chen, H.C. Chen, Y. Bai, Y., Li, C. Li, W. Jia, eButton: a
wearable computer for health monitoring and personal assistance, in Proceedings of the 51st
Annual Design Automation Conference (2014), pp. 1–6
15. J. Shang, K. Sundara-Rajan, L. Lindsey, A. Mamishev, E. Johnson, A. Teredesai, A. Kristal, A
pervasive dietary data recording system, in 2011 IEEE International Conference on Pervasive
Computing and Communications Workshops (PERCOM Workshops), (IEEE, 2011) pp. 307–
309
16. Y. Dong, Tracking wrist motion to detect and measure the eating intake of free-living humans
(Univ. Clemson, Clemson, 2012)
17. L. Liu, E. Johns, L. Atallah, C. Pettitt, B. Lo, G. Frost, G.Z. Yang, An intelligent food-
intake monitoring system using wearable sensors. in 2012 Ninth International Conference
on Wearable and Implantable Body Sensor Networks, (IEEE, 2012), pp. 154–160
18. S. Sen, V. Subbaraju, A. Misra, R.K. Balan, Y. Lee, The case for smartwatch-based diet moni-
toring, in 2015 IEEE International Conference on pervasive computing and communication
workshops (PerCom workshops), (IEEE 2015), pp. 585–590
19. J.M. Fontana, M. Farooq, E. Sazonov, Automatic ingestion monitor: a novel wearable device
for monitoring of ingestive behavior. IEEE Trans. Biomed. Eng. 61(6), 1772–1779 (2014)
20. X. Lai, Q. Liu, X. Wei, W. Wang, G. Zhou, G. Han, A survey of body sensor networks. Sensors
13(5), 5406–5447 (2013)
Intruder Insinuation and Smart
Surveillance Using Face Detection
and Mask Detection
M. Rajesh Khanna, G. Prakash Raj, S. Prem Kumar, and N. S. Vignesh Raaj
Abstract A surveillance system uses Raspberry Pi, PIR sensor and mobile devices
to provide security at nighttimes against any intrusion and theft in shopping malls,
jewelry shops and other places which holds valuable items. The main objective of
the proposed system is to provide security to shops in the closed times. The proposed
system uses Haar cascade algorithm for face detection to detect the unknown faces
and own object detection method to detect masked faces. The proposed system
comprises PIR sensor, GSM equipment and any type of good-quality camera. GSM
system will send alert messages to the owner and nearby police station when there is
any intrusion detected. The PIR sensor and camera, all of the equipment are controlled
by the Raspberry Pi system. The camera captures the image, whenever any motion
is detected through remote sensing by PIR sensor. When any human face is detected
through the captured image, the camera will stay turned on; otherwise, it will go to
standby mode. If any unknown and unauthorized face is detected from the image,
camera will continuously capture images. Finally, if the system has detected only
unknown faces, then the owner, security and nearby police station will receive the
alert messages.
Keywords Raspberry Pi · PIR sensor · GSM · Buzzer · USB camera · SD card
1 Introduction
This system consists of both hardware and software requirements. The hardware
requirements are Raspberry Pi model B, PIR sensor, power cable, keyboard, mouse,
SD card, webcam. The software requirements are Raspbian OS, Python programming
language, OpenCV, Win 32 Disk Imager, SD card formatter. Through USB port of
Raspberry Pi, the USB camera is connected. The Raspbian OS is installed in the PC.
Raspberry Pi works only on the Raspbian OS. Raspbian is a free operating system
M. Rajesh Khanna (B) · G. Prakash Raj · S. Prem Kumar · N. S. Vignesh Raaj

Department of Information Technology, Vel Tech Multi Tech Dr.Rangarajan Dr.Sakunthala
Engineering College, Avadi, Chennai, India
e-mail: rajeshkhanna@veltechmultitech.org
544 M. Rajesh Khanna et al.
Fig. 1 System architecture
based on Debian and is optimized for the Raspberry Pi hardware. This operating
system has the set of basic programs and utilities that makes our Raspberry Pi run.
Raspbian has more than 35,000 packages; pre-available software bundles can be
installed in Raspberry Pi. Putty configuration and VNC viewer are necessary to
install Raspbian OS. Putty configuration is SSH and Telnet client. It is an open-source
software which is available with source code.
2 System Architecture
See Fig. 1.
3 Prevailing Technology
In the existing system, the video footage from the camera is used, where the frames
in every 5 s intervals are considered as input to the system. The system will analyze
these frames for facial details. If any facial information is found, then the system
will analyze the face detected with preregistered authorized faces [1]. If the detected
information is different from the preregistered information, then the system will
trigger the alert message.
3.1 Drawbacks
(i) The major drawback of this existing system is that it will not be able to detect
the faces which are covered with any sorts of masks and other things.
Intruder Insinuation and Smart Surveillance … 545
(ii) The system will be active for all the times and that it will not be a power efficient
system; thus, it requires continuous processing of the frames in the fixed interval.
4 Enhanced System
Surveillance and tracking analysis use Raspberry Pi and PIR sensor with the use of
mobile devices to provide our system with critical protection and theft detection. The
proposed system for the detection of theft combines camera and PIR detector sensor,
GSM equipment to provide the owner with the warning via a message through SMS.
Both the PIR sensor and camera are operated and controlled by Raspberry Pi. When
motion is detected by PIR sensor, the camera will be turned on, captures the image
and then stores it for future replay. Finally, if camera detects unknown face means,
warning message will go to owner, security and police. The camera will also be
turned on when any movement is detected; otherwise, it will stay in standby. Also,
if unknown face is detected from camera, the Taser gun is activated or LED will be
turned on. At the same time, camera will continuously capture human faces. In our
project if any theft occurs, a message with alert and address (where theft occurred)
will be automatically sent to nearby (or) closest police station around the theft site
(shopping mall) and owner.
5 Haar Cascade Classifier for Face Detection
5.1 Feature Extraction
Normally, face will have many dark and bright regions. For example, the eye region
will be dark when compared to the cheeks and bright nose region when compared
with eyes. Based on the color density, the pixel values range from 0 to 255. The
features in the image can be calculated by summing the pixel values of black area
and subtracting them with white pixel area.

Rectangle Feature = (Dark pixel) − (White pixel)
5.2 Integral Image
The rectangle feature has coordinates (x, y). Then, integral image of the pixel sum is
calculated by using the formula below
Fig. 2 Converting image

into integral image

ii(x y) = x ≤ x y ≤ yi(x , y )
where ii(x, y) denotes the integral image and i(x, y) denotes original image [2].
These pair of recurrences will be found, in the process of calculating the integral
image (Fig. 2) [2].
s(x, y) = s(x, y − 1) + (x, y)

ii(x, y = ii(x − 1, y) + s(x, y).
5.3 AdaBoosting
AdaBoosting is the process of removing the unwanted features or features that are
irrelevant in the image. The ultimate aim is to gather the features of face. The sub-
window which does not have any features matching a face will be eliminated.
5.4 Cascade Classifier
Cascade classifier identifies the region in the frame which has more facial features
like eyes, nose, etc. It combines the above all techniques to reject the sub-windows
that do not contain any reference for the face and chooses the best region that contains
face. The features are extracted and saved for face recognition.
5.5 Algorithm for Face Detection
Face detection working process is as follows:

1. Extract the features from the image using Haar feature extraction.
2. Convert the original image into integral image for faster calculation of Haar
feature.
3. Remove unwanted features using weak classifier (AdaBoosting).

4. Cascade classifier combines all the above techniques to produce the best region
that has facial features and rejects all other sub-windows.
5.6 Implementation Details
The dataset which contains the collection of images is given as input. Extract the
features from the image and store it separately for comparison or recognition of face.
Then, the frame which contains face from the live video feed is given as input for
face detection. The features are compared and then the program checks whether it is
an authorized person or an unauthorized person.
6 Mask Detection Using Own Object Detection
The own object detector needs some images or dataset. Collect the required dataset
from either Internet or by taking pictures through good-quality camera.
6.1 Labeling the Data
Labeling the data helps us better in detecting the object. A various labeling software
is available, and make use of these software for labeling the images (Fig. 3).
Fig. 3 Labeling the image

6.2 Training the Object Detector
Training is the process which gathers or educes the features from labeled objects.
After training, the features are saved in terms of mathematical values. These
mathematical values should match when testing is done.
6.3 Testing the Object Detector
Testing process evaluates the features educed from the input image with trained data.
The testing process checks whether any labeled objects are present in the camera
window.
6.4 Algorithm for Mask Detection
1. Gather the dataset for mask detection.

2. Labeling the data which contains masked face region.
3. Train the dataset by using own object detection.
4. Test the system using sample images from webcam and evaluate results.
6.5 Implementation Details
The dataset for masked faces is collected and stored in the system. The masked face
regions are labeled from the dataset images. Then, the system is trained using the
labeled dataset images by own object detection method. Finally, sample images from
webcam are taken and tested. After the evaluation, the result is produced.
7 Implementation Result of Face and Mask Detection
Face detection showing unauthorized person has entered, and then SMS is sent to
the authorized person and nearby police station via GSM network. If an authorized
person has entered, then SMS is sent only to the authorized person (Figs. 4, 5 and 6).
Fig. 4 Unauthorized person

recognition
Fig. 5 Authorized person is

recognized with name
Fig. 6 Thief detection
8 Performance Analysis
The above diagram shows the comparison of other existing face detection algorithms’
aspects with Haar algorithm (Figs. 7 and 8).
Fig. 7 Comparison of proposed with other face detection techniques [2–5]
Fig. 8 Accuracy level of face detection
The above figure shows the accuracy levels of the various face detection algorithms
in comparison with the Haar face detection algorithm (Table 1).
Proposed Method is Haar Cascade Algorithm.
Table 1 Accuracy table for the algorithm

Method Accuracy Sensitivity Specificity Dataset images
HAAR 0.905 0.975 0.987 100 images
HOG 0.821 0.761 0.803 100 images
CNN 0.812 0.752 0.791 100 images
LBP 0.854 0.805 0.837 100 images
9 Conclusion
The project “Intruder Insinuation and Smart Surveillance for Face Detection and
Mask Detection” using Raspberry Pi has demonstrated how to get a fully functional
embedded product that can be used for surveillance and security purposes. The system
uses face detection and mask detection for detecting the intruders from camera feed
when any intrusion is detected via embedded motion detection sensors used in the
system. This has included the configuration of embedded Linux, cross-compilation
and deployment of essential libraries with the GSM system. The algorithms used for
face detection and masked face detection are implemented in easier way possible by
the Python packages which provides higher performance, accuracy and other features
with good scale values. The system is power efficient as it starts working, only when
any movement is detected. The developed system is suitable for surveillance at closed
times for places such as malls, jewelry shops, bank locker rooms and many other
places which holds valuable items. The main advantages of this project are it has
easy implementation and low cost with high quality.
References
1. N. Jamil, S. Iqbal, N. Iqbal, Face recognition using neural networks, in Proceeding of IEEE 21st
Century International. Multi Topic Conferences (INMIC), (2001), pp. 277–281
2. V., Paul, M.J Jones, Rapid object detection using a boosted cascade of simple features, in
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern
Recognition vol. 1, pp. 511–518 (2001)
3. M.A. Abuzneid, A. Mahmood, Enhanced human face recognition using LBPH Descriptor, multi-
KNN, and back- propagation neural network. IEEE Access 6, 2064120651 (2018)
4. X.-Y. Li, Z.-X. Lin, Face recognition based on HOG and fast PCA algorithm, in Proceedings
of the Fourth Euro-China Conference on Intelligent Data Analysis and Applications, (Springer,
Malaga, Spain Oct. 2017)
5. N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in IEEE Computer
Society Conference on Computer Vision and Pattern Recognition (2005)
Misinformation Analysis During
Covid-19 Pandemic
Shubhangi Rastogi and Divya Bansal
Abstract Online diffusion of misinformation has gained extreme attention in the

research from past few years. Moreover, during ongoing Covid-19 pandemic, the
proliferation of misinformation became more prominent. In this paper, a compar-
ison of two feature engineering techniques, namely term frequency–inverse doc-
ument frequency (tf-idf) and word embeddings (doc2vec), is done over different
machine learning classifiers. A Web scraper is developed for fact-checking Web site,
Snopes.com, to extract labeled news related to Covid-19. Although the size of dataset
is less, the body content under headlines contains large amount of text. Therefore, the
model works well with both the feature engineering techniques and machine learning
algorithms. Apparently, we obtained best accuracy of 95.38% with tf-idf on decision
tree and same accuracy of 90.77% using doc2vec on support vector machine and
logistic regression machine learning classifier.
Keywords Misinformation · Snopes · Covid-19 · Word embedding · tf-idf
1 Introduction
The phenomenon of misinformation was best ever highlighted during US presidential

election in 2016 and became an active research area [3]. Misinformation is defined
as false information including rumors, hoaxes, fake news, etc. However, it gained
more interest during Covid-19 pandemic [13]. The whole world is suffering from
the contagious Coronavirus disease, and the diffusion of misinformation presents an
additional to human health. Due to lack of vaccine, a lot of misinformation related to
home remedies is circulated online. For instance, gargling with saltwater is helpful to
kill Coronavirus. Machine learning algorithms provide a valuable solution to decide
S. Rastogi (B) · D. Bansal

Punjab Engineering College (Deemed to be University), Chandigarh, India
e-mail: shubhangi05.rastogi@gmail.com
D. Bansal
e-mail: divya@pec.edu.in
554 S. Rastogi and D. Bansal
Table 1 Fact-checking Web sites in different domains

Online rumors and hoaxes Political claims Controversies
Snopes PolitiFact TruthBeTold
Hoax-Slayer ThefactChecker StopFake
Viralgranskaren-Metro FactCheck #RefugeeCheck
TruthOrFiction OpenSecrets ClimateFeedback
HoaxBusters CNNRealityCheck Brown-moses.blogspot
whether any piece of content contains misinformation or not. Although, literature

gives a rich introduction of various techniques applied by researchers to solve the
problem [15]. However, at initial stage of crisis, lack of labeled data is a big chal-
lenge to researchers. Providentially, the fact-checking Web sites play an important
role in collecting verified labeled data [8]. The reporter’s laboratory at Duke Univer-
sity maintains a database which consists of fact-checking Web sites worldwide [17].
The list of fact-checking Web sites in different areas of concern was extracted from
the database as given in Table 1. Therefore, the work in this paper is interested in
solving three research questions:
RQ1: What are the alternatives to obtain labeled dataset to avoid efforts and time
required in manual expert-based labeling?
RQ2: Which feature engineering approach gives high performance in misinforma-
tion detection?
RQ3: Which machine learning classifier is most precise in the domain of misinfor-
mation analysis?
The remaining paper is organized as Sect. 2 highlights the related work done in the
domain of misinformation classification. The methodology is discussed in Sect. 3
with obtained results. Section 4 concludes the paper.
2 Related Work
Fake news detection is an active research area. In this section, various closely related
research works have been examined based on detection of online misinformation.
Table 2 gives a summarization of related work done in this domain based on the
dataset used, purpose of the research, and feature engineering approach. It has been
analyzed that most of the researchers have found linguistic features and word embed-
dings suitable in this field. The studies deliver future direction to be followed for
effective research. No work has been explored in the direction of doc2vec. More-
over, researchers have used existing datasets and randomly crawled datasets. In this
study, we analyzed misinformation related to the current topic Covid-19.
Misinformation Analysis During Covid-19 Pandemic 555
Table 2 Summarization of literature review

References Dataset Purpose Feature engineering
approach
Bulgarian news To filter fake news and Word embedding:
[9] articles clickbait articles using Word2vec
hierarchical attention
deep neural network
FNC dataset Fake news detection Term
[12] using stance frequency–inverse
evaluation approach document frequency
(tf-idf)
Tweets from Developed a URL doc2vec
[16] fact-checking Web recommendation
sites model to debunk fake
news
Social news articles Fake news tf-idf and word2vec
[20] identification using
SVM
Data crawled from Fake news Linguistic (n-gram),
[7] online sources classification using credibility-related, and
logistic regression semantic features
(embeddings)
Different existing To classify fake news Linguistic and word
[6] datasets using ensemble embeddings
algorithms and SVM combinedly
Facebook data Developed a novel LDA, tf-idf, and
[18] approach to classify document similarity
misleading news using Jaccard
similarity
Twitter data related to Developed a classifier LDA and expert-led
[4] Zika virus for the top ten features information retrieval
of IG and GBE
techniques with three
machine learning
algorithms
Existing Chinese Detection of abnormal doc2vec
[2] comment dataset comments based on
semantic similarity
Satirical Web sites Detection of satirical n-grams, tfidf, and
[19] data data using SVM doc2vec
3 Methodology
In this paper, two feature engineering techniques, tf-idf and word embeddings,
have been experimented over machine learning algorithms, and Fig. 3 provides the
overview of the framework for detection of misinformation related to Covid-19 pan-
demic.
3.1 Data Collection and Preprocessing
The data has been collected from Snopes fact-checking Web sites using keywords
like, Covid-19 and Coronavirus. In this paper, Snopes has been selected from the first
category in Table 1 because it targets online rumors. Other services like HoaxBusters
and HoaxSlayer also aim similar. Nonetheless, Snopes is accomplished by a small
volunteer party that has arisen from a single-person initiative and funded through
advertising revenue [1]. A Web scraper has been developed to extract 500 news
headline, text body, and respective labels (True or False) from Snopes fact-checking
Web site. Figure 1 shows the sample of data extracted from Snopes related to Covid-
19. Furthermore, headline and body are merged into a single attribute, text. A series
of data cleaning steps have been applied on text including removal of punctuation
marks, special characters, white spaces, and converting upper case letters to lower
case. A sample of cleaned data is shown in Fig. 2. Moreover, Fig. 4 depicts word
clouds generated for fake and real news articles.
Fig. 1 Sample of data extracted from snopes
Fig. 2 Sample of cleaned data

Fig. 3 Framework of fake news detection model
Fig. 4 Visualization of data using word clouds a fake data, b real data
3.2 Feature Engineering
There exist different feature engineering techniques to convert data into feature
vectors and preparing data to perform better using machine learning classification
algorithms. According to the literature, word counts, term frequency–inverse doc-
ument frequency (TFIDF), word embeddings, etc., are popular feature engineering
approaches.
3.2.1 TFIDF
Essentially, TFIDF or tf-idf is used in information retrieval and text mining and
provides a weight which is used to assess importance of word to the document in
a corpus. TFIDF is the product of two statistics, term frequency, and inverse docu-
ment frequency [14]. Term frequency: The weight assigned to words based on their
occurrence in the document. Inverse document frequency: The weight assigned to
words based on their importance in the document. The ‘term frequency’ (tf) statistics
considers each term or word equally. However, there are frequently occurring terms
like is, the, of, that, and many more which are less important. Therefore, idf scales
up the rare terms and down frequent terms.
3.2.2 Word Embedding
Word embeddings are a kind of word representation that provides similar words with
similar meaning, a similar representation. Word2vec is not a single algorithm but
works in the combination of two methods—continuous bag of words (CBOW) and
skip-gram model. Both the methods are based on shallow neural networks which
map word to target variable or word. Moreover, learned weights by both the tech-
niques are termed as vector representations [11]. Word embedding protects the word
relationships in such a way that words with similar context will have similar vec-
tor representation or embeddings. In this paper, doc2vec have been used which is a
small extension of CBOW [10]. Doc2vec improves by adding another feature vector,
document-id instead of using words to predict the next word. Moreover, Doc2Vec
model is used to create a vectorized representation of a set of words taken together as
a single unit and trained with word vectors [21]. In this paper, word embeddings have
been used to train 300-dimensional domain-specific word embeddings on Snopes’
news articles extracted, and details are given in Table 3. The labeled news headlines
and body content combinedly contain an enormous amount of text of 2526038 unique
words. We applied doc2vec using genism for text representation, and as an input to
the machine learning classifiers.
3.3 Machine Learning Classifiers
Performed experiments with the following established classification methods: Logis-

tic regression (LR), Support vector machines (SVM), Decision trees (DT), Random
forest (RF), K-nearest neighbors (KNN), and Naïve Bayes (NB) [5]. The training
size is selected as 80% of the total dataset, while test size is 20% of the dataset.
Table 3 Word embeddings configuration

Embedding technique Doc2vec using gensim
Window size 5
Vector dimension 300
Epochs 10
Distributed memory 1
3.4 Results
The results have been obtained using two feature engineering techniques, namely
tf-idf and word embeddings, and the performance of machine learning classifiers
was examined using accuracy, precision, recall, and f-measure metrics (as shown
in Fig. 5). Tables 4 and 5 show the performance of each machine learning classifier
using tf-idf and word embedding techniques, respectively.
Fig. 5 Comparative analysis of tf-idf and doc2vec over six classifiers
Table 4 Performance metrics over different machine learning classifiers using tf-idf
Classifier Accuracy Precision Recall F-measure
LR 87.69 88 99 93
SVM 92.31 88 99 93
KNN 87.69 93 98 96
DT 95.38 95 98 97
RF 87.69 95 98 97
NB 87.69 87 100 93
Table 5 Performance metrics over different machine learning classifiers using doc2vec word
embedding
Classifier Accuracy Precision Recall F-measure
LR 90.77 91 100 95
SVM 90.77 91 100 95
KNN 87.69 92 95 93
DT 80.00 94 83 88
RF 86.15 92 93 92
NB 58.47 94 58 72
During Covid-19 pandemic, the diffusion of misinformation became as harmful as

the virus. In this paper, a Web scraper in Python programming environment for
Snopes.com has been developed to extract Covid-19 related news headlines and body
with respective labels true and false. Thereafter, two popular feature engineering
approaches were applied, namely tf-idf and doc2vec word embedding, to represent
text into vectors because tf-idf is simplest and effective technique in text classifica-
tion while word embedding is an improvement over this technique by including the
semantic relationship between the words. Moreover, the vectors are trained sequen-
tially over six machine learning classifiers. In this paper, it has been observed that
tf-idf delivered best accuracy with decision tree classifier of 95.38%, while doc2vec
performed well with support vector machine and logistic regression with same accu-
racy of 90.77%. The paper highlights the description of doc2vec and tf-idf feature
engineering techniques. Since the paper covers the current topic, there exists a paucity
of available data on fact-checking Web sites. Covid-19 is an ongoing pandemic, and
a lot of information is floating over online sources. Therefore, in future, we are
focussed to work on this limitation of small dataset by collecting more data related
to Covid-19 to train the model in a better way.
References
1. P.B. Brandtzaeg, A. Følstad, M.Á. Chaparro Domínguez, How journalists and social media
users perceive online fact-checking and verification services. Journalism Pract 12(9), 1109–
1129 (2018)
2. W. Chang, Z. Xu, S. Zhou, W. Cao, Research on detection methods based on doc2vec abnormal
comments. Future Gener. Comput. Syst. 86, 656–662 (2018)
3. K. Clayton, S. Blair, J.A. Busam, S. Forstner, J. Glance, G. Green, A. Kawata, A. Kovvuri, J.
Martin, E. Morgan, et al., Real solutions for fake news? measuring the effectiveness of general
warnings and fact-check tags in reducing belief in false stories on social media. Polit. Behav.,
pp. 1–23
4. A. Ghenai, Y. Mejova, Catching Zika Fever: Application of Crowdsourcing and Machine Learn-
ing for Tracking Health Misinformation on Twitter. arXiv preprint arXiv:1707.03778 (2017)
5. S. Gilda, Evaluating machine learning algorithms for fake news detection, in 2017 IEEE 15th
Student Conference on Research and Development (SCOReD) (IEEE, 2017), pp. 110–115
6. G. Gravanis, A. Vakali, K. Diamantaras, P. Karadais, Behind the cues: a benchmarking study
for fake news detection. Expert Syst. Appl. 128, 201–213 (2019)
7. M. Hardalov, I. Koychev, P. Nakov, In search of credible news, in International Conference on
Artificial Intelligence: Methodology, Systems, and Applications (Springer, Berlin, 2016), pp.
172–180
8. S. Jiang, C. Wilson, Linguistic signals under misinformation and fact-checking: evidence from
user comments on social media. Proc. ACM Hum.-Comput. Interac. 2(CSCW), 1–23 (2018)
9. G. Karadzhov, P. Gencheva, P. Nakov, I. Koychev, We Built a Fake News & Click-Bait Filter:
What Happened Next will Blow Your Mind! arXiv preprint arXiv:1803.03786 (2018)
10. Q. Le, T. Mikolov, Distributed representations of sentences and documents, in International
Conference on Machine Learning, pp. 1188–1196 (2014)
11. T. Mikolov, I. Sutskever, K. Chen, G.S. Corrado, J. Dean, Distributed representations of words
and phrases and their compositionality, in Advances in Neural Information Processing Systems,
pp. 3111–3119 (2013)
12. B. Riedel, I. Augenstein, G.P. Spithourakis, S. Riedel, A Simple But Tough-to-Beat Baseline
for the Fake News Challenge Stance Detection Task. arXiv preprint arXiv:1707.03264 (2017)
13. K. Sharma, S. Seo, C. Meng, S. Rambhatla, Y. Liu, Covid-19 on Social Media: Analyzing
Misinformation in Twitter Conversations. arXiv preprint arXiv:2003.12309 (2020)
14. C. Shi, C. Xu, X. Yang, Study of tfidf algorithm. J. Comput. Appl. 29(6), 167–170 (2009)
15. K. Shu, A. Sliva, S. Wang, J. Tang, H. Liu, Fake news detection on social media: a data mining
perspective. ACM SIGKDD Explor. Newsl. 19(1), 22–36 (2017)
16. N. Vo, K. Lee, The rise of guardians: Fact-checking url recommendation to combat fake news,
in The 41st International ACM SIGIR Conference on Research & Development in Information
Retrieval, pp. 275–284 (2018)
17. Y. Wu, P.K. Agarwal, C. Li, J. Yang, C. Yu, Toward computational fact-checking. Proc. VLDB
Endowment 7(7), 589–600 (2014)
18. K. Xu, F. Wang, H. Wang, B. Yang, A first step towards combating fake news over online
social media, in International Conference on Wireless Algorithms, Systems, and Applications
(Springer, 2018), pp. 521–531
19. F. Yang, A. Mukherjee, E. Dragut, Satirical News Detection and Analysis Using Attention
Mechanism and Linguistic Features. arXiv preprint arXiv:1709.01189 (2017)
20. S. Zhang, Y. Wang, C. Tan, Research on text classification for identifying fake news, in 2018
International Conference on Security, Pattern Analysis, and Cybernetics (SPAC) (IEEE, 2018),
pp. 178–181
21. X. Zhang, A.A. Ghorbani, An overview of online fake news: characterization, detection, and
discussion. Inf. Process. Manage. 57(2), 102025 (2020)
An Agent-Based Cloud Service
Negotiation in Hybrid Cloud Computing
Saurabh Deochake and Debajyoti Mukhopadhyay
Abstract With the advent of evolution of cloud computing, large organizations

have been scaling the on-premise IT infrastructure to the cloud. Although this being
a popular practice, it lacks comprehensive efforts to study the aspects of automated
negotiation of resources among cloud customers and providers. This paper pro-
poses a full-fledged framework for the multi-party, multi-issue negotiation system
for cloud resources. It introduces a robust cloud marketplace system to buy and sell
cloud resources. The belief–desire–intention (BDI) model-based cloud customer and
provider agents concurrently negotiate on multiple issues, pursuing a hybrid tactic
of time- and resource-based dynamic deadline algorithms to generate offers and
counter-offers. The cloud marketplace-based system is further augmented with the
assignment of behavior norm score and Reputation Index to the agents to establish
trust among them.
Keywords Automated negotiation · Intelligent systems · Cloud computing ·

Multi-agent systems
1 Introduction
Negotiation is a mechanism in which different agents concede to an agreement based

on a joint future venture [1]. For a technology organization, expanding the on-premise
infrastructure is no longer limited to expanding the hardware footprint in the data
centers. The organization may benefit from expanding their infrastructure into a
multi-tenant system like public cloud, embracing a hybrid cloud infrastructure model.
The recent studies of e-negotiation can be applied to such unique scenarios of hybrid
S. Deochake (B)
Twitter, Inc., Boulder, CO 80302, USA
e-mail: sdeochake@twitter.com
D. Mukhopadhyay
WIDiCoReL Research Lab, Mumbai University, Mumbai 400032, India
e-mail: debajyoti.mukhopadhyay@gmail.com
564 S. Deochake and D. Mukhopadhyay
cloud computing. A well-known example is Amazon AWS EC2 Marketplace that

allows customers to bid for EC2 Spot Instances virtual machines. To implement
a robust cloud marketplace system, the transactions involving exchange of offers
and counter-offers conceive the alliances that could be either ad-hoc or permanent.
This agenda necessitates a strict coherence to an appropriate set of protocols and
a threshold for every negotiation agreement between a consumer and a provider
agents [2]. An agent is an encapsulated computer system based on an unknown
external environment with the ability of elasticity and self-governing actions in that
environment in order to meet its designed objectives. The belief–desire–intention
(BDI) [3], a software model designed to program intelligent agents, is perhaps that
most studied model for bounded rational systems. BDI model closely associates
with practical human thinking and abstract logical reasoning enabling rational BDI
agents perform complex tasks in a dynamic environment using real-time reasoning
frameworks like procedural reasoning systems (PRS).
The negotiation agenda possesses a multitude of issues that need to be fulfilled in
order to reach an agreement. In the scenario of hybrid cloud computing negotiation,
the agenda may consist of a single object such as price of a virtual machine resource
or the issues may be multiple such as offered memory size, disk storage size, and the
operating system image for a virtual machine. The transactions become even more
multifaceted when they involve issues related to non-functional attributes like cloud
service-level agreement (SLA). A cloud customer may concurrently negotiate with
multiple cloud providers on the issues for a product that are the most important to the
customer. However, before commencing a negotiation round, all stakeholders must
agree upon the issues in the negotiation to form an alliance. Alliances are formed
when multiple parties join hands together in pursuit of a command business goal.
The paper presents a multi-party multi-issue negotiation system for BDI based
agents that adapt to advancements from each round of negotiation and make appro-
priate counter-offers in a cloud marketplace to reach an agreement. The paper is
structured as follows: Sect. 2 discusses previous work, Sect. 3 showcases the model,
Sect. 4 presents the algorithmic approach, Sect. 5 presents future work, and finally,
we conclude this paper in Sect. 6.
2 Related Work
BDI based programming paradigm [3] is permeated with philosophical attitudes viz
beliefs, desires, and intentions. Beliefs represent the agent’s current knowledge of
the outside world. Desires represent the objectives that the agent may carry out based
on the current beliefs. Finally, intentions are the ruminative state of the agent. A few
shortcomings of original BDI based agents such as lack of learning and adaption,
and multi-agent competence as showcased in [4] were attempted to alleviate, for
example, in [5] that showcases distributed multi-agent reasoning system (dMARS)
An Agent-Based Cloud Service Negotiation in Hybrid … 565
architecture that couples PRS with BDI architecture to enable distributed BDI agents
to learn and adapt. Additionally, the work in [6] extends the original BDI agent
architecture to facilitate multi-agent paradigm with shared beliefs using distributed
shared memory and network messages. With the work that has been done in the
area of learning and multi-agent system capability, BDI agents can be used to solve
complex problems with limited information operating in a dynamic environment
[7]. Our paper relies on BDI agents’ competence to adapt to constantly changing
environment when negotiating complex cloud service issues.
With the advent of evolution in cloud computing [8], an organization looking
to expand their IT infrastructure to cloud has multiple choices for cloud service
providers. In order to choose the best deal, it must negotiate with multiple cloud
providers to select the best possible outcome. To the authors’ best information, most
of the study has been focused on negotiating the SLAs in cloud computing [9–
11]. The paper in [9] discusses negotiating cloud service SLA for grid computing
where the process owner specifies service-level requirements from business process
perspectives and selects a cloud provider. While this approach works well nego-
tiating non-functional parameters like SLAs from a pre-elected opponent, it does
not discuss negotiating functional parameters like cloud service’s specifications with
multiple parties in a cloud marketplace scenario. To the authors’ best belief, cloud
agency-based approach such as mOSAIC [10] adopts a software-based multi-cloud
negotiation system. It proposes a middleware-like cloud agency model that nego-
tiates on behalf of a cloud customer. However, this approach does not consider a
fine-grained control such as provisioning a concurrent multi-party multi-issue nego-
tiation. Our approach proposes a marketplace-based negotiation system where each
BDI-based agent sets a fine-grained control on the issues that it values the most in
form of weights.
Traditional negotiation exhibits predefined strategy that works well with an oppo-
nent. For instance, in [2, 12], the agent selects a predefined negotiation strategy that
works well with a specific negotiation scenario using the game theory principles
seeking a win-win outcome. However, this approach forces both buyers and sellers
to lose their utility in reaching the agreement. In our work, BDI agents adapt to each
negotiation round, selecting appropriate tactic to reach an agreement within a fixed
time period. While [13] discusses a bilateral multi-issue negotiation system where
a buyer and seller negotiate on multiple issues concurrently, buyer and seller must
reach an agreement on every issue for a consensus. This leads to over-negotiating
between a buyer and seller that potentially continues for extended period of time.
While belief–goal–plan (BGP)-based multi-strategy approach [14] makes it easier
for an agent to select the best-working negotiation strategy to counter the oppo-
nent, it does not take into account multilateral concurrent negotiation on multiple
issues. Furthermore, negotiation may not terminate within a fixed time period if both
buyer and seller have a set of mutually exclusive issues. Our approach is inspired by
Faratin’s time and behavior-dependent negotiation strategy, selecting a set of issues
with least associated weight to concede to the opponent. A similar attempt has been
performed in [15]. Allowing negotiation between two parties on mutually exclusive
issues would inundate the negotiation system with quotes with few interested oppo-
nent agents. In our approach, our framework allows each party to select the issues
that they deem the most important, and the opponent is matched for negotiation based
on the shared mutual interests to reduce the overhead on the negotiation engine.
3 Proposed System Architecture
As showcased in Fig. 1, our proposed system architecture is composed of a cloud

marketplace and cloud agents. The cloud marketplace facilitates negotiation among
multiple agents on a certain subset of issues. It also provides the marketplace for buyer
and sellers of cloud services to advertise their products and procure the services.
3.1 Cloud Marketplace
The cloud marketplace is a central distributed system where cloud providers and
customers submit their advertisements. The negotiation system and the associated
Web services are implemented using Web Service Description Language (WSDL),
XML and Simple Object Access Protocol (SOAP). The cloud marketplace back-
end components can be interacted via a RESTful Marketplace API service after a
successful authentication.
Advertisement repository stores the agent and product information sent by cloud
providers and customers. It assists in agent and product discovery by querying the
parameters like agent ID, product ID, and corresponding issues. The cloud customer
Fig. 1 Proposed system architecture—an agent’s view

agent that is looking to procure a service from a cloud provider will send a request
for quote (RFQ) with its information, product information, and a certain criteria such
as Reputation Index for an agent. Other services poll advertisement repository for
match-making and statistical purposes.
Cloud provider and customer may have their own agenda on how important they
deem individual issues of a product. Expensive compute resources in distributed
multi-threaded transactions [16] necessitate the provisioning of a feature that would
limit the futile concurrent negotiation on issues that are mutually exclusive for both
agents. Alliance Engine is essentially a match-maker between a cloud provider agent
and a customer agent that consistently polls the advertisement repository database
for issues of mutual interest and other parameters in RFQ like Reputation Index for
a match. This service then forwards a successful match to the involved agents via
Negotiation Engine with a message to commence the negotiation.
Negotiation Engine is a pivotal service where all transactions between the two
agents occur. After an alliance is formed, a commenceNegotiation message,
along with a timestamp in UTC, is sent to both the agents. The cloud provider agent
starts the negotiation by sending an offer to a customer agent. The rounds of offers
and counter-offers continue between the agents for a threshold of time. Negotiation
system also keeps track of all the transactions between agents and gathers statistics
about the negotiation process and agents. These statistics are then fed into Behavior
Watchdog service that assigns behavior and reputation indices for the agents.
A cloud customer agent may prefer negotiating with a cloud provider that provides
better SLAs, cost-effective resources, and faster agreements with its customers. A
cloud provider may have similar expectations. Based on the statistical data gathered
by Negotiation Engine, Behavior Watchdog service calculates Reputation Index, a
rational number value—R, where 0 ≤ R ≤ 1, stored in a database named Reputation
Index Archive. More trustworthy and reputable agents possess higher value and tend
to generate more agreements than lesser reputable agents. Similarly, an agent could
use the information about common behavior of its opponent to draft a negotiation
strategy. Based on the negotiation messages, Behavior Watchdog also curates a gen-
eralized Behavior Norm index, B, for an agent that indicates how an agent generally
behaves in negotiations. Behavior Norm index is also a rational number value that
decides the nature of the agent as mentioned in (1) for the strategy S.
⎧
⎪
⎨ Ba < 1, where agent a tends to be headstrong
S = Ba = 1, where agent a tends to be linear (1)
⎪
⎩
Ba > 1, where agent a tends to be conceder
3.2 Cloud Agent
Each BDI based cloud agent has three relational database systems that store agent’s
beliefs, goals, and plans. An agent’s beliefs, stored in Beliefset, are the knowledge
about the current negotiation round on an issue i and are constantly updated after
each round. The agent assigns weights to every issue in the negotiation, indicating
n
the importance of that issue, such that i=1 Wi = 1. A goal can be an activity of
trying to reach an agreement for an utility of an issue based on the belief gained from
the previous round of negotiation. Like beliefs, goals are updated every round of
negotiation. In terms of negotiation, a plan would indicate that the agent has started
to execute an offer, a counter-offer, or termination of the negotiation. Frequently
occurring scenarios of plan creation are stored in a database named Plan Library.
A negotiation between two agents requires a strict adherence to the agreed issues
and a time threshold. Receiving a counter-offer from an opponent after the time
threshold that is passed would be an illegal transaction. Additionally, to secure an
agent from a malicious attempt to disrupt its negotiation with other agents, it becomes
crucial to accept only those transactions that are valid. Proxy Filter filters illegal
messages pertaining to a negotiation if a transaction counter-offer is outside of current
negotiation space.
An agent may enter negotiation to buy or sell its product with multiple agents at
once reaching an agreement with an opponent that offers a higher utility. It becomes
essential for an agent to be aware of every negotiation space that it is participating.
Alliance Engine service is responsible for curating the information about a negoti-
ation space with opponent agents, storing the metadata of all active negotiations in
agenda DB.
Negotiation Engine service is the heart of the agent software. This service per-
forms essential transactions such as generating offers and counter-offers, evaluating
those offers and selecting appropriate negotiation strategy. It is also responsible for
submitting advertisements and RFQs for the issues to the Cloud Marketplace API.
After each round of negotiation, Negotiation Engine service also updates Beliefset,
Goal Repository, and Plan Library for the agent.
Resource Monitor constantly polls the on-premise resource availability by means
of a monitoring service. Based on the dynamic availability of the resources, the agent
will ensure that it selects an appropriate negotiation strategy. If Resource Monitor
alerts the agent that the resource it is looking to purchase is running short, the agent
will choose to be aggressive in reaching an agreement as soon as possible.
4 Negotiation Process
Before the start of the actual negotiation process, a cloud agent submits its desires
such as weight of an issue (W i ), a range of acceptable cost value (C i ) to each issue, i
as mentioned in (1), and a threshold value for the negotiation time period (tmax ) such
that tmini ≤ ti ≤ tmaxi . The utility value for each issue submitted by the agent is then
calculated as per (2).

n
n
U= Ci Wi where Wi = 1 (2)
i=1 i=1
When the negotiation rounds start, there are multiple ways of how an agent records
its response, R. For a transaction between agents a and b, an agent may acquire the
t
offer for an issue i (Oin−1 ) if the measured utility of that offer is greater than or equal
to the utility of the counter-offer. Otherwise, the agent sends a counter-offer with a
modified cost value for an issue i. As per the protocol, if an counter-offer is received
beyond the threshold value of the negotiation time period, the agent would terminate
the negotiation round as per (3).
⎧ tn−1 tn−1
⎪ tn
⎨acquire(a, b, Oi ), Ui (Oi a→b ) ≤ Ui (Oi b→a )
R = terminate(a, b), tco > tmax (3)
⎪
⎩ tn
counter(a, b, Oi a→b ), otherwise
Once an agent enters negotiation process, the goal of the agent is maximizing the
utility for all the issues in the product that the agent deems important. In order to do
so, the agent should emphasize on counter-offer generating strategies. As showcased
in [17, 18], the agent may use dynamic time-dependent tactic where offers vary as
the deadline approaches, based on a parameterized time function. On the other hand,
dynamic resource-dependent tactic enables agents vary their offers based on the
quantity of resources. In our model, cloud customer and cloud provider agents use
a hybrid time and resource-dependent deadline-based tactics to generate counter-
offers. As mentioned in [17], time-dependent tactic is governed by a time-based
function f(t) where an agent must conclude the negotiation within the threshold
value of tmax such that 0 < t ≤ tmax . A counter-offer O for issue i from cloud agent
a to b at time tn is measured as per (4).

tn mini a + f i a (tn )(maxi a − mini a ), if cost decreases
Oi a→b = (4)
mini a + (1 − f i a (tn ))(maxi a − mini a ), if cost increases
Furthermore as showcased in [17], the time-based function that depicts the agent
behavior is based on fulfillment of following constraints, L as per (5).
⎧
⎪
⎨0 ≤ f i a (tn ) ≤ 1, offer range
L = f i a (0) = Ci a (0), initial cost belief at time 0 (5)
⎪
⎩
f i a (tmax ) = 1, at the time threshold
After a round of negotiation, if a counter-offer sent by opponent agent does not fall
within the acceptable utility range, a new offer is generated based on a concession
rate, λ for the cost value of the issue as shown in (6).
t
Oi tb→a
n
− Oi b→a
n−1
λ= t t (6)
n−1
Oi b→a − Oi b→a
n−2
In the hybrid time-resource-dependent counter-offer tactic, a hardheaded tactic is

applied near the negotiation threshold or if opponent’s resources run short. That is,
concessions are made near the end of the time threshold, tmax or resource threshold,
rmax , whichever is smaller. On the other hand, a conceder tactic is applied if an agent
is desperate to reach a deal quickly by applying concessions to its offers starting early
in the negotiation rounds. An agent may start conceding in the early rounds of the
negotiation, and then depending on the resources, it may choose to alter the tactic to
become more hardheaded. If the agent finds that the opponent is practising a linear
conceding tactic, then the agent will try to match it, changing the counter-offers only
slightly.
To calculate the concession rate, λ, an agent must wait until the third round of
negotiation, if it finds a better utility offer before third round, the agent will accept
that offer. Beyond third round, negotiation engine service calculates the concession
rate by evaluating past offers as per (6). Concession rate of λ > 1 signifies opponent’s
conceding strategy, and the agent will imitate opponent tactic. On the other hand,
concession rate of λ < 1 conveys opponents headstrong tactic, and the agent chooses
conceding tactic to facilitate faster agreement. Agent tactic is unchanged for the
linear concession rate of λ = 1.
5 Future Work
The paper discussed a negotiation model ensuring concurrent negotiation transactions

with multiple opponent agents. The paper focuses on a single owner of IT infrastruc-
ture looking to scale their infrastructure into the cloud. However, the authors believe
that this is not indicative of all organizations and their infrastructure setup. We fur-
ther intend to examine more complex use cases allowing individual microservices in
the infrastructure to automatically negotiate with public cloud providers and scale
themselves. We plan to study the possibility of assigning an intelligent agent to an
individual service and explore how those multiple agents together form multi-agent
system and scale the overall infrastructure. Another aspect of future work refers to
extending this negotiation system for multi-agent behavior [19] and communication
[20] using paradigms like distributed shared memory and law-governed interaction,
respectively.
6 Conclusion
This paper presented an enhanced cloud marketplace-based automated negotiation

system for hybrid cloud computing that exhibits multi-party, multi-issue negotia-
tions. To reduce the compute overhead incurred in parallel negotiation, it introduces
forming the alliances with converging agenda among BDI agents. The cloud mar-
ketplace system ensures that an agent may choose to transact with a certain set of
reputable agents. The model also enables agents to craft their negotiation tactic using
hybrid time-resource deadline algorithm. The paper paves way toward an automated
scaling of an on-premise infrastructure into public cloud. While the current approach
presented in the paper focuses on a single agent, further study is underway to aug-
ment the current work with the provisions of distributed learning and interaction in
a multi-agent system.
References
1. N. Jennings, P. Faratin, A. Lomuscio, S. Parsons, M. Wooldridge, C. Sierra, Automated nego-

tiation: prospects, methods and challenges. Group Decis. Negotiat. 10, 199–215 (2001)
2. S. Deochake, S. Sarode, S. Kanth, V. Potdar, S. Chakraborty, D. Mukhopadhyay, HENRI: high
efficiency negotiation-based robust interface for multi-party multi-issue negotiation over the
internet, in Proceedings of the CUBE International Information Technology Conference, pp.
647-652 (2012). https://doi.org/10.1145/2381716.2381840
3. M. Bratman, Intention, Plans, and Practical Reason, vol. 10 (Harvard University Press, Cam-
bridge, MA, 1987)
4. M. Georgeff, B. Pell, M. Pollack, M. Tambe, M. Wooldridge, The belief-desire-intention model
of agency, in Intelligent Agents V: Agents Theories, Architectures, and Languages, pp. 1–10
(1999)
5. M. D’Inverno, M. Luck, M. Georgeff, D. Kinny, M. Wooldridge, The dMARS architecture: a
specification of the distributed multi-agent reasoning system. Auton. Agents Multi-Agent Syst.
9, 5–53 (2004)
6. M. Barbieri, V. Mascardi, Hive-BDI: extending jason with shared beliefs and stigmergy, in
ICAART 2011 - Proceedings of the 3rd International Conference on Agents and Artificial
Intelligence, vol. 2, pp. 479–482 (2011)
7. M.S.S. Rao, J. An, BDI: applications and architectures. Int. J. Eng. Res. Technol. (IJERT) 2
(2013)
8. R. Buyya, C. Yeo, S. Venugopal, J. Broberg, I. Brandic, Cloud computing and emerging IT
platforms: vsion, hype, and reality for delivering computing as the 5th utility. Future Gener.
Comput. Syst. 25, 599–616 (2009)
9. V. Stantchev, C. Schröpfer, Negotiating and enforcing QoS and SLAs in grid and cloud com-
puting, in Advances in Grid and Pervasive Computing, pp. 25–35 (2009)
10. S. Venticinque, R. Aversa, B. Di Martino, M. Rak, D. Petcu, A cloud agency for SLA negotiation
and management, Euro-Par 2010 Parallel Processing Workshops., pp. 587–594 (2011)
11. B. Shojaiemehr, A. Rahmani, N. Qader, A three-phase process for SLA negotiation of composite
cloud services. Comput. Stan. Interfaces 64, 85–95 (2019)
12. D. Mukhopadhyay, S. Deochake, S. Kanth, S. Chakraborty, S. Sarode, MAINWAVE: multi
agents and issues negotiation for web using alliance virtual engine. Smart Comput. Rev. 2,
308–317 (2012). ISSN: 2234-4624
13. S. Vij, M. Patrikar, D. Mukhopadhyay, A. Agrawal, A smart and automated negotiation system
based on linear programming in a multilateral environment. Smart Comput. Rev. 5, 540–552
(2015). ISSN 2234-4624
14. S. Awasthi, S. Vij, D. Mukhopadhyay, A. Agrawal, Multi-strategy based automated negotiation:
BGP based architecture, 2016 International Conference on Computing, Communication and
Automation (ICCCA) (2016)
15. M. Cao, Dynamic time-dependent strategy model for predicting human’s offer in E-commerce
negotiation. Open J. Soc. Sci. 04, 64–69 (2016)
16. S. Ravoor, J. Wong, Multithreaded transaction processing in distributed systems. J. Syst. Softw.
38, 107–117 (1997)
17. H. Cardoso, M. Schaefer, E. Oliveira, A multi-agent system for electronic commerce including
adaptive strategic behaviours, in Progress in Artificial Intelligence, pp. 252–266 (1999)
18. P. Faratin, C. Sierra, N. Jennings, Negotiation decision functions for autonomous agents. Robot.
Auton. Syst. 24, 159–182 (1998)
19. T. Gonzaga, C. Bentes, R. Farias, M. de Castro, A. Garcia, Using distributed-shared mem-
ory mechanisms for agents communication in a distributed system, in Seventh International
Conference on Intelligent Systems Design and Applications (ISDA 2007) (2007)
20. N. Minsky, Law-governed multi-agent systems. ACM SIGSOFT Softw. Eng. Notes 30, 1–1
(2005)
IoT-Based Smart Farming Application
for Sustainable Agriculture
Sanjeevakumar M. Hatture and Pallavi V. Yankati
Abstract The farmers are struggling to obtain higher rate of yield due to lack of
poor knowledge about the soil and water nutrients and suitability of the crop for
the soil. To overcome these issues and to promote digital agriculture concept, we
propose an IoT-enabled sensor system for monitoring soil nutrient [NPK] and pH
of irrigation water to reduce the manual laboratory method of testing and get the
results via mobile application. Smart contract farming-based mobile application will
further process these nutrients value to predict and suggest the suitable crop to grow
and the usage of the appropriate amount of fertilizer to increase the soil fertility.
The mobile application also helps in establishing the connection between farmers
and Agricultural Produce Market Committee (APMC) in order to avoid fragmen-
tation of profit shares and attain pricing uncertainty and marketing of the yields by
avoiding the middleman. APMC is a state government body which ensures safeguard
to the farmers from exploitation by large retailers. Thus, the proposed work helps the
farmers in adopting stress-free farming practice by signing the contract offered from
APMC to select the suitable crop for their agriculture land, dispensing the required
amount of fertilizer to the soil based on soil and water nutrients analysis using IoT-
enabled sensor, funding/insurance to the crops in case of occurrence of unpredictable
natural disaster in future and direct marketing facility without middleman.
Keywords Sustainable agriculture · Smart contract farming · Soil and water

nutrients · IoT · APMC · Crop insurance
S. M. Hatture (B) · P. V. Yankati

Basaveshwar Engineering College (Autonomous), Bagalkot 587103, Karnataka, India
e-mail: smhatture@gmail.com
P. V. Yankati
e-mail: pallavivijay3575@gmail.com
574 S. M. Hatture and P. V. Yankati
1 Introduction
India is a country where agriculture is given a lot of importance as it is the backbone

of the Indian economy. It provides employment opportunity to more than half of the
workforce of the country and also contributes 16.5% to total gross domestic product
(GDP) and also. Earlier days, there was no scarcity for food and water as it was
used in efficient manner. But nowadays, there is scarcity of food and water due to
increase in population growth of the country. To satisfy their needs, fertilizers are
used to increase crop production. Fertilizer is any substance which is added to soil
for supplying nutrients externally which are crucial to plant growth. Too much use
of fertilizers shatters the growth of crops and little usage will result in less yield;
therefore, it is necessary for farmers to understand the correct usage of fertilizer to
increase the soil fertility.
Nutrients are important factors for plant growth and are divided into micro and
macronutrients. Totally, soil consists sixteen nutrients, and among them the three
primary nutrients are nitrogen (N), phosphorus (P) and potassium (K). Micronutrients
are required in less quantity; whereas, macronutrients are required in more quantity.
Plants consume necessary elements from the soil through their roots and from the air
through their leaves. These consumed nutrients are moved within plants to where they
are required most. Figure 1 shows the classification and sources of plant nutrients.
Fig. 1 Plant nutrients sources and classification

IoT-Based Smart Farming Application for Sustainable Agriculture 575
Plants require all these nutrients in right quantity at right time during cropping
either naturally from soil or from the use of fertilizers. Crop yield depends upon
various soil parameters such as humidity, temperature, moisture and amount of
nutrient present in the soil and pH value of the irrigated water, to get rich quality crops
farmers need to be aware of the soil nutrients and optimum dosage fertilizers. The
existing method of laboratory soil testing incorporates gathering soil samples from
different spots of field and then forwards the samples to the laboratory for testing
which is tedious task and also consumes more time and cost more money.
To automate the testing process and to promote digital agriculture, a part of Digital
India campaign launched by the Government of India, a new innovative technology
called Internet of things (IOT) is used. IoT is a system of interrelated computing
devices which are provided with unique identifiers (UIDs) and the ability to transfer
data over a network without requiring human-to-human interaction. The adoption
of IoT solutions for agriculture serves as the better platform for the sustainable
agriculture. IoT uses digital sensors coupled with the microcontroller to read the
nutrients of the soil and water. This paper proposes a system to monitor the humidity,
moisture content and major nutrients [NPK] of soil and pH value of the irrigated water
using IoT-enabled sensor for sustainable agriculture to get better crop yield.
Even though farmer is rich enough to serve the food for every individual of the
society but poor himself due to improper pricing for their cultivated crop due to
the involvement of the mediator in marketing the agriculture products. To overcome
this issue, we propose a smart contract farming service which involves production
and marketing of agriculture yield based on the agreement between buyer and farm
producers. To implement this service, we establish a communication link between the
farmers who are producers and Agricultural Produce Market Committee (APMC)
who are buyers in order to attain pricing uncertainty and marketing of the yields
by avoiding the middleman. The registered farmers in the mobile application get
the facility to share the soil and water nutrients reports to the APMC authorities.
Further based on the reports, the APMC authorities can offer an agreement proposal
to the farmers which involves suitable crop and appropriate fertilizer suggestion
based on soil analysis, and the purchasing price of yield is fixed before farmers
start cultivating crops and provide insurance for their yield in case of any natural
calamities occur in future for the farmers to provide tension-free farming to promote
smart contract farming and direct market facility. The registered farmers can also
utilize the other facilitates offered by the APMC authorities in full or partial extent
to their lands and make their efforts in better quality yields without worries for
the marketing and profit/pricing of their yields. Thus, the proposed work helps the
farmers in selecting the suitable crop for their agriculture land based on soil and
water nutrients, funding/insurance to the crops and direct marketing facility without
middleman.
The rest of the paper is organized as follows: Sect. 2 briefs the latest advances in
smart farming and sustainable agriculture. Section 3 describes the proposed model
of smart farming for sustainable agriculture using water and soil nutrients. Section 4
provides the experimental results and analysis of the proposed method. Finally, Sect. 5
concludes the work and enlists the future directions.
2 Literature Survey
The role of information and communication technology plays a major role in strength-
ening the economy of the country through smart farming and sustainable agriculture
practices. Several works by the researchers are carried to achieve digital agriculture
by employing the technology in the traditional agriculture practices. In the recent
past years, a lot of effort has been made to determine the soil primary nutrients and
pH value of the irrigation water of the agriculture field to achieve sustainable agri-
culture. Different approaches are studied in literature survey and few of the research
work in agriculture applications are summarized below.
In present scenario of climate changing environment, water plays an important
role, the smart irrigation system is employed during drought condition, and smart
suction pump is used which will suck excess water during flooding condition is
explained in [1]. IoT-enabled solution for monitoring the soil conditions for effective
growth of plant is presented in [2], and necessary action is taken based on the IoT
information to cultivate land to create awareness about optimum usage of chemical
fertilizer and to avoid its negative impacts like damage to soil and groundwater
contamination on agriculture field. The authors in [3] proposed technology-based
solution to improvise agriculture yield and at the same time decrease the fertilizer
usage. To resolve this issue, authors in [4] proposed method to create a model with
various machine learning algorithms to precisely predict the crop yield and deliver the
end user with proper fertilizer recommendations ratio. To meet the food production
of growing population, authors in [5] presented the IoT-based technological solution
to determine the important parameter of soil such as humidity, moisture and primary
nutrients (NPK) of soil for predicting best crop to cultivate and also to recommend
the appropriate fertilizers to farmers with a aim of minimizing the farming cost;
whereas, authors in [6] used fuzzy-based model to predict the soil nutrients. The
primary nutrients of soil are determined using pH value of the soil. Authors in [7,
8] proposed a model for accurately predicting the crop yield and also assisted the
farmer about the usage of required fertilizer.
A soil is the foundation for any agriculture, and selection of crops to grow depends
on the type of soil present in agriculture field, the nutrients it contains, the level of
water it requires, etc. The literature review explores the requirement of the quality,
type and nutrients of the soil for sustainable agriculture. The smart farming by
adopting the state-of-the-art technologies like IoT, cloud computing, e-commerce
and mobile applications. Hence, a smart farming for sustainable agriculture using
water and soil nutrients is proposed in this work.
3 Proposed Model
To overcome the problems of the existing system and to lower the farming cost,
a smart farming-based android mobile application is proposed to provide several
service like IoT-based soil and water nutrient testing, promoting smart contract
farming service between buyers (APMC) and farm producers (farmers) by facili-
tating an interface to communicate to APMC to view the suitable crop to grow based
on uploaded soil analysis report, agreement for purchasing the yields and insur-
ance to crops. Hence, avoiding fragmentation of profit shares due to the presence
of middleman by promoting direct selling of their crops to APMC. And app also
provides the farmers about the awareness of different government funding agency
and usage of ICT-enabled tools and technology to encourage e-agriculture practice
for sustainable agriculture.
A proposed mobile application is developed to consolidate all the automated
functionalities of farmers and APMC officials to ease farming and selling process
of crops. The mobile app is developed in multilingual to support different parts of
farmers of the country. The farmers are having the choice of selecting app language.
The developed mobile app is used by both farmers and APMC officials, based on
login credentials and requirement of the services. Figure 2 shows the framework of
Fig. 2 Framework of proposed model

Fig. 3 Block diagram of soil NPK and pH of irrigation water analysis
several service proposed in method. Farmers are given with a choice of selecting a
nearest APMC shop to register through the app. APMC admin will verify and approve
the farmers to grant the access permission for the app. IoT-enabled color sensor is
used to determine nutrients present in the soil sample which is based on the principle
of absorption of color by solution. Figure 3 shows the hardware setup consists of
NPK and pH testing chemicals, color sensors, Arduino mega microcontroller and
an android application to test the soil nutrients. The system is embedded with the
three sensors, namely moisture sensor, pH sensor, color sensor, with the help of these
sensors, the required values are collected from the soil. The pH value of irrigation
water is measured using pH level sensor, the correct and considered value of pH of
irrigation water lies between 5.0 and 7.0, water with pH below 7.0 is acidic and above
7.0 is basic. The four LED lights present on the color sensor emit the light source,
and the soil sample is placed in such a way that it absorbs the light from the LED
and remaining light is reflected back to sensor which is detected by the photodiode.
This photodiode converts light energy into current which is then converted to digital
display readings using current to frequency converter. The soil moisture sensor gives
water contained in the soil sample. Finally, the results after testing are sent to the
farmers mobile using GSM module. Thus, this soil data form a repository for further
agriculture operation such as recommending fertilizer and predicting the suitable
crop.
The app also provides the facility for farmers to share their soil test reports with
the APMC. Further APMC admin uses this report to predict suitable crop and recom-
mend the appropriate fertilizer to the respective farmers mobile along with agreement
proposal. The proposed work uses support vector classifier machine learning algo-
rithm to predict suitable crop. Thus, developed smart farming mobile app will help in
promoting smart contract farming service between buyer and producer to maintain
sustainable economy among the both party for sustainable agriculture. Along with
this, the app also provides information and description about the usage of new inno-
vative ICT-enabled tools for agriculture operations such as huge harvesting of crops
and automatic weeder. Video facility is provided for farmers to learn the working and
operating of these farm machineries, thus the developed mobile app creates aware-
ness about usage of various ICT-enabled tools and technology for making farmers
life better.
4 Experimental Result
The working of the proposed system of a smart farming for sustainable agriculture
using water and soil nutrients is divided into two steps; first step is to determine the
humidity, moisture and NPK value of the soil using IoT-enabled respective sensors.
And second step is to predict crop based on soil analysis done in the first step and
provide fertilizer suggestion to the farmer. The experimentation is carried out by
selecting twenty-five soil samples from different fields located in villages of two
districts of Karnataka state, namely Bagalkot and Vijayapur. IoT-enabled sensors are
used to determine the NPK value of the soil and water nutrients. Further the crop
suggestion and usage of fertilizer to the farmers from the growth of crop to yield
are described. Interfacing of farmer to APMC module and their associated activities
are tested. The functionality of each module and analysis of the experimentation are
enlisted in the following.
4.1 Soil Sampling and NPK Detection
Soil sampling is done by dividing field into homogenous units, soil is taken from
each unit, and collected samples are mixed thoroughly and removed foreign particles
like roots and stones. Soil and water are mixed and leave the mixture for soil to get
settle down. After that solution is taken in the sampling tube, NPK testing chemicals
have been added to respective sampling tubes and mixed properly till the chemicals
get dissolved in it. Then sample has been left for 15 min for color to be changed. To
determine moisture and humidity of soil, moisture sensor and humidity sensor are
placed in the field, then hardware kit has been switched on so that all sensors have
been activated and started to sense data. Finally, all sensed data have been sent to
mobile phone for further processing as shown in Figs. 4 and 5.
4.2 Crop Prediction
Crop prediction is done using machine learning algorithm, and the prediction is
based on the past output data of the crop’s and involves two phases: training phase
and testing phase. Dataset collected consists of the moisture, humidity and NPK
values of soil, pH value of irrigation water and crop type. Nearly, I collected dataset
Fig. 4 NPK and pH testing

chemicals
Fig. 5 Readings from the

sensor
consisting of 200 records from Bagalkot and Vijapur district. Table 1 shows the
sample dataset value. Among that two-third of the dataset is used to train the model,
rest one-third of the dataset is used to test the models. Using support vector classifier
machine learning algorithm, we can predict the suitable crop for entered soil and
Table 1 Crop prediction dataset

N P K pH of irrigation water Humidity Moisture Crop type Fertilizer name
37 45 34 5.23 52 38 Maize Urea
12 56 36 6.09 52 45 Sugarcane DAP
35 22 24 5.0 54 46 Paddy 20–20
41 56 45 6.8 64 50 Wheat 28–28
13 19 40 6.9 52 33 Oil seeds 14–35–14
21 14 18 6.7 60 42 Millets 10–26–26
38 34 14 6.8 64 34 Pulses Urea
water parameter. Figures 6 and 7 show the result of crop and fertilizer suggestion
based on soil NPK.
Fig. 6 Prediction of crop by

APMC admin
Fig. 7 View crop prediction

by farmer
5 Conclusion
The proposed smart contract farming system acts as an agro-advisory mobile applica-
tion to the farmers in generating quick soil analysis report, and prediction of suitable
crop provides the crop and fertilizer suggestion based on the analysis of different
parameters of soil. Thus, the IoT-enabled sensor kit helps the farmers to get the soil
testing services at the doorstep. It also provides awareness and usage of tools and
technology. Hence, overall mobile app is used to automate the agricultural process.
Future Enhancement The project can be expanded to implement on large fields and skill-based
trainings on usage and monitoring of sensor module given to the farmers which helps them to ease
the farming practice. Further the developed smart farming-based mobile application can also be
expanded to implement the government initiative scheme for “One District One Crop” functionality
to augment crop diversity.
References
1. S. Gupta, V. Malhotra, V. Vashisht, Water irrigation and flood prevention using IOT, in 10th Inter-
national Conference on Cloud Computing, Data Science & Engineering, Indonesia-Banyuwangi
(2020)
2. M.S.D. Abhiram, J. Kuppili, N.A. Manga.: Smart farming system using IoT for efficient crop
growth, in IEEE International Students Conference on Electrical, Electronics and Computer
Science, Bhopal, India (2020), https://doi.org/10.1109/sceecs48394.2020.147
3. S.Y. Chaganti, P. Ainapur, M. Singh et al, Prediction based smart farming, in 2nd International
Conference of Computer and Informatics Engineering (IC2IE), Indonesia-Banyuwangi (2019),
https://doi.org/10.1109/ic2ie47452.2019.8940834
4. S. Bhanumathi, M. Vineeth, N. Rohit, Crop yield prediction and efficient use of fertilizers, in
International Conference on Communication and Signal Processing, Chennai, India (2019),
https://doi.org/10.1109/iccsp.2019.8698087
5. N. Kulkarni, A. Thakur, T. Rajwal, Smart soil nutrients analysis and prediction of the level
of nutrients using a bot, in 3rd International Conference on Recent Developments in Control,
Automation & Power Engineering (2019)
6. R. Singh, R. Kaur, A. Singla, Two phase fuzzy based prediction model to predict Soil nutrient,
in 5th IEEE International Conference on Parallel, Distributed and Grid Computing (PDGC),
Solan, India (2018)
7. S.M. Hatture, S. Naik, Agro Guardian: a smart agriculture framework for precision farming.
Modern Techniques for Agricultural Disease Management and Crop Yield Prediction, Vol. 8,
pp. 179–202, IGI Global, (2019)
8. S.M. Hatture, S. Naik, Krishi-Suvidha: a framework for smart agriculture, in First IEEE Inter-
national Conference on Advances in Information Technology (ICAIT), pp. 1–6, 25–27 July, AIT,
Chikkamagaluru, (2019)
Follow Me: A Human Following Robot
Using Wi-Fi Received Signal Strength
Indicator
V. Geetha, Sanket Salvi, Gurdeep Saini, Naveen Yadav,

and Rudra Pratap Singh Tomar
Abstract Modernization has brought up significant changes in the way humans used
to communicate with devices which aid in reducing strenuous work. Robotics has
made life simpler by automating some specific tasks, one such application of robotics
is explored in this paper. Human Following Robot (HFR) can be used to carry heavy
goods or simply follow its user to act as an interface between IoT connected devices.
In this paper, a Low-Cost Protoytpe of Human Follower Robot is proposed which
uses Received Signal Strength Indicator of Wi-Fi to establish its task.
Keywords Wi-Fi signal · RSSI value · Serial communication · Human following

robot
1 Introduction
For developing countries like India, where manual labor is one of the major employ-
ment option, the amount of money earned is comparatively very less considering
the amount of work carried out. This type of work majorly involves manual loading
of goods and transport. Similar, situation can also be observed in military, where
ammunition and food material have to be carried over longer distances via difficult
V. Geetha (B) · S. Salvi · G. Saini · N. Yadav · R. P. Singh Tomar

Department of Information Technology, National Institute of Technology,
Surathkal, Karnataka, India
e-mail: geethav@nitk.edu.in
S. Salvi
e-mail: sanketsalvi.salvi@gmail.com
G. Saini
e-mail: gsaini.192it006@nitk.edu.in
N. Yadav
e-mail: naveen.192it011@nitk.edu.in
R. P. Singh Tomar
e-mail: rudra.192it018@nitk.edu.in
586 V. Geetha et al.
terrains. Thus, there is need of providing some additional smart tools and mechanisms
which could aid transportation system or communication system without the need for
changing existing work structure. The growing industry of Internet of Things (IoT)
based devices can be considered as a possible solution which provided its users with
features such as remote control and monitoring of tasks by performing data analysis
and thereby supporting automation of the tasks. Thus, introduction of fleet of load
carrying robots could significantly offload the arduous tasks of manual labour and
soldiers and supports to develop humanity in the world. To achieve this, the technical
aspects such as optimum route selection, obstacle avoidance, etc, must be addressed
while designing the system. Addressing technical issues also adds cost to the manu-
facturing of the same. Thus, there is need to build robots that could accomplish same
task with lesser computation at lower cost. One logical solution to this problem is
to let human take the intelligent decision of route selection and obstacle avoidance
while the robot can simply follow the human.
2 Related Work
In this section, the work done by various authors in the domain of Human Following
Robot is highlighted.
In paper [1], a low cost, compact antenna for Wi-Fi is built to remotely control
a robot. This work has provided basic understanding of how the modern days Wi-
Fi modules work in principle. The necessary parameters need to be considered for
estimating directions in robots are explained well in this paper.
A vision based localization and navigation technique for robot is proposed in
paper [2] and A distributed system based mobile robot is proposed in paper [3].
Both these works, highlight the major task in human following robot which is to
distribute the compute intensive tasks such as localization, path planning and speech
recognition so that the delay in the response of the robot can be minimised as much
as possible.
In paper [4, 5], a SONAR based distance sensing approach is used for designing
a Human Following Robot. The proposed approach continuously finds the distance
between the sensor and human and tracks the change in the motion within its range.
Thus, by moving the robot in the direction of change in motion, the robot end up
following the human. However, this system fails when multiple moving objects are
present. To address this issue only one user can be tagged with the robot. Further
work on minimization of computation and use of readily available components to
achieve similar results with lower cost is essential.
In the paper [7], RFID is used for tracking the autonomous identity such as
human or robot within the indoor environment. The paper clearly show the difference
between locating the indoor identity using GPS and RFID, and results shows that
RFID worked well for close environment. Paper [8] proposed basic path planning
Follow Me: A Human Following Robot Using Wi-Fi … 587
algorithm so that the robot can follow the human and can avoid the obstruction by
choosing the new path automatically. The robot is tested in virtual robot experiment
platform and results were satisfactory.
3 Proposed System
In this section, the design and working of the proposed system is discussed. The main
motivation for our proposed work is to design a low cost HFR system. Therefore,
usage of existing easily available components for building HFR is a major task as
it directly effects on cost of the system. Compared to earlier work done where the
major power consumption depends on complex image processing and extra hard-
ware components such as additional sensors and Infrared camera [4], our proposed
work uses less computational power as power consumption mainly depends only for
calculation of left and right RSSI values and its comparison. Our proposed system
uses mobile phone of user and two Wi-Fi modules with onboard micro controller for
decision making.
Figure 1 shows the basic block diagram of the proposed system. The mobile phone
of the person to be followed is connected to Human Following Robot (HFR) via Wi-
Fi. The HFR subsystem has components such as Left Wi-Fi Module and Right Wi-Fi
Module which is connected to Mobile Phone at one end and Micro-controller on the
other end. The micro-controller controls motor movement via motor driver.
Fig. 1 Block diagram of proposed system

Fig. 2 Basic working diagram of follow me robot
It was observed that, the received signal strength of Wi-Fi module was sensitive
to the change in distance between module and its connected Wi-Fi router/hotspot.
By leveraging to this fact, it was decided to use RSSI value for designing HFR. A
pair Wi-Fi connected microcontroller can be used to achieve this task. Once the pair
of Wi-Fi module is connected to the desired user’s mobile hotspot, the controller will
decide upon the direction and distance between the user based on the observed RSSI
value. The concept is explained in detail as shown in Fig. 2.
Figure 2 highlights the basic working flow of our proposed system. As the HFR is
powered, the Wi-Fi modules gets connected to a pre-defined Mobile Wi-Fi Hotspot
which is presumably of the user behind which the HFR has to follow.
These, Wi-Fi modules are placed over the HFR in such a way that both are on the
extremes of each side. Once the HFR Wi-Fi modules are connected, the proposed
system continuously calls for checking the RSSI value at respective modules. These
values change based on the distance between Hotspot hosting Mobile Device and the
HFR. The Microcontroller connected to HFR Wi-Fi modules continuously monitors
this change and based on the coded logic it makes appropriate decisions. The micro
controller is also connected to a motor driver which is used to control speed and
direction of rotation of connected wheels. Based on the programmed logic, the wheels
will be rotated. Thus, if the RSSI value of one side of HFR Wi-Fi module is more
compared to other, it implies that the Mobile phone is closer to one side, thus the other
side motor will move fast in order to balance the RSSI at both Wi-Fi modules. Also,
if the RSSI value at both HFR Wi-Fi modules is equal but less than the threshold,
threshold is used to measure the distance from the human if it is greater the limit
then HFR will stop then, the HFR will move forward to achieve the desired RSSI
value. Same principle is used when the mobile device gets too close to HFR to move
it backwards.
4 Requirement Analysis
Table 1 highlights the hardware and software requirement for the implementation of
our proposed system.
Table 1 Hardware and software requirements

Component name Usage
Arduino UNO An Opensource programmable board for Atmega microcontroller
family. This The board provides set of analog and digital pins which
can be connected to take input from or give output to other devices
NodeMCU esp8266 It is esp8266 family programmable board with onboard Wi-Fi
capability. We have used two nodemcu to calculate the RSSI value at
the right side and left side and after calculating these values, node mcu
send it to the arduino using serial communication
L298D Motor Driver L298D is a 16-pin Integrated Circuit allows the DC motor to drive in
both the direction according to our requirement. It handles the
movement of the 2 DC motors at a time irrespective of the direction.
Arduino supply the signal to the L298D then it send the power to the
DC motor according to signal
DC Motor 12 V DC Motor are used to rotate the wheels of the robot car whose
speed is 300 RPM and rated torque value is 1.2 Kg-cm, four such type
of motors are used to drive the HFR
Arduino IDE A programming tool for Arduino and NodeMCU
Fig. 3 Circuit diagram of human following robot
5 Implementation
The section highlights the implementation of proposed system along with the detailed
discussion on circuit diagram, connection details of components and algorithmic
implementation of the system. Figure 3 shows the connections of overall components
involved in building proposed system. Here, the NodeMCUs are connected with
Arduino UNO via software serial ports. This enables Arduino to read the RSSI
values printed at serial ports of each NodeMCU. A L298D motor driver is externally
powered by 12 V battery and it is controlled by digital PWM pins of Arduino. This
allows to control the power supplied from battery to motors and thus control speed
of HFR. Arduino can vary the direction of rotation of these wheels. NodeMCUs and
Arduino are also externally powered by battery.
Algorithm 1 Algorithm used in Arduino for Direction Detection

1: countL = RSSI value received from left nodemcu
2: countR = RSSI value received from right nodemcu
3: if(countL<countR)
4: then rotate right()
5: if(countL>countR)
6: then rotate left()
7: if(countL == countR and countL≤ threshold)
8: then move forward();
9: if(countR≥threshold and countL≥threshold )
10: then stop();
In Algorithm 1, countR is RSSI value detected by right nodemcu and countL is

RSSI value detected by left nodemcu. The motion of the human is detected based on
RSSI values of and comparing the same with threshold value. If the both the RSSI
value is greater than certain threshold value, then HFR will stop. High RSSI value
indicate that human is very near to robot. This algorithm is plays major role in the
HFR as it performs the major decisions of HFR such as decision of left turn, right
turn, forward move and backward move.
Algorithm 2 Algorithm used in left nodemcu

1: connect to wifi using password and signal name
2: countL = WiFi.RSSI();
3: StaticJsonBuffer<1000> jsonBuffer;
4: JsonObject& rootL=jsonBuffer.createObject();
5: rootL["countL"] = countL;
6: send this value to arduino using serial communication
In Algorithm 2, our objective is to send the RSSI value from Nodemcu to the
microcontroller (Arduino). From the left nodemcu, we get the RSSI value. To send
RSSI value to Arduino, a JSON object called as rootL is created, which takes value
from the serial monitor and stores in the object by using it’s countL attribute. The
algorithm for right nodecu is similar to the algorithm given in Algorithm 2.
The performance of the proposed system is tested using a manually marked surface
by constructing different region of 15 cm square. The RSSI values at both side of HFR
is noted for the manually marked surface region. Table 2 shows the observed values
in each of the manually marked regions. HFR is placed at origin and Mobile Phone
is kept in different regions to observe the RSSI values. In Table 2, L and R denotes
the left and right side of the HFR respectively. The first column denote the forward
straight distance in centimeter from the robot and we are measuring the RSSI value
after the 15 cm distance gap. We observed that as the distance increases, the RSSI
value decreasing generally. But we have also observed that in some cases where the
row and column of Table 2 are (2, 4), (2, 5), (3, 4), (3, 5), (4, 4), (4, 5), (5, 4), and
(5, 5), the values are getting opposite which means that left RSSI should be greater
than the right RSSI, but it is opposite because of increase distance of human and
weak wifi signal strength. For example lets take (5, 4) = (fifth row, fourth column)
it means the box which present at 70 cm straight and 15 cm Left and at (5, 4) values
are LN = −60, RN = −52.
In Fig. 4, the RSSI values of Left node and same for right node with respect
to the displacement of the hot-spot device as describe in Table 2 is plotted. The
displacement is calculated as d = x 2 + y 2 where d is the displacement, x and y
are the coordinates of hot-spot device considering the centre of HFR as origin. The
values of x and y can be obtained from Table 2. Consider the coordinates (4, 1) the
Table 2 RSSI value detected

Distance in cm RSSI value observed
−28 cm R −15 cm R 0 15 cm L 28 cm L
15 LN = −44, LN = −43, LN = −34, LN = −35, LN = −39,
RN = −31 RN = −28 RN = −34 RN = −41 RN = −48
30 LN = −44, LN = −41, LN = −39, LN = −51, LN = −50,
RN = −35 RN = −23 RN = −40 RN = −48 RN = −44
45 LN = −44, LN = −41, LN = −39, LN = −51, LN = −50,
RN = −35 RN = −23 RN = −40 RN = −48 RN = −44
60 LN = −44, LN = −44, LN = −42, LN = −57, LN = −43,
RN = −40 RN = −42 RN = −42 RN = −53 RN = −42
70 LN = −68, LN = −50, LN = −46, LN = −60, LN = −51,
RN = −55 RN = −52 RN = −44 RN = −52 RN = −47
Fig. 4 Bar graph for the RSSI values with respect to displacement of hot-spot device
x coordinate is −28 cm and y coordinate is 60 cm. So, the displacement calculated

by the equation is 66.21 cm and the clock is in the left side of origin so we are
considering it as negative and RSSI values of left node is −44 and same for right
node is −40. The right node is colored as orange color and left node is colored as
blue as mentioned in the graph as well. The hardware components of the prototype
design are as shown in Fig. 5.
7 Conclusion
The proposed HFR using the low cost components is implemented and its working
is demonstrated successfully. Through experiment we have observed that HFR suc-
Fig. 5 Side and top view of HFR
cessfully followed the human up to a significant mark. Proposed HFR follows the
direction in which human/mobile device is moving, which means the HFR moves
in the direction of hot-spot device. We have also observed that as the distance of
the mobile hot spot increases from the HFR. HFR fails to detect the correct RSSI
values on the left and right nodemcu. This is majorly due to increased interference
and weak wifi signal strength. The limitation in proposed system is that the weak
strength of Wi-Fi signas results in smaller operational distance. Hence only RSSI
based solution is not enough and the work can be extended by using filters for more
precise and stable values.
References
1. M.B. Perotoni, B.E. Garibello, S.E. Barbin, An IEEE802.11 Low Cost Planner Antenna for a
Mobile Robot (2006)
2. H.-M. Gross, H.-J. Boehme, T. Wilhelm, Contribution to vision based localization, tracking
and navigation methods for an interactive mobile service-robot, in Proceedings of the IEEE
International Conference on Systems, Man, and Cybernetics (2001), pp. 672–677
3. A.A. Loukianov, H. Kimura, M. Sugisaka, Implementing distributed control system for intelli-
gent mobile robot, in 8th International Symposium on Artificial Life and Robotics, Oita, Japan,
24–26 Jan 2003
4. Q.K. Dang, Y.S. Suh, Human following robot using Infrared Camera, 2011 11th International
Conference on Control, Automation and Systems, BGyeonggi-do, South Korea (2011)
5. W. Peng, J. Wang, W. Chen, Tracking control of human-following robot with sonar sensors,
Intelligent Autonomous Systems 14. IAS. Advances in Intelligent Systems and Computing, vol.
531 (Springer, Cham, 2016)
6. C.-W. Chen, S.-P. Tseng, Y.-T. Hsu, J.-F. Wang, Design and implementation of human follow-
ing for separable omnidirectional mobile system of smart home robot, in 2017 International
Conference on Orange Technologies (ICOT) (IEEE, Singapore, Singapore, 2017)
7. R. Tesoriero, J.A. Gallud, M.D. Lozano, V.M.R. Penichet, Tracking autonomous entities using
RFID technology. IEEE Trans. Consum. Electron. 55(2) (2009)
8. S. Mirai, M. Nobuto, Simulation of a human following robot with object avoidance function,
in 12th South East Asian Technical University Consortium, 2018
9. S. Kaoru, Human characterization by a following robot using a depth sensor (I), in IEEE/SICE
International Symposium on System Integration, 2017
Implementation of Arithmetic Unit
for RNS Using 2n + 3 as Base
Nagaraj R. Aiholli , Uday V. Wali , and Rashmi Rachh
Abstract Choice of a good modulo base number is an important factor in strength

of Modular arithmetic used in Cryptography, Digital signal processing (DSP) and
in other areas and therefore needs special attention. Existing literature abundantly
describes operations with 2n ± 1 as modulo base. The numbers 2n ± 3 are also
prime and offer a set of new base numbers for modular arithmetic. In this paper,
implementation of modulo 2n + 3 addition based on shifting and distributive law
are described. Modulo operation is performed till the final residue is less than 2n +
3. The results are compared with 2n − 3 architectures described in recent literature.
The proposed method avoids mapping of bits reported in the earlier work and hence
provides an improved architecture for implementation of modular operations.
Keywords Modulo architecture · Residue number system · Squarer ·

Cryptography
1 Introduction
Modulo arithmetic systems using various base numbers have been discussed in recent
literature. These systems find applications in sophisticated and critical applications
like cryptography, signal processing and neural computations [1]. Design of hardware
for modular arithmetic depends on a prime number to be used as a base. Various
authors have reported implementations using base numbers like 2n , 2n − 1 and 2n +
1 that are prime e.g., Szabo and Tanaka [2], Soderstrand et al. [3], Ananda Mohan
N. R. Aiholli (B)
KLE College of Engineering and Technology, Belagavi 590008, India
e-mail: nagarajaiholli1978@gmail.com
U. V. Wali
S.G. Balekundri Institute of Technology, Belagavi 590010, India
e-mail: udaywali@gmail.com
R. Rachh
VTU Post Graduate Studies, Belagavi 590018, India
e-mail: rashmirachh@gmail.com
596 N. R. Aiholli et al.
[4] and Omondi [5] among others. Few authors Ananda Mohan [6] and Huihua et al.
[7] have worked on another possible set of modulo bases like 2n − 3 and 2n +
3 but have received much less attention due to the complexity of implementation.
Some such attempts have also been reported by Sheu et al. [8], Muralidharan and
Chang [9], and others. In particular, the forward and reverse converters that transform
given numbers between residue number system and binary number system have been
investigated by Adamdis and Vergos [10], Spyrou et al. [11], Strollo and Caro [12].
Majority of the existing literature by authors such as Govekar et al. [13] and
Sinthura et al. [14] and Sharma [15] describes implementation of multipliers.
However, use of squarers introduced by authors Aiholli et al. [16], Muralidharan
et al. [17] and Vergos [18] can speed up some common operations and deserves
specific attention. In this paper, a novel implementation of squarer based on distribu-
tive property of modulo multiplication has been reported. The work describes a
generic method to implement modulo 2n + 3 arithmetic. Existing literature uses a
conventional method of mapping of bits. The method proposed in this work is based
on distributive property of modulo multiplication and avoids the mapping of bits.
This novel approach leads to a better performance. The results obtained have been
compared with the existing works.
Representing 2n + 3 will require n + 1 bits. Therefore, using n-bit hardware allows
implementation of a 2n−1 + 3 modulo architecture. Implementing any modulo 2n is
easy as the higher order bits can be simply truncated. The implementation becomes
harder when we consider numbers other than 2n . For modulo bases that are close to
2n , it is interesting to see how the operations can be implemented in terms of a 2n
modulo operation. It is easy to observe that the difference between the two bases will
also behave like a modulo number, gradually increasing and dropping to zero after
few cycles. Implementation schemes can exploit this particular feature.
Design of a modulo 2n + 3 squarer has been presented in this paper. Design
of various modules is discussed in Sect. 2. The approach suggested in this work
is general enough to extend it to other prime number systems as well. Section 3
discusses an implementation of squarers based on 2n + 3 base. Section 4 concludes
the paper.
2 Novel Squarer for Modulus 2n + 3
Multiplication of two n-bit numbers yields a result of 2n bits with a maximum value
of 22n −2(2n −1). Several authors map the product bits beyond n bits by folding the
result, n bit wide LSB as normal and then folding the higher bits to match the modulo
base. Matutino [19] models his work around the equation
(A × B)mod(2n + 3)=|P0 − 3P1 + 9P2|mod 2n + 3 (1)
where P0 represents least significant n-bits (LSB) and P1 are remaining most signif-
icant bits (MSB) of the product. The partial product 3P1 is implemented as sum of
Implementation of Arithmetic Unit for RNS Using 2n + 3 as Base 597
P1 and 2P1, by shifting one bit to left. The P1 bits are mapped to least n-bits, once
normal (to bits n−1, n−2, … 0) and then with one bit left shift. This is continued
until the final product is of n bits. P2 is a single bit vector.
Generally, the approach taken to compute a modulo arithmetic is to use the
distributive property: if y = a + b, then y mod n =(a + b) mod n = ((a mod n) +
(b mod n)) mod n, which allows successive reduction in size of number. To facili-
tate easy implementation, a is taken a multiple of 2k which allows easy bifurcation
of the original number. Further, such implementation should be faster than a direct
computation or remainder using an integer division.
In a recent paper implementing modulo multiplication (squaring), such mapped
partial products have been used to compute the final modulo product [20]. These
mapped partial products are converted into their corresponding decimal number.
Then the decimal number modulo 2n − 3 is performed. This is equivalent to the
result obtained with conventional method described in literature.
This paper provides an improvement over the method presented in [20]. It is
not required to perform mapping of bits for the proposed implementation. For
implementing the modulo 2n +3 squarer operation, following observations can be
made.
(1) As 2n +3 > 2n −1, we need n + 1 bits to represent the input number.
(2) Adding 2’s complement of the number, which implements a subtraction, may
be used to implement the modulo operation.
(3) Partial products in computation of squared number are generated by shifting
the number to the left, multiplying the number by a factor of 2. This allows
the difference between mod 2n and mod 2n + 3 to be precomputed. This is the
essential step in reducing the computation of modulo operation.
Figure 1 illustrates the third point for n = 4. For every multiplication, the difference
between the mod 2n and 2n +3 can be precomputed. For the first, second and third
partial products, the differences are 3, 6, and 12 respectively. For n > 4, other partial
products may be similarly computed. Implementing a 2n modulo is a truncation
operation. Therefore, the required modulo 2n +3 operation can be implemented by a
truncation followed by an addition. This step is the essence of fast modulo squaring
operation. It is easy to see that the method may also be used for multiplication.
Example 1 Considering 3 bit input number the architecture is illustrated as shown

below.
Fig. 1 Range for modulo 2n + 3 operation with n = 4 bits

Input number = (a2 a1 a0)2 = (5)10 = (1 0 1)2

We need to find square of this number
a2 a1 a0 = 0 1 0 1
a2 a1 a0 X 0 1 0 1
As per the observations made n + 1 bits (3 + 1=4) are used to represent the input
number. Hence input number is written as (0 1 0 1)2 = (5)10 , i.e.,
0 a2 a1 a0 0 1 0 1
x 0 a2 a1 a0 X 0 1 0 1
a2a0 a1a0 a0 = x1 1 0 1 = 5
a2a1 a1 a1a0 + = x2 0 0 0 + = 0
a2 a2a1 a2a0 + + x3 1 0 1 + + = 20
0 0 0 + + + x4 0 0 0 + + + = 0
x1 Mod 11 = y1 5 Mod 11 = 5
x2 Mod 11 = y2 0 Mod 11 = 0
x3 Mod 11 = y3 20 Mod 11 = 9
x4 Mod 11 = y4 0 Mod 11 = 0
(y1 + y2 + y3 + y4) mod 11 = z (5 + 0 + 9 + 0)mod 11 =

14 mod 11 = 3
Square of input number = 52 = 25, and base number is 23 + 3 = 11. Therefore,
the result is 25 mod 11 = 3
First partial product row represents the input number. Second row is bits with all
zeros shifted by 1 bit left. Third row is input number shifted by two bits left. Fourth
row is bits with all zeros left shifted by three bits. As per the distributive law of
modular operation
a(b + c + d) Mod 11 = (a × b) Mod 11

+ (a × c) Mod 11
(a × d) Mod 11
x Mod 11
It is important to note that if the terms b, c and d are multiples of 2k , 0 ≤ k < n,

i.e., they can be chosen to be the partial products of the bits in order from right to left,
it is possible to pre-compute part of the modulo operation, thereby, speeding up the
modulo computation. The modulo operation is performed by successive subtraction.
Two’s complement of 0 × 01011 is 0 × 10101. Therefore we can add the numbers
as follows:
5×4 1 0 1 0 0 20
mod 11 + 1 0 1 0 1 – 11
= 0 1 0 0 1 = 9
+5 × 1 + 0 0 1 0 1 + 5
= 0 1 1 1 0 = 14
mod 11 + 1 0 1 0 1 – 11
0 0 0 1 1 = 3 (x mod 11)
Example 2 For the second example, let us consider a 4 bit squarer using modulo 2n
+ 3 base
Input number = (0 c3 c2 c1 c0)2 = (c)10 . Square of the input number (c)2 = (d)10 .
Consider Input number = (13)10 = (0 1 1 0 1)2
0 1 1 0 1
X 0 1 1 0 1
0 1 1 0 1 = 13 mod 19
0 0 0 0 0 = 0 mod 19
0 1 1 0 1 0 0 = 52 mod 19
0 1 1 0 1 0 0 0 = 104 mod 19
1. 13 mod 19 = 13
2. 0 mod 19 = 0
3. 52 mod 19. This is performed by first obtaining 52 mod 32 as 52 falls in the range
of 32 to 64. 52 mod 32 = 20. As 32 = 25 , it may be implemented by truncating
the left side bits beyond the fifth bit.
4. Then the difference (2 × 2n + 3 − 2 × 2n ) is obtained from Table 1 as 6. Subtract
(add 2’s complement) from the previous number: 20 − 6 = 14
5. Next partial product 104 mod 19 can be computed as 104 mod 64 = 40, followed
by subtracting 12 from the result. Hence 40 − 12 = 28. As 28 > 19, 28 mod 19
is performed as step 3) 28 mod 16 = 12, step 4) 12 − 3 = 9.
6. Exit loop when partial product is less than 2n + 3.
7. All the partial products are added. If it is more than 2n + 3, the procedure of
truncation followed by subtraction may be used as one final step.
Table 1 Difference between modulo 2n and 2n + 3 for the partial products, n = 4

k 0 1 2 3
Range 2n+k 16 32 64 128
2n+k + 3 19 38 76 152
Difference 3 6 12 24
13 mod 19 = 13
0 mod 19 = 0
52 mod 19 = 14
104 mod 19 = 9
+ 36
Same steps 3 to 5 may be repeated to obtain final result.

36 mod 19 = (26 + 10) mod 19
= {(26 mod 19) + (10 mod 19)}mod 19.
26 mod 16 = 10, followed by subtraction 10 − 3 = 7
∴ 36 mod 19 = 7 + 10 mod 19 = 17
Note that when any partial product is larger than 2n + 3, then the procedure
of truncation followed by subtracting the difference can be applied till final partial
product is obtained. Final result matches with the conventional squarer method. A
flowchart with the steps of algorithm is illustrated in Fig. 2.
3 Comparison with Earlier Architectures
In the previous paper [20], the authors have worked on modulo 2n − 3 squarer
architecture implementation for residue number system. Figure 3a illustrates the
partial product bits after multiplication. Bits beyond n bits are mapped once normal
and then with one bit left shift. Figure 3b shows the mapped partial product bits.
Initially the bits in the matrix are arranged as per the partial product bits obtained.
Bits in the subsequent mapped partial product matrix are rearranged such that there
is no vacant space left in between the consecutive bits as compared to the previous
matrix bits. Each row is converted into its equivalent decimal number expressed in
terms of bit vector as the method employs operation on bits rather than on integer.
Modulo reduction operation is performed on each row. Obtained bits are added with
a CSA (Carry save adder) tree. Top three rows are added in the first row and then
one row at a time. The column is stopped at that level, if no bits are remaining to
be added. CPA (Carry propagate adder) is used to add final sum and carry vectors.
The sum obtained is subjected to modulo reduction. Final value is equivalent to that
obtained from conventional approach.
As compared to the previous architecture, present architecture proposed is on
modulo 2n + 3 squarer implementation for residue number system using the same
simulation tool set as the previous architecture [20]. There is certainly comparable
difference between the two in terms of delay requirement and hardware with the
present one having an edge over that of the previous one [20]. This is illustrated in
the result section that follows.
Fig. 2 Flowchart of squarer
4 Hardware Simulation and Testing
Hardware can be constructed directly from the algorithm presented above in Sect. 2
(Fig. 4).
Xilinx ISE design suite 14.7 has been used as the design tool with Verilog program-
ming. Proposed algorithm is synthesized using the same tool as that for the previous
architecture of the author [20]. Xilinx FPGA viz. Virtex 6 XC6VLX75T in FF484
Fig. 3 Mapping the partial products and rearranging for modulo computation
Fig. 4 Block diagram showing steps used in proposed squarer design
package and -3 speed grade is the target implementation device under Xilinx 14.7
working environment. The simulator used is Integrated SIM (ISIM) for behavioral
verification. The speed measured in terms of delay and areas in terms of the number
of LUTs are considered as parameters of measurement respectively.
Simulation waveform is as shown in Fig. 5a. Multiplicand and multiplier values
used for simulation are:
multiPlicand[7:0] = 10101010 = 17010
multiPlier[7:0] = 10101010 = 17010
Expected result of the squarer is 170×170 = 28900 mod 259 = 151, which is
observed in Fig. 5a above as resf [8:0] = 10010111 = 151(decimal).
The RTL schematic obtained during synthesis is represented as shown in Fig. 5b;
(Tables 2, 3 and 4).
The device utilization summary report gives number of LUTs utilized, number of
LUT-FF pairs and IO blocks used. Offset out after for clock delay is 0.562 ns. As
compared to the delay of the proposed work for squarer, delay of the 8 bit squarer
of paper [20] is 22.188 ns and that of Jaberipur [21] is 6.5 ns. Jaberipur’s work is
on custom IC design. But, the proposed work is on FPGA architecture. Still results
of the work carried out in this paper are better than that of paper [20] and Jaberipur
[21] in terms of delay. The power estimated using Xpower analyzer is 1.293 W.
Fig. 5 a Simulation results of proposed algorithm b RTL schematic of 8 bit proposed squarer
Table 2 Device utilization

Logic utilization Used Available Utilization (%)
summary of the proposed
work of 8 bit squarer Number of slice 1055 46,560 2%
LUTs
Number of used 67 1109 6%
LUT-FF pairs
Number of bonded 26 240 10%
IOBs

Logic utilization Used Available Utilization
summary of the proposed
work of 12 bit squarer Number of slice LUTs 3066 46,560 6%
Number of used LUT-FF pairs 145 3155 4%
Number of bonded IOBs 38 240 15%

Logic utilization Used Available Utilization
summary of [20] of 8 bit
squarer Number of slice LUTs 484 46,560 1%
Number of used LUT-FF pairs 0 484 0%
Number of bonded IOBs 28 240 11%
5 Conclusion
Squarer implementation for modulo 2n + 3 is the main focus of this paper. Design
for 4, 8 bit and 12 bit units is demonstrated here, but is expected to also work for
higher bits. The future work may be to implement the design on Virtex 6 low power
and Virtex 7 FPGA’s. The main objective of this paper was to show that new method
is comparable to the conventional architectures in terms of delay requirement.
All the Verilog code used for simulation is available from github project https://
github.com/nagarajaiholli.
References
1. S. Mayannavar, U. Wali, Design of hardware accelerator for artificial neural networks

using multi-operand adder, in Proceedings of the International Conference on Information,
Communication and Computing Technology, ICICCT (2019)
2. N. Szabo, R. Tanaka, Residue arithmetic and its applications in computer technology, McGraw
Hill (1967)
3. M. Soderstrand, G. Jullien, W. Jenkins, F. Taylor, (eds.) Residue Number System Arithmetic:
Modern Applications in Digital Signal Processing IEEE Press (1986)
4. A.P. Mohan, Residue number systems: algorithms and architectures, Kluwer (2002)
5. A. Omondi, A. Premkumar, Residue number systems: theory and implementation, (Imperial
College Press, 2007)
6. P.A. Mohan, New Reverse converters for the moduli set {2n − 3, 2n + 1, 2n -1, 2n + 3}”, AEU,
2008, vol. 62, pp 643–658 (2008)
7. L. Huihua, L. Lei, Z. Wanting, High-speed and high-efficient modulo (2n − 3) multipliers, in
International Conference on Education, Management, Computer and Society (EMCS 2016)
8. M. Sheu, S. Lin, C. Chen, S. Yang, An efficient VLSI design for a residue to binary converter
for general Balance moduli (2n - 3, 2n -1, 2n + 1, 2n + 3), IEEE Trans. Circ. Syst. Express
Briefs 51, 52–55 (2004)
9. R. Muralidharan, C. Chang, C. Jong, A low complexity modulo 2n + 1 squarer design, in
Proceeding of IEEE Asia Pacific conference on Circuits and Systems (2008), pp. 1296–1299
10. D. Adamidis, H. Vergos, RNS multiplication/sum-of-squares units. IET computers Digit Tech
1, 38–48 (2007)
11. A. Spyrou, D. Bakalis, H. Vergos, Efficient architectures for modulo 2n − 1 squarers, in
Proceeding of IEEE International Conference on DSP 2009, (2009), pp. 1–6
12. A. Strollo, D. Caro, Booth Folding encoding for high performance squarer circuits. IEEE Trans
CAS, Part II 50, 250–254 (2003)
13. D. Govekar, A. Amonkar, Design and implementation of high speed modified booth multi-
plier using hybrid adder, in Proceedings of the IEEE International Conference on Computing
Methodologies and Communication, ICCMC (2017)
14. S. Sinthura, A. Begum, et al, Implementation and analysis of different 32-bit multipliers on
aspects of power, speed and area, in International Conference on Trends in Electronics and
Informatics (ICOEI 2018)
15. M. Sharma, R. Verma, Disposition (reduction) of (negative) partial product for radix 4 booth’s
algorithm, in World Congress on Information and Communication Technologies (2011)
16. N. Aiholli, R. Rachh, U. Wali, Design of arithmetic unit for RNS using 2n − 3 as base, in
Proceedings of the IEEE International Conference on Electrical, Electronics, Communication,
Computer Technologies and Optimization Techniques ICEECCOT (2018)
17. R. Muralidharan, C. Chang, Fixed and variable multi-modulus Squarer architectures for triple
moduli base of RNS, in Proceeding of IEEE ISCAS (2009), pp. 441–444
18. H. Vergos, C. Efstathiou, Efficient modulo 2n + 1 squarers, in Proceeding XXI Conference on
Design of Circuits and Integrated systems, DCIS (2006)
19. P. Matutino, R. Chaves, L. Sousa, Arithmetic units for RNS moduli {2n − 3} and {2n + 3}
operations, in 13th Euromicro conference on Digital System Design: Architectures, methods
and Tools (2010)
20. N. Aiholli, U. Wali, R. Rachh, Implementation of arithmetic unit for RNS using 2n − 3 as base.
Int. J. Innov. Technol Exploring Eng. 9, 254–258 (2019)
21. H. Ahmadifar, G. Jaberipur, Improved modulo—(2n + 3) multipliers, in International
Symposium on Computer Architecture and Digital Systems, (CADS), (2013), pp. 31–35
Hybridization of Artificial Bee Colony
Algorithm with Estimation of
Distribution Algorithm for Minimum
Weight Dominating Set Problem
Sima Shetgaonkar and Alok Singh
Abstract In this paper, we have proposed a hybrid approach combining Artificial

Bee Colony Algorithm (ABC) with Estimation of Distribution Algorithm (EDA) for
Minimum Weight Dominating Set (MWDS) problem. By combining ABC algorithm
with EDA, we are able to guide the search process toward more promising solutions in
comparison with any of the two constituent approaches. We have also used a solution
improvement method which further trims down the solution in order to minimize its
fitness value. Computational results on standard benchmark instances of the MWDS
problem show that the hybrid ABC-EDA performs as good as or better than existing
approaches in terms of solution quality on most of the instances.
Keywords Minimum Weight Dominating Set · Artificial Bee Colony Algorithm ·

Estimation of Distribution Algorithm · Swarm intelligence
1 Introduction
For an undirected graph, a dominating set is a subset of vertices such that each
vertex of the graph is either a member of the dominating set or connected by an
edge to a vertex in the dominating set. The vertices belonging to the dominating
set are called dominating vertices, and the rest of vertices is called dominated or
non-dominating vertices. Given an undirected graph, the Minimum Dominating Set
(MDS) problem seeks a dominating set of minimum cardinality on this graph. The
Minimum Weight Dominating Set (MWDS) problem is a generalization of MDS
problem. Given a vertex-weighted undirected graph, the MWDS problem seeks a
dominating set with minimum sum of weights of its constituent vertices. Both MDS
S. Shetgaonkar (B) · A. Singh

School of Computer and Information Sciences,
University of Hyderabad, Hyderabad 500046, India
e-mail: sima.shet@gmail.com
A. Singh
e-mail: alokcs@uohyd.ernet.in
608 S. Shetgaonkar and A. Singh
and MWDS problems are N P-hard [3]. Throughout this paper, we will use node
and vertex interchangeably.
Both MDS and MWDS find real-world applications in various domains from text
summarization to formation of a virtual backbone for routing in wireless network
[7]. Detection and prevention of attacks over a network can be done effectively using
the concept of dominating set [6]. Gateway placement in wireless mobile networks
is another application [1]. All these applications involve computing a MDS or a
MWDS depending on whether the underlying graph consists of homogeneous or
heterogeneous nodes.
A number of approximation algorithms, heuristics, and metaheuristics have been
proposed in the literature. Due to paucity of space, we will survey only a few of them.
Two hybrid approaches based on Ant Colony Optimization (ACO) and Genetic Algo-
rithm (GA) have been presented in [5]. An Artificial Bee Colony (ABC) algorithm is
presented in [4] which outperformed the hybrid approaches of [5]. An evolutionary
algorithm with guided mutation (EA/G) approach is presented in [2]. EA/G com-
bines Estimation of Distribution Algorithm (EDA) with a mutation operator [8]. This
EA/G approach also outperformed the two hybrid approaches of [5].
Motivated by the success of ABC and EA/G approaches for MWDS, we have
proposed a hybrid approach combining ABC algorithm with EDA and which we will
refer to as ABC-EDA. The ABC algorithm searches through the entire search space
for an optimal solution utilizing the location information of a population of solutions
and does not exploit the global statistical information about the search space which
is available to it as byproduct of search process. On the other hand, an EDA utilizes
the global statistical information about the search space to produce offspring instead
of utilizing the location information of solutions. Location information of a solution
uniquely identifies that solution in the search space. For MWDS, set of dominating
vertices uniquely identifies a solution. The ABC algorithm uses this information to
produce a neighboring solution. An EDA on the other hand makes use of a probability
model representing the distribution of promising solutions in the search space, and
offspring are produced by sampling this model. This probability model is updated
at each generation according to the population of solutions in that generation. By
combining ABC algorithm with an EDA, we can exploit location information of
Algorithm 1: Hybridized ABC-EDA Algorithm

1 Initialize Population () ;
2 Bestglobal = Calculate Best Solution amongst the Population () ;
3 Initialize EDA () ;
4 repeat
5 Employed Bee Phase ();
6 Onlooker Bee Phase ();
7 Update EDA ();
8 if Bestiteration is better than Bestglobal then
9 Bestglobal := Bestiteration ;
10 until termination condition is satisfied;
11 return Bestglobal ;
Hybridization of Artificial Bee Colony Algorithm … 609
population of solutions as well as global statistical information to produce a new

solution. This may guide the search process toward more promising solutions. Our
proposed approach can be considered as a cross between ABC approach & EA/G
approach and utilizes several features of both. Computational results on standard
benchmark instances of the problem show the effectiveness of ABC-EDA.
The remainder of this paper is organized as follows: Sect. 2 describes our ABC-
EDA approach. Computational results and their analysis will be presented in Sect. 3.
At the end, Sect. 4 outlines some concluding remarks.
Algorithm 2: Initialize Population
1 Population size N p
2 Solutions S0 , S1 , S2 , . . . , S N p −1
3 Fitness of solutions f 0 , f 1 , . . . , f N p −1
4 S0 ← ∅;
5 Repair (S0 );
6 Calculate fitness (S0 );
7 S1 ← Include all the vertices in the graph;
8 Improve (S1 );
9 Calculate Fitness (S1 );
10 for i := 2 to N p − 1 do
11 Init(Si )
Algorithm 3: Employed bee phase

1 n e is the number of employed bees.
2 for i := 0 to n e − 1 do
3 Ŝ ← Generate neighboring solution of Si using GM operator;
4 Repair and improve Ŝ;
5 if Ŝ is better than Si then
6 Si ← Ŝ;
7 else if Si has not changed over the last Ntrials iterations then
8 Scout Bee Phase (Si );
9 if Si is better than best then
10 best ← Si
Algorithm 4: Onlooker bee phase

1 n o is the number of onlooker bees.
2 for i := 1 to n o do
3 Pi ← Select and return Index S0 , S1 , S2 ...Sn o ;
4 Oi ← Generate neighboring solution of S Pi using GM operator;
5 Repair and improve Oi ;
6 if Oi is better than best then
7 best ← Oi ;
8 for i := 1 to n o do
9 if Oi is better than S Pi then
10 S Pi ← Oi
Algorithm 5: Scout Bee Phase (Si )

1 Randomly generate a number rand1 in [0, 1];
2 if rand1 < μ1 then
3 Init (Si )
4 else
5 Si = ∅ foreach v ∈ best do
6 Randomly generate a number rand2 in [0, 1] if rand2 < μ2 then
7 Si = Si ∪ v
8 Repair Si
2 Proposed Hybridized Approach ABC-EDA for MWDS

Problem
The proposed algorithm has the components of the ABC algorithm along with the
features of EDA which helps in directing the search toward the promising solutions.
The ABC searches through the search space for an optimal solution utilizing a popu-
lation of solutions. The EDA builds a probabilistic model using the global statistical
information and directs the search toward the more promising solutions. We have
also used a solution improvement method which further trims down the solution in
order to minimize its fitness value.
In Algorithm 1, we have explained the working of the ABC-EDA algorithm. The
algorithm starts by initializing a population of solutions. The solutions are repre-
sented using subset encoding and binary encoding both as such a policy facilitates
efficient implementation of various methods. The termination condition is the max-
imum number of iterations. In Algorithm 2, we explain how we generate the initial
population. Solution 0 and 1 are generated by using the repair and improve operators,
respectively. These operators will be described subsequently. These two solutions are
added since they may represent some of the best solutions in the solution space and
it would be ideal to explore the region around them for the optimal solution. The
Init function generates remaining solutions randomly by adding nodes to the solution
until it becomes a dominating set. The fitness of each solution is calculated by adding
the weight associated with each node present in the solution.
2.1 ABC Algorithm
The ABC algorithm has been used for a number of optimization problems. Its sim-
plicity and self organizing nature make it desirable while solving complex problems.
It basically consist of three phases, i.e., employed, onlooker, and scout bee phase.
Employed Bee Phase: In this phase, each employed bee is associated with a
solution. Each employed bee generates a neighboring solution using guided mutation
which is described subsequently in Sect. 2.4. If the neighboring solution is infeasible,
then it repairs it by adding nodes until it becomes a dominating set as described in
Sect. 2.3. It then improves it by eliminating nodes from the solution which do not
affect its feasibility as explained in Sect. 2.5. If the original solution does not improve
over a certain period, then it calls the scout bee phase as shown in Algorithm 3.
Onlooker Bee Phase: In Algorithm 4, each onlooker bee randomly selects a
solution associated with the employed bee and explores it even further. It generates a
neighboring solution using GM, which is then repaired and improved. If the solution
is better than the best solution, then it is replaced. There is a chance that the same
solution is selected by more than one onlooker bee.
Scout Bee Phase: It either initializes the solution completely randomly or takes
part of the best solution depending on some defined probability, i.e., μ1 . The param-
eter μ2 controls contribution of the best solution in the newly generated solution as
shown in Algorithm 5.
2.2 EDA
The EDA algorithm is a population-based algorithm. In this approach, it exploits

the global statistical information of the population provided by the search process
performed by the ABC algorithm.
Initialize EDA: This method initializes the probability vectors using the initial
population as shown in Algorithm 6. It is based on the idea that if a node is present
in many of the good solutions, then there is a high possibility that it might be present
in the optimal solution. Some percentage of best solutions are selected from the
population, and the probability vector is initialized.
Update EDA: After every iteration, this method is run in-order to update the
probability vectors based on the new population as shown in Algorithm 7.
Algorithm 6: Initialize EDA
Np
1 Select best solutions from the population;
d
2 Pi is the probability vector for node i;
3 for i := 1 to V do
4 Ni ← Count of best solutions containing node i;
5 if Ni = 0 then
Ni
6 Pi ← N p ;
(d )
7 else
8 Initialize Pi to some small value.
Algorithm 7: Update EDA

Np
1 Select best solutions from the population;
d
2 Pi is the probability vector for node i;
3 for i := 1 to V do
4 Ni ← Count of best solutions containing node i;
Ni
5 Pi ← (1 − λ)Pi + λ N p ;
(d )
2.3 Repair
This methods converts an infeasible solution into a feasible one. Our repair method is
same as the one used in [4]. It keeps on adding a new node until the solution becomes
a dominating set. The node to be added is determined by using one of the heuristics
mentioned below:
d(u)
v ← arg maxu∈V −Si (1)
w(u)
W (u)
w(u)
W (u) × d(u)
w(u)
where w(u) is the weight of the node u, d(u) is the degree of the node u, W (u) is
the sum of the weights of the non-dominated neighbors of u, and V is the complete
set of nodes. The heuristic to be applied is selected randomly. These heuristics are
introduced in [5].
2.4 Guided Mutation
Algorithm 8 shows how the neighboring solution is generated in the ABC algorithm.
The neighboring solutions are generated partly by taking elements from the existing
solution and partly by using the probability vector.
2.5 Improve
The improve algorithm is explained in Algorithm 9. If a solution remains a dominat-

ing set even after the elimination of a node from the solution, then the node is said to
be a redundant node. Our improve algorithm is an iterative algorithm where during
each iteration it determines a set of redundant nodes R and selects the node from R
to be eliminated from solution using the heuristic given below:
w(u)
v ← arg maxu∈R (4)
d(u)
where w(u) is the weight of the node u and d(u) is the degree of the node u. Thus,
it selects a node which has higher weight and lower connectivity. This heuristic is
introduced in [5] and is also used in [2, 4] for redundant node removal. The selected
node is eliminated, and the next iteration begins. This process is repeated as long as
the set of redundant nodes is non-empty.
Algorithm 8: Guided mutation (Si )
1 G ← ∅;
2 foreach v ∈ V do
4 if rand1 < β then
6 if rand2 < Pv then
7 G ← G ∪ {v};
8 else
9 if v ∈ Si then
10 G ← G ∪ {v};
11 return Si ;
Algorithm 9: Improve (Si )

1 R ← ∅;
2 foreach v ∈ Si do
3 if Si \ v = DS then
4 R ← R ∪ v;
5 while R = ∅ do
6 v ← argmaxu∈R w(u)
d(u)
;
7 Si ← Si \ v;
8 R ← ∅;
9 foreach v ∈ Si do
10 if Si \ v = DS then
11 R ← R ∪ v;
3 Computational Results
3.1 Dataset
We have used the same dataset as used in [4, 5] to test our approach. This dataset
consist of three types of graphs, viz Type I, Type II, and UDG. The number of nodes
varies from 50 to 1000 in these graphs. Each combination of nodes and edges has ten
instances. The solutions are computed for all the ten instances of a combination, and
then, the average value is reported for that combination. In Type I graphs, the weight
of every node is between [20, 70]. In Type II graphs, the weight of each node v is
between [1, d(v)2 ], where d(v) is the degree of the node v. UDG instances consists
of randomly distributed nodes in a space of 1000 × 1000 units. The transmission
range of every node is defined as 150 or 200 units, and there is an edge between two
nodes if they are within transmission range of one another.
3.2 Parameters
All parameter values were chosen empirically. The number of iterations is taken as
700 for Type I and II, and 90 for UDG; number of trials for the scout bee phase Ntrials
as 25 for Type I and II, and 30 for UDG; number of employed bees = 50, number of
onlooker bees = 20; number of solutions selected to generate the probability vector
d = 5; learning rate for probability vectors λ = 0.6; contribution of the probability
vectors β = 0.2; probability that the solution will be generated randomly or using
best solution μ1 as 0.5 for Type I and II, and 0.85 for UDG; probability which decides
the contribution of best solution while generating a solution in scout bee phase as
0.5 in Type I and II, and 0.25 in UDG.
3.3 Results
We have implemented our ABC-EDA algorithm in C and executed it on a Linux

system with 3GHz Intel Core 2 Duo CPU with 4 GB RAM. We have not included
HGA, ACO-LS, and ACO-PP-LS approaches of [5] in our comparisons since the
approaches EA/G, EA/G-IR proposed in [2] and ABC in [4] are superior. Tables 1,
2 and 3 present the results of our approach along with EA/G, EA/G-IR, and ABC
approaches. Best results are presented in bold font. ‘*’ and ‘**’ indicate the solution
value obtained by ABC-EDA is equal to ABC and better than ABC, respectively. Data
for EA/G, EA/G-IR, and ABC approaches are taken from their respective papers.
Tables 1, 2 and 3 were analyzed, and following observations were made based
on the results. For Type I and Type II instances, the ABC-EDA performs better than
EA/G-IR, EA/G, and ABC algorithm in most of the cases. It gives worse results
than the best known result for some cases by a very small margin. As the number
of nodes increase, we can see a pattern emerging. We see that the ABC-EDA gives
worse values as the density of the edges increases, but with very small margins.
On the other hand, the improvement provided by the approach for other instances
is much more for many of the cases, e.g., in Table 1, for instance 500 × 1000 the
solution improved from 3695.2 to 3652.1. The degree of improvement increases as
the number of nodes increases. It provides better results than the known best values
for all the instances with 800 nodes and 1000 nodes in case of Type I, and 1000 nodes
for Type II. The ABC-EDA obtained the known best results for all the instances of
UDG dataset except for one instance where it improves the known best result by
a small margin. For Type I and II instances, the ABC-EDA gives best results in
most of the cases at the expense of higher execution time. For UDG instances, the
computation time given by ABC-EDA is better than that of EA/G-IR except for one
instance. Except for the first two small instances, ABC-EDA outperforms the ABC
algorithm in terms of execution time.
In this paper, we have proposed a hybrid approach ABC-EDA combining ABC algo-
rithm with EDA for MWDS problem. We have compared the ABC-EDA approach
with EA/G, EA/G-IR, and ABC approaches available in the literature on standard
benchmark instances for the MWDS problem. For a few instances, ABC-EDA algo-
rithm produces slightly worse results than the best known values with some small
margins. But for most of the large instances, it provides better results than the known
best results. ABC-EDA performs better than ABC algorithm in most of the cases,
Table 1 Results for Type I instances

Nodes Edges EA/G-IR EA/G ABC ABC-EDA
Avg Time Avg Time Avg Time Avg Time
50 50 532.9 0.21 531.8 0.15 534 0 531.3** 0.41
50 100 371.5 0.23 371.3 0.15 371.2 0 370.9** 0.39
50 250 175.7 0.18 175.7 0.12 175.7 0 175.7* 0.33
50 500 94.9 0.16 94.9 0.13 94.9 0 94.9* 0.32
50 750 63.3 0.1 63.3 0.13 63.1 0 63.1* 0.31
50 1000 41.5 0.1 41.5 0.11 41.5 0 41.5* 0.23
100 100 1065.5 0.76 1065.4 0.45 1077.7 1.3 1061.4** 1.49
100 250 620 0.72 620.8 0.39 621.6 1 621.1** 1.29
100 500 355.9 0.6 357 0.33 356.4 0 355.8** 1.1
100 750 256.7 0.52 257.9 0.31 255.9 0 258 1.07
100 1000 203.6 0.49 204.8 0.32 203.6 0 205.2 1.01
100 2000 108.1 0.45 107.9 0.35 108.2 0 107.9** 0.93
150 150 1587.4 1.64 1594.6 0.81 1607.9 4.1 1583.8** 3.63
150 250 1224.5 1.71 1240.3 0.78 1231.2 2.8 1223.2** 3.43
150 500 755.3 1.45 762.3 0.67 752.1 1.8 747.5** 2.73
150 750 550.8 1.27 552.7 0.61 549.3 1 549.2** 2.35
150 1000 435.2 1.11 439.3 0.56 435.1 1 434.8** 2.21
150 2000 241.5 0.88 242.1 0.55 242.2 1 243.7 2.02
150 3000 168.1 0.82 168 0.59 167.8 1 169.3 2.08
200 250 1924.1 3.68 1951.7 1.26 1941.1 6.6 1916.7** 7.24
200 500 1251.3 3.3 1269.5 1.11 1246.9 4 1243.4** 6
200 750 927.3 2.78 935.3 1.01 923.7 3 924.2 5.26
200 1000 731.1 2.39 736.5 0.94 730.4 2.8 729** 4.56
200 2000 417 1.68 422.4 0.82 417.6 2 417.9 3.63
200 3000 294.7 1.42 297.8 0.81 294.4 2 298.1 3.48
250 250 2653.7 4.65 2695.5 1.85 2685.7 15.7 2647** 12.78
250 500 1835.3 4.66 1909.3 1.67 1836 8 1817.9** 11.33
250 750 1399.2 4.25 1436.9 1.52 1391.9 5.9 1382.6** 10.04
250 1000 1114.9 3.69 1137.3 1.42 1115.3 5 1107.7** 8.77
250 2000 637.5 2.65 657.4 1.22 630.5 3.8 628.6** 6.16
250 3000 456.3 2.16 459.5 1.15 454.9 3 457.1 5.81
250 5000 291.8 1.85 295.5 1.16 292.4 3 297.4 5.39
300 300 3213.7 8.73 3288.3 2.38 3240.7 23.6 3193** 20.53
300 500 2474.8 7.21 2574.9 2.18 2484.6 13.9 2461.3** 18.99
300 750 1896.3 6.48 1977.6 1.93 1901.4 10.1 1876** 16.91
300 1000 1531 5.7 1604.3 1.8 1523.4 8.1 1516.7** 15.16
300 2000 880.1 4.03 904.4 1.46 875.5 5.9 878.8 10.33
300 3000 638.2 3.27 645.6 1.34 635.3 5 636.8 8.55
300 5000 415.7 2.59 419.6 1.3 411 4.4 417.7 7.32
500 500 5380.1 37.52 5630 6.02 5480.1 78.1 5370.7** 82.83
500 1000 3695.2 26.36 3990.2 5.11 3707.6 38.7 3652.1** 71
500 2000 2264.3 25.54 2427.3 4.05 2266.1 22.8 2225** 53.8
(continued)
Table 1 (continued)
500 5000 1083.5 9.02 1128 3.06 1070.9 15.3 1073.5 29.41
500 10,000 606.8 6.08 625.7 2.77 596 13.1 605.8 21.37
800 1000 7792.2 129.94 8426.6 14.63 7907.3 208 7755.7** 310.5
800 2000 5160.7 102.09 5641.4 11.76 5193.2 95.1 5079.1** 269.85
800 5000 2561.9 53.01 2702.4 7.11 2548.6 50.9 2492.2** 140.94
800 10,000 1497 31.17 1559.6 5.12 1471.7 40.6 1440** 81.02
1000 1000 10771.7 249.82 11599.1 23.76 10992.4 589 10766.4** 585.73
1000 5000 3876.3 107.41 4163.5 13.03 3853.7 95.8 3757.4** 331.43
1000 10,000 2265.1 63.22 2365.8 9.02 2215.9 69.8 2180.8** 180.46
1000 15,000 1629.4 45.86 1682.6 7.44 1603.2 64 1573.6** 131.36
1000 20,000 1299.9 36.35 1333.7 6.57 1259.5 60.4 1242.1** 113.69
Table 2 Results for Type II instances

50 50 60.8 0.19 60.8 0.16 60.8 0 60.8* 0.42
50 100 90.3 0.27 90.3 0.15 90.3 0 90.3* 0.42
50 250 146.7 0.23 146.7 0.14 146.7 0 146.7* 0.34
50 500 179.9 0.09 179.9 0.14 179.9 0 179.9* 0.3
50 750 171.1 0.07 171.1 0.14 171.1 0 171.1* 0.27
50 1000 146.5 0.06 146.5 0.12 146.5 0 146.5* 0.23
100 100 123.5 0.86 123.7 0.46 124.4 2 123.5** 1.64
100 250 209.2 0.92 209.6 0.43 209.6 1.1 209.6* 1.42
100 500 305.7 0.78 305.7 0.38 305.8 1 305.7** 1.11
100 750 384.5 0.7 384.5 0.36 384.5 0.3 384.5* 0.99
100 1000 427.3 0.67 427.3 0.35 427.3 0 427.3* 0.93
100 2000 550.6 0.54 550.6 0.39 550.6 0 550.6* 0.87
150 150 184.5 1.85 185.4 0.82 185.9 6.1 184.5** 4.02
150 250 232.8 2.03 233.5 0.82 233.4 5 233.1** 4.01
150 500 349.7 1.95 351.3 0.75 349.5 3.3 349.5* 2.96
150 750 452.4 1.78 453.6 0.7 453.7 2.4 452.8** 2.51
150 1000 548.20 1.61 551 0.68 547.8 2 547.2** 2.29
150 2000 720.1 1.2 720.1 0.64 720.1 1 720.1* 1.82
150 3000 792.4 1.07 793.2 0.68 793.2 1 792.4** 1.74
200 250 272.3 4.38 274.8 1.31 273.5 11.9 271.7** 8.22
200 500 388.4 4.51 389 1.22 387.6 8.6 386.8** 6.76
200 750 497.2 4.18 501.7 1.14 498.5 6.5 497.3** 5.58
200 1000 598.2 3.89 599.2 1.08 599.3 5.2 597.3** 4.88
200 2000 885.8 2.78 885.8 0.98 885.5 3.5 884.6** 3.46
200 3000 1019.7 2.16 1023.8 0.95 1021.3 2.8 1022.4 2.98
(continued)
Table 2 (continued)
250 250 306.5 5.26 312.5 1.9 308.6 23.7 306.1** 14.54
250 500 441.6 6.1 448.5 1.8 442.6 18 440.8** 13.16
250 750 569.2 6.09 579.9 1.7 569.9 13.7 568.2** 11.35
250 1000 671.7 5.89 677.4 1.61 670.3 11.5 669.5** 9.77
250 2000 1010.3 4.23 1019.6 1.41 1010.4 6.9 1010.2** 6.45
250 3000 1250.6 3.5 1253.8 1.35 1251.3 5.6 1250.6** 5.25
250 5000 1464.2 2.59 1470.3 1.34 1464.7 4 1464.2** 4.4
300 300 370.5 9.01 380 2.41 373.5 41.1 369.9** 23.38
300 500 480 8.83 491 2.32 481.6 28.2 478.4** 21.7
300 750 613.8 7.57 631.5 2.17 617.6 26.1 613.9** 19.06
300 1000 742.2 8.96 767.1 2.06 743.6 20.2 738.1** 16.9
300 2000 1094.9 6.67 1107.7 1.72 1095.9 11.6 1095.2** 10.83
300 3000 1359.5 5.41 1371.9 1.58 1361.7 9.1 1361.3** 8.22
300 5000 1683.6 4.02 1686.2 1.5 1682.7 6.9 1682.7* 6.55
500 500 625.8 31.06 650.9 6.36 630.4 162.3 623.7** 98.25
500 1000 906 28.27 955.7 5.69 906.7 108.3 901.6** 83.89
500 2000 1376.7 23.41 1441.3 4.58 1383.6 66 1367.3** 59.56
500 5000 2340.3 17.36 2412.5 3.17 2337.9 31 2342.1 28.71
500 10,000 3216.4 10.8 3236.9 2.38 3211.5 21.1 3214.5 19.32
800 1000 1107.9 132.36 1187.6 14.58 1119.2 505 1104.5** 357.18
800 2000 1641.7 111.84 1797.5 12.49 1656.4 338.9 1636.1** 306.02
800 5000 2939.3 68.14 3200.4 8.64 2917.4 127.3 2888.1** 154.42
800 10,000 4155.1 40.15 4353.5 6.02 4121.3 70.9 4121.5 80.72
1000 1000 1240.8 202.08 1333.1 23.57 1256.2 1173.7 1239.4** 745.57
1000 5000 3222 132.94 3586.2 15.05 3240.7 303.6 3194.4** 387.94
1000 10,000 4798.6 84.82 5247 10.91 4781.2 150.6 4748.1** 190.3
1000 15,000 5958.1 61.64 6364.4 8.96 5934 111.9 5897.6** 132.03
1000 20,000 6775.8 59.2 7066.8 7.78 6729 75.3 6691.5** 107.38
but at the expense of higher computational time. ABC-EDA is better than all the
approaches considered in comparison with most of the instances in terms of the
quality of the solution.
As a future work, we intend to explore new tie-breaking rules for existing construc-
tive heuristics and altogether new constructive heuristics to improve the performance
of our approach. Similar approaches combining ABC algorithm with EDA can be
developed for other combinatorial optimization problems.
Table 3 Results for UDG instances

Nodes Range EA/G-IR EA/G ABC ABC-EDA
50 150 393.9 0.23 393.9 0.13 393.9 0.01 393.9* 0.05
50 200 247.8 0.2 247.8 0.12 247.8 0.01 247.8* 0.05
100 150 449.7 0.47 449.7 0.36 449.7 0.3 449.7* 0.19
100 200 216 0.47 216 0.31 216 0.22 216* 0.15
250 150 293.7 2.3 293.9 1.16 293.9 2.6 293.7** 0.98
250 200 118.7 1.77 118.7 1.03 119 2.1 118.7** 0.67
500 150 169.9 3.43 170.5 2.48 170.7 6.7 169.9** 3.36
500 200 68 2.34 68 2.51 68 5.9 68* 2.3
800 150 113.9 10.23 114.5 2.99 114 13.9 113.9** 8.4
800 200 40.8 7.86 40.9 5.42 40.8 12.56 40.8* 6.36
1000 150 121.1 13.76 121.5 6.71 121.4 28.4 121** 13.16
1000 200 45.9 5.42 46.8 3.45 46.1 23.5 45.9** 10.4
References
1. B. Aoun, R. Boutaba, Y. Iraqi, G. Kenward, Gateway placement optimization in wireless mesh

networks with QoS constraints. IEEE J. Sel. Areas Commun. 24(11), 2127–2136 (2006)
2. S.N. Chaurasia, A. Singh, A hybrid evolutionary algorithm with guided mutation for minimum
weight dominating set. Appl. Intell. 43(3), 512–529 (2015)
3. M.R. Garey, D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP-
Completeness (W. H. Freeman, New York, NY, 1979)
4. C.G. Nitash, A. Singh, An artificial bee colony algorithm for minimum weight dominating set,
in 2014 IEEE Symposium on Swarm Intelligence (SIS 2014) (IEEE, 2014), pp. 313–319
5. A. Potluri, A. Singh, Hybrid metaheuristic algorithms for minimum weight dominating set.
Appl. Soft Comput. 13(1), 76–88 (2013)
6. D. Subhadrabandhu, S. Sarkar, F. Anjum, Efficacy of misuse detection in ad hoc networks, in
2004 First Annual IEEE Communications Society Conference on Sensor and Ad Hoc Commu-
nications and Networks, 2004. IEEE SECON 2004 (IEEE, 2004), pp. 97–107
7. Y. Wang, W. Wang, X.-Y. Li, Distributed low-cost backbone formation for wireless ad hoc net-
works, in Proceedings of the 6th ACM International Symposium on Mobile Ad Hoc Networking
and Computing (ACM, 2005), pp. 2–13
8. Q. Zhang, J. Sun, E. Tsang, An evolutionary algorithm with guided mutation for the maximum
clique problem. IEEE Trans. Evol. Comput. 9(2), 192–200 (2005)
Privacy Preservation for Data Sharing
with Multi-party Access Control
in Cloud Computing: A Review Paper
Shubhangi Tanaji Khot, Amrita A. Manjrekar, and Rashmi J. Deshmukh
Abstract The cloud computing is the advancement to shared volume of information

through the network. There are lots of techniques that are used to providing security
for data in cloud. But current techniques are not as better related to the ciphertext.
So here, we propose information gathering, sharing and restrictive distribution plan
with multi-owner privacy preserving in cloud. Here, data owner can impart private
information to group of clients through cloud in secure manner with identity-based
broadcast encryption (IBBE) technique.
Keywords Data sharing · Multi-owner · Identity-based encryption · Privacy

conflict
1 Introduction
Cloud computing is the combination of various services like servers, storage,

databases, networking, over the Internet to giving fastest intelligence, flexible
resources and economies of scale. Cloud computing offers pervasive and on-demand
access to the fully featured applications.
Cloud computing is web-based processing so the data will be always available
to the client. It means cloud computing is done by storing local data of user into
cloud server so that user can access their own data whenever they needed [1]. Cloud
computing is brilliant condition for industry world regarding cost and giving services.
It is for quite a while envisioned vision of handling as an utility where data owners can
S. T. Khot (B) · A. A. Manjrekar · R. J. Deshmukh

Department of Technology, Computer Science and Technology, Shivaji University, Kolhapur,
Maharashtra, India
e-mail: shubhangikhot521@gmail.com
A. A. Manjrekar
e-mail: aam_tech@unishivaji.ac.in
R. J. Deshmukh
e-mail: rvm_tech@unishivaji.ac.in
622 S. T. Khot et al.
remotely store their data in a cloud to acknowledge on-demand amazing applications

and organizations from a pooled configurable registering assets.
The cloud computing is very famous because of its rich storage resources and
immediate access at anyplace [2]. It adds up to the advantages of preparing structure
and after that gives on-demand benefits over the Internet. Many popular organiza-
tions are as of now giving open cloud authority, for example, Amazon, Google and
Alibaba. These authorities enable individual clients and enterprise clients to upload
information (e.g. photographs, documents and videos) to cloud specialist as cloud
service provider (CSP), to get to the information whenever in anyplace and distribute
the information to others, especially with different types of searching of data over
the cloud [3] and [21]. To ensure the security of clients related to data, generally
cloud administrations accomplish access control by keeping up access control list
(ACL). Along these lines, clients can decide to either distribute their information to
anybody or award access rights simply to their affirmed individuals. However, the
security dangers have brought worries up in individuals, because of the information
is put away in plaintext structure by the CSP. When the information is presented on
the CSP, it is out of the data owner’s control [4].
However, combining privacy selection of data owner and co-owner is a big task,
so that privacy conflicts occur. Privacy conflict happens because of owner and co-
owner has opposite permissions of accessing the data, so anyone can easily access
that data. All existing systems were based on plaintext. To deal with this problem,
we propose system with multi-party access control mechanism.
The proposed system supports group sharing of secured data and conditional
distribution of data with multi-owner in cloud. This system will also support multi-
party access control mechanism. The system will be based o group sharing with
respect to ciphertext among various users and collect all the functionality of multi-
party access control.
In this paper, we discuss about brief literature review with some key parame-
ters. The review focus is on also different group sharing cryptographic schemes and
comparison between them.
2 Literature Survey
Cloud computing is a very secure storage which gives customers very easy and
simple access towards their data. It allows customers to distribute data to another
by functioning as collecting the data. It faces the challange as maintaing integrity
and furthermore privacy conflicts. Several schemes which ensure the integrity of the
shared group data with multi-owner cloud environment are reviewed here.
In the paper, creator proposed a few private information sharing strategies in
cloud computing [2]. In this paper, they characterized that information proprietor
sends private information to the cloud provider by giving record of customers, along
these lines just approved customers in the record can get decryption key and further
interpret the private data.
Privacy Preservation for Data Sharing with Multi-Party Access … 623
The author gives in-depth analysis about various security attacks and threats
related to the data integrity with proper solution [5]. They proposed an independent
and data secrecy scheme in cloud. This technique shows that attribute-based encryp-
tion with ciphertext policy (CP-ABE) scheme is less secure than RBAC-CPABE
because access authorization is only achieved by data itself.
The structure discussed as MONA in [1] accomplishes fine grained get to control
and which user has not priority to access the shared private information are come in
category of revoked users and will not have permission to share again.
The framework is exibited to accomplish secured and adaptable information bunch
transfer in [6]. In this paper, they utilize the method as quality and planned discharge
proxy re-encryption to guarantee that private information provider who satisfies the
entrance strategy of encoded information and just exchange it to different gatherings
in the wake of liberating time by dispensing a re-encryption key to cloud server.
Here creator utilized the paper in which framework permits a customer to send a
scrambled message to a recipent by utilizing cloud storage server [7]. The sender just
has to know the personality of the beneficiary; however, no other data (e.g. public key
or its certificate). The recipient needs to have two things so as to decode the ciphertext.
The primary thing is his/her secret key put away in the PC. The subsequent thing is
an extraordinary individual security gadget which interfaces with the PC.
In paper, they proposed the system which has two sides: one is cloud and other is
data owner side in cloud storage [8]. This is also combined with the different attacks
and provide resource consumption accounting.
In paper, creator presented new idea as privacy preserving and effective informa-
tion sharing help on cloud [9]. In this plan, creator joined two methods to be specific
CP-ABE and IBE where client opens key which is tight with client secret key to
acknowledge completely impact secure and protection saving. This is secure and
related with execution, and P2E has low overhead and high effectiveness.
Table 1 is based on the criteria of above summary of literature review where
it focused on some major papers. This table contains information which is used
to impart the private data to group of users through cloud with secure manner. It
Table 1 Summary of literature survey

Criteria Data security Data efficiency Data sharing Data integity
techniques
Yan [2] Yes Yes ABE and proxy No
re-encryption (PRE)
Lang [5] Yes Yes CP-ABE No
Boyang Wang [10] Yes No Multisignature Yes
Huang [6] Yes No ABE and time Yes
relevant—conditional
PRE (TR-CPRE)
Liu [1] No Yes Dynamic broadcast Yes
encryption (DBE)
also gives the information about various data sharing techniques over the years are
used. The main techniques illustrated in the following Table 1 are data security, data
efficiency, data sharing techniques and data integrity. Data security model is used
for secure private data, only authenticated users can show it, data efficiency model
is suitable for flexible data access, data sharing techniques model shows what are
the techniques used to communicate in the cloud over the years, and data integrity
model is checking mechanism on data owner side.
3 Different Cryptographic Scheme
Figure 1 shows different cryptographic schemes are developed over the time. The
first scheme is introduced in 2006, and here we analyze the ABE and IBBE scheme.
For sharing the data in the group, IBBE has access control mechanism to achieve
secure data sharing.
3.1 Attribute-Based Encryption (ABE) Scheme
Attribute-based encryption scheme is utilized for secure and fine-grained information

sharing. In this plan, both the client secret key and their separate ciphertext are relating
to set of qualities.This scheme uses access policies for encrypted data with decryption
key. This scheme focuses on one to many encryption methods. The real-time example:
Consider the College has attributes: {“Student”, “Teacher”, “HOD”, “Admin”}.
If the user A is a student, he has the attributes{“Teacher” ∧ “HOD”}. If admin
encrypts the file with attributes{“Colllege” ∧ (“Teacher” ∨ “HOD”)}, the student
will not be allowed to decrypt the data. If the attributes are {“Colllege” ∧ (“Teacher”
∨ “HOD” ∨ “Student”)}, the students are allowed to decrypt the data.
Fig. 1 Different 2010

cryptographic schemes
2008
Year
2006 Different
Cryptographic
2004 Schemes
3.2 Proxy Re-Encryption(PRE) Scheme
The PRE scheme is same as traditional symmetric and asymmetric encryption

scheme. The asymmetric PRE comes with two basic function as unidirectional
forwarding and bidirection forwarding.
(1) Unidirection forwarding: This is one-way communication. For example,
message can be re-encrypted from teacher to student but not the reverse.
(2) Bidirection forwarding: This is two-way communication where message can
be re-encrypted from teacher to student and vice versa.
This plan is utilized to accomplish secure information dissemination in distributed
computing with the assistance of re-encryption key related to the new recipients to
the Cloud Service Provider (CSP). But it disseminates a particular document.
3.3 Conditional Proxy Re-Encryption (C-PRE) Scheme
In the PRE scheme, data dissmination depends on sender’s and receiver’s public key.
So the new procedure is presented as C-PRE plot with bilinear paring. In C-PRE,
intermediary additionally needs to have the correct condition to move the ciphertext.
In this plan, information proprietor can be re-scrambled information with comparing
re-encryption key which fulfilling explicit condition.
3.4 Identity-Based Broadcast Encryption (IBBE) Scheme
In this plan, client can impart their scrambled information to different recipients one
after another and open key of beneficiary can be viewed as any substantial strings.
It will take minimal effort key administration to broadcasting information to explicit
collectors in cloud computing. This scheme is known as semi-structured because
ciphertext is not modifying uploaded by data owners.
3.5 Comparison of ABE and IBBE Algorithms
The below Table 2 shows the comparison between ABE and IBBE algorithm based
on following parameters as Setup, Key Generation, Encryption and Decryption. In
ABE scheme, client secret key and their separate ciphertext are relating to the set of
qualities. So this technique takes high cost for key management and access policies
as compared to IBBE technique. The Table 3 shows the notations of algorithm.
Table 2 Comparison of ABE and IBBE algorithm

S. No. Parameters Algorithms
ABE IBBE
1 Setup (MKa , PKa ) (MKi , PKi )
2 Key Generation (MKa , PKa , Sa ) (MKi , PKi , IDi )
3 Encryption (Ta , PKa , Ma ) (IDi , PKi , Mi )
4 Decryption (AKa , PKa , CTa ) (SKi , PKi , CTi )
Table 3 Notations
S. No. Notations Description
1 MKa , MKi Master key of ABE and IBBE
2 PKa , PKi Public parameter key of ABE and IBBE
3 Sa A set of attributes of ABE
4 Ta Access policies of ABE
5 M a , Mi Message of ABE and IBBE
6 AKa Attribute set of ABE
7 CTa , CTi Ciphertext of ABE and IBBE
8 IDi Identity key of IBBE
9 SKi Secret key of IBBE
4 Proposed Work: Privacy Preservation for Data Sharing

with Multi-party Access Control in Cloud Computing
As cloud becomes an efficient way for the data storage and sharing services, people
are cooperating as a gathering to impart the information to one another. Cloud is the
best solution for sharing, maintaining and handling data. In the recent year, cloud is
used in large system to store data and retrieve data online. Using the cloud computing,
the user accesses the data anywhere, anytime and any device. Cloud computing is
a most recent and quickly developing innovation which depends on the utilization
of assets. These resources are provided dynamically via the Internet. The service
provided by service provider is used by multiple customers. The customer cannot
fully trust for storage service provided by cloud service provider; hence, instead of
storing data as it is, customer prefers to store encrypted data. The problem arised due
to this is searching required data from encrypted data.
Existing research work manages security issue through encryption and access
control model to the redistributed information. These benefits are only suitable for
single owner.
In single-owner framework, only group admin has priority of read, write and
update any task. But multi-owner is more adaptable than single owner. In multi-
owner, each group member has read, write and update rights. The current member of
group can deny the group at any time. All the group members have their own unique
identity in group so data sharing is done by a secure manner.
5 Conclusion
There are different cryptographic techniques available in cloud computing.

These cryptographic techniques are used for sharing data in group. This paper
provides a review of attribute-based encryption (ABE), proxy re-encryption (PRE),
conditional proxy re-encryption (C-PRE) and identity-based broadcast encryption
(IBBE) techniques. The comparison table shows that IBBE technique is better than
ABE. It is easy to implement. IBBE scheme always preferred the users who can
share their encrypted data with multiple receivers. This scheme allows low-cost key
management and also takes less complexity for encryption and decryption than ABE.
In IBBE, approved client can perform searching legitimately on encoded information
without sharing mystery key. The information proprietor just needs to scramble one
ciphertext for every approved client.
References
1. X. Liu, Y. Zhang, B. Wang, J. Yan, Mona: secure multi-owner data sharing for dynamic groups
in the cloud. IEEE Trans. Parallel Distrib. Syst. (TPDS’13) 24(6), 1182–1191 (2013)
2. Z. Yan, X. Li, M. Wang, A.V. Vasilakos, Flexible data access control based on trust and
reputation in cloud computing. IEEE Trans. Cloud Comput. 5(3), 485–498 (2017)
3. P.A. Tamgave, A.A. Manjrekar, S. Khot, Regular language search over encrypted cloud data
using token generation: a review paper, in: Innovative Data Communication Technologies and
Application, ICIDCA 2019, ed. by J. Raj, A. Bashar, S. Ramson (eds). Lecture Notes on Data
Engineering and Communications Technologies, vol 46. (Springer, Cham, 2020)
4. A.S. Anila, M.J. Manoj, Multi-owner data sharing in cloud. IJCSMC 4(6), 36–48 (2015)
5. B. Lang, J. Wang, Y. Liu, Achieving flexible and self-contained data protection in cloud
computing. IEEE Access 5, 1510–1523 (2017)
6. Q. Huang, Y. Yang, J. Fu, Secure data group sharing and dissemination with attribute and time
conditions in Public Clouds. IEEE Trans. Serv. Comput. (2018)
7. F. Wang, J. Mickens, N. Zeldovich, V. Vaikuntanathan, Sieve: cryptographically enforced access
control for user data in untrusted clouds, in NSDI, pp. 611–626 (2016)
8. K. Xue, W. Chen, W. Li, J. Hong, P. Hong, Combining data owner-side and cloud-side access
control for encrypted cloud storage. IEEE Trans. Inf. Forensics Secur. 13(8), 2062–2074 (2018)
9. A. Jhari, S. Fernandes, Techniques for secure multi-owner data sharing in cloud. Int. J. Eng.
Sci. Comput. (2017)
10. B. Wang, H. Li, X. Liu, F. Li, X. Li, Efficient public verification on the integrity of multi-owner
data in the cloud. IEEE J. Comm. Netw. 16(6), 2014
11. B. Wang, S.S.M. Chow, M. Li, H. Li, Storing shared data on the cloud via security-mediator,
in IEEE 33rd Internation Conference on Distributed Computing System, 2013
12. C. Delerablée, Identity-based broadcast encryption with constant size ciphertexts and private
keys, in Proceeding of International Conference on the Theory and Application of Cryptology
and Information Security (ASIACRYPT ‘2007), pp. 200–215 (2007)
13. Z. Yan, X. Li, M. Wang, A.V. Vasilakos, Flexible data access control based on trust and
reputation in cloud computing. IEEE Trans. Cloud Comput. (2015)
14. X. Dongt, J. Yut, Y. Luot, P2E: privacy-preserving and effective cloud data sharing service, in
IEEE Global Communications Conference (GLOBECOM), 2013
15. B. Wang, H. Li, Efficeint public verification on the integrity of multi-owner data in the cloud.
J. Commun. Netw. 16(6) 2014
16. L. Li, Y. Zheng, X. Li, KeyD: secure key de-duplication with identity-based broadcast
encryption. J. Latex Files 14(8) 2015
17. J. Bethencourt, A. Sahai, B. Waters, Ciphertext-policy attribute based encryption, in IEEE
Symposium on Security and Privacy (SP), (IEEE, 2007), pp. 321–334
18. K. Ren, C. Wang, Q. Wang, Security challenges for the public cloud. IEEE Internet Comput.
16(1), 69–73 (2012)
19. J. Hong, K. Xue, W. Li, Y. Xue, TAFC: time and attribute factors combined access control
on time-sensitive data in public cloud, in Proceeding of 2015 IEEE Global Communications
Conference (GLOBECOM 2015), pp. 1–6 (2015)
20. M. Sepehri, S. Cimato, E. Damiani, C. Yeuny, Data sharing on the cloud: a scalable proxy-based
protocol for privacy-preserving queries, in Proceedings of 14th IEEE International Conference
on Trust, Security and Privacy in Computing and Communications (TrustCom 2015)
21. A. Hanumant Bhosale, A.A. MAnjrekar, Attribute-based storage control with smart de-
depulication filter using hybrid cloud, in 2018 Fourth International Conference on Computing
Communication Control and Automation (ICCUBEA) (Pune, India, 2018), pp. 1–6. https://doi.
org/10.1109/ICCUBEA.2018.8697399.
Pixel-Based Attack on ODENet
Classifiers
Nitheesh Chandra Yaratapalli, Reethesh Venkataraman, Abhishek Dinesan,

Ashni Manish Bhagvandas, and Padmamala Sriram
Abstract With rapid improvements in deep learning, one of the challenging open
issues is the exposure of neural systems to adversarial attack. Many competitions
explored various attacks and prevention strategies against these attacks. Another
worldview of profound neural systems with persistent inward activations was pre-
sented at NeurIPS 2018 that likewise acquired the best paper grant. With the neural
ODE block, all discrete transformations can be parametrized as a continuous trans-
formation that can be solved using a ODE solver. We break down picture/content
classifiers with ODE Nets, upon subject to ill-disposed assaults and contrast it with
standard profound models. Neural ODEs are more robust towards black-box attacks.
Some of their intrinsic properties, such as accuracy–efficiency trade-off and versatile
calculation cost, allow us to build powerful models further.
Keywords Ordinary differential equations · Adversarial attacks · Deep model

learning · One-pixel attack · ODEnet
N. C. Yaratapalli (B) · R. Venkataraman · A. Dinesan · A. M. Bhagvandas · P. Sriram

Department of Computer Science and Engineering,
Amrita Vishwa Vidyapeetham, Amritapuri, India
e-mail: nitheesh.my@gmail.com
R. Venkataraman
e-mail: reetheshv.rv@gmail.com
A. Dinesan
e-mail: abhishekdinesh3@gmail.com
A. M. Bhagvandas
e-mail: ashnibhagvandas16@gmail.com
P. Sriram
e-mail: padmamala@am.amrita.edu
630 N. C. Yaratapalli et al.
1 Introduction
The adversarial attack is a strategy utilized in the domain of machine learning/deep

learning which tries to trick a model through malicious input made by subtle modifi-
cations negligible to a human to the input data. This technique is commonly used to
attack detection/classification models. The real-life implications of such attacks can
be severe. For example, modification of reviews of a topic in social media can trick
content moderation systems. This could lead to display and spread of misinforma-
tion or copyrighted data onto these platforms [1]. Currently, no approach, however,
is capable of a completely robust approach towards adversarial inputs.
Today’s widely used models consist of discrete layers, which use gradient-based
methods to train itself. The transformations of data happen in layers in a layered
approach. At each layer, features will be transformed and passed to the next layer.
Generally, deeper the model, more the training data required to achieve accept-
able results. Deep learning has widespread applications such as text classification,
autonomous vehicles, speech to text models, anomaly detection, decision-making
and breast cancer detection [2, 3]. With improvements in technologies like GANs,
models are even able to generate life-like faces and perform a style transfer.
The investigation into the adversarial examples on trained models is actively being
worked upon by the research community [4]. The characterization of the techniques
provides various insights to identify and differentiate adversarial examples from
authentic inputs. Just recognizing the “cause and effect” is good enough to validate
the usage and regulations of these models in critical systems.
The goal of our work is to pick up insights from a paperback, which lately proposed
a new family of deep learning architectures based on and derived from the similarity of
the parametrize equation of ResNets with ordinary differential equations (ODEs) [5].
Instead of a set of discrete transformations in traditional neural networks that follow
a layered approach, neural ODEs have continuous transformations in the direction of
the output. Any adaptive ODE solver like variants of RK4 [6–9]. Most of the solvers
provide attributes which are used to handle trade-off between speed and precision of
the system. Due to the generality and independence of ODE nets, they can easily be
used in existing models and integrated in new applications. Neural ODE networks
are still differential models meaning they use gradient descent. Adversarial sample
generation algorithms [10, 11], which are used in the existing differential models,
can be applied to neural ODEs.
Current work includes the following contributions: Comparison of the following
architectures—a ResNet and two ODE-based models, along with a variant WRNs—
and their robustness to adversarial attacks on data sets: Kaggle CIFAR-10 [12]
and ImageNet [13]. Along with stability against adversarial attacks with tweaks
to precision–cost trade-off, which is made possible by adaptive ODE solvers hav-
ing tolerance parameters, decreasing the solver threshold brings about a decline in
the success percentage of the attack with only small degradation in classification
accuracy of the model.
Pixel-Based Attack on ODENet Classifiers 631
2 Related Work
2.1 Adversarial Attacks on Deep Learning Models in Natural

Language Processing: A Survey
The paper presents a survey on a plethora of techniques used for attacking NLP-
based deep learning models. The works presented in this paper date back to 2017
when the concept was first proposed. Unlike CV-based models, NLP presents a lot of
challenges because of the intrinsic differences between CV models like CNNs and
NLP models. The paper selects, summarizes, discusses and analyses 40 representative
works in a comprehensive way along with covering a lot of preliminary knowledge
required to get into the field [14].
2.2 Ensemble Adversarial Training: Attacks and Defences
In this paper, Tramer et al. (2017) find a novel black-box attack and a defence mecha-
nism for black-box attacks in general. They claim that overfitting on the training loss
of the model is more common when models with adversarial training are one-shot
attacked. They also experimentally proved that the network changes the gradient a
little locally to confuse the attacker [15].
2.3 Spatially Transformed Adversarial Examples
The stAdv attack [16] aims to prove that perturbation deviation metrics (such as L 2 )
are not effective. Instead of pixel value alteration, the spatial position is minimally
altered. The loss function is same as the one used in Carlini and Wagner [11].
2.4 Adversarial Training Methods for Semi-Supervised Text

Classification
The adversarial and virtual adversarial methods proposed in this document perform
better than previous models not only in terms of text classification tasks but in terms
of word-embedding training. This can provide valuable insights into many other
research topics such as machine translation and question answering systems. The
proposed methods could also be used for other sequential tasks such as voice and
video. One of the records is the large movie review record, often referred to as the
IMDB record. 88.89% accuracy was reached [17].
3 Robustness of NeuralODEs
Neural ODE-based classifiers are networks using an ODE block. We studied the
robustness of classifiers with an ODE block against different variations. We consid-
ered three architectures: with baseline as regular ResNet (RES), a combination of
residual and ODE block which we refer to as a (MIX) mixed model and a pure ODE
block-based model (ODE-only network, OON). All models are trained on public
data sets with adversarial attacks on test sets of those corresponding data sets. Attack
success rate is the metric being used for comparison.
3.1 Model Architectures
The model architectures we used (Fig. 1) are as described.
3.1.1 Residual Network (RES)
A series of ResNet blocks, where input will be added to outputs of layers, is also
with skip connections. This was used as baseline for the results. This model has a
Fig. 1 Model structures [stride]; padding = 1. K = 256

kernel size of 3 × 3 with 64-filter convolutional layers as ResNets [18, 19]. Model
consists of eight ResNets with one convolution skip connection.
3.1.2 Wide Residual Network (WRN)
Similar to a regular ResNet, a number of layers between skip connections are more,
thus becoming higher-order ResNet blocks. Using a model similar to RES, but double
the width, model consists of eight ResNets with two convolution skip connections
[19, 20].
3.1.3 Mixed ResNet-ODE Network (MIX)
In both RES and WRN, excluding the first two ResNet blocks, replace all others
combined with one ODE Block, akin to the ODENet architecture defined by Chen
et al. [5].
3.1.4 ODE-Only Network (OON)
All the ResNet blocks combined have been replaced with a single ODE Block. This
allows for a drastic drop in model parameters.
3.2 One-Pixel Attack
Differential evolution(DE) method, by encoding perturbation to an array with com-

binations of values, where each combination has a fixed perturbation count, is with
elements: (x, y, r, g, b). x-y coordinates with pixel RGB value. Each perturbation
only modifies one pixel. Each perturbation will be following the equations:
xi (g + 1) = xr 1 (g) + F(xr 2 (g) − xr 3 (g)) (1)
r1 = r2 = r3 , (2)
here xi is the candidate element solution, and r1 , r2 , r3 are the results, scale parameter
F using 0.5 and index of generation g [21].
Robustness with adversarial examples on classifiers can be calculated using the
attack success rate, which can be analysed with various attack parameter configura-
tions.
4 Experimental Set-up
Evaluate the success of one-pixel attacks on the neural networks models and variants
mentioned in the previous section.
4.1 Data Sets
We have used the data sets most common with low resolution images: ImageNet [13]
and CIFAR-10 [12]. CIFAR-10 consists of around 60,000 images of size 32 × 32
spread across ten different classes [22]. ImageNet is similar but significantly larger
in number.
4.2 Training Details
Number of filters in each convolution block represented by K is set to 256. A dropout

probability of 0.5 has been used for fully connected layers, and SGD momentum
equals 0.9 with α(learning rate) equals 0.1, along with batch size of 256 [23].
Dormand–Prince fifth-order Runge–Kutta ODE solver version [24] was used for
models with ODE Blocks, with tolerance τ taken as 10−3 . if τ is less than step error,
then integration will be ignored and the solver continues to the next step which is
used package provided by [5] https://github.com/rtqichen/torchdiffeq.
In Table 1, the classification accuracy for each model and data set is mentioned
along with total parameters. ODE-based models are tested with various τ . τ is
inversely proportional to precision of the integration. This causes reduction in accu-
racy, but with only a small effect on overall accuracy. The worst deterioration occurs
for the ODE-only network where accuracy difference is noticeable. Decrease in
number of parameters can also be observed.
4.3 Adversarial Attack Details
No consideration was given to naturally misclassified images. In one-pixel attack,

an evolutionary technique black-box attack, we have used with different amount
of modified pixels 1, 3, 5. Iterations will be stopped as soon as a misclassification
occurs [23]. Also verified for targeted attacks, i.e. trying to misclassify as a targeted
category. source reference https://github.com/Hyperparticle/one-pixel-attack-keras.
Table 1 Classification accuracy and number of trainable parameters on test data sets of Kaggle
CIFAR-10 and ImageNet for (RES)ResNet, (MIX)mixed and (OON) ODE-only net. The ODE
solver tolerance τ is also used for architectures with ODE blocks
ImageNet CIFAR-10(Kaggle)
τ RES (RES) MIX (RES) OON RES (RES) MIX (RES) OON
10−4 73.54% 73.28% 72.98% 77.42% 75.33% 74.16%
10−3 – 73.15% 73.12% – 75.33% 74.16%
10−2 - 72.96% 72.92% – 75.29% 74.03%
10−1 - 72.70% 72.03% – 75.16% 73.56%
100 – 72.43% 71.84% – 75.03% 73.25%
101 – 72.17% 71.42% – 74.91% 73.01%
Params 270K 250K 240K 190,154 177,845 142,458
ImageNet CIFAR-10
τ WRN (WRN)MIX (WRN)OON WRN (WRN)MIX (WRN)OON
10−4 76% 75% 75% 79% 77.33% 77%
10−3 – 75% 74% – 77% 76%
10−2 – 74% 74% – 77% 76%
10−1 – 74% 73% – 77% 75%
100 – 73% 73% – 76% 75%
101 – 73% 72% – 76% 75%
Params 3.4M 3.1M 2.9M 414,978 378,945 326,503
5 Results
The observations for the models with different configurations are as follows (Table 2).
Our results imply that neural ODE-based models have an attack success rate
consistently lower or equal in comparison with regular residual or wide residual
networks. For similar changes to the original image, the attack success rate was
almost halved.
When the τ tolerance of the ODE solver is set to 0.1, attack success rate was
almost halved while the classification accuracy only changed by 2%. Allowing higher
significance in feature extraction process makes the decision counts flexible, thus
bringing down the effects of adversarial inputs with neural ODE integration.
When classifying a perturbed or pristine image with neural ODE, analysis of the
internal states is as follows: let the continuous hidden state trajectory for normal
image be h(t) and by an adversarial sample of the same image be hadv (t).
h(t) = |h(t) − hadv (t)|2 (3)
The differences are averaged across all test samples, using same τ tolerance for each
segment in adversarial examples measured by h(t). By setting a higher tolerance
τ , the adversarial effect weakens in respect of internal state changes.
Table 2 Pixel attack success rate for different models with number of pixels to be modified for
untargeted and targeted cases
Untargeted Targeted
Model 1 pixel 3 pixel 5 pixel 1 pixel 3 pixel 5 pixel
(%) (%) (%) (%) (%) (%)
ResNet 34 79 79 14 21 22
Wide ResNet 19 58 65 1 18 22
MIX(Res)τ = 0.1 27 46 56 11 17 20
MIX(Res)τ = 0.01 24 3 48 9 15 20
MIX(WRN)τ = 0.1 21 46 52 0.3 15 19
MIX(WRN)τ = 0.01 21 40 44 0.3 13 18
OON(Res)τ = 0.1 23 43 49 9 15 19
OON(Res)τ = 0.01 22 42 47 9 13 17
OON(WRN)τ = 0.1 19 43 48 0.1 12 17
OON(WRN)τ = 0.01 19 39 42 0.1 9 15
6 Conclusion
We proved the robustness of neural ODEs, a neural architecture in comparison with

three types of variations on two different architectures (ResNet, WRN), using the
ImageNet, Kaggle CIFAR-10 data sets. We concluded that neural ODEs are robust
against pixel attacks with 1, 3 and 5 pixels. Furthermore, ODE solver’s error toler-
ance parameter has minimal effect on classification accuracy. Robustness towards
adversarial attacks has been increased, i.e. lower success rate for adversarial attacks.
Different variations of ResNets with ODE blocks and pure ODE-based networks
can be extended to any larger real-world high resolution image or video data sets.
They also can be extended to integrate ODEP-based networks into advanced models.
As our work focuses on black-box attacks, future work can apply white-box attacks,
which can analyse the internal states of attacks. Robustness of the models is yet to
be verified in the white-box attack context and can also focus on text-based models.
References
1. R. Vinayakumar, M. Alazab, K.P. Soman, P. Poornachandran, A. Al-Nemrat, S. Venkatraman,

Deep learning approach for intelligent intrusion detection system. IEEE Access 7, 41525–
41550 (2019)
2. M.A. Anupama, V. Sowmya, K.P. Soman, Breast cancer classification using capsule network
with preprocessed histology images, in 2019 International Conference on Communication and
Signal Processing (ICCSP), pp. 0143–0147 (2019)
3. N.S. Rani, P. Rao, P. Clinton, Visual recognition and classification of videos using deep con-
volutional neural networks. Int. J. Eng. Technol. (UAE) 7, 85–88 (2018)
4. S. Sriram, K. Simran, R. Vinayakumar, S. Akarsh, K.P. Soman, Towards evaluating the robust-
ness of deep intrusion detection models in adversarial environment, in Security in Computing
and Communications, ed. by S.M. Thampi, G. Martinez Perez, R. Ko, D.B. Rawat (Springer
Singapore, Singapore, 2020), pp. 111–120
5. T.Q. Chen, Y. Rubanova, J. Bettencourt, D. Duvenaud, Neural Ordinary Differential Equations.
CoRR abs/1806.07366, http://arxiv.org/abs/1806.07366 (2018)
6. E. Süli, D.F. Mayers, An Introduction to Numerical Analysis (Cambridge University Press,
Cambridge, 2003)
7. F. Lu, Two implicit runge-kutta methods for stochastic differential equation. Appl. Math. 3,
1103–1108 (2012)
8. N.A. Ahamad, S. Charan, Study of numerical solution of fourth order ordinary differential
equations by fifth order runge-kutta method. Int. J. Sci. Res. Sci. Eng. Technol. 6, 230–237
(2019)
9. Wikipedia: Runge–Kutta methods—Wikipedia, the free encyclopedia. http://en.wikipedia.
org/w/index.php?title=Runge%E2%80%93Kutta%20methods&oldid=956618157 (2020),
Online; Accessed 30 May2020
10. I.J. Goodfellow, J. Shlens, C. Szegedy, Explaining and Harnessing Adversarial Examples
(2014)
11. N. Carlini, D. Wagner, Towards Evaluating the Robustness of Neural Networks (2016)
12. A. Krizhevsky, Learning Multiple Layers of Features from Tiny Images (University of Toronto,
Toronto, 2012)
13. J. Deng, W. Dong, R. Socher, L.J. Li, K. Li, L. Fei-Fei, ImageNet: a large-scale hierarchical
image database, in CVPR09 (2009)
14. W.E. Zhang, Q.Z. Sheng, A. Alhazmi, C. Li, Adversarial Attacks on Deep Learning Models in
Natural Language Processing: A Survey (2019)
15. C. Xiao, B. Li, J.Y. Zhu, W. He, M. Liu, D.X. Song, Generating adversarial examples with
adversarial networks, in IJCAI (2018)
16. C. Xiao, J.Y. Zhu, B. Li, W. He, M. Liu, D. Song, Spatially Transformed Adversarial Examples
(2018)
17. T. Miyato, A.M. Dai, I. Goodfellow, Adversarial Training Methods for Semi-supervised Text
Classification (2016)
18. X. Yu, Z. Yu, S. Ramalingam, Learning strict identity mappings in deep residual networks,
in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4432–4440
(2018)
19. I. Kim, W.,Baek, S. Kim, Spatially Attentive Output Layer for Image Classification (2020)
20. S. Zagoruyko, N. Komodakis, Wide Residual Networks (2016)
21. J. Su, D.V. Vargas, K. Sakurai, One pixel attack for fooling deep neural networks. IEEE Trans.
Evol. Comput. 23(5), 828–841 (2019). https://doi.org/10.1109/TEVC.2019.2890858
22. Q. Shi, J. Wan, Y. Feng, C. Fang, Z. Chen, Deepgini: Prioritizing Massive Tests to Reduce
Labeling Cost (2019)
23. F. Carrara, R. Caldelli, F. Falchi, G. Amato, On the robustness to adversarial examples of
neural ode image classifiers, in 2019 IEEE International Workshop on Information Forensics
and Security (WIFS), pp. 1–6 (2019)
24. J. Dormand, P. Prince, A family of embedded runge-kutta formulae. J. Comput. Appl. Math.
6(1), 19–26 (1980). https://doi.org/10.1016/0771-050X(80)90013-3
Survey on Advanced Spectrum Sharing
Using Cognitive Radio Technique
Mohamed Hassan, Manwinder Singh, and Khaild Hamid
Abstract Wireless communications from generation to generation up to the fifth

depend on the increase band width output and lower latency, which have been
achieved through the development and addition of new technologies on the dedi-
cated channels. The fluctuating nature of the available range and the varied quality
of service (QoS) requirements for different applications, limited spectrum, and the
spectrum resource allocation policies present challenges for radio networks. Cogni-
tive radio (CR) is one of the keyways to optimize the spectrum to the fullest degree
possible; this paper describes CR, analyzes key functions, takes measures to share
spectrum, discussing the effect of CR when combined with non-orthogonal multiple
access (NOMA) at 5G and its impact on improving the performance of satellite
communications.
Index Terms Quality of service (QoS) · Cognitive radio (CR) · Fifth-generation

(5G) · Non-orthogonal multiple access (NOMA)
1 Introduction
The way people live and communicate across wireless networks has changed and
this was either positive or negative. This area is still a constant example of many
applications that can be developed and developed. A fast survey of the growth and
complexity of mobile communications was made of the first 1G wave, distinguished
by a single sound service, FDMA (Frequency Division Multiple Access), the channel
M. Hassan (B) · M. Singh

Department of Wireless Communication, Lovely Professional University, Phagwara 144411, India
e-mail: mhbe4321@gmail.com
M. Singh
e-mail: manwinder.25231@lpu.co.in
K. Hamid
Department of Electronics and Communication, University of Science and Technology,
Omdurman 22110, Sudan
e-mail: khalidhamidk9@gmail.com
640 M. Hassan et al.
control network with a maximum speed of 1G, 2.4 kbps. The first remarkable devel-
opment is achieved when mobile phones have switched from 1G to 2G. The principal
differences between mobile systems (1G and 2G) are that radio signals from the 1G
network are analog, and that the 2G are digital. A secure and reliable communica-
tion system, TDMA (Timing Division Multiple Access) as a channel access method
and a 144-kbps data interchange rate were the main drivers of this generation. The
third-generation sets the standards for WCDMA (Wideband Code Division Multiple
Access), whereby Web browsing, e-mail, video downloading, photo sharing, and
other smart phone technologies are possible. In 4G aims to give users high speed,
high quality, and high capacity, while enhancing security and reducing the rates
of speech and service, multimedia and the Internet via an IP possible and existing
applications include updated mobile web access and IP telephony, gaming, high defi-
nition mobile TV, video conferencing, 3D TV and cloud computing based on OFDM
(Orthogonal Frequency Division Multiplexing), maximum speed approximately 1
Gbps.
5G technology seeks to improve the 4G network by increasing data rates even
more efficiently, rising connections density, and dramatically reducing latency by,
among other steps, improving data speeds. Increasing battery consumption and
enhanced wireless coverage by adding a device to device communication func-
tions. The software targets a maximum bandwidth of between 5 and 10 Gbit/s
and uses non-orthogonal multiple access (NOMA) radio access technology [1].
The fifth-generation technology will make important changes in this field and rein-
force other technologies such as the Internet of Things, all technologies focus on
increasing performance and reducing latency, wireless communication technologies
are developing rapidly and restricting the limited radio spectrum resources.
In accordance with existing spectrum regulations, wireless service providers,
including telecommunications, television, and the satellite, have allocated or licensed
a large portion of the spectrum available to them as licensed services (or primary
service providers).
Statistics from the Federal Communications Board (FCC) show that the average
use of authorized bands varies between 15 and 85%. The FCC also considered that
only 5.2% of the spectrum is used for less than 3 GHz in any location in the USA [2].
Many multiplex technologies such as the frequency division multi-access (FDMA),
the time division multi-access (TDMA), and the code division multi-access (CDMA)
are being used to maximum effect in order to improve licensed spectrum and continue
to support other users with limited spectrum resources.
In this article, we provide a survey the basic concepts of cognitive radio (CR)
definition, essential characteristics, functions, steps that follow to share the spectrum,
explain the effect of CR when used in 5G and satellite communication.
Survey on Advanced Spectrum Sharing Using … 641
2 Cognitive Radio Definition
Dr. Metulla presented a modern idea of maximizing the gentle level at the end of
the last century, and renamed it to cerebral radio CR. It was defined as an intelli-
gent communication technique that knows the surroundings considerably and uses
a method of awareness by making it possible to impart from the atmosphere and
adjust its internal cases to inspire statistical differences within the incoming radio
frequency by making corresponding adjustments to the working parameters in real
time [3].
2.1 Attribtes of Cognitive Radio
CR’s core characteristics are classified into two groups.

a. Cognitive capabilities
CR has the potential to identify the dominant domain by utilizing specialized
techniques on the radio spectrum and to test correct parameters. The CR jobs are
often referred to as the cognitive process and have three key phases.
• Observe spectrum: CR is trying to gather information on wasted radio
spectrum, to identify the spectrum within the cavity.
• Spectrum analytic: frequency observation and evaluation of signal cavity
characteristics.
• Spectrum resolution: selection of the best possible fixed line contact channel
focused on choosing the optimal unused fixed-band link channel according to
the customer specifications and features assessed [4, 5].
b. Reconfiguration option
It is the ability to acclimatize changing operating parameters during radio opera-
tion without any component modulation of the equipment [4, 5]. The parameters
used to reconfigure are:
• Frequency use.
• Switch methodology.
• Energy transfer.
• Technology of communication.
2.2 CR Functions
The primary role of CR is to monitor, sensitize, and collect information on the envi-
ronment. The judgment was then taken, and the conditions were established [6]. Spec-
trum sensing, spectrum management, spectrum division, and spectrum movement are
the core features.
Spectrum sensing: To take cognitive radio operation into account, intelligent

sensing and intelligent decision-making methods are needed. Perhaps one of the
most critical mechanisms for cerebral radios is frequency identification and one
of the most important functions of cognitive radio systems is also the method of
frequency detection [7].
Spectrum management: The best available signal frequency, transmitting capacity,
consumer numbers, interference, energy efficiency and QoS specifications are
selected as well. This method also includes spectrum exchange mechanisms for the
best distribution of channels and power and other features apply to the key functions
of the layer protocols of medium access controls. It can therefore be included in the
MAC layer. However, efficient spectrum sharing challenges are associated with time
synchronization and distributed power allocation [7].
Division of spectrum: Access to spectrum with other users is arranged. Switch
spectrum: The secondary user exit and switch into another frequency band opening
when the first user returns or appears [8].
3 Spectrum Sharing
The CR technology allows the reuse of important resources in the spectrum without
altering the existing policy. CR’s key concept is to ensure spectrum allocation by
dynamic access to the spectrum so secondary users (Sus) can use an idle spectrum
of primary users (Pus). The distribution of spectrum usually takes four steps.
3.1 The Detection of Spectrum
The first procedure includes collecting information on the available spectral

frequency bands, the primary use of existing bands, and the finding of spectrum
gabs [9]. It includes various techniques like detection of power, detection of charac-
teristics, consistent discovery, and filtering. It can be divided into white, gray, and
black regions according to the status and condition of the frequency spectrum. Appli-
cants for secondary users are gray and white region [10]. We provide a general idea
of some of these approaches through the great development in the area of wireless
communication and the development of new methods of spectrum sensors.
• First non-cooperative sensor frequency range
In this model, each user selects the best way to increase the use of available
frequencies, to communicate egoistically and in a way that is so different from
ideal [6].
• The second frequency range sensor full duplex (FD)

In this model, both primary and secondary users simultaneously take advantage
of the spectrum, while ensuring optimum use of the frequencies available [3] in
the number of collisions with the principal user.
• Third spectrum database frequency range sensor
It is used to save old data and to create schedules, which provide information on
the currently available frequencies based on frequency setting at the time.
• Fourth sensor cooperative range frequency
The model is used for sensitive applications, which does not enable delays,
for example, in industrial areas, where knowledge sharing between users helps
accelerate procedures and distribute opportunities.
• Fifth spectrum of compression
This model is primarily intended for Internet of things technology, as it is aimed
at large populations who exchange information in limited amounts but constantly
over and above their low energy consumption [3].
3.2 Allotment of Spectrum
It is a vital mechanism that avoids conflict between cognitive radio and primary
licensing, enabling wireless radio frequency to have more functionality. Two
approaches to do this.
a. Spectrum sharing cooperative
A data collection center and a control channel between them, or over distributed
networks, communicate with each other in this frequency range. The first method
is more accurate, but the time is much longer, the implementation is very complex,
and energy usage is greater than the second method. The key downside of the
second method is its inaccuracy [11].
b. Spectrum sharing non-cooperative
This does not allow a kind of user-to-user information exchange, yet a very
useful method of communication is available for a small number of cognitive
user networks [12].
3.3 Access Types of Spectrum
The first type, divided into two types, is called static and assigns the frequency
range for a company approved by a government agency. This company is called a
service provider and is called primary users for subscribers. The second type, named
dynamic, does not have a license to use frequency bandwidth, but it tries to classify
the gaps in the allocated frequencies by using CR methods and, on the other hand, its
inefficiency is divided into three types, namely: The second type, named dynamics.
(1) Commons, (2) Models of common use and (3) Models of exclusive use [4]. It
will cover every model in some detail.
a. Commons: There are no parties allocated to any 2 frequency bands 2.4–5 GHz,
so that all users with a specific interest can use them [3, 9].
b. Shared usage type: This type allows frequency distribution between primary
and secondary users, but guidelines for the second type of user are given. Three
methods have been developed to hit the frequency bands for secondary users:
(1) Method of interweaving:
These methods allow secondary users to use the frequency band gaps when
they are not used by the primary user. If no financial expenses are involved,
secondary users are therefore required to leave the bandwidth immediately
upon arrival of primary users in search of another domain to convert to this;
the difficulty of this process is constantly changing from one bandwidth to
another [2].
(2) Access method for spectrum underlay
This method allows the secondary user to transmit at the same time with the
primary user, according to the main condition of the interference from the
secondary users does not exceed to reasonable limit. Therefore, they send
very low power, and this affects the coverage area negatively.
(3) Type of exposure to spectrum overlays
The secondary user searches the most assigned frequency bands with modi-
fied methods for knowing the unused places and using them by synchro-
nizing sending with the primary users. If the secondary user knows the
presence of the primary users, the features of signal and transmission are
changed to avoid any interference [2].
c. Layout for exclusive use
The principal frequency service providers would then be leased or loaned by
primary service providers, and secondary service providers tend to be the new
name, and secondary users, like primary users, have full access rights to the
frequency band according to the contract agreement [2].
3.4 Hand-off Spectrum
Due to the various spectrum parameters, it is referred to as the handout method and
consists of three cases: passive, effective, and hybrid [3] where the secondary users
are transferred from one channel to another according to their time and place or
returns.
This is the following description of the proposed spectrum hand-off policy. For
devices with the planned spectrum hand-off, two radio spectrum bands are called. The
radio spectrum band which is higher by average off time is allocated to machines to
send data packets according to historical details in the radio spectrum band databases.
Radio frequency energy can be collected by machines from both spectrum bands. At
the beginning of a multi-frame, a machine gets details of both radio bands about the
busy state or idle state. Then, with higher percent of mean off time depending upon
the idle or busy state of the band [12].
4 Game Theory
This is a method for simulation and evaluating nodes. It will forecast the positions and
actions of all participants and then determine for all participants the best possible
choices. One of the best practices, including many of the advantages of the best
bandwidth and frequency utilization, is the Nash method or the equilibrium point
[11, 13].
5 Commercial Application
5.1 Networks of 5G
Cognitive radio has many problems since the emergence of technologies of the fifth
generation, contributing to a huge and immense growth in the usage of broad and wide
frequency ranges, which ensures that information collection and sensing become
more complicated and measured [14]. The NOMA techniques are expected to be
integrated with the CR networks to enhance spectral efficiency and serve the number
of users. The Noma techniques are useful to secondary users, to cooperate effectively
and to gain access to the Primary User (PU). The effectiveness of the SU and PU can be
increased simultaneously by implementing a satisfactory CR network coordination
system with NOMA [15].
5.2 Satellite Communications
Since satellite communication has a wide reach, seamless connectivity and easy
Internet access for foreign users, it is considered a key component of future wireless
communications [16].
The recent development of satellite communications has seen a major interest
in them, since they have been working simultaneously or close to the high-
performance cellular communications and have reduced the gap between them.
Cognitive radio was suggested to optimize the width of the band assigned to this
type of communication, with a downlink speed of 2.4 GHz and uplink speed of
2 GHz [17].
6 Conclusions
By examining the basic concepts of cognitive radio, it became clear to us that this is
one of the best ways to solve the issues of limited frequency range. CR technique is
a paradigm that enables the frequency band to be utilized by secondary consumers
in a proactive way by detecting the unused frequency spectrum, sometimes known
as strength gabs. Once the primary user enters, the domain would be empty without
the usage of the hierarchical series triggering conflict with other applications. For
future research, we would recommend NOMA in a commercial registry defined 5G
network that can access frequency bands in all primary consumer cases and undertake
a comprehensive device efficiency analysis taking into account the quality of service
parameters.
References
1. D. Cohen, S. Tsiper, Y. Eldar, Analog-to-digital cognitive radio sampling, detection, and

hardware IEEE Sig. Process. Mag. 137–167 (2018)
2. M. Hassan, G. Karmakar, Exclusive use spectrum access trading models in cognitive radio
networks: a survey. IEEE Commun. Surv. Tutorials 19, 2018–2025 (2017)
3. F. Hu, B. Chen, A. Kunzhu, Full spectrum sharing in cognitive radio networks toward 5G: a
survey. IEEE 3(3), 15754–15776 (2018)
4. J. Thomas, P. Menon, A survey on spectrum handoff in cognitive radio networks, in Inter-
national Conference on Innovations in Information, Embedded and Communication Systems
(ICIIECS), Coimbatore (2017), pp. 1–4
5. J. Agarkhed, V. Gatate, Survey on spectrum sensing techniques in cognitive radio networks,
in IEEE, International Conference on Intelligent Computing, Instrumentation and Control
Technologies (ICICICT) (2017)
6. M. Manesh, M. Apu, N. Kaabouch, W. Chen Hu, Performance evaluation of spectrum sensingm
techniques for cognitive radio systems, in IEEE Conference (UEMCON) (2016)
7. E.U. Ogbodo, D. Dorrell, A.M. Abu-Mahfouz, Cognitive radio based sensor network in smart
grid: architectures, applications and communication technologies. IEEE Access 5, 19084–
19098 (2017)
8. N. Muchandi, R. Khanai, Cognitive radio spectrum sensing: a survey, in IEEE, International
Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT) (2016)
9. D. Alias, G. Ragesh, Cognitive radio networks: a survey, in IEEE WiSPNET Conference (2016)
10. S. Sharma, S. Chatzinotas, L. Bao, Cognitive radio techniques under practical imperfections:
a survey. IEEE Commun. Surv. Tutorials 17 (2015)
11. A. Khan, M. Rehmani, M. Reisslein, Cognitive radio for smart grids: survey of architectures,
spectrum sensing mechanisms, and networking protocols. IEEE Commun. Surv. Tutorials 18
(2015)
12. S. Tzeng, Y. Lin, L. Wei, Spectrum hand-off for RF energy harvesting in cognitive radio based
machine to machine communication networks, in 2018 International Conference on System
Science and Engineering (ICSSE), New Taipei, pp. 1–4 (2018)
13. J. Zheng, Y. Wu, Dynamic computation offloading for mobile cloud computing: a stochastic
game-theoretic approach. IEEE Trans, Mob. Comput. 18, 757–770 (2018)
14. L. Zhang, M. Xiao, G. Wu, M. Alam, Y. Liang, S. Li, A survey of advanced techniques for
spectrum sharing in 5G networks. IEEE Wirel. Commun. 24 (2017)
15. F. Zhou, Y. Wu, Y. Liang, Z. Li, Y. Wang, K. Wong, State of the art, taxonomy, and open issues
on cognitive radio networks with NOMA. IEEE Wirel. Commun. 25(2), 100–108 (2018)
16. T. Liang, K. An, S. Shi, Statistical modeling-based deployment issue in cognitive satellite
terrestrial networks. IEEE Wirel. Commun. Lett. 7(2), 202–205 (2018)
17. M. Jia, X. Gu, Q. Guo, W. Xiang, N. Zhang, Broadband hybrid satellite-terrestrial communi-
cation systems based on cognitive radio toward 5G, in IEEE Wireless Communications, vol.
23, December 2016
An AI-Based Pedagogical Tool
for Creating Sketched Representation
of Emotive Product Forms
in the Conceptual Design Stages
Sunny Prakash Prajapati , Rahul Bhaumik , Tarun Kumar ,

and Unais Sait
Abstract This paper presents an AI-based pedagogical solution to aid designers

in an appropriate analysis of existing product forms, and to provide feedback to
them in the ideation of new forms pertaining to specific emotions and attributes.
Sketching is often used as a tool for ideation and communication, especially by
artists and designers. It is the non-ambiguity in the sketched representation that is
crucial for the effective communication of ideas amongst the stakeholders involved in
the product development process. Sketching, therefore, becomes the critical skill that
an aspiring designer should acquire in design schools or during the learning phases
of their careers. Hence, a pedagogical support system to assist novice designers in
developing their ideas unambiguously through sketched representations is necessary.
In this study, an image classification-based machine learning (ML) model is trained
on a dataset containing images of lines. Further, this model is validated on a set of
concept sketches, and a smartphone-based learning application for novice designers
is proposed as a pedagogical feedback tool for the conceptual design stage.
Keywords Pedagogy · Design sketching · Emotive forms · Form generation ·

Machine learning · Image classification
S. P. Prajapati (B) · R. Bhaumik · U. Sait

PES University, Bangalore, India
e-mail: sunnyprakashp@pes.edu
R. Bhaumik
e-mail: rahulbhaumik@pes.edu
U. Sait
e-mail: unaissait@pes.edu
T. Kumar
Indian Institute of Science, Bangalore, India
e-mail: tarunkumar@iisc.ac.in
650 S. P. Prajapati et al.
1 Introduction
1.1 Emotions and Artworks
In the context of user experience, emotion is a complex phenomenon that emerges

because of the interactions between a person (user) and an artefact (object of use). All
creative forms of work—be it music, visual art, or literature—elicit certain emotions
in the perceiver. In the context of written works, there have been attempts to describe
its relationship with the emotions through frameworks (and models) which can be
dimensional or categorical [1]. The categorical frameworks represent emotions, as a
set composed of individual primaries which amalgamate in various combinations to
form complex emotions. Plutchik’s theory is a prime example wherein all complex
emotions are based out of eight fundamental emotions [2]. The dimensional model
presents emotions by mapping them against specific key parameters on continuous
scales. For example, emotional states have been mapped to the following parameters:
valence (positivity or negativity of emotion) and arousal (alertness and calmness of
emotional state), in Russell’s circumplex model [3].
Looking at the categorical framework connecting emotions and words, it is
observed that a set of words is highly effective in describing the emotional response
of an individual. This connection can be perceived as a relational map between
the emotions and the word set, R: Emotion → {words}. The critical aspect of this
mapping is the mutual exclusivity amongst the categories of emotions and overlaps
in the word sets describing the emotion. This limits it from being studied as a simple,
functional model between sets of words and emotions. There have been attempts to
formalise this complicated relationship which can be insightful for the present case
[4, 5].
The relationship between shapes and emotions is less coherent as compared to
words and emotions. Collier’s experiment on matching a word to a shape provides a
significant correlation between simple line drawings and universal emotional percep-
tion [6]. This experiment is further extended to 3D shapes and emotions [7]. Taking
insights from the works of Plutchik, Russel, and Collier [2, 3, 6], a relationship that
connects forms to emotions through words can be constructed. This construct is
detailed out for the creation of a pedagogical support system, as discussed later.
1.2 Design and Sketching
In the process of new form development (in product design), sketching is one of the
primary tools used by the designers. Product form development is a part of a larger
process which is the design thinking process. Although there are various models of
the ‘design thinking’ process, a generalised version has four stages: identification
of problem and requirements; ideation; consolidation and refinement of solutions;
and testing and selection of solutions [8]. Sketching is not only a mere tool for
An AI-Based Pedagogical Tool for Creating Sketched … 651
representation but also an important activity during the ideation phase of the design
thinking process. Sketching in product design differs from the artistic sketching
process, which involves a mere replication of an object or a scene in observation.
The design sketching is analogous to an iterative enquiry that helps the designer in
bringing about the intrinsic characteristics of the product form [9], which holds good
even in the field of architectural form design [10].
Designing products that evoke an emotional response in the user is an essential
skill of a designer. The emotional perception of the product can be achieved through
modification of the characteristics of the product elements [11]. One of the essential
elements which need to be focused is the form or the overall geometry of the product,
the others being the colour–material–finish(CMF) of the product, packaging, and
functionality. Design sketching is one of the most direct, quick, and cost-effective
ways to represent and ideate forms. Hence, it becomes a vital skill for the designer
that needs to be acquired in the initial years of design training.
A designer initiates the development of a product form through sketches, and
iteratively adds ‘meaning’ to it guided by metaphors to establish an emotional connect
with the end-user. This activity is usually supervised by an experienced designer
or a mentor in institutions offering design education. Feedback of the mentors on
the student’s work is an essential aspect of learning, which is not always available
at every step of the training process. Considering the promising results shown by
the AI-based technologies in 3D form generation [12, 13], their incorporation in
the ideation phase, where the form exploration happens through sketching, can be
beneficial for the designers. The ubiquitousness of smartphone-based and easily
deployable machine learning (ML) codes (https://www.tensorflow.org/lite/models)
allowed the development of this pilot project to aid novice designers in the process of
developing sketched representations for emotive forms. Since the tool is built on the
digital platform, it can potentially reduce the need for the constant physical presence
of a mentor and can perhaps enhance the assessment due to its agile nature [14].
2 Methodology
Attributes of a product are determined by its form (to a large extent), and hence,
exploration in the early design phase is necessary for the effective incorporation of
an emotion (which is described by the attributes). Sketching iteratively during the
ideation helps the designer to accomplish this feat. The initial ideation happens on the
profile of the form which sets the shape in a 2D format. This shape is again composed
of many line elements that have their characteristic features, which aggregates to an
overall attribute of the shape.
So, for the present study, an image classifier was built by retraining an existing
CNN model MobileNets [15], on our generated dataset that contained line elements.
The lines had different geometrical characteristics that mapped to specific attributes
,which were representative of the underlying emotions. Two mutually exclusive
attributes were selected to characterise the line segments, which were then used
Fig. 1 Dataset preparation for the image classification model
to train the model. The retrained model was further tested against the form profiles
for the preliminary validation. After that, it was deployed as a smartphone-based
application to be used by the novice designers to get feedback on their ideation
sketches (see Fig. 3).
2.1 Preparation of the Dataset
The data set for the model was obtained through images sampled from the video that
captured the line elements drawn on a paper. For sampling, the ‘FFmpeg’ application
was used, which gave more than 2000 images from the video [16].
The images were further processed on an image manipulation software, GIMP to
minimise the background noise and improve the contrast. They were then resized to
256 px by 256 px size using a batch processing plugin (‘BIMP’) [17] (see Fig. 1). After
this preprocessing, the images were manually categorised into two mutually exclusive
attributes (labels)—‘cute’ and ‘rugged’, for training the classification model. Any
inconsistencies and errors in the processed images were manually removed.
2.2 Training and Testing of the Model
The image classifier was developed by retraining the CNN model MobileNets [15]
using a Python script from the TensorFlow [18] library on the dataset containing the
line elements categorised into two attributes (2.1). The data was split into training
and testing set, and the model was trained with a step size of 500. The cross-entropy
loss was used as a metric for classification accuracy.
In order to obtain the attribute of the form profile, the image was split, and the
attribute scores for every part were predicted using the image classifier. The splitting
was performed using the image slicer library in Python [19], which split the image into
four parts of equal pixel size. The scores were averaged for the categories ‘cute’ and
‘rugged’, then the attribute with the higher average score was assigned to the image.
The attribute assigned using this method was then compared to the one assigned by
the experienced designer for the sketches to test this approach.
2.3 Developing the Smartphone Application
An Android-based smartphone application has been developed for implementing the

trained classification model for the novice designers, which would act as a continuous
feedback provider for their works during product form generation. The interface for
the application has been developed by considering the prime functions and task flows
undertaken by the app users (i.e. novice designers and learners in this case). This
application is focused on the early design phases of iterative product development. In
the present case, the application has manually drawn sketches or scanned sketches as
input, but in the future, it can be integrated into digital sketching software platforms
(Fig. 2).
Fig. 2 The proposed solution—A mobile application SketchmoE

Fig. 3 Methodology for the development of an AI-based pedagogical tool
3.1 Model Accuracy
During the training of the MobileNets model, the cross-entropy decreased over time
(see Fig. 4). Since this is a cost function, its minimisation implied better association
between the class labels and the image features. The training and the test accuracies
were 100% and 79.8%, respectively.
3.2 Workflow
The application takes concept sketches of forms (made by a designer during the
exploration phase) as input (through a camera). The image is then processed (size,
resolution and colour correction) to make it suitable for the classifier which outputs
the attributes and their respective scores. This feedback can then be taken up by the
designer to strengthen the desired attribute in the form (see Fig. 5).
Fig. 4 Cross-entropy loss during training (initial value of 0.087 is omitted)
Fig. 5 Workflow of the smartphone application
3.3 Preliminary Validation of the Model
The attribute scores for the eleven different sketches obtained by the model are
compared against the attributes assigned by the designers. These sketches represent
a closed form (of approximately the same proportions) with no intersecting line
elements. The attributes predicted by the model (through maximum average score)
are the same when compared to the attributes assigned by an experienced designer
for each image. The results are summarised in Fig. 6.
Fig. 6 A figure depicting the score of the rugged attribute and the cute attribute of different forms
sketched out for preliminary validation as indicated by the circular labels with their sizes indicating
the score
3.4 Discussion
Though the model correctly predicted attributes for the above cases, it has few limi-
tations in its current form. Limitations can range from incorporating a fixed number
of attributes in the model to its applicability to a certain kind of sketches (those
with intersecting line elements are currently not supported). Moreover, factors such
as attributes emerging from the sketch, as opposed to the combined attribute of the
parts of the sketch, are not captured by the present model. Also preparing a high-
quality dataset with non-mutually exclusive attributes is vital for the development of
a reliable system of this kind.
This tool can be used by the designers while sketching during the product ideation
phase. The mobile application will allow the user to capture the image of the sketch
and provide scores for the attributes in real time. Apart from presenting the current
attribute for a sketch, the app will also provide suggestive feedback to the user to
improve upon the sketch to ‘enhance’ an attribute. The inbuilt library feature will
provide more information to the app user regarding the selected attribute.
The systemic diagram (see Fig. 7) represents the positioning of the AI-based
learning tool in the broader context of the ‘design thinking’ process. The proposed
pedagogic tool is useful not only in the preliminary ideation stage but also in the idea
refinement and consolidation stage, leading to the finalisation of the concept. The
systemic diagram also elucidates briefly about the influence of the pedagogic tool
Fig. 7 A systemic diagram representing the proposed pedagogic tool in a broader context
in the student’s learning process. The continual feedback from the mobile appli-
cation helps the users in decision making while ideating on product forms (by
sketching), through suggestions on the desired attributes and possible improvement
in the features of the sketched forms. Over time, this can have a potential effect on the
learner’s perception to create desirable emotive forms. The diagram also indicates
the role of the human mentor (who is usually an experienced human designer) in
the overall learning process of the novice designer. The mentor can provide exten-
sive feedback or information and evaluate the progress of a learner in periods of a
longer time interval as opposed to the continual involvement of the learner with the
pedagogic tool (the mobile application in our case). Moreover, the human mentor
can play an active part in the development of the AI-based pedagogic tool based on
the changing design trends and the performances of the pedagogic tool users over
longer periods of time.
4 Conclusion
The proposed smartphone application based on the classification model is analogous

to a mentor who is continuously providing feedback to the learner at every instance
of his conceptual design ideation phase. The solution leverages the power of ML
to provide insights from a large set of data (that maps product forms to emotions)
during the form exploration. At present, the feedbacks are on the shape features (like
angularity or curvature) which influences the desired attribute in the product form.
The designer can then factor in the recommendations to make the form more aligned
towards the attribute. The solution also removes the need for the mentor or expert to
be physically present while providing feedback to the learners.
Since AI is making its mark in almost every domain of human interest, product
design and design education is no exception; and the proposed solution effectively
generates and communicates the feedback on the learner’s work, which aids in
continuously training them to make them more seasoned in their profession.
In the future, a real-time AI-based application based on digital sketching platforms
could be derived from the present mobile application. Further, after reinforcement
learning by the model over a period (with the same designer), the application may be
useful for suggesting design ideas in terms of emotive form exploration in the iterative
design exercise. This application can be an integral part of the digital sketching
software platforms providing real-time continuous feedback, which can also learn
from the designers’ concept generation workflow history. Moreover, the solution
provides a roadmap to data-driven form design, which can be a useful technique for
sustainable product development.
References
1. R.A. Calvo, S. Mac Kim, Emotions in text: dimensional and categorical models. Comput. Intell.
29, 527–543 (2013)
2. R. Plutchik, The Emotions. (University Press of America, 1991)
3. J.A. Russell, A circumplex model of affect. J. Pers. Soc. Psychol. 39, 1161 (1980)
4. S.M. Mohammad, P.D. Turney, Nrc Emotion Lexicon (National Research Council, Canada,
2013)
5. P.P.J. Mothersill, The Form of Emotive Design (2014)
6. G.L. Collier, Affective synesthesia: Extracting emotion space from simple perceptual stimuli.
Motiv. Emot. 20, 1–32 (1996)
7. M. Anft, This is your brain on art. https://magazine.jhu.edu/2010/03/06/this-is-your-brain-on-
art/
8. R. Bhaumik, A. Bhatt, M.C. Kumari, S.R. Menon, A. Chakrabarti, A gamified model of design
thinking for fostering learning in children, in Research into Design for a Connected World
(Springer, 2019), pp. 1023–1036.
9. K. Eissen, R. Steur, Sketching (5th printing): Drawing Techniques for Product Designers (Art
Des. Ser, Bis, 2008)
10. G. Goldschmidt, The dialectics of sketching. Creat. Res. J. 4, 123–143 (1991)
11. S.J. Westerman, P.H. Gardner, E.J. Sutherland, T. White, K. Jordan, D. Watts, S. Wells, Product
design: preference for rounded versus angular design elements. Psychol. Mark. 29, 595–605
(2012)
12. S. Chaudhuri, E. Kalogerakis, S. Giguere, T. Funkhouser, Attribit: content creation with
semantic attributes, in Proceedings of the 26th Annual ACM Symposium on User Interface
Software and Technology (2013), pp. 193–202
13. M.E. Yumer, S. Chaudhuri, J.K. Hodgins, L.B. Kara, Semantic shape editing using deformation
handles. ACM Trans. Graph. 34, 1–12 (2015)
14. A.N. Bhatt, R. Bhaumik, K.M. Chandran, A. Chakrabarti, IISC DBox: a game for children
to learn design thinking, in International Design Engineering Technical Conferences and
Computers and Information in Engineering Conference (American Society of Mechanical
Engineers, 2019), p. V003T04A010
15. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H.
Adam, Mobilenets: efficient convolutional neural networks for mobile vision applications.
arXiv Prepr. arXiv1704.04861 (2017)
16. FFmpeg, https://ffmpeg.org/
17. A. Francesconi, BIMP, https://alessandrofrancesconi.it/projects/bimp/
18. M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G.S. Corrado, A. Davis,
J. Dean, M. Devin, Tensorflow: large-scale machine learning on heterogeneous distributed
systems. arXiv Preprint arXiv:1603.04467 (2016)
19. image-slicer 0.3.0, https://pypi.org/project/image-slicer/
Social Interaction-Enabled Industrial
Internet of Things for Predictive
Maintenance
M. S. Roopa, B. Pallavi, Rajkumar Buyya, K. R. Venugopal, S. S. Iyengar,

and L. M. Patnaik
Abstract As Internet of Things (IoT) advances from interacting objects to smart

objects. Further, adopting social networking concepts into IoT leads to social objects.
Following this conception, the Social Internet of Things (SIoT) comprising the
amalgamation of social aspects into IoT was incised. In the recent past, the use
of IoT technologies in the industrial sectors and applications has given birth to a new
research area termed as the Industrial Internet of Things (IIoT) which has attained
notable attention extensively. In machinery-based industries, elements such as bear-
ings, axles, brackets, hangers, etc., are the essential machine element that serves a
variety of purposes in the production environment. Thus, any severity related to it
causing failure is unfavorable. This paper introduces the SIoT concept into the man-
ufacturing industry for predicting the failure of the machine element before it could
occur. This work proposes the construction of the ontological model that assists in the
implementation of the prognostics and health management to predict the remaining
useful life (RUL) of machine element, which is essential in minimizing maintenance
cost and machine downtime, and further limiting disasters.
Keywords Bearing failure · Industrial Internet of Things · Ontological model ·

Predictive maintenance · RUL prediction · Social Internet of Things
M. S. Roopa (B) · B. Pallavi

IoT Lab, University Visvesvaraya College of Engineering, Bengaluru, India
e-mail: roopams22@gmail.com
R. Buyya
University of Melbourne, Melbourne, Australia
K. R. Venugopal
Bangalore University, Bengaluru, India
S. S. Iyengar
Florida International University, Miami, USA
L. M. Patnaik
National Institute of Advanced Studies, Bengaluru, India
662 M. S. Roopa et al.
1 Introduction
The Industrial Internet of Things (IIoT) is an industrial framework where machines

or devices are linked and synchronized utilizing software development tools and
third platform technologies in a machine-to-machine (M2M) and Internet of Things
(IoT) context. IIoT applies unified protocols and architectures for the development
of devices that has better control, seamless communication, and improved interoper-
ability between the machines. It exhibits quality control, asset tracking, sustainable
and green practices, supply chain effectiveness, and supply chain traceability. All this
leads to predictive management in the industrial settings where data from embedded
sensors and connected devices are gathered and exchanged intelligently. In industries
with a high risk of human error, IIoT comes as a boon, where the precision level that
is achieved through IIoT plays a prominent role in industrial predictive maintenance
[1].
Predictive maintenance (PdM) is a strategy that deals with maintenance issues,
given the increasing need to reduce the maintenance cost and machine downtime.
Predictive maintenance relies mainly on sensors to monitor the machine and detect
when a critical piece of machinery is close to failure. In machinery domain, machine
elements plays a crucial role in the production environment, hence predicting its
remaining useful life (RUL) to avoid the severity of failure benefits [2]. To predict
the RUL of a machine element, it is still a significant challenge task, as it depends
on several factors that cannot be measured and analyzed quantitatively. To enhance
the accuracy in the prediction of machine elements, RUL and the factors that affect
its failures become a critical issue and have attracted a large number of prognostic
researchers [3].
The Social Internet of Things (SIoT) is a thriving research field that emerged as a
result of the integration of social networking concepts with IoT. SIoT is an interaction
paradigm where the objects establish social relations with various objects and users to
achieve a common goal [4–6]. Integrating SIoT concept into the Industrial Internet of
Things advances the machines with the formation of social relationships supporting
the human–machine communications and machine–machine interactions. The idea
of utilizing SIoT in industry is named as Social Internet of Industrial Things (SIoIT).
SIoIT conceptualizes the application of social behavior to intelligent assets (refers
to machines or machine parts, the term asset, machine, and resources that are used
interchangeably) that are connected with actuators, sensors, or any identification
technology [7]. In SIoIT, relationships are built among the industrial assets according
to the rules set by the asset relationships.
1.1 Motivations
Predictive maintenance is the recent research topic that determines the probability
of failure of a machine before it could occur and allows industries to progress from
Social Interaction-Enabled Industrial Internet of Things … 663
a repair and replacement model to a predict and fix maintenance model. For a long
time, we have been tending issues in the machinery when they crop up instead of
thwarting them, which immensely affects the expenses and endeavors to fix the assets.
If the prediction of these issues are made based on the historical data, there can be
mitigation in the costs and finer allocation of resources related to the unplanned
maintenance. With the introduction of the Industrial Internet of Things (IIoT), it
enables us to enhance efficiency and production quality.
The proposed PdM in SIoIT is based on the number of social relationships defined
between assets and the actors (human operators). These relationships are used to
predict the failure of the machine element and notify the prediction to the maintenance
or repair personnel. Meanwhile, the sensor readings are examined over time to learn
the relationship of the machine element.
1.2 Contributions
The main contributions of this paper include the following:
1. PdM-SIoIT Architecture: An overview of predictive maintenance-based SIoIT

architecture to detect the failure condition and predict the remaining life of the
assets in industries.
2. Social Relationship Ontology for Predictive Maintenance: We construct an onto-
logical model to describe various relationships in the prediction of multiple factors
that might cause the failure of machine elements based on the obtained data points
collected over time through different sensors. The RUL prediction and the health
management of machine elements are based on continuous monitoring, and if any
discrepancies are observed in the specified threshold, it sends alerts conforming
to predefined rules.
1.3 Organization
The rest of the paper is organized as follows. A brief summary of related works in
Industrial IoT, Social IoT, and ontology model is discussed in Sect. 2. Predictive
maintenance-based SIoIT architecture to detect the failure condition and predict the
remaining life of the machine assets is explained in Sect. 3. The ontology model for
predictive maintenance in SIoIT is discussed in Sect. 4. Section 5, emphasizes on the
implementation of the proposed predictive maintenance-based SIoIT with the aid of
a scenario, simulation setup, and results. The concluding remarks are presented in
Sect. 6.
2 Literature Survey
In this section, we present the state-of-the-art research works in the field of predictive
maintenance in IIoT, SIoT, and ontology-based fault diagnosis methods. Vianna et al.
[8] have developed a predictive line maintenance optimization strategies subjected to
multiple wear conditions in aeronautic systems. It utilizes a Kalman filter technique
to calculate trends of degradation and wear values in future. A case study is carried
out considering field prognostics data of various hydraulic systems for prediction.
The limitation is that the states and variances of the model cannot be combined and
shared among different models.
Wu et al. [9] have presented hidden Markov model based on cluster K-PdM for
RUL prediction and estimation of machinery deterioration supported by multiple key
performance indicators (KPIs). An experimental approach is applied to exhibit its
application on the PHM08 dataset.
Yan et al. [10] have proposed a framework to predict the RUL of key components
in a CNC machine. It exploits single and multiple source data models and employs
a vibration signal to estimate the performance and forecast the remaining life. The
limitation is deep learning techniques to examine big industrial data that were not
adopted.
Wang et al. [11] have explored a cloud-based model for predictive maintenance
using a mobile agent approach. It enables appropriate information utilization, dis-
tribution, and acquisition for improved accuracy and reliability in the prediction
of remaining service life, fault diagnosis, and maintenance scheduling of various
induction motors.
Swamy et al. [12] have developed a health-centric tourism recommender system.
It is developed on an ontology framework, which can advise the availability of food
based on climate attributes using user’s personal favorite and nutritive value. Seman-
tic ontologies connect the difference between various user-profile information. The
real-time IoT based healthcare support system is used to evaluate different food
recommendations.
Cao et al. [13] have proposed an approach that combines fuzzy clustering and
semantic technologies to understand the significance of a machine’s failure based
on the obtained classical data. A case study on a real-world industrial dataset is
developed to estimate the suggested method’s utility and effectiveness.
Wen et al. [14] have provided an ontology model for fault in the mechanical diag-
nosis system. It is based on the resource space model (RSM). Protégé 4 is used to
make an ontological model for diagnosing AC motor defects. However, for mechan-
ical fault diagnosis, this ontological model cannot be used.
Xu et al. [15] have proposed an ontology-based failure diagnosis technique that
overcomes the complexity of learning a sophisticated fault diagnosis information.
It additionally provides a comprehensive procedure for fault diagnosis of loaders.
The ontology model is highly expansible; because the components of the model like
classes, individuals, and properties in the failure diagnosis can be greatly enhanced
and updated.
Jung et al. [16] have proposed a model that captures the spatio-temporal attributes
along with assorted co-usage data varieties to predict the relationship between differ-
ent objects. It explains the entropy and distance-based social strength computations
among objects. It portrays the structure of the SIoT model to IoT that offers navigation
of networks, which includes adaptation and effective objects discovery and services
like human-centric social networks. However, it does not include how it has to deal
with the strengthening of social connections as well as time-related deterioration.
Further, numerous domains’ large datasets can also be used.
3 SIoIT Architecture for Predictive Maintenance

(PdM-SIoIT)
In this section, we present the SIoIT architecture for predictive maintenance and
design an RUL predictor engine, which can be applied to a production environment
in the industries to examine the assets existing condition and predict its remaining
life before entering into practical failure.
The proposed PdM-SIoIT architecture follows the three-layer system constructed
from the base layer, network layer, and application layer, as shown in Fig. 1. The
base layer includes manufacturing assets embedded with identification tags, sensors,
and actuators. The second layer is the network layer, which is used for connectivity
to cellular networks, GPS, WLANs, or Internet as required. The application layer
consists of two sub-layers: interface sub-layer and SIoIT component sub-layer.
The primary constituents of PdM-SIoIT is positioned in the SIoIT component
sub-layer of the application layer. It includes RUL predictor engine, social agents,
Fig. 1 PdM-SIoIT Application Layer

architecture Interface Sub-Layer
Industrial Machine
Applications Interfaces
Service
APIs
SIoIT Component Sub-Layer
Social Relationship Profile
Agents Management Management
Context Performance
Management Management
Data/Metadata RUL Predictor

Management Ontologies Engine
Network Layer
Cellular IEEE
Networks
WLAN ZigBee …… Internet
802.11G
Base Layer
Manufacturing Identification
Sensors …… Actuators
Assets Tags
PrevenƟve
Provides
AcƟon Remaining
Machine Allows Allows
Element use PdM
Threshold PredicƟon
value Allows
monitoring
PredicƟve
Techniques Deep models Predicted
Learning o
Feed
failure
feed Data
analysis Allows
Monitors Health
Maintenance management
Fig. 2 Taxonomy of remaining useful life (RUL) predictor
relationship management, profile management, security and access control, perfor-

mance management, semantic engines, ontologies, and data/metadata repositories.
The interface sub-layer of the application layer consists of industrial applications,
machine interface, human interface, and service APIs to access the IIoT applica-
tions using different devices. The RUL predictor engine is an essential component
of PdM-SIoIT architecture that predicts the remaining life of the machine assets in
industries. The taxonomy of RUL predictor is illustrated in Fig. 2. The operation of
the assets varies for different materials that provide preventive action and its thresh-
old limit to predict the remaining useful life. The predictive methods are applied to
collect and analyze the data from sensors embedded on the machine elements. The
data is collected based on different parameters such as temperature, vibration, load,
and lubrication level. When the parameters exceed the specified threshold limit, the
machine’s failure is predicted in the mere future. Once a prediction is made, the
necessary action will be taken to avoid failure in the future. The health management
of the assets are performed when a failure is recognized; if the health state indicates
a practical failure, then request for an operation-based maintenance (OBM) is sched-
uled. Ultimately, when the maintenance interruption is finished, new data obtained
in such interruption is added or updated in the assets data and in the parameters of
the predictive techniques to proceed with the prognostics health management cycle.
4 Ontology model for PdM-SIoIT
The ontological model is a structure that defines relationships between classes and
subclasses existing in a domain. It analyzes data to transfer information for granting
decisions regarding the RUL of a mechanical element, before failure occurs. Each
asset in the industry is associated with software agents that fulfill several computing
essentials of the machine elements, such as managing relationships among the assets
and the human operators [17]. We define various relationships that predict the failure
of the machine and notify the prediction to the actors like maintenance or repair
personnel who interact with the devices. Various relationships can be established
among the assets and actors, which are enumerated as follows.
1. Coercion object relationship (COR) enables machines to intimate the maintenance
personnel that machine element has crossed its specified fatigue limit, which might
cause fatigue failure.
2. Torrid object relationship (TOR) is established between a temperature sensor and
the maintenance team to take immediate action when the temperature exceeds its
maximum limit, to avoid overheating of machine elements.
3. Corrosion object relationship (CROR): When the machine elements are exposed
to corrosive fluids or corrosive atmosphere, it forms a CROR with saddle-type
eddy current sensor, and then the information is transferred to the technician.
4. False Brinell object relationship (FBOR): When excessive external vibration
occurs, a false Brinell object relationship is established between the vibration
sensor and the maintenance team to avoid false brinelling on the machine ele-
ments.
5. True Brinell object relationship (TBOR): When the load exceeds the elastic limit
of the ring material, a true Brinell object relationship is established between the
vibration sensor and the maintenance team to avoid true brinelling on the machine
elements.
6. Spalling object relationship (SOR) is formed between the UV sensor and techni-
cian when there is a marked increase in the vibrations generated by the machine
elements.
7. Back freight object relationship (BFOR): When the load is induced in the opposite
direction, a BFOR is established between an infrared sensor and maintenance
personnel to avoid early failure of machine elements.
8. Withered object relationship (WOR) is established between smart oil sensors and
lubrication experts when the required lubrication is not supplied to the machine
element, which might cause excessive wear of balls and cages.
9. Defile object relationship (DOR): When there is contamination in the operating
area, a DOR is formed between the revolutions per minute (RPM) sensor and the
maintenance team to avoid machine element dents.
The ontological model for predictive maintenance and its components is as shown
in Fig. 3. The components of the model includes machines, sensors, machine ele-
ment, actors, analyzer, and RUL predictor. The model is constructed in the industrial
domain, studying machine elements such as bearings, brackets, axles, hangers, etc.
The RUL predictor utilizes the information from other components and thereby pre-
dicts the remaining life of key machine element much ahead the actual failure occurs.
In the ontology model, the object properties represent the links between components;
and the data properties describe the values of the components. The ontology elements
such as object and data properties for predictive maintenance is detailed in Table 1.
Effect Mode Specification Data Team

Measurement
:hasRelationship
Actors
Type Relationships
COR
T
Sensor Machine
isEmbeddedWith FBOR TBOR
Alert
Message
Machine RUL
Element Analyzer :hasRelationship
Predictor
Minimum Maximum Measurement Threshold

Analyzed Data
Value Value unit Value
Fig. 3 Ontology model for predictive maintenance
5 Implementation of PdM-SIoIT
In this section, we present an overview of the feasible implementation of the designed

PdM-SIoIT architecture by an illustration under multiple wear conditions of assets
and the simulation setup.
5.1 A Scenario
A bearing is a machine component that minimizes friction among moving parts,

forcing relative motion to merely the desired motion. Bearing facilitates free rotation
around the fixed axis of the moving parts or a free linear movement. The bearing
usually does a pretty excellent job of managing the moving parts; however, when
a bearing fails, the outcomes can be catastrophic [18]. Bearing fails unexpectedly
and untimely, despite planning and maintenance. The primary reasons of damage
and unexpected bearing failure for all bearing types are overheating, overloading,
improper lubrication, excessive external vibration, contamination, etc. [19], and they
are discussed as follows:
1. Overloading: When the bearing is at a speed of 2200 rpm, the maximum load
it can accept is approximately 47,620 N; if the load exceeds, there will be the
formation of irregular spiderweb-like cracks and craters in the overlay surface
which causes bearing failure.
Table 1 Object and data properties for ontology model

Properties Domain Description
Object properties hasMode Machine Represents the mode of machinery operation
hasEffect Machine Identifies the effect of failure
hasMeasurement Sensor Measures the threshold values
hasCause Machine element Determines the cause of an element failure
failure
hasType Sensor Represents different types of sensors used
isPartOf Machine/machine Identifies the part that might get affected
element
isIdentifiedWith Symptoms/effects Rectifies the variations in normal routine
operations of a machine
hasSpecification Machine Provides the machine specifications
hasAnalyzedData Analyzer Analyzes the data obtained from sensor
hasData Actor Represents the available predicted data
hasTeam Actor Represents the different types of actors
isEmbeddedWith Machine Represents that the machines are embedded
with different types of sensors
hasRelationship Machine, actors, Represents the relationships between actors and
and analyser machine elements
dependsOn Machine Machine depends on predictor for RUL
prediction
monitors RUL predictor Monitors the threshold value for RUL prediction
Data properties Minimalvalue Machine element The minimum value under which an element
can operate normally
MaximumValue Machine element Represents the maximum value which causes
failure of a machine element
MeasurementUnit Machine element Every measurement has different units
AlertMessage Sensor Alerts when specific threshold is reached
FatigueLimit Machine element The load an element can accept
MaximumTemperature Machine element Maximum temperature condition an element
can reach
VibrationValue Machine element The speed at which an element can operate
ContaminationLevel Machine element The limit of contamination which does not
affect an element operation
LubricationLevel Machine element The minimum lubrication level an element
should maintain
Loadtype Machine element The type of load on machine element
BrinellEffect Machine element The effect of different loads on an element is
measured
Threshold value RUL predictor Life span of a machine element
2. Excessive external vibration: When bearings in a machine experiences vibration

above 3200 rpm, the bearing might face true and false Brinel, which might cause
bearing failure due to excessive vibrations.
3. Contamination: A bearing that is typically contaminated beyond 1141 MPa causes
denting of the bearing, and balls emerged from high vibration and wore, leading
to bearing damage.
4. Overheating: Temperature above 400 ◦ F anneals the ring and ball materials of
bearings. Excessive heat can cause serious problems, resulting in machine failure
and hours of downtime.
5. Improper lubrication: Minimum lubrication level must be maintained at 1.97 µm,
and the bearing is serviced either according to a schedule or when the lubricant
goes below the specified minimum level.
6. Life span: The life of a bearing depends on the load and the speed at which it
operates. Most of the bearings last approximately 8–12 years.
5.2 Simulation Setup and Results
As IoT enabled industrial setting has not been extensively implemented to date,
and the experiment of such an ecosystem is deployed on Node-RED [20], an open-
source simulation, and modeling tool. Table 2 presents the complete information of
the varieties of sensors deployed along with the threshold limit exerted for each of
the bearing failure types. The performance of the bearing is monitored continuously;
and if any discrepancy is observed in the specified threshold, the system sends alerts
to the dedicated maintenance personnel regarding the fundamental specification of
action. Figure 4 depicts the example of alerts generated in the event of any exceptional
behavior with the help of NodeRED simulator.
Simultaneously, besides creating alerts, the system also regularly predicts the RUL
of the bearing, so that the maintenance can be planned in advance. The dataset used
in this experiment is from prognostics and health management (PHM08) [21], which
includes time series data. The dataset is separated into training and test datasets.
It comprises twenty-one sensor measurements and three operational settings of an
Table 2 Sensor description for bearing threshold monitoring

Sensor Description Unit Threshold limit
Temperature sensor Temperature value Fahrenheit (◦ F) 400
Pressure senors Fatigue limit Newton (N) 47,620
Smart oil sensor Lubrication level Micrometer (µm) 1.97 (min)
Vibration sensor External vibration Revolutions per 3600
minute (rpm)
RPM sensor Contamination MegaPascal (MPa) 1141
Fig. 4 Alerts generated using NodeRED simulation tool
Table 3 Results
Model Precision Recall F1-Score Accuracy
Deep learning 0.960000 0.96 0.960000 0.978495
model
Template model 0.952381 0.80 0.869565 0.940000
engine with varying degrees of initial wear and manufacturing variation. We develop
a data-driven method to predict the RUL of a machine that deteriorates from an
unknown primary state to failure. The proposed method uses the data from the train-
ing set, and then, the performance of the RUL prediction using testing data is eval-
uated. The solution employs an advanced, recurrent neural network to improve the
prediction accuracy of the bearings remaining life. To reliably evaluate the fault
diagnosis of a bearing, along with accuracy, we compare the other metrics such as
precision, recall, and F1-Score as given in Table 3.
We see that the deep learning model results are better than the template when
analyzing the above test results in the predictive maintenance template. It should be
noted that the dataset used here is minimal, and deep learning models are known to
offer excellent results with large datasets, so more massive datasets should be used
for a more fair comparison.
6 Conclusions
In this paper, the inclusion of social interactions into the industry settings for pre-
dictive maintenance is envisioned. A novel architecture for predictive maintenance
in SIoIT systems is introduced to monitor future failure by keeping the companies
prepared and allows them to plan even before the failure occurs. An ontological
model is constructed to describe various relationships in the prediction of multiple
factors that might cause an asset failure. To implement the proposed PdM-SIoIT
architecture, the failure condition of an essential machine element, i.e., a bearing

is discussed as an example. The simulation model is developed to detect the failure
condition and predict the remaining life of a bearing. Further, we intend to implement
the recommended scheme in the actual manufacturing settings and exploit the other
deep learning techniques to improve prediction accuracy.
References
1. F. Civerchia, S. Bocchino, C. Salvadori, E. Rossi, L. Maggiani, M. Petracca, Industrial Internet

of Things monitoring solution for advanced predictive maintenance applications. J. Ind. Inf.
Integr. 7, 4–12 (2017)
2. K. Zhu, T. Liu, Online tool wear monitoring via hidden semi-Markov model with dependent
durations. IEEE Trans. Ind. Inform. 14(1), 69–78 (2017)
3. Y. Wang, Y. Peng, Y. Zi, X. Jin, K.L. Tsui, A two-stage data-driven-based prognostic approach
for bearing degradation problem. IEEE Trans. Ind. Inform. 12(3), 924–932 (2016)
4. M.S. Roopa, S. Pattar, R. Buyya, K.R. Venugopal, S.S. Iyengar, L.M. Patnaik, Social Internet
of Things (SIoT): foundations, thrust areas, systematic review and future directions. Comput.
Commun. 139, 32–57 (2019)
5. M.S. Roopa, D. Valla, R. Buyya, K.R. Venugopal, S.S. Iyengar, L.M. Patnaik, SSSSS: search
for social similar smart objects in SIoT, in 2018 Fourteenth International Conference on Infor-
mation Processing (ICINPRO) (2018), pp. 1–6
6. M.S. Roopa, A. Siddiq, R. Buyya, K.R. Venugopal, S.S. Iyengar, L.M. Patnaik, Dynamic
management of traffic signals through social IoT. Procedia Comput. Sci. 171, 1908–1916
(2020)
7. H. Li, A. Parlikad, Social Internet of Industrial Things for Industrial and Manufacturing Assets
(2016)
8. W.O.L. Vianna, T. Yoneyama, Predictive maintenance optimization for aircraft redundant sys-
tems subjected to multiple wear profiles. IEEE Syst. J. 12(2), 1170–1181 (2017)
9. Z. Wu, H. Luo, Y. Yang, P. Lv, X. Zhu, Y. Ji, B. Wu, K-PdM: KPI-oriented machinery deteri-
oration estimation framework for predictive maintenance using cluster-based hidden Markov
model. IEEE Access 6, 41676–41687 (2018)
10. J. Yan, Y. Meng, L. Lu, L. Li, Industrial big data in an industry 4.0 environment: challenges,
schemes, and applications for predictive maintenance. IEEE Access 5, 23484–23491 (2017)
11. J. Wang, L. Zhang, L. Duan, R.X. Gao, A new paradigm of cloud-based predictive maintenance
for intelligent manufacturing. J. Intell. Manuf. 28(5), 1125–1137 (2017)
12. V. Subramaniyaswamy, G. Manogaran, R. Logesh, V. Vijayakumar, N. Chilamkurti, D. Malathi,
N. Senthilselvan, An ontology-driven personalized food recommendation in IoT-based health-
care system. J. Supercomput. 75(6), 3184–3216 (2019)
13. Q. Cao, A. Samet, C. Zanni-Merk, F.D.B. de Beuvron, C. Reich, An ontology-based approach
for failure classification in predictive maintenance using fuzzy C-means and SWRL rules.
Procedia Comput. Sci. 159, 630–639 (2019)
14. H. Wen, Y. Zhen, H. Zhang, A. Chen, D. Liu, An ontology modeling method of mechanical fault
diagnosis system based on RSM, in Proceedings of the 2009 Fifth International Conference
on Semantics, Knowledge and Grid (2009), pp. 408–409
15. F. Xu, X. Liu, W. Chen, C. Zhou, B. Cao, Ontology-based method for fault diagnosis of loaders.
Sensors 18(3), 729 (2018)
16. J. Jung, S. Chun, X. Jin, K.H. Lee, Quantitative computation of social strength in Social Internet
of Things. IEEE Internet Things J. 5(5), 4066–4075 (2018)
17. N. Gulati, P.D. Kaur, Towards socially enabled internet of industrial things: architecture, seman-
tic model and relationship management. Ad Hoc Netw. 91, 101869 (2019)
18. A. Harnoy, Bearing Design in Machinery: Engineering Tribology and Lubrication (CRC Press,
2002)
19. F.T. Barwell, D. Scott, Paper 10: the investigation of unusual bearing failures. Proc. Inst. Mech.
Eng. 180(11), 277–317 (1965)
20. A. Javed, Complex flows: node-RED, in Building Arduino Projects for the Internet of Things,
Apress, Berkeley, CA (2016), pp. 51–73
21. F.O. Heimes, Recurrent neural networks for remaining useful life estimation, in Proceedings
of the 2008 International Conference on Prognostics and Health Management (IEEE, 2008),
pp. 1–6
Emerging Trends in the Marketing
of Financially Engineered Insurance
Products
Sunitha Ratnakaram, Venkamaraju Chakravaram, Nitin Simha Vihari,

and G. Vidyasagar Rao
Abstract Insurance plays a fundamental role in the economies of nations world-

wide. Protecting the future earnings of individuals, companies from uncertainties,
and risks fastens a nation’s gross domestic product (GDP) [1]. The second decade
(2D) of the current century (2D/twenty-first century) has witnessed a revolutionary
disruptive technological advancement, innovative strategies, and financially engi-
neered models in the marketing of insurance products across the insurance sector
worldwide. Some of these technologies and advancements are named as “FinTech,
InsurTech, Blockchain Technology, Cryptocurrency, Robotics, Cloud Technologies,
Data Science, Data Analytics, Big Data, Financial Engineering, ICTs (Informa-
tion and Communication Technology), IoT (Internet of Things), AI (Artificial Intel-
ligence), Machine Learning, Mobile Phones, Social Media sites, and Drown Tech-
nologies” [2]. To list and to explain these major disruptive technologies of the sector
is the main purpose of this research work. The main focus is to study the specially
designed financial engineering models in the process of insurance marketing. To
conclude, the detailed role of financial engineering applications in the design, devel-
opment, and execution of marketing models in the insurance sector is discussed.
Descriptive-cum-exploratory research methodology is used.
Keywords Emerging trends · FinTech · Financial engineering · Insurance ·

InsurTech)
S. Ratnakaram (B) · V. Chakravaram

Jindal Global Business School, O. P. Jindal Global University, Haryana 131001, India
e-mail: sratnakaram@jgu.edu.in
V. Chakravaram
e-mail: vchakravaram@jgu.edu.in
N. S. Vihari
BITS Pilani, Dubai Campus, Dubai, UAE
e-mail: nitinvihari@dubai.bits-pilani.ac.in
G. Vidyasagar Rao
UCCBM, Osmania University, Hyderabad, India
e-mail: gvsraog@gmail.com
676 S. Ratnakaram et al.
1 Introduction
Technologies are driving most of the sectors all over the world through their innova-
tive apps, models, and online portals. Particularly, in banking and insurance sectors,
technologies are disrupting for a decade. Every year, new things and new and inno-
vative style of operations are taking place in these sectors in the form of FinTech and
InsurTech.
2 Financial Engineering
Financial engineering involves the design, the development and the implementation of inno-
vative financial instruments and processes, and the formulation of creative solutions to
problems in finance.–John Finnerty
Financial engineering (FE) is a process of creating innovative financial models

and products or solving the existing financial problems of the sectors by using tools
and formulas from econometrics, economics, financial mathematics, statistics, and
computer sciences. The insurance sector is using FE applications in the creation of
innovative products and models. The given Image 1 is showing the same combina-
tion of the listed sciences which is financial engineering. This sector is also using
financial engineering applications in the process of marketing of insurance policies.
Particularly in the designing process of customized insurance policies, customized
riders with innovative features like short time coverages, specific event or specific
purpose coverages etc. [3].
Image 1 Ref: https://sofe.iiqf.org/pages/what-fe.html

Emerging Trends in the Marketing … 677
3 Objectives and Hypotheses of Research
1. What are the emerging trends in the marketing of financially engineered insurance
products?
2. What are evidence exists on the benefits of using financial engineering instru-
ments as tools for the quick and satisfactory services rendered to the customers
The following hypothesis is proposed:

• H0: There are positive emerging trends in the marketing of financially engineered
products in the insurance sector.
Image 2 shows various steps in the designing and development process of finan-
cially engineering products. In this process, instruments or statistics, financial math-
ematics, econometrics, financial management models, computer sciences, informa-
tion and communication technologies (ICTs), big data, blockchain technology, data
analytics, data science, Internet of things (IoTs,), robotics, automation technologies,
and drones are used.
Image 3 shows the first generation of the financially engineering process in the
insurance business. This image explains how the premium is being distributed among
various portfolios to maximize the financial benefits to the customers. This typical
Financial Engineering Process (1)
Process Identification of need (2)
Design of Insurance Product (3)
Complex Model of Building Exercise (4)
Testing of Insurance Product (5)
A perfect product based on the entire Exercise (6)
Pricing of the Product (7)
Restructuring of the insurance product (8)
Test Marketing (9)
Launching of Financially Engineered Insurance Product (10)
Image 2 Financial engineering process in the design, development, and launching of insurance
products
Image 3 Financial engineering in life insurance business in 1950
image or financial engineering process in the life insurance business belongs to the
period between 1950 and 2000.
Image 4 shows the financial engineering process in the life insurance business after
the privatization of the insurance sector in India. We can take an image for the first
decade (2000-2010) of the current century. Insurance companies started participating
in the share markets. The major portion shifting to the marketing investments in the
Image 4 Financial engineering in life insurance business in India–LPG era (2000)

Image 5 Financial engineering in life insurance business in India–ULIPs era (2010)
form of various securities, bonds, and NAVs has been started during this period only.
So many customers who invested in Unit Linked Insurance Policies (ULIPs) incurred
heavy losses since the returns depend on the share market conditions.
Image 5 represents the present era of the financial engineering process in the
insurance business. See in the picture once, how the paid premium by the customer
is being divided, shifted among multiple portfolios, investment zones, risk coverage
premiums, and as rider premiums.
4 Emerging Trends in the Marketing of Insurance Products
As per the recent Accenture report, 86% of insurance companies worldwide are
focusing on the use of technologies in every stage of the insurance business. They
believe that only they can be in a competitive edge in this technology world [4].
4.1 Artificial Intelligence (AI)
AI technologies are using in the marketing of financially engineered products while

issuing the policies to the customers. They are using AI technologies for facial
recognition which will help the insurers in their underwriting process of the newly
issuing policies. This AI-enabled app is also helping the insurers in the underwriting
process to estimate the life span of the prospect applicant by detecting his/her habits
using gamification techniques. Hence, they can easily calculate the risk levels and
threats of the prospect to determine his/her required premium payable to the company
(Pragya Saxena 2019).
4.2 Robotic Process Automation (RPA)
Particularly, RPA is happening in the claim settlement process of insurance busi-

ness and data management of the insurance companies. When the companies are
achieving the better claim settlement ratios during business by using RPA automa-
tion technologies, goodwill of the company will increase and indirectly helps the
marketing team to market their products to the prospect customers. RPA also helps
insurers by simplifying their data management processes at high-speed computer
servers and systems (Pragya Saxena 2019).
4.3 Machine Learning
Underwriting is one of the important activities in the marketing process of insurance.

To give exact risk coverage, riders, etc., consider the prospect’s profiles. Machine
learning technologies have improved this underwriting process in the companies to
generate an exact quotation of premium by considering the detailed profile of the
prospects. These quotations include premium to be paid, duration of risk coverage,
sum assurance value, maturity benefits, additional interim benefits, interim and
maturity bonuses etc. Machine learning technologies also help insurers to list the
lapsed policies and give strategies in renewing the process of that policies. Hence,
in the prediction of premiums, risk coverages, and the process of claim settlement
processes, machine learning technologies are playing a greater role in the sector
(Pragya Saxena 2019).
4.4 Augmented Reality/Virtual Reality (AR/VR)
These technologies are helping both clients and customers in their claims settling
process. Particularly, when the risk/event occurred in a remote village or place, these
AR/VR technologies will help both the stakeholders. By reconstruction of the scene
of accident using the augmented reality. Hence, this kind of live video clips will help
the underwriters to estimate the loss to be stele to the beneficiary (Pragya Saxena
2019).
4.5 Big Data Predictive Analytics (BDPA)
Analytical insights, risk chances, and risk coverage productions are the outcome of
predictive analytics. Big data predictive analytics helps insurers for their smooth and
effective marketing decisions. BDPA helps the marketing team to understand and to
analyze the consumer behaviors also to identify their patterns. BDPA also helps the
underwriters and claim management team (Pragya Saxena 2019).
4.6 Blockchain Technology (BCT)
Insurance companies started using this BCT while analyzing the clients profile to
decide his/her risk factors, insurance coverage to be allowed, risk riders to be allowed,
etc. With the effective use of BCT, insurers can easily stop the fraudulent practices of
some fraudulent customers and organizations particularly in the claims management
process. Since the data is available in the decentralized ledger systems (BCT) about
the customer, no need to submit the common documents again and again to avoid
duplication and to minimize the repetition while applying for a new policy or while
applying for surrender or the policy or while submitting application for the claim
settlement process (Betsy McLeod 2019).
5 Other Emerging Technologies Using by the Marketing

Team
The marketing teams always try to present differently about their products and compa-
nies to grab the attention of prospects. They differentiate the business comparing
with other service providers to make him/her accept their proposals. Teams mostly
follow or use the following techniques (Betsy McLeod 2019). Communication of
brand through videos: Marketing professionals are keeping the video testimonials
of the existing customers. Companies also try to convey their commitment through
these videos to customers and introduce their company Web site(s) or blogs to the
customers [5]. They are habituating the prospects to visit the Web sites, to down-
load their required information. Insurers are investing in search engine optimization
(SEO) with mobile-friendly features and also with mail-friendly features [6].
Content should be in a quality manner, and it should be in crisp and easy to
the user in understandable simple language and simple sentences. Companies are
keeping contact information in all the folders of the Web site and pages of the Web
site. Insurers are explaining the path and their action plan by explaining every step
clearly in case if customers wish to submit their online applications then and there.
Insurers are keeping chat boxes to guide the online prospects in their application
submissions process or their decision-making process. Companies also use sensible

language based on the age, gender, and educational levels of the customers [7].
Some insurers are using colors and images to convey the information while they are
communicating with the customers in the chat boxes. Marketing teams of insurers
are creating blogs, explaining the past performance of those policies, with FAQs,
other analytical statements showing the yields and financial benefits that they are
going to offer their prospects [8].
Marketing professionals are keeping reviews of the Web site visitors and
customers also on the Web site. This will help those who wish to see the reviews
of the company. Here, professionals are using statistical tools and analytical tools to
display their reviews in different variations (Betsy McLeod 2019).
6 Social Media
Marketing professionals are using effectively various social media platforms to intro-
duce their products and features. Particular special marketing teams are disseminating
information in the form of brochures and success stories. Also announcing their
special offers, special plans through their benefit analysis statements using social
media platforms. Some of them are WhatsApp, Facebook, Twitter apps, etc. Using
these social media platforms, marketing professionals are easily doing new busi-
ness with their existing customers [9]. Also, they are using these platforms to remind
their renewal date, maturity date, claims procedure, etc., which also causes the health
premium turnover (Betsy McLeod 2019).
7 Discussion
Based on the above discussion, it is proved that the following listed technologies
are playing a major role in the emerging trends in the marketing of financially engi-
neered insurance products. These are trendsetters and associated with the financial
engineering process, and it is proved that there are many instruments from the listed
sciences that are involved in the design, development, and launching (marketing)
process of financially engineered products in the insurance business. Hence, it has
been declared that there is a significant role of these instruments in financial engi-
neering. All these instruments are from statistics, financial mathematics, economet-
rics, Internet of things (IoT), Information and communication technologies (ICTs).
Also from FinTech, InsurTech, Blockchain technology, Data sciences, Data analytics,
Big data, Automotive technologies, Robots, Drones etc.
8 Limitations and Scope for Further Research
Since the research study used is descriptive-cum-exploratory research methodology,

we could not get time and space to go for the collection of primary data from the
key stakeholders of the topic. All the articles and research papers reviewed are up
to 100, and we took 15 most appropriate references to compile this research output.
There is a huge scope in this research field. Since technologies are playing a greater
role in the banking and insurance sectors, particularly while designing the innova-
tive and creative financially engineered financial and insurance models, companies
are using these latest technologies. Upcoming researchers can study the role of the
following technologies into their consideration. Role of blockchain technologies in
the financial engineering process of banking and insurance sector products. Role of
cryptography in the financial engineering process of the same sectors. Role of social
media sights and their analysis in the marketing of financially engineered products.
Role of consumer feedback sights on the marketing process of financially engineered
insurance products.
9 Conclusion
Based on the above discussions and the listed facts, researchers are concluding that
emerging trends in the marketing of financially engineered insurance products are
doing business progressively when compared with traditional business or marketing
activities. These emerging trends are laid the strong research path to the upcoming
researchers as we listed above. Hence, this research study has been concluded with
this above, findings, discussions, suggestions, and scope for further research.
References
1. CII Report, Competing in a New Age of Insurance: How India is Adopting Emerging Technologies
(June 2019)
2. Insurance Industry; 12 trends (May 2019). https://www.oneincsystems.com/blog/insurance-ind
ustry-12-trends-for-2019
3. School of financial engineering, what is financial engineering, https://sofe.iiqf.org/pages/what-
fe.html. Detailed and minor references available on request
4. Marek Seretny, Sustainable marketing-a new era in the responsible marketing development
(January 2012)
5. Recent trends in insurance sector, https://nios.ac.in/media/documents/vocinsservices/m1-1f.pdf
6. Scale Focus, 9 insurance industry technology trends in 2020 [Updated] (October 2018). https://
www.scalefocus.com
7. Insurance business overview and the move towards the digital era by digital ventures (November
2018). http://www.dv.co.th/blog-en/digitizing-ins-view/
8. S. K. Chaudhury, S.K. Das, Trends in marketing of new insurance schemes and distribution: an
empirical study on Indian life insurance sector (December 2014)
9. Where do insurers stand as they enter 2019? 2019 insurance outlook: growing economy bolsters
insurers, but longer-term trends may require transformation. https://www2.deloitte.com/us/env/
pages/financial-services/articles/insurance-industry-outlook.html
Attitude of the International Students
Towards Integrating ICT in Foreign
Language Learning: A Case Study
Khushboo Kuddus and Nafis Mahmud Khan
Abstract The study investigated the attitude and the awareness of the foreign
students about the role of Information and Communication Technology (ICT) in
learning English and in overcoming the language barriers. The purpose of the study
was to explore the use of ICT in enhancing the English language skills among the
international students at a university of India. In order to collect the data, a survey
was conducted and questions based on the objectives of the study were asked to
the international students. The findings of the study suggest that students integrate
both synchronous and asynchronous mediums like mobile applications, watching
movies with subtitles, games, and YouTube to enhance their language proficiency.
Moreover, the students prefer learning the language passively at their own pace and
space without being actively involved in the learning process. Further, the interna-
tional students at the university have positive attitude towards the use of technology
to overcome foreign language barrier.
Keywords Information and communication technology · Technology-enhanced

language learning · Social networking sites · Mobile-assisted language learning
1 Introduction
The remarkable growth of Information and Communication Technology (ICT) has

made English, a global language. English is the language of technology and the
language of the Internet. The excessive use of the Internet in almost all the fields has
expanded the demand for English manifolds and at the same time has also made the
native English language easily accessible across the world. It plays a significant role
in being a lingua franca and a link language for the people sitting at the different
K. Kuddus (B) · N. M. Khan

School of Humanities (English), KIIT Deemed to Be University,
Bhubaneswar 751024, Odisha, India
e-mail: khushboo3133@gmail.com
N. M. Khan
Division of Research, Daffodil International University, Dhaka, Bangladesh
686 K. Kuddus and N. M. Khan
corners of the world. Further, with the growth and access of the Internet, people
have become more exposed to the real, authentic and native English language which
earlier was quite difficult.
A significant integration of ICT in education and specifically in language teaching
and learning has also been observed in the recent years. According to the studies,
it has been found that the integration of technology increases learners’ motivation
and at the same time engages the learners in the learning process. It provides the
language learners with an authentic learning environment and exposes the learners
to the authentic native English language. It encourages autonomous learning which
supports learning at one’s own pace and space. Technology-enhanced language
learning can provide learners a rich source of information and excellent materials to
develop language proficiency.
Though technology has a significant impact on language learning, does it suffice
to overcome the foreign language learning barriers of the non-native speakers of
English? As per the best possible information of authors, no work has been recorded
in open literature about the attitude of international students towards the use of ICT
in enhancing the foreign language skills. This motivates us to study the awareness of
ICT among international students of the university and to integrate it in overcoming
the foreign language barriers.
It has been observed that the foreign students from the Middle East, Ethiopia,
Syria, Nepal, and others who come to study in India, find it difficult to communicate
with the other native and foreign students due to the lack of proficiency in English.
Observing the benefits of integrating technology in language learning and the diffi-
culties faced by the international students, the paper attempts to examine the attitude
of the international students towards using ICT for overcoming the language barrier
and its integration in enhancing language skills.
2 Literature Review
Technology is playing a significant role in language learning across the world

at different levels. Undoubtedly, the exponential expansion of Information and
Communication Technology (ICT) has made English language learning easier and
autonomous. The incorporation of technology in language learning makes the
teaching and learning of foreign language easy, active and interesting [1]. In this
context, Warschauer asserts that one significant advantage of incorporating tech-
nology is to increase the learners’ motivation as they are likely to spend more
time on tasks when doing it on computer [2]. He further states that Information
and Communication Technology provides the language learners with an authentic
learning environment and fosters the enhancement of language skills like listening,
reading, writing and speaking at the same time [3].
Undoubtedly, the tech savvy generation of today is certainly aligned towards
technology enhanced language learning as they find it quite interesting and moti-
vating. The technology-enhanced language learning (TELL) and mobile-assisted
Attitude of the International Students Towards Integrating ICT … 687
language learning (MALL) foster active learning in the learners, and they also enjoy
the liberty of learning at their own time, pace and space [4–6]. Clark [7] states that
using social networks can generate meaningful output and stimulate learners’ interest
through the utmost encouragement. In addition to this, interaction through Internet-
mediated intercultural foreign language education tools such as tandem learning,
pen-pal communication, collaboration, excessively gets a tremendous response from
the language learner [8]. Hence, using social networking sites (SNS), learners can
develop their language proficiency by interacting with multi-communities, through
socialization process [9].
Moreover, using social networking sites in English language learning, students
get more exposed to the target language and its native speakers, learn the authentic
language collaboratively, become more motivated, autonomous and more accus-
tomed to native speakers’ accent and pronunciation as they learn the target language
in real and authentic situation [10].
Mobile-assisted language learning (MALL)—an emerging advanced educational
technology that provides personal and learner-centred learning opportunities through
flexible practices—has become an important trend in EFL learning [11]. For devel-
oping foreign language speaking skills, mobile phone application with automatic
speech recognition could be the best possible way as it helps the learners to improve
pronunciation skills through practice [12]. It has become the most popular and easily
affordable device for learning the English language as it is widely used not only
among younger generation but by individuals of all age groups and of all finan-
cial backgrounds. There are a number of mobile applications for learning English
vocabulary, grammar and downloadable dictionaries like Merriam Webster with
embedded pronunciation [10]. In addition, online games can support and improve
various vocabulary fields and also give valuable feedback [13]. Yang 2013 adds that
students increase their language awareness by using on-site games and by discus-
sion in different social and cultural contexts [14]. Gamification-game element to
non-game context language learning is a standout platform for learners’ preferences
among other mobile language learning applications [15]. Telegram apps and movie
clips invariably play a decisive role in the enhancement of foreign language learning
[16].
The computer-assisted language learning (CALL) program offers students with
a variety of items for learning the English language such as grammar, vocabulary,
pronunciation, etc. Designed teaching and learning practice videos incorporated with
pedagogical contour are another great wonder of technology that urges learners to be
diligent in using materials that are authentic, memorable, and realistic [15]. Video-
based instruction on YouTube is also an outstanding platform for language learning.
A study shows that YouTube enhances students’ ability to engage themselves in
learning and getting contentment in higher education [17, 18]
A study was done on “Exploring the Effects of Telegram Messenger and Movie
Clips on Idiom learning of Iranian EFL learners” where 59 students were assigned
to two experimental and one control group. Finally, the results revealed that both
the telegram and movie clips groups showed outstanding performances and they
can be considered as effective educational tools for learning idioms [19]. Another
study on “Social networking for language learners: Creating meaningful output with
Web 2.0 tools” shows that how technology-based social networking sites provide
language learners with vast resources of authentic written, audio, and video materials
to supplement lessons and how the advanced technology learners of a language can
easily interact with their peers through a meaningful practice that foster language
acquisition and motivation [20].
A study on students’ attitudes towards use of technology in language teaching and
learning in the UK was conducted and found that students were highly interested
in using email and internet for study purposes [21]. Investigation on “Students’
Critical Thinking Skills, Attitude to ICT and Perception of ICT Classroom Learning
Environments Under the ICT School Pilot Project in Thailand” showed that many
students improved their critical thinking capability and developed the degree of
attitude towards technology [22].
From the above review, it is believed that the study is appropriate enough with
the time and the context as it focuses on the attitude of international students of 15
different countries toward technology and their awareness of the effectiveness of
using technology to overcome the foreign language barrier and to develop language
proficiency.
3 Methodology
A questionnaire was designed after conducting a comprehensive review of the related

literature. A survey of the international students at a university of India was conducted
to examine the awareness of using technology in language learning among the foreign
students from 15 different countries, viz. Ethiopia, Yemen, Colombia, Somalia, Iran,
Syria, Nepal, Bhutan, Afghanistan, Togo, Chad, Maldives, Vietnam, Ghana and Saudi
Arabia.
3.1 Participants
For the purpose of the study, 100 foreign students from the academic years 2017–
2018, 2018–2019 and 2019–2020 were invited to participate in the survey. Total 88
responses were obtained out of which 58 were B. Tech., 15 PG and 15 Ph.D. students.
3.2 Researching Tools
This study applies a quantitative approach with close-ended questionnaire. For

sections II and III, a five-point Likert scale was designed for foreign students to
find out their attitude, opinions and the integration of technology in learning English
language. The questionnaire consisted of 20 items that showed the extent of agree-
ment of students, from 1. strongly disagree to 5. strongly agree. To elicit data for
the study, the questionnaire was divided into three sections: (a) participants’ back-
ground, (b) attitude to technology and (c) functional language learning experience
using technology.
3.3 Procedure
This study was conducted by survey method. The invitation clearly explained the
objectives of the research study and also guaranteed the anonymity of the participants.
The participants had the right to not participate or withdraw from the study at any time.
The survey was conducted offline and so the questionnaires were given in person
to the respondents. The questionnaire was administered among 100 international
students of the university’s international boys hostel. Out of 100 samples for the study
in hand, 88 responses were obtained. The data of the responses were quantified and
were analysed by SPSS 20 program. Data calculated with the SPSS program were
expected to determine the mean and standard deviation which were used to interpret
the data and to draw conclusion.
4 Findings
The obtained data from the questionnaire was examined in order to investigate the
awareness of use of technology in English language learning. Since the young learners
are not only tech savvy, but they are also using technology for education and for
language learning in particular; therefore, it was considered important to analyse the
awareness and the attitude of the foreign students about the incorporation and impact
of technology in language learning.
4.1 Students Attitude Towards Technology in Language

Learning
Table 1 represents the “Attitude of the international students towards using tech-
nology in learning English language”. It is interesting to find from the table that
91% of the students have positive attitude towards the use of technology in learning
English language, whereas 4.5% of the students are neutral, and the rest 4.5% did
not use technology to learn the English language. The score obtained from item
3 indicates that most of the students believe that technology can improve English
language skills. Hence, the table below shows that 80–90% of the students have
Table 1 Distribution and frequency of attitude of international students towards technology in

language learning
Attitude to technology
Sl. Items Strongly Agree Neutral Disagree Strongly Mean Standard
No. agree (%) (%) (%) (%) disagree deviations
(%)
1 I enjoy 47 33 4 1 3 4.36 0.899
using ICT 53.4 37.5 4.5 1.1 3.4
for
language
learning
2 I prefer all 38 34 13 3 4.22 0.823
the medium 43.2 38.6 14.8 3.4
of
instruction
in English
3 I believe 37 37 9 3 2 4.18 0.917
that using 42 42 10.2 3.4 2.3
technology
can
improve
my English
language
speaking
skills
4 I believe 32 46 7 1 2 4.19 0.814
that 36.4 52.3 8.0 1.1 2.3
knowing
and
utilizing
technology
properly for
language
learning is
more useful
5 I like to use 39 32 9 6 2 4.14 1.008
technology 44.3 36.4 10.2 6.8 2.3
whenever
face any
language
barrier
shown positive attitude towards usage of technology and have also agreed that usage
of the technology helps in building their English language skills. Only a few students
who range from only 3–9% have shown their disagreement towards the usefulness of
technology in learning which is again a very minimal percentage of the sample. The
conclusive finding of Table 1 has been illustrated in Fig. 1.
Fig. 1 Attitude to
technology
4.2 Students Attitude to Functional Language Learning

Experience Using Technology
Table 2 exhibits that majority of the international participants showed their awareness
about use of technology in language learning. Out of the total population surveyed,
76.1% agreed to the use of mobile phones to improve language proficiency and 67%
international students showed their interest in learning English language through
use of various media of technology like chat, email or call over phones. The score
obtained for item 7 clearly shows that 56.8% students watched English movies with
subtitles to enhance their language skills (speaking, listening and reading). Further,
by observing item 5 and 8, it can be understood that, students like to learn language
in a passive way like reading e-books and games.
From the above discussion, it can be inferred that around 60–75% of the students
agree that they use various technological tools such as YouTube, mobile phones
and apps to improve their language skills and prefer using English language while
chatting, emailing or even talking with friends. Moreover, watching English movies
with subtitles helps develop language proficiency.
5 Discussion
From the study, it can be concluded that technology bears a great testimony for
learning English language. A great percentage of students are showing positive
response for using technology in overcoming English language barrier. Table 1 clearly
exhibits that the international students are aware of the fact that technology plays
a significant role in improving English language proficiency skills and they believe
that if they utilize technology properly they can overcome language barrier easily.
Most of the students agreed with using mobile phone and different apps to enrich
their vocabulary skills and developing proficiency skills. They also use chat, email
and text to friends in English. International students show their fondness for watching
English movies with subtitles for developing their language proficiency skills.
Table 2 Distribution and frequency of functional language learning experience using technology
Distribution and frequency of functional language learning experience using technology
Sl. Items Strongly Agree Neutral Disagree Strongly Mean Standard
No. agree (%) (%) (%) (%) disagree deviations
(%)
1 I use Youtube 16 39 22 8 3 3.75 0.861
to develop 18.2 44.7 25.0 9.1 3.3
my speaking
2 I use mobile 20 47 14 6 1 3.90 0.827
phone to 22.7 53.4 15.9 6.8 1.11
improve my
language
proficiency
3 I use 23 35 21 7 2 3.80 0.996
different 26.1 39.8 23.9 8.0 2.3
apps to
enrich my
vocabulary
and
proficiency
skills
4 I use English 28 31 17 10 2 3.83 1.074
language 31.8 35.2 19.3 11.4 2.3
when I chat,
email, or talk
to friends
over phone
5 I use e-books 14 29 25 14 6 3.35 1.135
15.9 33.0 28.4 15.9 6.8
6 I record my 6 8 19 33 22 2.35 1.155
video 6.8 9.1 21.6 37.5 25.0
conversation
when I make
dialogue in
English
7 I like 22 28 18 15 5 3.53 1.203
watching 25.0 31.8 20.5 17.0 5.7
English
movie with
subtitle
8 I like to play 10 28 23 18 8 3.16 1.160
language 11.4 31.8 26.1 20.5 9.1
related
games
Fig. 2 Functional language

learning experience using
technology
Regarding the use of technology for language learning, students give their utmost
positive response which is shown in Fig. 2, where more than 55% agreed to learning
through technological tools for overcoming language barrier and 23% of the partic-
ipants were neutral. So from this analysis, it can easily be understood that students
are very much aware of using technology for language learning.
6 Limitations
The generalizability of the study has the following limitations. Firstly, the sample
consisted of only male international students as the data was collected from the
international boys’ hostel of the university. Secondly, the sample for the survey was
selected from only one university of India, and hence, further research can be done
considering the international students from different educational institutes as well.
Finally, with a small number of participants completing the survey, the likelihood
of generalizing the findings of the study is limited. Therefore, investigation with a
larger sample size is necessary.
7 Conclusion
The study shows that the students are aware of the benefits of Information and
Communication Technology in overcoming language barrier and also about its signif-
icant role in enhancing their proficiency of English language. But in some cases, we
see that students learn the language skills passively by reading ebooks and playing
games. Although most of the international students showed positive attitude towards
the use of technology in overcoming language barrier purposes, there are some
students who either have a neutral view or disagree with the point. It is because
they might be hesitant in using ICT for language learning or might not be very tech
savvy. In such cases, several trainings or workshops can be provided to the students
so that they can be exposed to the benefits of integrating technology in learning the
English language. Moreover, teachers can also suggest the students some available
websites which are useful for developing students English language proficiency skills
to overcome language barrier.
References
1. B. Skinner, R. Austin, Computer conferencing: does it motivate EFL students? ELT J. 53(4),
270–279 (1999)
2. M. Warschauer, Of digital divides and social multipliers: combining language and technology
for human development. in Information and Communication Technologies in the Teaching and
Learning of Foreign Languages: State of the Art, Needs and Perspectives (UNESCO Institute
for Information Technologies in Education, Moscow, 2004), pp. 46–52
3. M. Warschauer, Computer-assisted language learning: an introduction, in Multimedia
Language Teaching, ed. by S. Fotos (Logos International, Tokyo, 1996), pp. 3–20
4. A. Dash, K. Kuddus, Leveraging the benefits of ICT usage in teaching of english language
and literature, in Smart Intelligent Computing and Applications. Smart Innovation, Systems
and Technologies, vol. 160, ed. by S. Satapathy, V. Bhateja, J. Mohanty, S. Udgata (Springer,
Singapore, 2020)
5. K. Kuddus, Emerging technologies and the evolving roles of language teachers: an overview.
Lang. India 18(6), 81–86 (2018)
6. B. Chatterjee, K. Kuddus, Mass media approach to second language acquisition. J. Engl. Stud.
10(1), 10–16 (2015)
7. C. Clark, P. Gruba, The use of social networking sites for foreign language learning: an
autoethnographic study of Livemocha, in Proceedings Ascilite Sydney (2010), pp. 164–173
8. S.L. Thorne, The intercultural turn and language learning in the crucible of new media, in
Telecollaboration 2.0 for Language and Intercultural Learning, ed. by F. Helm, S. Guth (Peter
Lang, Bern, 2010), pp. 139–164
9. T.Y. Ahn, S.M. Lee, User experience of a mobile speaking application with automatic speech
recognition for EFL learning. Br. J. Edu. Technol. 47, 778–786 (2016)
10. B. Chatterjee, K. Kuddus, Second language acquisition through technology: a need for under-
developed regions like Jharkhand. Res. Sch. Int. Ref. E-J. Literacy Explor. 2(2), 252–259
(2014)
11. H. Kellam, C.J. MacDonald, D. Archibald, D. Puddester, Designing digital video to support
learner outcomes: a study in an online learning resource for healthcare professionals and
students. Int. J. Online Pedagogy Course Des. (IJOPCD) 2, 45–66 (2012)
12. H. Hamzeh, The effects of two technological tools on idiom learning of Iranian EFL learners:
a MALL perspective. MLSJ 4(1), 1–12 (2017)
13. M. Prensky, Digital Game-Based Learning (McGraw-Hill, Tradeberg, New York, 2002)
14. Y.F. Yang, Exploring students’ language awareness through intercultural communication in
computer-supported collaborative learning. Edu. Technol. Soc. 16(2), 325–342 (2013)
15. B. More, A. Nicole, An examination of undergraduate students’ perceptions and predilections
of the use of YouTube in the teaching and learning process. Interdisc. J. e-Skills Lifelong Learn.
10, 17–32 (2014)
16. The media of social media, http://tristantreadwell.wordpress.com/tag/grahl/
17. B.G. İlter, How does Technology affect language learning process at an early age? Procedia
Soc. Behav. Sci. 199, 311–316 (2015)
18. K. Kuddus, Web 2.0 technology in teaching and learning english as a second language. Int. J.
Engl. Lang. Lit. 1(4), 292–302 (2013)
19. R. Chartrand, Social networking for language learners: creating meaningful output with Web
2.0 tools. Knowl. Manag. E-learn. Int. J. 4, 97–101 (2012)
20. D.A. Castaneda, M.H. Cho, Use of a game-like application on a mobile device to improve
accuracy in conjugating Spanish verbs. Comput. Assist. Lang. Learn. 29, 1195–1204 (2016)
21. G. Gay, S. Mahon, D. Devonish, P. Alleyne, P. Alleyne, Perceptions of information and commu-
nication technology among undergraduate management students in Barbados. Int. J. Edu. Dev.
ICT 2, 6–17 (2006)
22. M.W. Rumpagaporn, Student’s Critical Thinking Skills, Attitudes to ICT and Perceptions of ICT
Classroom Learning Environment Under the ICT Schools Pilot Project in Thailand (School of
Education, University of Adelaide, Bangkok, 2007)
Limiting Check Algorithm: Primality
Test
Gunjan Sethi and Harsh
Abstract This algorithm is designed to check the primality of a large number in a

very different manner by combining the research of Manindra Agrawal, Neeraj Kayal,
Nitin Saxena [1] and Greek mathematics. The time complexity of the algorithm is
shown in graphical and tabular way by implementing it in various ranges of numbers.
Keywords Prime · Semi-prime · Euclid elements · Complexity · Prime ideals ·

Primality testing · Matplotlib
1 Introduction
Mathematics, the actual science behind the daily life, performs an important role
in our day-to-day decisions and activities and nothing wrong to say that even all
the decisions in our day-to-day life requires some mathematics behind them either
directly or indirectly. Today, all the industries, education institutes, our household
applications are running because of mathematics. If there would be no mathematics,
there would not be electricity and there would not be any kind of machines. Even the
topic of today’s generation, machine learning and artificial intelligence, which made
the machines learn easily and let them make decisions by themselves with lesser or
even negligible assistance by a human being, is totally grounded on mathematics.
Concerning the prime numbers, they are one of the major parts of mathematics. As
stated in Wikipedia [2], prime numbers are those numbers which yields to reminder
0 only by dividing them with 1 or itself. Prime numbers, due to their uniqueness,
perform a vital role in various cryptography techniques [3]. With the evolution of
usage of prime numbers in various situations, a new jargon came up, i.e., semi-prime
numbers [4] which are also derived from prime numbers to provide more flexibility
G. Sethi (B)
Delhi Technical Campus, Greater Noida, Uttar Pradesh, India
e-mail: gunjan.ctit@gmail.com
Harsh
Cult Skills, Jalandhar, Punjab, India
e-mail: aravsharma798@gmail.com
698 G. Sethi and Harsh
to the usage of these numbers. Prime numbers are topic of research since Greek
mathematicians in 200 BC and there after a long break the next research came up
in seventeenth centuries. Also, Euclid’s elements enlighten the subject in a very
convenient way than someone else does that.
2 Base Algorithm
The very first algorithm to check the primality of a number were designed by Greek
Eratosthenes in 200 BC as stated in [5]. This algorithm was designed in such a way
that it just declare every multiple of a prime number as unprimed. Nowadays, the
algorithm taught in theory to check whether a number is prime or not is that the
number itself is divided with all the numbers from 2 to the number preceding the
actual number and their remainders are checked. If all the numbers failed to lead the
reminder of 0, the number is considered as a prime number. But if any number in
2 to number preceding the actual number leads to a reminder of 0, then the number
is not a prime number, it is a composite number. Then, multiple algorithms came
up and some of them were found to be failed and some others to be successful but
when they got implemented in the real-world machines they took ample of time to
execute that algorithm. Then a new research of Manindra Agrawal, Neeraj Kayal,
Nitin Saxena of IIT Kanpur came up cited here [5] that says that every composite
number must have a factor equals to or less than its square root.
3 Limiting Check Algorithm
In our limiting check algorithm, we have combined the results of both the research
cited above, i.e., Sieve of Eratosthenes [4] and the research by IIT Kanpur, Primes
is in P [5]. Here in this algorithm number is checked for its primality only with the
prime numbers lesser than or equal to its square root. If the number has no prime
divisor before or equal to its square root, then the number is prime number else the
number is a composite number. This is the conclusion of two researches as mentioned
previously—one saying a number (say n)√needs to have at least one factor less than
the square root of the original number ( n). This incorporates the usage of Greek
algorithm that for every composite number at least one divisor must be prime and the
other algorithm which says that for every composite number it must have a divisor
before its square root. In this algorithm, the number is checked before its square root
and only by all the prime numbers within that square root which would reduce the
complexity by a factor of more than square root of the number.
Limiting Check Algorithm: Primality Test 699
4 Results
The result of this algorithm can be well understood by graphical representations of

numbers with different orders and the time taken by the algorithm to implement that
check on those numbers. It is of no use to design algorithm to check whether ‘13’
is a prime number or not or ‘23’ is a prime number or not. The algorithm must give
correct results for these numbers, but the main focus is to check the numbers of higher
orders which cannot be checked so quickly in the casual algorithm. For 2, 3 or 4 digits,
numbers any algorithm will lead to almost similar kind of results. We have here plots
for our algorithm for various 6-digit numbers, 8-digit numbers, 10-digit numbers
and 12-digit numbers. Timing patterns may also vary with language and IDE used
for the execution of the algorithm. The algorithm can be well implemented in any
of the programming language, with any possible machine and with any supporting
development environment. All the graphs are implemented in Matplotlib library of
Python using PyCharm IDE and the machine which leads to this time patterns is
made up of i5-3320 2.60 GHz processor, 4 GB RAM and hard disk of 5400 rpm.
The timing patterns may vary in different machines as per the configurations. The
reason for using python here is just to make the code crisp and short and easily
readable to anyone with little or no knowledge of programming. The X—axis of
the graph denotes the numbers, and the Y —axis is the time taken (in milliseconds)
by the algorithm to test the number. The blue dots on the graph signify the relation
between the number, and the time it has taken to accomplish the test on that number.
The numbers on the X—axis is shortened in some cases by Matplotlib in the form
of logarithm and is mentioned below with ‘le’ notation (Fig. 1).
Fig. 1 Limiting check algorithm of 6-digit numbers against time taken (in milli-seconds)
Table 1 Performance
Prime number Limiting check algorithm SymPy algorithm
comparison on 6-digit number
235,919 5.722e–05 8.296e–05
441,461 5.865e–05 7.748e–05
767,857 5.197e–05 8.630e–05
Figure 1 is the plot of 10 numbers of 6-digits along with the time taken by the
algorithm to test their primality. As we can see, there are three different slabs 0, 1
and 2 ms (Table 1).
Table 1 describes the results of numbers of 6 digits with the time taken by the algo-
rithm to test their primality using SymPy algorithm (existing algorithm) and limiting
check algorithm. It can be clearly analyzed that using limiting check algorithm in
all the cases, the results are better that means the proposed algorithm consumes less
amount time for doing primality test (Fig. 2).
Figure 2 is the plot of 10 numbers of 8 digits along with the time taken by the
algorithm to test their primality. As we can see, there is a variety of time ranging
from around 6 to 25 ms (Table 2).
Table 2 Performance
comparison on 8-digit number
49,390,927 4.250e–03 9.525e–03
10,000,931 2.024e–03 7.202e–03
43,112,609 4.523e–03 1.630e–02
Table 2 describes the results of numbers of 8 digits with the time taken by the algo-
rithm to test their primality using SymPy algorithm (existing algorithm) and limiting
check algorithm. It can be clearly analyzed that using limiting check algorithm in
all the cases, the results are better that means the proposed algorithm consumes less
amount time for doing primality test (Fig. 3)
Figure 3 is the plot of 10 numbers of 10-digits along with the time taken by
the algorithm to test their primality. As we can see, there is a hierarchy of time
ranging from around 130 to 420 ms (approx.), but there are two clusters one around
200–250 ms and one around 280–320 ms (Table 3).
Table 3 describes the results of numbers of 10 digits with the time taken by
the algorithm to test their primality using SymPy algorithm (existing algorithm)
and limiting check algorithm. It can be clearly analyzed that using limiting check
algorithm in all the cases the results are better that means the proposed algorithm
consumes less amount time for doing primality test (Fig. 4).
Figure 4 is the plot of ten numbers of 12 digits along with the time taken by the
algorithm to test their primality. As we can see there is a hierarchy of time ranging
from around 2000 to 12,000 ms. The line above is a straight line, and clustering
of the points is almost negligible. Now if we consider the algorithm which check
Table 3 Performance
comparison on 10 digit
number 1,694,343,422 4.148e–01 7.409e–01
1,000,000,007 3.623e–01 1.671e–00
2,147,483,647 6.119e–01 2.235e–00
the number until its half that would take immense amount of time to achieve the
results of test and it is very hectic to work with this algorithm in cryptographies and
various other applications where getting the results in time is a big necessity. The
graphs of number and time taken by older stated algorithm to test the number are
also mentioned here (Fig. 5).
Figure 5 is the plot of ten numbers of 6 digits along with the time taken by the
algorithm to test their primality. As we can see there is a variety of time ranging
Fig. 5 Plot of 6-digit numbers against time taken (in milli-seconds)

from around 2 to 110 ms (approximately). Now if we compare both the plots for
6-digit numbers, then we will find that on one hand this algorithm is taking up to
100 ms, but limiting check algorithm would cost you only 2 ms at maximum for the
same number in the same machine. The difference is of the order of 50 or even larger
(Fig. 6).
The graph here is costing us ranging from 0 to 8000 ms, whereas the limiting
check algorithm was taking only up to 25 ms. The difference is of the order of 320
times (Fig. 7).
The graph here is costing us ranging from 0 to 1,500,000 ms (approx.), whereas
the limiting check algorithm was taking only up to 500 ms at its hard. This shows
the significant difference in both the algorithms when we consider a big integer. The
difference is of the order of 3000 times. We have also tried this algorithm to test the
primality on a 12 digit number, but it would not be able to test even a single number in
15 h of undisturbed CPU utilization. But on the other hand, limiting check algorithm
can execute the primality of number in 12,000 ms, i.e., 12 s. We can wait for more
time, but it is of no use to run a test of primality and wait for more than 15 h be to
executed.
5 Discussion
The graphical and tabular representations explain the time complexity of the algo-
rithm against various numbers. If we test for 13 to be prime or not limiting check
algorithm will run until base of square root of 13, i.e., 3 and eventually it has started
from 3, so the number of comparisons would be 1 only and in single comparison that
can make a decision, but in the regular theoretical checking algorithm the number of
comparisons would be 3. Now considering 31 as the number to check, then limiting
check algorithm would conclude the result only in three comparisons, but in the
regular theoretical checking algorithm the number of comparisons would be 14.
These numbers are not that different but when we try to implement the limiting
check algorithm for a big number like 999,999,000,001, then the casual algorithm
would end up doing 499,999,499,999 but limiting check algorithm would only check
for 664,579 comparisons only. Hence, the number of comparisons of limiting check
algorithm is decent good than the other discussed methods. Now in graph as well
we can see the algorithmic time complexity will increase by the size of number so
higher the number, greater will be the time taken to test that number.
6 Future Works
This algorithm is doing way better than some other algorithm, but this can also
be improved by doing some amendments. We can improve the time taken by the
algorithm by making the list of prime numbers up to square root in a recursive way.
This list can be created by calling a recursive function which will create a list of
all the primes in the range. Recursion can help us to check for primality of a factor
before considering it a valid divisor. For example, if we want to check the primality
of 101, then firstly base of its square root will be calculated which is 10 that confirms
that at least one factor needs to be between 10 (greater than 1), and other research
says that factor needs to be a prime number so we can check with recursiveness and
make a list of all valid divisors under 10 which will be [2, 3, 5–7] and this can be done
in a recursive manner. It is also possible to completely remove the list and check the
primality on the go without storing numbers in any list. This would reduce the space
complexity of the algorithm.
References
1. M. Agrawal, N. Kayal, N. Saxena, Primes is in P, in Annals of Mathematics (2004)

2. J.I. Ahmad, R. Din, M. Ahmad, Analysis review on public key cryptography algorithms.
Indonesian J. Electr. Eng. Comput. Sci. 12(2), 447–454 (2018)
3. E.W. Weisstein, Semiprime, http://mathworld.wolfram.com/semiprime.html
4. Prime numbers, http://www-groups.dcs.st-and.ac.uk/history/HistTopics/Prime_numbers.html
5. Math. Comput. 64(209), 397–405 (1995). Published by: American Mathematical Society, Article
Stable URL: http://www.jstor.org/stable/2153343
6. A.R.C. De Vas Gunasekara, A.A.C.A. Jayathilake, A.A.I. Perera, Survey on prime numbers.
Elixir Appl. Math. 88, 36296–36301 (2015)
7. K.R. Srinath, Python–the fastest growing programming language. Int. Res. J. Eng. Technol. 4
(2017)
Factors Affecting the Local Governance
Ngo Sy Trung, Do Huu Hai, Pham Van Tuan, and Nguyen Tien Long
Abstract This research focuses on an analysis of factors affecting the local gover-
nance, namely: (1) Intellectual level, (2) Participation level of citizens and social
organizations, (3) Responsibility for reporting and accountability, (4) Capacity of
officials and civil servants, (5) Information system. The authors survey practically
influence levels of these factors on the local governance of local government levels
in Vietnam (surveyed in 04 provinces representing three regions in Vietnam: Hanoi,
Thanh Hoa, Quang Nam, Ho Chi Minh City). Research results show that the capacity
of officials and civil servants has the greatest impact on the local governance, followed
by Responsibility for reporting and accountability, Information system, Participation
level of citizens, social organizations, and finally, Intellectual level.
Keyword Local governance
1 Overview of Factors Affecting the Local Governance
The term of “the local governance” has been studied in the world for a specific period
of time, has been associated with decentralization in many countries since the 1960s.
The connotation of the concept of the local governance, mentioned in many studies,
includes entities: local governments, social organizations and citizens, in which the
interactive relationship of these entities is the issue analyzed thoroughly such as:
“Local governments manage via legal instruments and policies; social organizations
N. S. Trung (B)
Hanoi University of Home Affairs, Hanoi, Vietnam
e-mail: ngosytrung01@yahoo.com
D. H. Hai
Ho Chi Minh City University of Food Industry, Ho Chi Minh, Vietnam
P. Van Tuan
National Economics University, Hanoi, Vietnam
N. T. Long
TNU—University of Economics and Business Administration, Thai Nguyen, Vietnam
708 N. S. Trung et al.
and citizens actively participate in policy decision-making” [1] or “The local gover-
nance implies the adoption and implementation of regulations and procedures of
organizations and institutions that allow citizens to express their concerns, exercise
their participation right in the local development” [2] or” The local governance is the
governance at the local level not only belongs to the government, but also belongs
to the community in general and the interactions between the community and local
authorities” [3], etc.
Thus, it is possible to identify the factors that affect the local governance from
its entities. For the government, it is the capacity of officials and civil servants
(the Capacity of officials and social organizations of officials, the practical profes-
sional capacity of civil servants), Responsibility for reporting and accountability,
and several of other factors of material facilities such as information system, tech-
niques, etc. For citizens, it is the intellectual level, self-consciousness level of their
participation right. For citizens and social organizations, it is their intellectual level,
participation level into the management process of the government.
1.1 Intellectual Level and Self-consciousness Level

of Citizenship (IL)
The intellectual level is the knowledge of citizens, expressed through the knowledge
about the fundamental issues of contemporary society they acquire such as: Economy,
culture, law, science, technology, etc., and even the issues of history or the predictive
issues of the future. Besides the scientific knowledge of the society, the intellectual
level is also understood as the understanding and awareness of citizens about their
citizen rights and responsibilities as well as the desire to exercise those rights and
responsibilities. Some basic content reflects the intellectual level: Legal knowledge
(TDDT1); Cultural-social knowledge (TDDT2); Economic and technical knowledge
(TDDT3); Knowledge of science and technology (TDDT4); Awareness of citizenship
(TDDT5); Awareness of civic responsibility (TDDT6). This is one of the basic and
important criteria to assess the development level of society, accordingly, the more
developed society is, the more improved the intellectual level is.
In the local governance, when the intellectual level develops, citizens in whom
participate promptly grasp the provisions of law relating to management fields of
the government and give the profound social criticism. Citizens are better aware of
their rights and responsibilities and will actively participate actively in checking and
supervising the implementation of grassroots democracy. This has a strong impact
on the local governance. Therefore, in the local governance, the authorities should
pay attention to the participation forms and means of citizens in accordance with the
intellectual level, in order to create a positive and effective interaction between the
authorities and citizens.
Factors Affecting the Local Governance 709
1.2 Participation Level of Citizens and Social Organizations

(SO)
Citizens and social organizations that are living and working in the area, under-
stand the specific conditions of the locality. Therefore, their participation will have
an impact on the management and administration of the government during the
implementation of the local governance tasks. However, the important issue is why
and how the participation level of citizens and these social organizations relate to
the policy mechanisms and the political views of government leaders. Some basic
content reflects the Participation level of citizens and social organizations: Positive
attitude towards election activities (TCXH1); Positive attitude towards government
policy decisions (TCXH2); Attention and criticism of government policies (TCXH3);
Abide by the law (TCXH4); Proactively propose ideas and policies to the government
(TCXH5); Join the opinion of citizens (TCXH6). Some impacts from the involvement
of citizens and social organizations on the local governance are:
• Firstly, the active participation of citizens and social organizations will give
the government the opportunity to mobilize more resources (finance, human
resources, information) from the community to archive easily the target of
socio-economic development of the locality.
• Secondly, the active participation of citizens and social organizations will reduce
barriers between government agencies and community to receive easily feed-
back from the community on policies, public services, and avoid the negative
consequences of the lack of referral information of service beneficiaries.
• Thirdly, the active participation of citizens and social organizations will form the
partnership in the local governance so that there can be joint initiatives between
the government and the community. This has important significance, establishing
a mutually beneficial interaction between the government and the community
through a monitoring mechanism that prevents manipulation and group interests;
At the same time, the restrictions of the governance and coercion type of traditional
state management are resolved.
1.3 Responsibility for Reporting and Accountability (RA)
Responsibility for reporting and accountability is the working modes clearly

mentioned in the legal system of many countries, whereby state agencies are required
to report and account the results of their assigned tasks to the higher level agencies,
citizens’ elected bodies at the same level and the citizens. In that sense, Responsi-
bility for reporting and accountability is also identified as one of the mechanisms for
controlling the performance of public authorities, particularly through Responsibility
for reporting and accountability, the superior government controls the activities of
subordinate government, the citizens and social organizations control the operation
of local government, etc. in a convenient way on the basis of compliance with the law;
At the same time, officials and civil servants will be more aware of their obligations
and responsibilities in the course of performing their duties and bear the conse-
quences when they fail to fulfill their responsibilities. Some basic content reflects the
Responsibility for reporting and accountability: Disclosure of legal information and
policy information (BCGT1); Publication of administrative procedures (BCGT2);
Publicizing the process of work settlement (BCGT3); Extensive policy consulta-
tions (BCGT4); Respect and listen to the people’s opinions (BCGT5); Responsible
for work performance results (BCGT6).
With interactions among the local governance’s entities, we can see that the local
governance is dominated by responsibility for reporting and accountability of officials
and civil servants. Namely:
• The good implementation of responsibility for reporting and accountability will
create a mechanism to control the activities of the local government from the citi-
zens and social organizations. In addition, a mechanism of interaction between the
government and the community is established, thus forming the political culture
in the local governance.
• The good implementation of responsibility for reporting and accountability will
contribute to reduce many negative issues such as: corruption, arbitrariness, sectar-
ianism, etc., through the deep self-awareness of the duty performance of officials
and civil servants, as well as the consequences that officials and civil servants have
to bear when they fail to fulfill their responsibilities in the course of performing
their duties..
1.4 Capacity of Officials and Civil Servants (OCS)
• Regarding the capacity of officials, it is the knowledge and the management skills
of leaders of local authorities.
So far, there have been many researches on the leadership, the leadership capacity as
well as the leadership results, the impact of leadership capacity on the leadership
results of businesses/organizations. Researchers have approached from a variety
of perspectives and addressed various research objectives, and the results of this
study are abundant and admitted. Cavazotte et al. [4]; Strohhecker and GroBler [5]
and Thun (Thesis) [6], in the studies show that the knowledge, the awareness, the
thinking capacity, the management skills are the basic factors forming the Capacity
of officials and social organizations of the leaders, having the directly influences
on the performance of leadership. Similarly, Northouse [7], through his study work,
he also find out that the thinking capacity of managers (intelligence, initiative, task
understanding, cognitive abilities) positively affects the leadership results of busi-
nesses/organizations. Quite consistent with the above-mentioned views, Hoa [8];
Su [9] also argues that the leader’s thinking capacity (knowledge, strategic vision)
and the management and executive skills have a great influence on the performance
of management. Therefore, for the local authorities, when leaders have the good
knowledge, the thinking capacity and the management skills, they will mobilize the
participation and support of the social forces in the area through the motivation-
making, trust-giving, encouragement, inspirational effort, the ability to propagate
the strategy, the ability to motivate and attract citizens to support the strategy and the
ability to link the system, etc., from that, helps them achieve the desired results in the
local governance. On the other hand, this is the factor which has a great influence on
the leadership and management results of the leaders of the authorities in the local
governance.
• Regarding the capacity of civil servants, that is the knowledge, skills of profes-
sional practice and the attitude of serving the people who directly perform the
tasks of the local governance. Local civil servants are the people who regularly
meet and address the needs of citizens and organizations. The level of satisfaction
of citizens and organizations is an important criterion for assessing the working
capacity of such civil servants. Therefore, civil servants with the knowledge, the
professional practice skills and the good service attitude will be the factors that
have a great impact on the performance of key tasks of the local governance,
meeting the expectations and requirements of the people about a government
serving in the context of the administrative reform.
Some basic content reflects the Capacity of officials and civil servants: Knowl-
edge and management skills of local government leaders (CBCC1); Knowledge and
practice skills of civil servants (CBCC2); Attitude to serve people of civil servants
(CBCC3).
Along with the development of society, as the intellectual level is improved, as
the science, technique and technology are widely used in all of social activities, it has
set new demands on the capacity of local officials and civil servants. This requires
the local officials and civil servants to constantly learn to accumulate the knowledge,
the professional practice skills and the management skills, along with having a good
attitude to serve the people in the best way. These are also regular requirements and
tasks of officials and civil servants in the process of carrying out the tasks of local
governance.
1.5 Information System (IS)
Information system is a necessary and important support of the local governance.

Building a good information system will create a premise for citizens’ active partic-
ipation and social organizations, create the favorable condition for implementing
the responsibility for accountability and public administration. Some basic content
reflects the Information system: Legal and policy documents are widely and system-
atically (HTTT1); Management and operating information is public (HTTT2); The
feedback mechanism of citizens is built and implement regularly and seriously
(HTTT3).
2 Study Model
In the above study model, the independent variables have the influence on the depen-
dent variable (the local governance: Decentralized work management (QTDP1);
Attracting the participation of people and social organizations (QTDP2); Control-
ling people’s participation and social organization (QTDP3)) at different levels. This
is the task that the research team focuses to clarify the relationship between these
variables.
Hypotheses:
H1: The intellectual level has a positive relationship with the elements of the local
governance.
H2: The participation level of citizens and social organizations has a negative
relationship with the elements of the local governance.
H3: The responsibility for reporting and accountability have a positive relationship
with the elements of the local governance.
H4: The capacity of officials and social organizations is have a positive relationship
H5: The information system has a positive relationship with the elements of the
local governance.
3 Research Methods
The study was conducted in two steps: the preliminary investigation and the
formal investigation. With the preliminary investigation, the authors used qualita-
tive research, conducted in-depth personal interviews and focused group interviews
with a number of experts and civil servants who participated in the local governance
to modify the research scale and complete the suitable questionnaire for Vietnam’s
context. Note that the questionnaire is based on the overview of the research, and
the scales which are checked by previous studies, so the authors only inherited and
adapted the words. The components in the model are measured with a 5-level Likert
scale, in which, 1 is strongly disagree and 5 is strongly agree. The last section is
demographic information, including questions about gender, age, intellectual level,
income and the number of years involved in the local governance, using the identifier
scale.
After that, the completed questionnaire was applied in the official quantitative
survey in 04 places representing for three regions of Vietnam, including Hanoi, Thanh
Hoa, Quang Nam and Ho Chi Minh City; the survey period is from 15th January 2020
to 31st Mar 2020. The research overall are the people who have participated in the
local governance local, have the management experience; they are of all ages, living
places, incomes, different levels. Sample subjects are those over 18 who have had the
participation experience in the local governance. The measurement model consists
of 27 observation variables and, according to the principle of at least 5 elements per
measurement variable [10]; therefore, the minimum sample size required is 27 * 5

= 135. However, in this research, the authors selected a sample size of 200, larger
than the minimal sample, ensuring the science of sample selection. Questionnaires
are sent directly to respondents via the Internet with random sampling method and
convenient sampling method. The results were 177 satisfactory, with a response rate
of 88.5%. The characteristics of the sample in Table 1 are 78% for females, 65.5%
for age from 18 to 36 and 66.7% for males and 36.5% for years involved in the
local governance less than 2 years. Data processing method is done through the
following steps: (1) checking the scale and the reliability of measurement variables
by Cronbach’s alpha coefficient and the value by the Exploratory Factor Analysis
(EFA); (2) analyzing the multiple linear regression for model and hypothesis testing.
Table 1 Characteristics of research samples

Sample size (n = 177) Frequency Ratio (%)
Gender Male 39 22.0
Female 138 78.0
Age From 18 to 25 49 27.7
From 26 to 30 26 14.7
From 31 to 36 41 23.2
Over 36 61 34.5
Average income per month Less than 3 million 36 20.3
Between 3 and 6 million 81 45.8
Between 6 and 10 million 45 25.4
Over 10 million 15 8.5
Academic level High school graduation 11 6.2
Graduated college/professional 20 11.3
high school
Graduate 83 46.9
Post graduate 59 33.3
Other 4 2.3
Years involved in the local Less than 1 year 67 37.9
governance From 1 to 2 years 51 28.8
Over 2 years 59 33.3
Source The author’s reseach
4 Study Findings
4.1 Scale Testing
Using the SPSS 23.0 software, the author drew the results of the Cronbach’s alpha and
the Exploratory Factor Analysis (EFA), suggested to remove and merger observation
variables to help the scale to assess the concepts accurately.
The first step is scale testing by the Cronbach’s alpha reliability coefficients. This
step is used to exclude garbage items avoiding the case of garbage items make up the
dummy factor when analyzing the EFA [10]. Testing standard is Cronbach’s alpha
coefficient which must be greater than 0.6 and the correlation coefficient of each
scale must be greater than 0.3 [10]. The analysis results in Table 2 show that all
measures of the factors are satisfactory. Thus, all scales of factors are both reliable
and are used to analyze the next factor.
After examining Cronbach’s alpha, the next step is Exploratory Factor Anal-
ysis (EFA) for a preliminary evaluation of the uniqueness, convergence value, and
discriminant value of the scale. The EFA was performed using the Principal Compo-
nent Analysis method and the Varimax rotation method to group the factors. With a
sample size of 177, the load factor coefficient of observed variables must be greater
than 0.5; The variables converge on the same factor and distinguish it from other
factors [10]. In addition, the KMO coefficient must be between 0.5 and 1.
The analysis results in Table 3 show that the load factor coefficient of all the
observed variables are greater than 0.5; Bartlett test with significance level Sig.
Table 2 Summary of reliability and relative minimum variables of scales

Scale Number of Reliability coefficient The correlation
observation variables (Cronbach’s alpha) coefficient variation
smallest total
Intellectual level (IS) 6 0.843 0.605
Participation level of 6 0.842 0.659
citizens and social
organizations (SO)
Responsibility for 6 0.824 0.664
reporting and
accountability (RA)
Capacity of officials 3 0.801 0.571
and civil servants
(OCS)
Information system 3 0.778 0.579
(IS)
Local governance 3 0.753 0.569
(LG)
Table 3 EFA factor analysis of the factors affecting the local governance
Observed variables Factors
1 2 3 4 5
Intellectual level (IL) TDDT2 0.852
TDDT3 0.806
TDDT5 0.726
TDDT1 0.710
TDDT4 0.685
TDDT6 0.684
Participation level of citizens and social TCXH5 0.783
organizations (SO) TCXH3 0.781
TCXH2 0.745
TCXH4 0.743
TCXH1 0.710
TCXH6 0.709
Responsibility for reporting and BCGT4 0.773
accountability (RA) BCGT3 0.759
BCGT2 0.754
BCGT6 0.737
BCGT5 0.676
BCGT1 0.655
Capacity of officials and civil servants CBCC3 0.876
(OCS) CBCC2 0.859
CBCC1 0.763
Information system (IS) HTTT2 0.863
HTTT3 0.828
HTTT1 0.787
= 0.000 with KMO coefficient = 0.797. All 24 variables after EFA analysis are
divided into 5 factors with eigenvalues greater than 1 and deviation more than 50%.
The research model changes into 5 independent variables, and 1 dependent variable,
used for the linear regression analysis and next hypothesis testing.
4.2 Regression Analysis and Hypothesis Testing
First, using Pearson correlation analysis analyzes the correlation between quantita-
tive variables. At a significance level of 5%, the correlation coefficients show that
the relationship between the dependent variable and the independent variables is
statistically significant (Sig value is less than 0.05). The magnitude of the correlation
coefficients ensures that there are no hyperbolic phenomena. Therefore, it is possible
to use other statistics to test the relationship between variables.
Next, analyzing the multiple linear regression about the relationship between
factors belonging to the local governance affects the local governance. According to
Table 4, the R2 coefficient is 0.722, which shows that the multiple linear regression
models were constructed in accordance with the data set of 72.2%. The results of the
regression analysis showed that the 5 independent variables were IL, SO, RA, OCS,
and IS (Table 3) with non-normalized Beta coefficients of −0.095; 0.316; 0.363;
0.049; 0.062 with significance level less than 0.05. Thus, the hypotheses H1, H2,
H3, H4, H5 are accepted.
Based on B coefficient (Table 4), it can be concluded that Responsibility for
reporting and accountability has the strongest influence on the local governance;
followed by the Participation level of citizens and social organizations; Information
Table 4 Regression results of model of factors affecting the local governance

Model Non-normalized Normalized Value t Significance Collinear tatistics
regression regression (Sig.)
coefficient coefficient
B Standard Beta Tolerance VIF
error
(Constant) 1.471 0.015 100.268 0.000
Intellectual −0.095 0.015 −0.162 −6.466 0.000 1.000 1.000
level (IL)
Participation 0.316 0.015 0.540 21.522 0.000 1.000 1.000
level of
citizens and
social
organizations
(SO)
Responsibility 0.363 0.015 0.621 24.754 0.000 1.000 1.000
for reporting
and
accountability
(RA)
Capacity of 0.049 0.015 0.083 3.304 0.001 1.000 1.000
officials and
social
organizations
(OCS)
Information 0.062 0.015 0.105 4.190 0.000 1.000 1.000
system (IS)
Dependent variable: Local governance (LG)
R2 adjusted = 0.722
Fig. 1 Study model. Source The authors offer
system; Capacity of officials and social organizations. The intellectual level is the
least influential factor in the local governance. The linear regression model obtained
is as follows:
LG = 1.471−0.095 ∗ IL + 0.316 ∗ SO + 0.363 ∗ RA

+ 0.049 ∗ OCS + 0.062 ∗ IS + e
4.3 Testing of the Research Model
The SEM theoretical model results (Fig. 1) are shown in Fig. 2: Chi-square/df =
2,152; GFI = 0.900; TLI = 0.905; CFI = 0.916; RMSEA = 0.051, indicating a theo-
retical model suitable for market data. However, the estimation results (normalized)
of the relationship between the concept: SO → LG is not statistically significant at
90% .
4.4 Testing of Research Hypotheses
Using other methods in the study by software Amos 23 gave test results of formal
theory with 5 hypotheses: H1, H2, H3, H4, H5. The estimation results show that
the concept of Intellectual level, Responsibility for reporting and accountability,
Capacity of officials and social organizations and Information systems have the same
impact on the local governance and are statistically significant (P ≤ 0.05); Only the
Participation level of citizens and social organizations has the opposite effect on the
local governance (P > 0.05), indicating that:
Fig. 2 CFA results of factor affecting the local governance (normalized). Source The author’s
reseach
• The intellectual level has a positive relationship with the elements of the local
governance.
• The participation level of citizens and social organizations has a negative
relationship with the elements of the local governance.
• The responsibility for reporting and accountability have a positive relationship
• The capacity of officials and social organizations is have a positive relationship
• The information system has a positive relationship with the elements of the local
governance.
It means that the hypotheses: H1, H2, H3, H4, H5 are accepted. Only hypothesis
H2 is different from the result using SPSS software for analysis and evaluation, the
difference is appropriate because the greater the participation level of citizens and
social organizations is, the less affected the local governance is.
5 Conclusions and Recommendations
Research results indicate that Capacity of officials and civil servants has the greatest
impact on the local governance, followed by Responsibility for reporting and account-
ability, Information system, Participation level of citizens and social organizations,
and finally Intellectual level. Based on the above research findings, some of following
solutions are recommended:
• In order to improve the effectiveness of the local governance, the capacity of
officials and civil servants (knowledge, management skills of leaders; the knowl-
edge, professional practice skills and the service attitude of civil servants) plays
an important role in the local governance. However, in reality, some officials and
civil servants, especially localities, do not have the necessary capacity to meet the
requirements of work. Therefore, the improvement of the capacity of officials and
civil servants of government agencies should be paid attention.
• Forming a board of Responsibility for reporting and accountability that adapts to
the conditions of each locality should be considered and discussed. In areas, there
is no public investment, it is not necessary to set up a Responsibility for reporting
and accountability board or the citizens’ Inspection Board can fulfill the tasks of
the Responsibility for reporting and accountability board.
• Information system must be considered properly. Besides previous propaganda
forms such as the mass media, posters, propaganda and propaganda slogans,
contests of election law, etc. should be held. Election-related materials may be
sent to every person, everyone can read whenever they have free time to promote
citizens’ mastery.
• Improving the intellectual level of citizens to be more aware of their participation
role in the activities of government should be focused. In addition, raising citizens’
awareness of the importance of implementing and promoting democracy in order
to promote their real ownership is very necessary. Because of understanding the
meaning and role of promoting democracy, citizens will actively use democratic
rights and fight for the protection of democratic rights. When citizens take the
initiative, implement the law on democracy, actively use and promote democratic
rights, then they have real control, overcome the state of formal democracy and
interests. Thank to this, the benefits of the state, the rights and interests of the new
citizens are guaranteed.
• Encouraging and creating favorable conditions for social organizations to actively
participate in the local governance such as: Establishing mechanisms for interac-
tion among the authorities with citizens and social organizations; supporting the
development of the governance capacity, etc., to both promote the democracy and
mobilize the resources from the community.
References
1. Le rapport national sur le développement humain (Bénin, 2000)

2. M. Dêdêgnon, Gouvernance locale (Local governance) (2013), http://www.awepa.org
3. United States Agency for International Development—USAID, Decentralization and demo-
cratic the local governance programming handbook (2000)
4. Flavia Cavazotte, Valter Moreno, Mateus Hickmann, Effect of leader intelligence, person-
ality and emotional intelligence on transformational leadership and managerial performance.
Leadersh. Quart. 23(2012), 443–455 (2012)
5. J. Strohhecker, A. GroBler, Do personal traits influence inventory management performance?
the case of intelligence, personality, interest and knowledge. Int. J. Prod. Econ. 142, 37–50
(2013)
6. N.B. Thun, Character strengths in leadership, Thesis, Doctor of Philosophyin Business Admin-
istration, The Departement of Management, Saint Mary’ s University, Halifax, Nova Scotia,
Canada, 2009
7. P.G. Northouse, Leadership—Theory and Practice, 5th ed. (Western Michigan University,
2004)
8. T.T. Van Hoa, Leadership Enhancement of Vietnamese Business Executives in the Process
of International Economic Integration (Ministerial Research Project, National Economics
University, 2011)
9. D.N. Su, Leadership Competencies of Small and Medium Enterprises in Vietnam, PhD thesis,
CIEM Institute for Economic Research, 2011
10. D.H. Hai, Corporate Culture—Peak of Knowledge (Transportation Publishing House, Hanoi,
Monograph, 2016)
A Multi-factorial Code Coverage Based
Test Case Selection and Prioritization
for Object Oriented Programs
Prashant Vats, Zunaid Aalam, Satnam Kaur, Amritpal Kaur,

and Saravjit Kaur
Abstract The selection and prioritization of test case generates a series of test cases
for Object Oriented Programs in a way that the selected test cases would have the
maximum possibility of fault detection by ensuring the maximum code coverage.
In this research paper we have put forward an algorithm for the test case selection
and their prioritization in such a way which is based on the Multi Factorial code
coverage for object oriented programs. The multiple code coverage factors would be
chosen on the account of their probability of finding errors and the code complexity
which has been introduced by these factors. For analysis and validation of our put
forward test case selection and prioritization algorithm we have considered few test
case studies using the Java Object Oriented Code. As a result of our case studies we
have found that the put forward algorithm provides minimized efforts for testing the
given Object Oriented Code with reduced time and cost.
Keywords Test case selection (TCS) · Test case prioritization (TCP) · Object
oriented programs (OOP) · Code coverage analysis · Multi factorial code coverage
P. Vats (B)
Banasthali Vidyapith, Vanasthali, Rajasthan, India
e-mail: prashantvats12345@gmail.com
Z. Aalam · S. Kaur · A. Kaur
S.G.T. University, Haryana, India
e-mail: zunaid.aalam6194@gmail.com
S. Kaur
e-mail: satnam_fet@sgtuniversity.org
A. Kaur
e-mail: amritpal_fet@sgtuniversity.org
S. Kaur
C. T. Institute of Engineering and Technology, I.K.G. Punjab Technical University, Kapurthala,
India
e-mail: sarvsaini@gmail.com
722 P. Vats et al.
1 Introduction
Testing OOP is an important part of software development process. Without software

testing Software can’t be successfully completed. It includes process of affirmation
and substantiating the software during the testing of a OOP. Thus the main objective
of OOP testing is to deliver and maintain the quality of Software product to while
installing the OOP at the client end. But a OOP testing involves an over budgeting of
testing cost in the form of efforts and timing. Due to timing and budgeting constraints
sometimes the exhaustive testing of OOP is not possible while executing the all set of
test cases. To reduce the efforts and costing of OOP testing up to some extent, some-
times the prioritization of the selected test cases becomes very necessary, so that while
executing theses selected more error prone software faults can be revealed while
executing these OOP. The selection of test cases and their prioritization becomes
crucial as for fault detection it increases the efficiency of testing of OOP. The conven-
tional TCP techniques which are being used for testing in OOP were not adequate as
they basically concern of either class coverage or coverage of object oriented class
method function calls. In our put forward work, an algorithm is being put forward for
the TCS and TCP which is based on multi-factorial code coverage based approach
for the test case selection and prioritization for OOP. This put forward algorithm uses
nine factors for the TCS and TCP for OOP by to quantifying these factors on the basis
of a test case selection and prioritization formula, a well optimized and prioritized
set of test suite would be obtained for testing the OOP. The put forward algorithm
for testing of OOP functions on the basis of the these factors which includes internal
input files, native method, method, external output files, input variable, conditional
statement, output variable, import packages and Line of Codes. For the affirmation
of the put forward algorithm for testing of OOP it has been considered by applying
on two testing case studies of OOP. The result shows the efficacy of the put forward
technique.
2 Related Work on Testing of OOP
Sultan et al. [1] have presented a regression TCP for OOP based on the craving
graph model of affected test code using genetic algorithm (GA). ESDG (Extended
System Dependency Graph) was put forward to find the class statement based level
changes in the source code. The identified object oriented changes are stored in a
file named changed and coverage information for each test case is generated from
source code. Then the GA is used to selected test cases are determined for their order
of preference.
Panigrahi and Mall [2] put forward a Class model based testing approach to
prioritize the regression test cases for OOP. They presented EOSDG (Extended OOS
Dependence Graph) to represent all closely connected OO features like inheritance,
association and aggregation of OO Classes, polymorphism among them. For the
A Multi-factorial Code Coverage Based Test Case Selection … 723
TCP they put forward M-RTP (Model based regression TCP). A forward test slice
of EOSDG is constructed w.r.t. each modified model elements to identify the main
model elements that are indirectly or directly affected.
Ashraf et al. [3] have put forward value based test case prioritization technique. An
algorithm which uses six factors: changes in requirement, customer priority, imple-
mentation complexity, execution time, requirement traceability, and fault impact of
requirement have been used to determine the priority of the system test cases.
Shahid and Ibrahim [4] put forward a new object code test case based TCP tech-
nique. They put forward a code coverage based test case prioritization algorithm.
The test case which covers the largest portion of code is put at higher priority.
Beena and Sarala [5] put forward coverage based TCS and TCP. They clustered
the test cases into three groups outdated, required and surplus. Then by using these
clusters test case selection algorithm (TCS) is put forward. Then the output obtained
from TCS is given as an input to test case prioritization (TCP).
Vedpal et al. [6] put forward a hierarchical TCP technique for OO software. Their
put forward an algorithm which works at two stages. The classes are prioritized at the
first stage and then test cases of prioritized classed are ordered at the second stage.
Kavitha and Kumar [7] put forward a TCP for multivariate analysis based on
severity. They put forward a TCP technique based on rate of fault detection and fault
impact. This algorithm finds severer fault at the early stage of testing process.
3 Proposed Work
The put forward test case selection and prioritization algorithm prioritizes the selected
suite of test cases by keeping in consideration the nine code coverage testing factors.
The test cases would be selected and will be put in ordered on the basis of these
fault detection coverage of inclusive of these factors. The considered code coverage
testing factors along weight assigned are as shown in Table 1 with having value
between 0 and 1. Every code coverage testing factor would be designated with a
weighing factor value which would denote the ability of the testing factor to detect
the presence of error in the OOP. The ability of a test suit to cover the maximum
number of the testing factors will show the number test cases being covered with
ensuring the maximum coverage of codes of the given OOP and will have the ability
to diagnose with detecting the maximum number of errors in an OOP.
After execution of these the Code Coverage Testing factors the selected test cases
among a test suite would be ordered and prioritized on the basis of the Test Code
Coverage Value (TCCV) metric. More the value of the TCCV, highest would be the
prioritized order of a test case to detect the probability of an error in OOP. The value
for the TCCV for test case would be given by the following Eq. 1,
⎧

n ⎨n
TCCV = lim Test Code Coverge Valuei j
n→∞⎩
i=0 j=1
724 P. Vats et al.
Table 1 Code coverage

S. No. Code coverage testing factors Assigned weights
testing factors with assigned
name
weights
1 Internal input logical file (ILF) 0.2
2 External output interface files 0.2
(EOIF)
3 Native input methods (NIF) 0.5
4 Import package files (IPF) 0.1
5 Input methods (IM) 0.5
6 Conditional logical statements 0.4
(CLS)
7 Line of code (LOC) 0.1
8 Class input variable (CIV) 0.3
9 Class output variable (COV) 0.2

n
∗ Weight Assigned to the Code Coevrge Testing Factor (1)
k=1
The algorithm for the put forward test case selection and prioritization technique
is as shown below:
For a given set of unprioritized test suite T
Set T- > empty, Repeat until T’ = empty.
Begin,
Determine the Code Coverage Testing factors covered by the test suit
Find value of TCCV for the test case.
Recursively call TestCase_Select ();
Arrange test cases in prioritized manner as per the code coverage.
End
4 Experimental Results and Their Analysis
For the experimental analysis and the validation of the put forward algorithm, it is
being explained using the two case studies. The here mentioned case studies and
results were being implemented in the Java language.
4.1 Case Study 1
In the first case study for the put forward algorithm a code in Java is
being implemented for a Class Student with its attributes like Student_Name,
Student_Roll_number and as follows in Fig. 1.
In the first case study a program written in Java is implemented as the OOP to
be tested using the put forward algorithm. This program written in Java feeds in
the information of student like Student name, Student Enrolment number, and their
marks in various subjects. Then the program finds the total of marks for a student
in his various subjects. Further it calculates the average marks for the individual
students. Table 2 shows the Code Coverage Testing Factors Covered by Test Cases
for OOP Fault Detection in Case Study 1. For the Case Study 1 finding the TCCV
for the test cases for the given OOP by applying the Eq. 1 values are given in Table 3,
the errors found in the OOP in the Case Study 1 by the test cases is given in Table 4.
The selected and prioritized order for the test cases for OOP in Case Study 1 is given
by applying the put forward algorithm is as TC3, TC5, TC2, TC1, TC4.
Fig. 1 To show the pseudo code for case study 1
Table 2 Code coverage testing factors covered by test cases for fault detection
S. No. Factors TCI TC2 TC3 TC4 TC5
1 ILF 0 0 0 0 0
2 EOIF 0 0 0 0 0
3 NIM 0 0 0 0 0
4 IPF 0 0 0 0 0
5 IM 3 10 10 2 11
6 CLS 0 1 0 0 0
7 LOC 7 19 17 1 1
8 CIV 0 1 1 0 0
9 COV 1 1 0 0 0
726 P. Vats et al.
Table 3 Values of test case code coverage value TCCV for the given test cases
S. No. Factors TCI TC2 TC3 TC4 TC5
1 ILF 0 0.4 0 0 0
2 EOIF 0.5 0.5 0.5 0.5 0.5
3 NIM 0.6 2 2 0.4 2.2
4 IPF 0 0.1 0.1 0 0
5 IM 2.1 5.7 5.1 1.8 7.5
6 CLS 3.2 8.7 7.7 2.7 10.2
7 LOC 0 0.1 0.1 0 0
8 CIV 0.5 0.5 0.5 0.5 0.5
9 COV 0.6 2 2 0.4 2.2
TCCV 7.5 20 18 6.3 23.1
Table 4 Errors detected by test case for case study 1

S. No. TC1 TC2 TC3 TC4 TC5
El *
E2 * * *
E3 * *
E4 *
E6 *
E7 * *
E8 *
E9 * *
E10 * *
For Case Study 1, The efficacy of the put forward algorithm for the fault detection
along with test case selection and prioritization is adjudicated by computing and
then comparing the value of Average Percentage of Fault Detected (APFD) and
using APFD graph drawn for the non-prioritized and prioritized test case suite is
given in Table 5 and using Fig. 2.
Table 5 Value for APFD for case study 1 for OOP

Test case selection and prioritization for case study 1 APFD metric value (%)
Random test case selection and prioritization approach 46
Proposed test case selection and prioritization approach 64
Fig. 2 To show the put

forward and random
approach for TCS and P for
case study 1
4.2 Case Study 2
In the case study 2 we have considered a program written in Java for various
kinds of banking related daily routines. This OOP written in mainly in Java for
a Banking organization maintains the banking customer related data like Costumer
name, his banking account number and type of Bank account as in Fig. 3. There
would be different object oriented classes that would deal with different type of
accounts like current account, Loan account, saving account etc. The said OOP code
would consists of various functions related to different types banking operations
like for depositing the money into a specific account, withdrawal of money from an
account and displaying account related to different types banking operations like for
depositing the money into a specific account, withdrawal of money from an account
and displaying account related information for different account holders and many
more. The pseudo code for the OOP for the Case Study 2 is shown in the Fig. 3.
Table 6 shows the Code Coverage Testing Factors Covered by Test Cases for OOP
Fault Detection in Case Study 2. For the Case Study 2finding the TCCV for the test
cases for the given OOP by applying the Eq. 1 values are given in Table 7, the errors
found in the OOP in the Case Study 2 by the test cases is given in Table 8. The selected
and prioritized order for the test cases for OOP in Case Study 2 is given by applying
the put forward algorithm is as TC8, TC4, TC5, TC7, TC3, TC2, TC1, TC6. For Case
Study 2, The efficacy of the put forward algorithm for the fault detection along with
test case selection and prioritization is adjudicated by computing and then comparing
the value of Average Percentage of Fault Detected (APFD) and using APFD graph
drawn for the non-prioritized and prioritized test case suite is given in Table 9 and
using Fig. 4.
728 P. Vats et al.
Fig. 3 To show the pseudo code details for the code for OOP for case study 2
5 Conclusion
In this research paper we have put forward a new multi-factorial based test case
selection and prioritization technique for testing the OOP based on the code coverage
of these testing which improvises the rate of fault detection during the early phases of
the Object Oriented Testing. We have also performed an investigative two case studies
for the put forward testing algorithm for OOP. After carrying out the experimental
study for both the case studies the put forward algorithm has provided us a much
improvised results in terms of the rate of error detection at the early phases of testing
in OOP. For the validation of the proposed technique approach we have applied it on
the two codes which were written in Java and using the APFD metric we have found
that the experimental results has shown the effectiveness of the put forward testing
Table 6 Code coverage testing factors covered by test cases for OOP fault detection in case study
2
S. No. Factors TC1 TC2 TC3 TC4 TC5 TC6 TC7 TC8
1 ILF 0 0 0 0 0 0 0 0
2 EOIF 0 2 1 2 0 2 1 1
3 NIM 0 0 0 0 0 0 0 0
4 IPF 1 0 1 1 1 0 1 0
5 IM 0 1 0 0 0 2 0 0
6 CLS 2 2 2 3 2 2 2 3
7 LOC 36 42 31 36 34 48 31 36
8 CIV 1 1 0 0 1 1 0 0
9 COV 2 2 2 3 2 2 2 3
Table 7 TCCV values of test cases for test cases in case study 2
S. No. Factors TC1 TC2 TC3 TC4 TC5 T6 T7 T8
1 ILF 1.2 0.4 0.2 0.4 0.1 0.9 0.2 0.3
2 EOIF 1.3 2.2 1.1 2.3 0.3 2.7 1.4 1.4
3 NIM 1.2 0.3 0.2 0.4 0.4 0.5 0.1 0.3
4 IPF 1.0 0.8 1.4 1.4 1.2 0.4 1.2 0.4
5 IM 0.2 1.3 0.3 0.3 0.1 2.3 0.2 0.2
6 CLS 2.3 2.2 2.4 3.2 2.2 2.2 2.6 3.3
7 LOC 3.6 4.2 3.1 3.6 3.4 4.8 3.1 3.6
8 CIV 1.9 1.9 0.5 0.9 1.1 1.2 0.6 0.9
9 COV 2.0 2.7 2.4 3.1 2.2 2.4 2.4 3.2
TCCV 14.7 16 11.6 15.6 11 17.4 11.8 13.6
Table 8 Errors detected by test case for case study 2

S. No. TC1 TC2 TC3 TC4 TC5 T6 T7 T8
E1 * * * * * *
E2 * * * * *
E3 * * * * *
E4 * * * * *
E5 * * * *
E6 * * *
E7 * *
E8 * * * * *
E9 * * *
E10 * * * * *
730 P. Vats et al.
Table 9 Value of APFD for case study 2 for OOP

Test case selection and prioritization for case study 2 APFD metric value (%)
Random test case selection and prioritization approach 57.5
Proposed test case selection and prioritization approach 71.2
Fig. 4 To show the put

forward and random
approach for test case
selection and prioritization
for case study 2
algorithm for error detection in OOP. As future work the proposed algorithm can be
applied for large chunks of codes and can be implemented as an automated tool to
find its application into the software industry.
References
1. A.B.M. Sultan, A.A.A. Ghani, S. Baharom, S. Musa, An evolutionary regression test case prior-
itization based on dependence graph and genetic algorithm for object oriented programs, in
2nd International conference on emerging trends in engineering and technology, May, 30–31,
London, UK (2014)
2. C.R. Panigrahi, R. Mall, Test case prioritization of object oriented programs. SETL abs Briefings
9(4), 31–40 (2011)
3. E. Ashraf, A. Rauf, K. Mahmood, Value based regression test case prioritization, in Proceedings
of World Congress on Engineering and Computer Science, October, vol. 1, San Francisco, USA,
pp. 24–26
4. M. Shahid, S. Ibrahim, A new code based test case prioritization technique. Int. J. Softw. Eng.
Appl. 8(6), 31–38 (2014)
5. R. Beena, S. Sarala, Code coverage based test case selection and prioritization. Int. J. Softw.
Eng. Appl. 4(6), 39–49 (2013)
6. Vedpal, N. Chauhan, H. Kumar, A hierarchical test case prioritization for object oriented
software, in International Conference on Contemporary Computing and Informatics (IC3I)
(2014)
7. R. Kavitha, N.S. Kumar, Test case prioritization for regression testing based on severity of fault.
Int. J. Comput. Sci. Eng. 2(5), 1462–1466 (2010)
New Horizons for the Application
of Microalgae in the National Economy
L. N. Medvedeva and O. Roiss
Abstract The abstract should summarize the contents of the paper in short terms,
i.e., 150–250 words. The research shows that green technologies, microalgae poten-
tial in particular, will probably become mandatory to the development of youth
entrepreneurship. It is still important to learn about young students’ intentions with
opening their own business. It is important to remember about the transfer of scientific
discoveries about using microalgae in healthy food production and in rehabilitation of
the environment. The hypothesis of scientific research has defined a toolkit, methods
and techniques in two different directions: an assessment of capacity for the develop-
ment of the youth entrepreneurship in the region and the inclusion of the microalgae
into the strategic asset of the national economy.
Keywords The global ecosystem · Imperatives of the national economy · Green

technologies · Potential of microalgae · Regional cluster · Entrepreneurial potential
1 Introduction
Innovative solutions based on environmentally friendly behaviour are united under

the generic name of «green technologies» [1]. There are such major applications
of the green technologies as environmental management, use of renewable energy
sources, reduction of harmful emissions into the atmosphere, development of envi-
ronmentally sustainable landscapes, search for new sources of nutrition and drinking
water [2]. According to one of the predictions, by the 2050 the population of the Earth
will rise to 9 billion people, which will require the intensification of entrepreneur-
ship and search for new sources [3]. The question is about the forming of humans
economical consumer behaviour and the search of intensives to develop innovative
L. N. Medvedeva (B)
Volgograd University, Volgograd 400005, Russian Federation
e-mail: Milena.medvedeva2012@yandex.ru
O. Roiss
GEO Roehren—Und Pumpenwerk Bauer GmbH, A 8570 Voitsberg, Austria
734 L. N. Medvedeva and O. Roiss
Fig. 1 Smart technologies which make it possible for people to monitor their health and balanced
nutrition
youth entrepreneurship. Today’s generation of young people is also known as millen-

nials or Generation Z. Since they were born they have lived in the digital world, none
of them can imagine their life without the Internet, a computer, a smartphone. The
hyperactivity disorder created by digital technologies makes it possible for young
people to interact with a tremendous amount of information and set up international
collectives. According to the Microsoft researches, Generation Z closely monitors the
world brands including personalization of food supplies and responsible consumer
behaviour. Smart devices allow them to programme healthy lifestyle, monitor the UF
radiation, measure the acid–base balance of cutaneous integument and determine the
level of physical activity and the number of burnt calories (Fig. 1) [4].
Wellness industry (deliberate healthy lifestyle) is one of the world top ones now;
the emerging technologies create new opportunities for balanced nutrition. Thus,
the popularity of the category “beauty is inside” is increasing all around the world,
consumers take internally products made of microalgae in addition to the external use
of cosmetics. At the present day, it is granted to eat the following species: Arthrospira
platensis, Arthrospira maxima, Chlorella vulgaris, Chlorella pyrenoidosa, Chlorella
sorokineana, Dunaliella salina, Nostoc pruniforme, Aphanizomenon flos-aquae,
Spirulina [5]. The hypothesis of the scientific research resolves itself into the ratio-
nale of using microalgae in economy on basis of youth entrepreneurship devel-
opment. Descriptive and theoretical methods were applied to the research, among
them are techniques for sociological researches (surveys, predictions modelling),
economic–mathematical modelling; these methods have made it possible to describe
the entrepreneurial approach in the microalgae potential rise. The novelty of the study
is that it is important for youth entrepreneurship to search for new natural resources
that can be transformed into innovations and establish their own production under
their own brand in various sectors of the national or world economy.
The researchers conducted by Volzhskiy Polytechnic Institute branch VSTU (2015–

2019) resulted in interviewing more than two thousand students and young people
of Volzhskiy (Volgograd region, Russia) and separating out groups of youngsters,
developing business in the basis of studying their intentions. As it is known, inten-
tions are the initial phase of becoming an entrepreneur. It often happens so that
New Horizons for the Application of Microalgae … 735
the time of intention—action gap—is quite long. The theory of planned behaviour
lets us distinguish the factors which contribute to one becoming an entrepreneur.
Among them are parents or relatives with their own business, educational attain-
ment, condition of supporting small-sized and medium-sized business infrastruc-
ture in the region. According to the theory of planned behaviour, the motivator of
a young person’s behaviour is their personal attitude to the entrepreneurship [6].
The research has shown that young people who intend to connect their lives with
entrepreneurship proceed with implementation of their intentions when they are still
in the university. They actively participate in scientific groups and speak at confer-
ences. Alongside with family and institute contexts, some external factors affect
students’ forming of entrepreneurship intentions: institutional environment (avail-
ability of SMB supporting programmes) and mentality in the existing society. The
question about how many people intend to move into small business stays important
for the local authorities. The studying produced isolation of three age groups with
diverse impact of external factors on their entrepreneurial intentions (see Table 1).
In Russian Federation, SMB support is carried out in five programmers: “The
improvement of business climate”, “The increase of SMB subjects’ access to finan-
cial resources”, “Acceleration of SMB subjects”, “Popularization of entrepreneur-
ship”, “Creation of “farmers support and rural cooperation” system”. Entrepreneurial
potential of the region can be assessed with the help of the following mathematical
model:
E entrepreneurtial = ρpersonal + ρlogistical + ρinvestment + ρinnovative (1)
in which the components of entrepreneurial potential are calculated by the following

formulas:
Table 1 Descriptive statistics on participation of diverse youth age groups in entrepreneurship

Age group/percent of a sample
The first from 18 to The second from 23 to The third from 27 to
23 years/19% 27 years/45% 35 years/36%
Mean value/Standard deviation
Actual shift from 0.7/0.1 3.2/1.4 4.4/1.9
intention to action
Entrepreneurial 4.6/1.62 5.1/1.7 4.8/1.5
intentions
Presence of family 1.2/0.9 3.2/2.0 5.4/2.1
business
Mentality of a 2.1/0.4 1.0/0.3 0.5/0.02
university
Regional business 0.8/0.1 6.8/2.1 5.1/2.11
supporting institutes
Lx
ρpersonal = W (2)
n
where
W the coefficient of validity based on expert assessments;

L x the elements forming personal entrepreneurial potential;
n the amount of elements.
Mx
ρlogistical = W , (3)
n
where
W the coefficient of validity based on expert assessments;

M x the elements forming personal entrepreneurial potential;
n the amount of elements.
Ix
ρinvestment = W , (4)
n
where
W the weighting coefficient based on expert assessments;

I x the elements that form the investment and business potential;
n number of elements.
In x
ρinvestment = W , (5)
n
where
Inx elements that form innovative business potential;
n number of elements.
The motivational component of an entrepreneur is calculated using the formula:
(L x + Mx + Ix + Inx )
ρmotivational = Wi , (6)
ni
where
800
700
600
500
400
300
200
100
0
2017 2018 2019 2020 2021 2022 2023 2024 2025
Fig. 2 Dynamics of the global biotechnology market, billion dollars. Source Global market insight

ni number of components of internal motives and incentives.
The capacity of the entrepreneurial potential of SMBs in the city of Volzhskiy in
the Volgograd region was calculated based on expert assessments and data from soci-
ological surveys on computer programs—11.87%. This indicates that the promotion
of entrepreneurship needs to be strengthened, especially among young people. It is
also important to help young people find business ideas. The increasing demand for
products produced by biotechnologies offers new possibilities for small businesses.
The global biotechnology market is expected to total $ 742 billion by 2025 (growth,
in comparison with 2019, by two times) (Fig. 2).
There is the potential of the microalgae Chlorella vulgaris among the green tech-
nologies. A microscopic monocellular organism, measuring from 2 to 10 microns
in diameter, is able to assimilate solar energy (the cell consists of chloroplasts)
and form organic substances and disengage oxygen [7]. Dry biomass of Chlorella
vulgaris contains 45% protein, including essential amino acids, 35% carbohydrates,
10% omega-3 fatty acids and microelements (iron, iodine, potassium, selenium,
etc.) [8]. Chlorella vulgaris has entered the ethnic cuisine of Japan, Malaysia and the
Philippines: it is used for making smoothies, cocktails, and added to bread products.
There is a huge industry of dietary supplements based on microalgae, mainly from
Chlorella vulgaris and Spirulina [9]. Scientists at the Harvard T.H. Chan School
of Public Health (USA) have developed a “healthy food Plate” in which proteins
occupy—one-fourth of the plate, complex carbohydrates—one-fourth of the plate,
(vegetables + fruits)—half of the plate. Scientists at Volgograd medical University
are using healthy food products to conduct research on metabolic correction in obese
Fig. 3 Structure of a “healthy food plate” with microalgae products included: dry powder and
suspension of Chlorella vulgaris
patients [10]. Inclusion of microalgae products in medical nutrition is only a matter

of time (Fig. 3) [11, 12].
Research conducted among students on the inclusion of microalgae products in the
“Healthy food Plate” showed that 12% of the respondents are ready to try microalgae
products (38% of the respondents preferred smoothies from Chlorella vulgaris); only
1.2% of the respondents are ready to eat regularly. Most students do not have strong
ideas about microalgae; they include kelp (algae). The most famous microalgae were
named: Chlorella vulgaris and Spirulina platensis (86% of the respondents).
Ideas about microalgae have entered the information space of modern youth,
but they are far from being actually included in the diet. One of the reasons is
the lack of Russian production of microalgae products and their promotion. Since
2006, scientists OF FGBNU VNIIOZ (Volgograd) in cooperation with scientific and
commercial organizations, agricultural and fish-breeding farms have been creating
an evidence base for the legality of the use of the strain of Chlorella vulgaris IFR no.
C-111 in the national economy. The laboratory of the Institute conducts research on
the use of the strain of Chlorella vulgaris IGF# C-111 in animal husbandry, aviculture,
and pond fish farming and for cleaning reservoirs [13, 14]. The BAUER GmbH group
of companies (Austria), realizing that the environment is very sensitive to climate
change, aims to take these factors into account when developing new machines and
making decisions [15]. The company has been working for more than 80 years in
the field of irrigation technologies. Today, 2.5 million hectares around the world are
watered by irrigation machines produced by the hands and skills of 675 employees,
irrigated by hand and skills, among which 64%. Chlorella vulgaris can be used to
restore soil fertility. Using the technology for irrigation and fertilizing made by the
BAUER Group, it is applied three times per season to the soil of agricultural fields
(Fig. 4).
Research on the use of Chlorella vulgaris in the economy is conducted in Germany,
the USA, Japan and China [16]. Designed by Pierre Calleja, street lights generate the
ability of microalgae to release energy during photosynthesis, while simultaneously
absorbing carbon dioxide and saturating urban air with oxygen. The “BIQ House”
project is a building with energy from microalgae, constructed by the design firm
Fig. 4 Watering and fertilizer application, manufactured by the BAUER Group, allows the
introduction of Chlorella vulgaris chlorella into the soil of agricultural fields
Fig. 5 Use of Chlorella vulgaris in economic sectors: street biolights, big house wall biopanels
«BIQ House», biofuel for cars
“Arup” in collaboration with the German company SSC Strategic Science Consul-
tants and the Austrian architectural Studio Splitterwerk. Research is underway in
Germany to obtain automobile fuel from Chlorella vulgaris (Fig. 5) [17].
3 Conclusion
Our research aims to expand the understanding of the intentions and conditions of
youth participation in business. No less important is the study of consumer behaviour
of a person, his desire to lead a healthy lifestyle and use environmentally friendly
products. The introduction of green technologies in the national economy is associ-
ated with the assessment of the entrepreneurial potential of territories and the search
for new natural resources. The demand for products obtained using biotechnologies,
including microalgae, is becoming a global trend. Research conducted at the FGBNU
VNIIOZ (Russia) reveals to young people—future entrepreneurs—the possibilities
of microalgae Chlorella vulgaris in food production and environmental health.
References
1. V.V. Melikhov, L.N. Medvedeva, A.A. Novikov, O.P. Komarova, Green technologies: the basis
for integration and clustering of subjects at the regional level of economy, in Integration
Clustering Sustain. Economic Growth (2017), pp. 365–82
2. M.K. Starovoytov, L.N. Medvedeva, K.Y. Kozenko, Y.M. Starovoytova, G.I. Lukyanov,
M.A. Timoshenko, Rational environmental management: the platform for integration and
optimization of resources, in Contributions to Economics (2017), pp. 347–63
3. Human development report 2011. Sustainable development and equal opportunities are a better
future for all 2011. Retrieved from: http://hdr.undp.org/sites/default/files/hdr_2011_ru_summ
ary.pdf. Last accessed 2020/05/12
4. K. Bogatyreva, G. Shirokova, From entrepreneurial aspirations to founding a business: the case
of Russian students. Foresight STI Gov. 11(3), 25–36 (2017)
5. N.I. Bogdanov, Environmental dimension aspects of planktonic Chlorella use, in Materials of
the II International Scientific Conference on Global Environmental Problems: Local Solution
(Pero, Moscow, 2019), pp. 8–11
6. I. Ajzen, The theory of planned behavior. Organ. Behav. Hum. Decis. Process. 50(2), 179–211
(1991)
7. H. Tang, M. Chen, K.Y. Simon, S.O. Salley, Continuous microalgae cultivation in a
photobioreactor. Biotechnol. Bioeng. 109(10), 2468–2474 (2012)
8. T. Granata, Dependency of microalgal production on biomass and the relationship to yield and
bioreactor scale-up for biofuels: a statistical analysis of 60+ years of algal bioreactor data.
BioEnerg. Res. 10(1), 267–87 (2017)
9. J.C. Bommer, Z.J. Sveida, B.F. Burnham, Further studies on the relationship between tetrapyr-
role structure and usefulness as photosensitizers, in Proceedings of 1st International Conference
on Clinical Applications of Photosensitization for Diagnosis and Treatment (1986), p. 129
10. C.C. Shalaeva, Possibilities of metabolic correction of obesity patients in outpatient practice,
in Materials of the VI Congress of Therapists of Siberia (Novosibirsk) (2018), pp. 80–89
11. V. Minina, M. Ivanova, E. Ganskau, Healthy nutrition in everyday life of Russians J. Sociol.
Soc. Anthropol. 21(4), 182–202 (2018)
12. A.L. Tumanova, Experimental studies on the effect of food suspension of microalgae Chlorella
vulgaris on the human body. Int. J. Appl. Basic Res. 9, 85–88 (2011)
13. L.N. Medvedeva, M.V. Frolova, M.V. Moskovets, A.V. Medvedev, Introduction of nature-
saving technologies is an environmental imperative in the development of regions. J. Volgograd
State Univ. Econ. 21(4), 126–140 (2019)
14. A.E. Novikov, V.V. Melikhov, M.I. Filimonov, M.I. Lamskova, A.G. Bolotin, A.A. Novikov, T.
Karengina, Pat. of the Russian Federation No 2015136412 appl. 27.08.2015, publ. 10.01.2017
15. BAUER GmbH. Retrieved from: https://www.bauer-at.com/ru/products/irrigation. Last
accessed 2020/05/02
16. L.N. Medvedeva, K.Y. Kozenko, O.P. Komarova, Environment quality management in green
cities. Eur. Res. Stud. J. 19(2), 34–45 (2016)
17. Working materials for the strategy of development of the biotechnological industry until
2020 year (project). Retrieved from: https://www.biorosinfo.ru/papers-society/Strategy_Bio
industry.pdf. Last accessed 2020/02/12
Context of the Ghanaian Government:
Social Media and Access to Information
Adasa Nkrumah Kofi Frimpong, Ping Li, Samuel Adu-Gyamfi,
and Sandra Chukwudumebi Obiora
Abstract Often overlooked–especially since Africa and technology are not usually
encountered together–such innovative thinking concerning the political use of social
media is rare in all its circumstances. In the middle of the nineteenth century, Conrad’s
famous description of the European colonial invasion still lingers in the minds of the
West. The paper investigates the extent of social media use by the Ghanaian govern-
ment and how it compares to other OECD countries. By using data from Twitter, the
results show that the Ghanaian government’s social media accounts are openly acces-
sible. The government leverages the power of Twitter as an important informational
and participatory medium in increasing people’s awareness and engagement in the
government. The extent of the government’s social media use is quite comparable
to other countries in the OECD region. Social media may not have enormous social
and economic value creation impact; developing countries like Ghana with limited
financial resources can make good use of the platform.
Keywords Government information access · Social media · Ghana · Information

and communications technology · E-government
1 Introduction
Mergel [1] seems to be the probable forerunner in a theoretical study of the role
of social media (SM) in government. Her research includes innovation and innova-
tiveness of SM in public circles. She has equally outlined a multistage procedure for
adopting SM in the contemporary government [2]. SM and the public ecosystem have
A. N. Kofi Frimpong (B) · P. Li · S. C. Obiora

School of Management and Economics–Center for West African Studies, University of Electronic
Science and Technology of China, No. 2006, Xiyuan Ave, West Hi-Tech Zone, Chengdu, Sichuan
611731, China
e-mail: adasankrumahkofi@gmail.com
S. Adu-Gyamfi
Department of Information Technology Education, University of Education, 1277 Kumasi,
Winneba, Ghana
742 A. N. Kofi Frimpong et al.
learned about participating and collaborating in government for transparency vis-à-

vis the interconnected world. Many researchers like [3] elaborate on the essential
definitions of SM and tools in contemporary governance and providing an under-
standing of their usefulness [3, 4, 1]. As such, we review the current state of the art
on SM in governance with regard to information access In Ghana. The country has
earned enviable respect in democracy following her successful elections in 1992,
1996, 2000, 2004, 2008, 2012, and 2016. All these elections, mainly the 2012 and
2016, saw extensive use of SM [4]. Given the apparent role it plays, growing numbers
of studies have sought to examine the influence of SM in the political sphere. For
example, the research of [5] shows that political media, rationality, soundbites,
and emotion have a positive effect on voter behavior and governance. In fact, [4]
studied the dynamics of Ghanaian youth on SM toward political participation, knowl-
edge, and internal efficacy. Imperatively, [4] showed that SM is promoting political
discourse and enhancing civic involvement. [6] established that Web sites and SM
platforms (Facebook and Twitter) have served as a contemporary form of informa-
tion dissemination between local government, including metropolitan, municipal,
and district assemblies (MMDAs) and citizens. The paper presents Ghana’s access
to information, and the government’s social media use context, based on documental
and secondary data analysis.
The paper presents the extent of political and social media use in the Ghanaian
context. We discuss the involvement of the media and how to help to disseminate
information to ordinary citizens on social media. The paper highlights how the inter-
play of the government, media, and citizens on social media can contribute to sustain-
able governance. By using data from Twitter, the paper analyzes the extent of the
political use of social media and how the usage leads to overturning governmental
decisions toward good governance. The studies also show the influence of Ghanaian
politicians on Twitter. The analysis shows that the government leverages the power
of Twitter as an important informational and participatory medium in increasing
people’s awareness and engagement in democracy. We believe that even though
social media may not have enormous social and economic value creation impact,
developing countries like Ghana with limited financial resources can make good use
of Twitter.
2 Media
Ghana has a robust media. The Ghana National Communications Authority [7] is
the official licensing and regulator of the Ghanaian media industry. NCA is respon-
sible for making sure that the media industry operates in a transparent manner that
promotes peace for national and sustained development. There are 128 certified TV
stations in Ghana [8]. Ghana Today Television (GTV) is the central station owned by
the Ghana government, which is operated by the Ghana Broadcasting Corporation
(GBC). As an agency of the Ghana Ministry of Information, GBC is the leading
information hub of the country. It is constitutionally mandated to ensure information
Context of the Ghanaian Government: Social Media … 743
broadcast [9]. Compared to TV, there are more radio stations in Ghana, of which
505 are authorized and 392 in full use [10]. There are 31 government radio stations,
five foreign, 81 community-based radio, 22 campus-based, and 366 fully private and
commercial stations. The traditional newspaper readership is massive in Ghana with
about 15 newspapers. The government-owned newspaper has an enormous audi-
ence of about 1.5 million [11]. Contemporary Internet publishing is also prevalent in
Ghana. There is an annual award scheme for blogging and SM [12]. The media is a
crucial disseminator of government information. Their activities are standardized by
the National Media Commission constitutionally tasked to ensure the highest stan-
dard of media reportage is maintained. They contribute massively to the Ghanaian
political discussion. Ghana has relative freedom of the media with no restriction
whatsoever [13].
2.1 Government’s Access to Information
The African Charter of Human and Peoples’ Rights by the Organization of African
Unity upholds freedom of information to a very high degree through its declara-
tion of Principles on Freedom of Expression in Africa [14]. The charter mandates
all member countries to enforce all legislative instruments to ensure by protecting
people’s rights with regard to accessing government information. To make the charter
fully functioning, there exist specific standards serving as member countries bench-
marks, including the African Principles of people’s Freedom of Expression and the
Model Law on Access to Public Information. These are all adopted by the African
Commission on Human and Peoples’ Rights. In Ghana, Article 21(1)(f) of Ghana’s
1992 Constitution gives full freedom to people and the media, including freedom of
information [15]. The act provides the Ghanaian people the constitutional right to
information held by the government and arms of government. Equally, as a demo-
cratic country, the law gives full exemptions that are necessary and consistent with
the protection of matters that are of public interest. In 2015, Ghana, through the
United Nations, committed to fully adopt and implement access to information laws
in her quest to achieving the UN’s Sustainable Development Goals [16].
All government SM accounts are openly accessible. There are available down-
loadable and exportable files in CSV, Xls, pdfs, and videos to make information
easily accessible. Interactivity on government Web sites cannot be said to be the
best, but one can customize some form of data (see [17]). The GSS maintains and
updates country statistics for all to access. Statistics, publications, databases, projects
via the National Data Archive [17] can all be accessed. One can also make a formal
data request either online or by post from several databases to compare, analyze, or
diagnose their unique situation.
2.2 Social Media and Elections in Ghana
The involvement of social media in disseminating information to the general public

by governments is on the ascendency of which Ghana is no exception. SM use in
Ghana has been reported by [18, 19, 12]. The Ghanaian government uses different
SM platforms to engage citizens in governmental activities, including the politicians
in opposition [19]. The government uses SM as an information hub to keep citizens
abreast of government activities. In this way, people can monitor and participate in
these activities. The Golden Jubilee House comprises the official seat and office of
the Government of Ghana, responsible for communicating all regulatory actions and
government information to the citizenry. Following, the data from Twitter, one of the
main communication instruments from Ghana’s office of the president, is analyzed.
2.3 Social Media Use by Ghanaian Politicians
In the following review, we dwell on data in its natural setting to do our analysis.
The data comes from the popular SM platform: Twitter. We considered only verified
Twitter accounts. Verified Twitter accounts are those that have been sealed by Twitter
as authentic and usually of public interest [20]. Ghana is host to about 10 million
Internet users from a population of about 30 million [21]. About 5.6 millions (19%
of Ghana’s population) are active SM users [22]. The most preferred platform is
WhatsApp, with an estimated 30% SM users trailed by Facebook (28%).
We used the NCapture extension in NVivo 12 to collect the SM data on Feb. 01,
2019. The program (NVivo 12) is the primary content analytical tool using queries
and comparison diagrams. Some of the information was exported and analyzed in
SPSS.
Starting from the year 2013, the government leveraged on the power of Twitter as
an important informational and participatory medium in increasing people’s aware-
ness and engagement in the government. The figure shows the timeline of tweets and
retweets emanating from the presidency. The analysis shows that from the period
beginning 2013, there were 89% of tweets compared to 11% retweets. The Ghanaian
government uses Twitter to disseminate government projects, services, both local
and foreign activities, laws and by-laws, different forms of governmental alerts, and
updates on diverse issues ranging from a currently passed to taxpayer notices. All
these take place in different transmission formats such as video, text, animations,
audio, and graphics. There are easy-to-see navigation buttons that link the official
Twitter account to the Ghana government Web sites. People can equally communicate
with the government or make service requests by tweeting to Twitter handler.
We share the opinion of [23] that developing countries like Ghana with limited
financial resources can make good use of Twitter. Indeed, the activities of the Ghana
information service through the Ministry of Information, the official public rela-
tions bureau of the Government of Ghana, have changed. Unlike the traditional
paper delivery of information on vans, the department has a vast SM presence

with verified Facebook, Twitter, and YouTube accounts. They are mandated by
the Ghanaian constitution as the department responsible for two-way government–
citizen communication to let people know about ongoing and future policies, activi-
ties, and programs, with a broader vision to promote Ghana as a brand in the interna-
tional market. The Ghana statistical service (GSS), the overseer of the Ghana National
Data Archive, Country Statistics, Open Data Initiative, and the National Reporting
Platform, also has significant social media presence. GSS was established in 1985
by the Ghana statistical service law (PNDC Law 135) with an official mandate
to conduct surveys/censuses by collecting, compiling, analyzing, abstracting, and
publishing Ghanaian statistical information [24]. Similar to other government portals,
GSS has data visualizations of industry statistics related to Ghana. There are year-
on-year inflation trends, GDP, and other fact sheets. All these have ensured massive
people participation in the Ghanaian government. For example, on Jan. 31, 2019,
The Ghanaian president made a statement on a campaign platform that people who
are complaining about his lack of achievement as president are either deaf or blind
[25].
This statement attracted so many social media comments, including the Ghana
Federation of Disability, that the president tended to disdain blind and deaf people as
socially unimportant. Finally, feedback came from the Ghanaian president through
an apology on his social media platforms. People seem active in policymaking. They
contribute to the general decision-making and service delivery processes on social
media. Khan [23] explains that comments and feedback are vital in this process.
The main opposition leader in Ghana equally has a vast social media presence. With
verified Twitter, Facebook, and other accounts, he uses these platforms tremendously.
As shown in Table 1, he has more followers than the president. The implication is that
he subscribes more to other people’s tweets and updates (tweets) more. Again, he
receives more direct messages (username references) than the president. However,
it shows that people usually mention and repost (retweet) the Ghanaian president’s
messages more than the main opposition leader. From which we consider, he is more
popular and influential. The factors that dwell on this need further exploration.
From our discussion so far, we agree with the notion [26] that a governments’
use of SM can create sustainable governance and political change (through voting
realignment) in two main categories: information and interaction. Governments can
use SM to deliver public services better and publicize political details to reach out
to citizens from all walks of life. This results in enhanced transparency in govern-
ment. Conversely, governments can use SM to improve their interaction with internal
and external stakeholders. To achieve, “we-government” requires improved commu-
nication/collaboration and public participation in governmental decision-making
processes as enshrined in “citizen coproduction” encompassed by “citizen sourcing.”
Several examples of information and government interactivity exist in Ghana. A
notable example is Afrobarometer [27], a watchdog organization. Through copro-
duction (citizen sourcing), the Afrobarometer has provided a citizen-based evaluation
system where Ghanaians periodically evaluate the government.
Table 1 Influence Ghanaian politicians on Twitter

Country/role Twitter account Owner Number of followers
Ghana’s president @NAkufoAddo Nana Akufo-Addo 884,910
Ghana’s opposition @JDMahama John Dramani Mahama 1,163,733
leader
Ghana @GhanaPresidency Office of the president 47,027
of the Republic of
Ghana
Estonia @EstonianGovt Estonian government 13,659
Finland @FinGovernment Finnish government 16,522
Malta @MaltaGov Government of Malta 16,965
Israel @Israel Israeli government 543,651
South Korea @TheBlueHouseENG The Office of president 22,056
Japan @JapanGov The Government of 265,594
Japan
Turkey @trpresidency Presidency of the 1,051,346
Republic of Turkey
As shown in Figs. 3 and 4 (and others), information from such systems has helped
to reduce information asymmetries and addressed the principal–agent problem. For
example, the most trusted institution in Ghana is the Ghana Armed Forces, while the
Ghana Police Service is the most corrupt [29]. The government can also monitor the
pattern of citizens’ behavior on such platforms. For example, Ghana’s parliament
dropped their decision to build a new parliamentary chamber after public outrage
following a video that went viral on SM (Table 2) on the arrests of three protesters after
they screamed: “drop that chamber” from the public gallery of the Ghana parliament
in protest against the decision [30]. Table 2 shows the extent of Twitter engagement
from the sampled stations after the announcement of the arrests in Jul. 07, 2019.
There are around 53 TV stations in Ghana. The radio stations are more numerate, of
which 505 are authorized, and 392 are in full use [10]. Table 2 suggests that, indeed,
Ghanaians have opportunities to participate in policy refinement and that crowd-
sourcing is a vital component of SG. This shows the structural configuration and
iterative nature of SG at recognizing and treating vulnerabilities while establishing
resilience. To further our understanding of Table 2, we compare it to other famous
political arrests around the world, as shown in Table 3. The result is quite remark-
able. We, therefore, infer that mining of the collective knowledge could produce
ideas and expertise that can be used to help improve the quality of governmental
decision making in Ghana. Stamati et al. [31] describe this participation as one of
the three principles of openness in SG. Openness is achieved when a government
discloses information in a manner that is easily accessible, coupled with widespread
citizen engagement through “frank and honest discussion about processes, policies,
and decisions” [31]. This requires three main principles: participation, collaboration,
and transparency [32]. Collaboration shows the extent to which governments estab-
lish partnerships with different stakeholder groups, including public agencies at all
governmental levels, private institutions, and individuals geared toward improving
government effectiveness. The extent of collaboration by the Ghanaian president, and
the opposition leader, seems significant (refer to Figs. 1 and 2). They collaborate with
public agencies at all levels (the electoral commission of Ghana), private institutions
(EAA Foundation), and individuals. Indeed, collaboration with Ghanaian political
leaders on SM for SG is achievable.
Transparency refers to the continuous distribution or flow of reliable informa-
tion (economic, social, and political) emanating from the government to all relevant
stakeholders in a timely fashion [33]. Transparency is a pivotal condition of govern-
ment accountability, and the two concepts go hand-in-hand, which, in combination,
contribute to sustainable governance [34]. Citizens must have the right to know what
the government is doing and why primarily to address the principal–agent problem.
As the principals (citizens), they have the right to know about the government’s
(agent’s) behavior to achieve inclusiveness. SM has proven to be a vital infome-
diary platform that translates and increases the meaning of government information,
providing accountability, in turn, leading to good governance [34].
Furthermore, the effectiveness of government information disclosure (trans-
parency) increases with active SM. Increased transparency can lead to improved
governance outcomes by considering the activities and actors that facilitate this
transformation, especially on SM. Table 2 contains examples of how government
information has led to the timely reversal of government decisions in Ghana. In this
Table 2 Twitter engagement following the broadcast of a viral video showing the arrests of three
protesters in parliament [30]
TV station Twitter Handle Comments Retweets Likes
Tv3 Network @tv3_ghana 184 542 19,000
GHOne TV @GHOneTV 369 12,000 42,000
United TV @utvghana 386 79 16,000
Table 3 Twitter engagement following the broadcast of famous political arrests around the world
TV station Twitter Handle Comments Retweets Likes Caption
CNN Africa @CNNAfrica 10 88 105
Dec. 10, 2019—Nigeria is facing backlash over the rearrest and detention of journalist and
activist Omoyele Sowore
309 296
Dec. 06, 2019—Wanzhou Meng, the CFO and deputy chairwoman of Huawei, has been
arrested in Vancouver and is facing potential extradition to the U.S.
Al Jazeera English @AJEnglish 39 545 330
October 7, 2018—Breaking news: Saudi journalist Jamal Khashoggi killed inside the Saudi
consulate in Istanbul, says Turkish police
Fig. 1 Frequency of social media use (Twitter) by Ghana’s political leadership in January 2019
8000
6798
7000
6000
5000
4000
3000 2109
2000 1191 1304 1048
1000 195 78 294 41 54
0
Username Popular Popular Tweets Retweets
References Hastag Ref @MenƟons
President OpposiƟon Leader
Fig. 2 Frequency of social media (Twitter) of Ghana’s leadership in February 2019
way, transparency holds the government to account for their actions. SM is pivotal
to making sure that the powers invested by citizens in the state are not abused or
misused. Transparency is vital for increasing the trust and democratic participation
of citizenship for sustainable governance [31].
We believe that there is a perfect relationship between political discussion on
SM and sustainable governance through voting realignment. The above mechanisms
(ideologies, national policies, leaders, systems, and structures within the political
Fig. 3 Level of corruption in Ghana 2012-2019 (Source [28]). Question In your opinion, over the
past year, has the level of corruption in this country increased, decreased, or stayed the same?
Fig. 4 Ghana’s economic condition 2002–2019 (Source [28]). Question How well or poorly would
you say the current government is handling the fight against corruption?
system) are vital for providing Ghanaian voters with the information they need to
decide whether to vote for one political party over another.
3 Conclusion
The paper presents the extent of political and social media use in the Ghanaian
context. We discuss the involvement of social media in disseminating information to
the general public by governments. We dwell on data in its natural setting to analyze
the extent of the political use of social media. The data comes from the popular
social media platform: Twitter to show the influence of Ghanaian politicians on
Twitter. The analysis shows that the government leverages the power of Twitter as an
important informational and participatory medium in increasing people’s awareness
and engagement in democracy (refer to Figs. 1 and 2). We believe that even though
social media may not have enormous social and economic value creation impact,
developing countries like Ghana with limited financial resources can make good use
of Twitter.
References
1. I. Mergel, Social media in the public sector: A guide to participation, collaboration, and
transparency in the networked world (Wiley, San Francisco, 2012)
2. I. Mergel, S.I. Bretschneider, A three-stage adoption process for social media use in
government. Public Adm. Rev. 73(3), 390–400 (2013). https://doi.org/10.1111/puar.12021
3. J.I. Criado, R. Sandoval-Almazan, J.R. Gil-Garcia, Government innovation through social
media. Govern. Inform. Q. (2013)
4. W.S. Dzisah, Social media and elections in Ghana: enhancing democratic participation. Afr. J.
Stud. 39(1), 27–47 (2018). https://doi.org/10.1080/23743670.2018.1452774
5. A.D. Kofi Preko, K. Agbanu, M. Feglo, Political marketing strategy: soundbites and voting
behaviour in contemporary Ghana. J. African Bus. 1–20 (2019). https://doi.org/10.1080/152
28916.2019.1641305
6. K. Asamoah, E-governance in Africa’s local governments: do district assemblies in Ghana
optimize the use of websites and social media? Electron. J. Inf. Syst. Dev. Ctries. 85(4). https://
doi.org/10.1002/isd2.12082
7. NCA, Ghana national communications authority, in National Communications Authority-
Communications for Development, 2019. https://www.nca.org.gh/the-nca/corporate-statem
ent/. Accessed March 09, 2019
8. NCA, List of Authorised TV Broadcasting Stations in Ghana as at Third Quarter of
2017, National Communications Authority, in Ghana National Communications Authority
News (2018). https://www.nca.org.gh/media-and-news/news/list-of-authorised-tv-broadcast
ing-stations-in-ghana-as-at-third-quarter-of-2017/. Accessed Mar. 09, 2019
9. S. Buckley, Ghana Broadcasting Study: A Report for the Government of Ghana and the World
Bank, Accra, 2005. [Online]. Available: http://siteresources.worldbank.org/INTCEERD/Res
ources/WBIGhanaBroadcasting.pdf
10. NCA, Authorised FM Radio Stations as of Second Quarter of 2017 , National Communi-
cations Authority (2018). https://www.nca.org.gh/media-and-news/news/authorised-fm-radio-
stations-as-at-second-quarter-of-2017/ (accessed Mar. 09, 2019)
11. K. Zurek, Daily Graphic leads print media readership in Ghana-GeoPoll-Graphic Online,
Graphic online (2018). https://www.graphic.com.gh/news/general-news/daily-graphic-leads-
print-media-readership-in-ghana-geopoll.html. Accessed Mar. 09, 2019.
12. J. Abdulai, Full list of winners: Ghana blogging and social media awards 2016|Circumspecte,
Circumspecte.com (2016). https://circumspecte.com/2016/05/winners-ghana-blogging-social-
media-awards-bloggers/. Accessed Mar. 09, 2019
13. BBC, Ghana profile-Media -BBC News (2017). https://www.bbc.com/news/world-africa-134
33793. Accessed Mar. 09, 2019
14. AU, “African Charter on Human and Peoples’ Rights| African Union,” African Charter on
Human and Peoples’ Rights, 1981. https://au.int/en/treaties/african-charter-human-and-peo
ples-rights (accessed Jan. 24, 2019)
15. CLD, Ghana: analysis of the right to information bill, Accra (2018). [Online]. Available https://
www.law-democracy.org/live/wp-content/uploads/2018/06/Ghana.FOI_.Jun18.pdf
16. AFIC, “Urgent Call on the Advancement of Citizens’ Access to Information in Ghana - IFEX,”
Africa Freedom of Information Centre (AFIC), 2017. https://www.ifex.org/ghana/2017/10/12/
access-to-information-ghana/ (accessed Jan. 24, 2019)
17. GSS, “Ghana Statical Service,” Ghana Statical Service - Statistics, Publications, Databases,
Projects, 2018. http://www2.statsghana.gov.gh/publications.html (accessed Mar. 09, 2019)
18. J.-A. Kuuire, “2017 Ghana Social Media Rankings,” Tech Nova, 2018. https://technovagh.
com/2018/02/20/2017-ghana-social-media-rankings-avance-media-cliqafrica-eazzy-social-
dream-ambassadors-foundation-gh/ (accessed Feb. 17, 2019)
19. C. Opara, “Behind The Scenes: How Social Media Played A Role In Ghana’s 2016 Elec-
tions,” Tech Nova, 2017. https://technovagh.com/2017/03/18/behind-the-scenes-how-social-
media-played-a-role-in-ghanas-2016-elections/ (accessed Feb. 17, 2019)
20. Twitter, “Verified accounts,” Twitter Help Center, 2019. https://help.twitter.com/en/managing-

your-account/about-twitter-verified-accounts (accessed Feb. 08, 2019)
21. Internetworldstats, “Africa Internet Users, 2019 Population and Facebook Statistics,” Internet
Users Statistics for Africa, 2019. https://www.internetworldstats.com/stats1.htm (accessed Apr.
20, 2019)
22. K. Zurek, “Over 10 million Ghanaians use the internet - Report - Graphic Online,”
Graphic Online, 2018. https://www.graphic.com.gh/news/general-news/over-10-million-gha
naians-using-the-internet-report.html (accessed Apr. 20, 2019)
23. G. F. Khan, Social Media for Government. A Practical Guide to Understanding, Implementing,
and Managing Social Media Tools in the Public Sphere. Singapore: Springer Singapore, 2017
24. GSS, “Ghana Statistical Service Legal Mandate,” Ghana Statistical Service, 2017. http://sta
tsghana.gov.gh/aboutgss.php?category=MjkwMzA1NjI0LjE0MTU=/webstats/oq43q9p651
(accessed Jan. 30, 2019)
25. Starrfmonline.com, “Akufo-Addo apologizes to disabled people over campaign comment,”
Ghaha Web, 2019. https://www.ghanaweb.com/GhanaHomePage/NewsArchive/Akufo-Addo-
apologises-to-disabled-people-over-campaign-comment-720478# (accessed Feb. 07, 2019)
26. L. Zheng, T. Zheng, Innovation through social media in the public sector: Information and
interactions. Gov. Inf. Q. 31, S106–S117 (2014). https://doi.org/10.1016/j.giq.2014.01.011
27. Afrobarometer, “Afrobarometer - Ghana,” Afrobarometer - Center for Democratic Develop-
ment (CDD-Ghana), 2019
28. J. Appiah-Nyamekye Sanny and S. Adusei Baaye, “Ghanaians voice dissatisfaction with living
conditions, government economic performance,” Accra, 332, 2019
29. MyJoyOnline, Fight against corruption: government scores damning marks in the latest report,
Joy Online (2019)
30. E. Appiah, Video: 3 protesters arrested after screaming ‘drop that chamber’, Joy Online (2019)
31. T. Stamati, T. Papadopoulos, D. Anagnostopoulos, Social media for openness and account-
ability in the public sector: cases in the Greek context. Gov. Inf. Q. 32(1), 12–29 (2015).
https://doi.org/10.1016/j.giq.2014.11.004
32. S.A. Chun, L.F.L. Reyes, Social media in government. Gov. Inf. Q. 29(4), 441–445 (2012).
https://doi.org/10.1016/j.giq.2012.07.003
33. D. Kaufmann, A. Bellver, Transparenting transparency: initial empirics and policy applications.
SSRN Electron. J. (2005). https://doi.org/10.2139/ssrn.808664
34. J. Fox, The uncertain relationship between transparency and accountability. Dev. Pract. 17(4–5),
663–671 (2007). https://doi.org/10.1080/09614520701469955
Convolutional Neural Network and Data
Augmentation for Behavioral-Based
Biometric User Identification
Sakorn Mekruksavanich and Anuchit Jitpattanakul
Abstract One classification problem that is especially challenging is biometric iden-

tification, which links cybersecurity to the analysis of human behavior. Biometric
data can be collected through the use of wearable devices, especially smartphones
that incorporate a variety of sensors, during the performance of activities by the users.
In recent research, numerous identification systems using machine learning classi-
fication algorithms have been proposed to provide solutions to this classification
problem. However, their ability to perform identification is limited to only suitable
selected features of time-series data from the biometric raw data. Therefore, in this
study, an architecture for the biometric user identification using walking patterns
that employs a convolutional neural network is proposed. Validation of the proposed
framework for biometric identification was accomplished through the generation
of synthetic data from datasets of samples from public walking. As a result, the
framework described in this study provides more effective outcomes in terms of the
accuracy of the model and additional metrics than the conventional machine learning
utilized for biometric user identification.
Keywords Biometric identification · Convolutional neural network · Data

augmentation
S. Mekruksavanich (B)
Department of Computer Engineering, School of Information and Communication Technology,
University of Phayao, Phayao, Thailand
e-mail: sakorn.me@up.ac.th
A. Jitpattanakul
Intelligent and Nonlinear Dynamic Innovations Research Center, Department of Mathematics,
Faculty of Applied Science, King Mongkut’s University of Technology North Bangkok, Bangkok,
Thailand
e-mail: anuchit.j@sci.kmutnb.ac.th
754 S. Mekruksavanich and A. Jitpattanakul
1 Introduction
The development of wearable technology has been continuous and a new type
of human–computer interaction has evolved with the accelerated development of
communication and information technologies. Many practical applications, such as
business transactions, fitness tracking, healthcare monitoring, and smart home moni-
toring, can be effectively conducted at any time through the use of wearable devices,
in particular smartphones that incorporate a broad range of sensors that are connected
to the Internet [2]. With regard to connectivity, the collection of biometric data,
including pulse rate, electrocardiogram (ECG), electromyography (EMG), move-
ment signals, etc., is enabled by wearable devices. Thus, a diverse range of new
personalized applications that can be employed as wearable sensors for the recording
of biometric information can be provided by these wearable devices [4–6].
Presently, biometrics user identification that connects cybersecurity with human
behavior analysis is one of the most challenging classification problems. User iden-
tification systems are now based on something that you are (biometrics) rather than
something that you possess (cards or keys) as cards, keys, or passwords can be discov-
ered, copied, lost, or stolen. As a result, the collection of biometric information by
wearable devices is focused on a person’s intrinsic characteristics that requires their
physical presence, in which the measurement of one or more biometric traits can be
applied to either authenticate or identify a subject [3]. In this way, the advantage of
identification systems utilizing physiological features, for example, ECG, EMG, and
other activity signals, is that the mobility of a user and a higher level of certainty of the
individual liveness are assured, as the pattern of walking activity is especially useful
for user identification. Therefore, many researchers use walking patterns collected
by wearable devices for biometric identification, as they provide features that are
unique to each person.
In a large number of identification systems that have been recently proposed, the
decision mechanism that is used involves machine learning (ML) classification algo-
rithms. The performance of an identification system is measured by the effectiveness
of its classification results. User studies in which participants are selected, multiple
measures of observables are conducted, and the datasets are constructed from the
measurements for a classification algorithm are often conducted to validate these
systems. However, human-crafted feature extraction and selection is a problem is
difficult for ML algorithms. In this study, the research focus is on biometric user
identification that employs walking patterns, and the results can primarily contribute
the following benefits:
1. The proposal of an architecture with convolutional neural networks for the
purpose of biometric user identification utilizing walking patterns.
2. Various data augmentation techniques that enhance the performance of the
proposed biometric user identification are studied.
3. That the proposed framework for the biometric user identification provides
better performance than the baseline performances is demonstrated; hence, the
establishment of a UIWADS dataset is completed.
Convolutional Neural Network and Data Augmentation … 755
The rest of the paper is organized as following. Section 2 presents the related works
of the biometric user identification. Section 3 describes the proposed framework of
data augmentation techniques for the biometric user identification. Section 4 shows
the experimental setting, dataset description, and results. Finally, Sect. 5 concludes
the results.
2 Biometric User Identification Using Sensor Data
Recently, many research studies have investigated the problems of biometric user
identification [11]. Biometric methods are divided into two types, which are behav-
ioral biometrics and physiological biometrics. The extraction of behavioral charac-
teristics from user actions, for example, handwaving, keystrokes, voice and walking
patterns, are the basis of behavioral biometrics. Physiological biometric methods
employ physiological traits, namely fingerprints, the iris, or facial features. In this
research, the focus is on behavioral biometric methods, as this work is concerned
with the data of walking patterns. Wearable devices such as smartwatches and smart-
phones can be used to measure many of these user actions, and walking is the activity
that is most commonly examined in behavioral biometrics.
The use of accelerometers has provided fairly good performance for gait iden-
tification methods in the past. In Pan et al. [7], five different accelerometers were
attached to the subjects, and it was shown that a combination of all of the data can
improve gait recognition. Although this approach provides highly accurate results for
a dataset with the multiple recording of 30 subjects, the need for an extensive setup
for the data collection diminishes the usefulness of the results. Similarly, several
sensors, one at the hip and one on the ankle, were attached to each participant by
Gafurov et al. [1] in order to collect the results for a dataset containing 50 subjects.
For the purpose of examining the ways that the gait pattern can be impacted, the
subjects were requested to walk normally both while carrying a backpack as well as
without it. Rong et al. [9] used a single sensor attached at the waist to obtain a dataset
with 21 subjects. The results from these methods are promising; however, special
accelerometer sensors built for this purpose must be used for the recording of the
data from the sample. Although the size of these sensors is generally quite small, it
is necessary to attach the sensor to the subject, usually with their knowledge.
The biometric user identification framework, which is proposed in this work, enables
the walking data captured from accelerometer sensor to identify the user performing
the walking activity. Figure 1 shows the proposed methodology to achieve the
research objective in this study.
Fig. 1 Proposed biometric user identification framework of this study
In this identification framework, raw activity data from the UIWADS dataset are
used as raw data. The 30% of raw data are spitted and keep separately as test data
for evaluating trained models. In the step involving data augmentation employing
four different techniques (e.g., jittering, scaling, rotation, and time warping), the
remaining data are used [10]. The purpose of these data augmentation methods is
the generation of synthetic training data. Following this, the augmented training data
are combined with the raw training data, which are then employed for the training of
the CNN model. Finally, the trained CNN model’s performance is evaluated by the
test data.
3.1 Description of the Dataset
For the evaluation of the data augmentation techniques that are proposed, a commonly
used benchmark dataset for user identification from walking activity known as the
walking activity dataset (UIWADS) was considered due to this dataset being required
to meet the criteria as follows: unique identifiers for each of the participants and a
sample size of over 20 participants. Initially, the UIWADS dataset was collected
for the performance of walking-based authentication with a two-staged approach,
in which the dataset authors used the data to infer the posture, such as standing or
walking, in the first stage, and then, a one-class classification to ensure that the posture
readings are matched to the authorized user was performed in the second stage.
Android smartphones located in the shirt pockets of the 22 participants provided the
data collected for the dataset. The triaxial accelerometer was captured with frequency
of 27 Hz. Time-step values range from 0 s to about 650 s depending on partici-
pant walking. This walking dataset has only four features of real type, which are
represented by: time-step, x-acceleration, y-acceleration, and z-acceleration. A short
sample from data is visualized in Fig. 2.
Fig. 2 A short sample from time-series data from the triaxial accelerometer sensor
3.2 Time-Series Data Augmentation
Synthetically generated time-series data that provides new data samples that can
be obtained using data augmentation techniques in cases where the classification
performances would be improved is explained in this section. The class labels of
the samples used for the augmentation are assigned to these samples of synthetic
time-series data. Overfitting can be prevented by the augmented data, which can also
provide improvement of the generalization ability of deep learning [8]. Mirroring,
scaling, cropping, rotating, etc., are permissible augmentation techniques for image
recognition applications because minor modifications resulting from these techniques
do not change the image’s label as could occur in real-world observation. However, for
the time-series sensory data, these transformations that preserve the labels are not able
to be intuitively recognized. Variability can be introduced without modification of the
labels of the time-series data with a number of factors that include sensor placements,
random noise, and the temporal properties of the activities. Four techniques for data
augmentation were selected for the time-series data in this study as follows:
• Jittering is applied as a method that simulates additive sensor noise. The mechan-
ical noise of each of the sensors is different. The training data’s robustness against
different sensor types and their additive and multiplicative noises is increased by
simulating random sensor noise. In this research, white Gaussian noise was used
to add jittering to the raw training data.
• Scaling is another method that is employed in data augmentation, which alters
the raw data’s magnitude while preserving the labels. This variation is adopted in
cases when there is a change in the dimensions of the equipment that the sensor is
attached to, for example, modifications in the excavator boom’s length, etc. The
implementation of scaling involves the multiplication of the raw training data
using a random scalar.
• Rotation is implemented for the introduction of the label-invariant variability of
the IMU sensor data in cases where the sensors are installed in the equipment
with various orientations.
• Time warping is a method that generates synthetic training data for a range of
temporal properties of the tools used for a specific task, such as a loading activity
that can be conducted using an excavator with a range of operating speeds. In
order to compensate for this variability, each activity was warped (i.e., shortened
or stretched) with different warping ratios.
3.3 Training CNN Model
A convolutional neural network (CNN) model works well for identifying simple
patterns within given data which will then be used to form more complex patterns
within higher layers. A one-dimensional convolutional neural network (1D-CNN)
is very effective when we expect to derive interesting features from fixed-length
segments of the overall dataset and where the location of the feature within the
segment is not of high relevance. Typically, the 1D-CNN is often applied to time-
series data, the convolutional output of which is one-dimensional. Structure of 1D-
CNN consists of three main layers that are an input layer, hidden layers, and an output
layer as shown in Fig. 3.
Fig. 3 Structure of 1D-convolutional neural network

4.1 Experiment Setup
In this study, Python’s Scikit-learn and Keras were used to conduct the experimental
implementation for each scenario tests consisted of the same 1D-CNN model. All
tests were run on Google Colab platform. We set up different five experiments as
following scenarios:
1. Scenario I—raw training data without any augmentation,
2. Scenario II—raw training data with augmented data using the jittering technique,
3. Scenario III—raw training data with augmented data using the scaling technique,
4. Scenario IV—raw training data with augmented data using the rotation technique,
5. Scenario V—raw training data with augmented data using the time-warping
technique.
4.2 Experimental Results
To evaluate performance of biometric user identification, we performed five exper-

iments on different scenarios as describe in Sect. 4.1 with the convolution neural
network model. Concerning FAR and FRR values, the scenario II had FAR metric
with lowest value than other scenarios, while the scenario V had FRR metric with
lowest value than others. Other evaluated metrics of the user identification are present
in Table 1.
Table 2 shows the experimental results for the different user identification
scenarios considering the UIWADS dataset with 22 subjects. Subjects in scenario II,
which uses raw training combined with augmented data by the jittering technique,
had all performance metrics with highest values than any other scenario.
Table 1 Values of mean FAR and FRR including standard deviation in each data augmentation
scenario
Performance metrics
Scenario FAR Standard deviation FRR Standard deviation
Scenario 1 0.003958 0.003571 0.107760 0.114872
Scenario 2 0.003030 0.003211 0.093484 0.099674
Scenario 3 0.003696 0.003970 0.119180 0.125588
Scenario 4 0.003379 0.003970 0.110670 0.115304
Scenario 5 0.003221 0.003100 0.088058 0.082485
Table 2 Values of mean accuracy, precision, recall, and F1-score in each data augmentation scenario
Performance metrics
Scenario Accuracy (%) Precision (%) Recal (%)l F1-score (%)
Scenario 1 91.90 92.00 91.90 91.80
Scenario 2 93.90 93.90 93.90 93.80
Scenario 3 92.50 92.60 92.50 92.40
Scenario 4 92.40 92.60 92.40 92.30
Scenario 5 93.40 93.50 93.40 93.40
5 Conclusion and Future Works
This work presented a data augmentation framework to improve the baseline of

performance metrics, which are accuracy, precision, recall, F1-score, FAR, and
FRR, of biometric user identification. We used walking activity data of 22 subjects
recorded by smartphone sensors from the UIWADS dataset. The user identifica-
tion task is successfully tackled using a four-layer one-dimensional convolutional
neural network. Without any data augmentation, the 4-layer 1D-CNN provided a high
accuracy value of 91.90% and a low FAR value of 0.003958 as baseline performance
metrics. After applying the proposed data augmentation framework, the combination
of raw training data and augmented data using the jittering technique improves the
baseline performance of accuracy metric from 91.90% to 93.90%. When concerning
the FAR value, the same scenario improves the baseline performance of FAR from
0.003958 down to 0.003030. For future implementation, we are planning to imple-
ment the proposed data augmentation framework with another deep learning models
such as long short-term memory recurrent neural network.
References
1. Gafurov, D., Snekkenes, E., Bours, P.: Gait authentication and identification using wear-
able accelerometer sensor. In: 2007 IEEE Workshop on Automatic Identification Advanced
Technologies, pp. 220-225 (2007)
2. Hnoohom, N., Mekruksavanich, S., Jitpattanakul, A.: Human activity recognition using
triaxial acceleration data from smartphone and ensemble learning. In: 2017 13th International
Conference on Signal-Image Technology Internet-Based Systems (SITIS), pp. 408-412 (2017)
3. Laghari, A., Waheed-ur-Rehman, Memon, Z.: Biometric authentication technique using smart-
phone sensor. In: 2016 13th International Bhurban Conference on Applied Sciences and
Technology (IBCAST), pp. 381-384 (2016)
4. Mekruksavanich, S., Hnoohom, N., Jitpattanakul, A.: Smartwatch-based sitting detection with
human activity recognition for office workers syndrome. In: 2018 International ECTI Northern
Section Conference on Electrical, Electronics, Computer and Telecommunications Engineering
(ECTI-NCON), pp. 160-164 (2018)
5. Mekruksavanich, S., Jitpattanakul, A.: Classification of gait pattern with wearable sensing data.
In: 2019 Joint International Conference on Digital Arts, Media and Technology with ECTI
Northern Section Conference on Electrical, Electronics, Computer and Telecommunications

Engineering (ECTI DAMT-NCON), pp. 137-141 (2019)
6. Mekruksavanich, S., Jitpattanakul, A.: Exercise activity recognition with surface electromyo-
graphy sensor using machine learning approach. In: 2020 Joint International Conference on
Digital Arts, Media and Technology with ECTI Northern Section Conference on Electrical,
Electronics, Computer and Telecommunications Engineering (ECTI DAMT-NCON), pp. 75-78
(2020)
7. G. Pan, Y. Zhang, Z. Wu, Accelerometer-based gait recognition via voting by signature points.
Electron. Lett. 45(22), 1116–1118 (2009)
8. K.M. Rashid, J. Louis, Times-series data augmentation and deep learning for construction
equipment activity recognition. Adv. Eng. Inform. 42, 100944 (2019)
9. Rong, L., Jianzhong, Z., Ming, L., Xiangfeng, H.: A wearable acceleration sensor system for
gait recognition. In: 2007 2nd IEEE Conference on Industrial Electronics and Applications,
pp. 2654-2659 (2007)
10. Um, T.T., Pfister, F.M.J., Pichler, D., Endo, S., Lang, M., Hirche, S., Fietzek, U., Kulić, D.: Data
augmentation of wearable sensor data for parkinson’s disease monitoring using convolutional
neural networks. In: Proceedings of the 19th ACM International Conference on Multimodal
Interaction, ICMI ‘17, pp. 216-220. Association for Computing Machinery, New York, NY,
USA (2017)
11. G.M. Weiss, K. Yoneda, T. Hayajneh, Smartphone and smartwatch-based biometrics using
activities of daily living. IEEE Access 7, 133190–133202 (2019)
Cybersecurity Attacks: Analysis
of “WannaCry” Attack and Proposing
Methods for Reducing or Preventing
Such Attacks in Future
Sumaiah Algarni
Abstract The technological revolution in the last few years has erased the commu-
nicational barriers and created a global network of users and data. At the same time,
they have also made it easy for malwares and other cybersecurity threats to spread
easily among users through networks. The threat of ransomwares in particularly has
been risen in the last few years, with 2017 witnessing one of the most destructive
ransomware attacks in the history of Internet. This attack known as the “WannaCry”
has infected computers across the globe and crippled both individuals and organiza-
tions. The essay provides an analysis of this attack including its definitions, methods
of infections, its impact on the society as well as discussion some important detection
and prevention techniques to avoid similar attacks in future.
Keywords Ransomware · WannaCry · Encrypt · Decrypt · Malware
1 Introduction
With globalization, technological advances, the advent of the digital era, and commu-
nication barriers between individuals and organizations have largely been erased.
Information ranging from an individual’s details to an organization’s data can be
found on the virtual network of interconnected computers called as “Internet” (Craig
2011). However, this has made communication and information access easier as
well as has made it easier for malicious software to spread rapidly from one place
to another through interconnected networks [1]. Indeed, one of the most impor-
tant security concerns in recent years has been the growing threat of cyberattacks,
specific from ransomwares [2–4]. According to Ledin [5], the perpetrators of such
cybersecurity threats are becoming increasing sophisticated with using methods like
mobile-based malware propagation, smart grid hackings, cloud-based attacks, and
GPS signal interferences. The “WannaCry” ransomware that combine a traditional
S. Algarni (B)
M.Sc. in Information Systems, School of Computing Sciences, University of East Anglia,
Norwich, UK
e-mail: sumaiah.garni91@gmail.com
764 S. Algarni
ransomware with worm-like self-propagation capabilities and wreak havoc world-

wide is an example of such attack that happened in 2017. This paper will provide
a detailed analysis of the “WannaCry” ransomware attack including its origin and
definition, modes of infection, its impact on the society, the motivation of perpetra-
tors as well as give some important techniques to detect and prevent similar attacks
in future.
2 Analysis of the “WannaCry” Ransomware Attack
Before analyzing “WannaCry” ransomware attacks, it is important to understand

what ransomware is and how it works. Ransomware is any piece of malicious soft-
ware or “malware” that blocks the victim’s access to his or her files by encrypting it
and the only way to regain access can be through paying a “ransom” [6, 7]. The first
ransomware attack was in 1989, known as the “AIDS Trojan” attack or “PC Cyborg,”
and was spread through floppy disk [8, 9]. Since then, hackers have developed many
more sophisticated ransomware programs that cripple user’s computers and force
them to pay money to regain control over their data. These ransomware programs
have two categories [10]. The first one is the “encrypting” types or the “cryptoran-
somwares” that include advanced encryption algorithms designed to block users’
files through encryption and the only way to unblock it through a decryption key
provide after paying a ransom [11]. The second one is the “locker” types that lock
the users out of their operating systems until payment is made [12]. CryptoLocker,
Locky, and CrytpoWall are examples of the former while Winlocker is an example of
the latter [1]. The “encrypting” types of ransomwares are more destructive and diffi-
cult to break because files cannot be decrypted without the key and it can change the
extension of various files to create confusion [1, 12]. In addition, they spread through
easily accessible ways such as phishing emails, suspicious software programs, unver-
ified downloads, malicious Web sites, or infected advertisements, and this makes it
extremely dangerous [9]. The losses caused by ransomwares attacks reach to several
billions of dollars every year and attackers becoming more and more global and
ambitious in their approach [3, 13].
2.1 Origins of “WannaCry”
The “WannaCry” ransomware attack happened on 12th May 2017 with an estimated
200 k + affected endpoints in organizations around the world (Ehrenfeld [14]. The
perpetrators requested from victims to pay $300 in bitcoins (Fox-Brewster 2017).
It reported that the amounts of money paid during the period reach to $50,000 and
can reach to 4 billion USD if all users paid ransom [15]. One of the most impor-
tant reasons for the rapidly spread of this ransomware was that unlike traditional
Cybersecurity Attacks: Analysis of “WannaCry” Attack … 765
malwares that spread usually through emails, the “WannaCry” exploited the vulner-
ability in Microsoft’s implementation of Server Message Block (SMB) protocol
called “EternalBlue” [16]. The first attacks were reported in the UK. However, it
then spread rapidly to other countries in Europe as well as USA, India, China,
Taiwan, Ukraine, and Russia [15]. The attacks not only affected individual users
but also businesses from health, travel, telecom, security services, manufacturing,
and several other sectors [17].
2.2 Anatomy of “WannaCry”
In order to understand the vulnerabilities that lead to this rapid propagation of the
“WannaCry” ransomware globally, it is important to understand the anatomy of the
attack. Actually, the ransomware searches and encrypts more than 200,000 computers
in 150 countries [16, 18]. It can infect computer in two ways. The first way is when
a hacker perform an Internet search for IP addresses that contain the “EternalBlue”
vulnerability, preys upon unpatched computers and attack them remotely through the
execution of arbitrary code [10, 17, 19]. The second one is the “worm method” in
which the ransomware spreads to other machines in the infected network through self-
propagation (ibid). Both methods have the same consequence of encrypting the files
of the victim’s computer and locking them out of it. The reason “WannaCry” spread
so rapidly was that it exploited the vulnerability to combine a traditional ransomware
with worm-like propagation capabilities [1, 20]. This is accomplished by following
a multi-step propagation pattern. In the first step, vulnerable systems with unpatched
port 445 endpoints are identified. Then, a backdoor payload named “DoublePulsar” is
uploaded that allow the “EternalBlue” exploit to control and inject code remotely into
the system. An additional payload is also injected into the system through the back-
door, which in case of “WannaCry” is a ransomware. A communication-anonymizing
software called “Tor-client” is installed on the affected system in order to prevent the
communication from being tracked. Finally, the system files are encrypted using an
AES algorithm that can only be decrypted using an RSA 2048 key, shadow copies are
deleted and the ransom message is display. The payload also include code to search
for additional systems within the network with open 445 ports. These systems will be
exploited with “EternalBlue” and the process repeats, thus creating an attack chain
(ibid). The overall process can be visualized as shown in Fig. 1 [1, p. 1939].
2.3 The Effects of Cyberattacks on Society
Apart from affecting individuals and organizations, cyberattacks can have a large
impact on the society. Factors such as identity theft and data loss are becoming very
common due to technological increase [21]. It is predicated that about 7.1 million
identity thefts occurred in the last eight years [3]. This issue is not only affecting
766 S. Algarni
Fig. 1 Propagation of “WannaCry” [1, p. 1939]
individuals, but can also harm the safety and security of the whole society because
cybercriminals can use data and identities to commit crimes [20]. Some attacks
can also impact the political climate and cause global unrest. As a result of these
attacks, the security costs and financial losses are increasing across the world, and
that affects the economy of the countries [22]. In addition to that, the human society
as a whole is becoming extremely dependent on digital technologies and as a result
“ It is becoming harder to identify the nodes and connection points whose protection
must be prioritized” [23, p. 6]. Indeed, the Internet and the digital technologies could
be great tools for hackers and cybercriminals because of digital anonymity and that
make it easy for committing crimes [21]. This issue is compounded by the widespread
use of new and sophisticated mobile, cloud and social media technologies. Cliff
[24] points out that while these technologies can make life easier, any failure can
cause “domino-effect chain reactions” leading to collapses. In fact, the “WannaCry”
ransomware attack is a case in point that demonstrates the dangers of increasing
dependence on technologies as whole.
2.4 The Motivation of Hackers
Given social, economical, and political risks related to cyberattacks, the need to
understand the motivation of hackers is important. While the financial motivation
behind attacks such as the WannaCry ransomware seem obvious, it is argued that
the motivation is far from straightforward as in many cases and the real motiva-
tion is hidden [25]. However, in order to understand the motivation of the hackers,
researchers have divided them into two categories “insiders” and “outsiders” [26,
27]. On the one hand, the “insiders” include aggrieved employees or family members
who motivated by revenge or financial gains [28]. It also includes those who facil-
itate cybercrimes unintentionally. The “outsiders” on the other hand are classified
into three subcategories: (a) organized attackers, (b) hackers, and (c) amateurs (ibid).
Organized attackers may be motivated by political, religious, or ideological agendas
who may want to inflict damages on their targets. Organized attackers may also
be organized criminals who motivated by power and financial gains. Hackers may
be motivated by the financial or antisocial reasons, or simply by the challenge of
breaking into systems. Amateurs’ motivations are generally benign like improve
their skills or curiosity (ibid).
Regardless, their actions often lead to damage. Dealing with these various groups
and understand their motivations is an extremely complex issue and should be given
high priorities by countries.
2.5 Prevention of “WannaCry” and Other Ransomware

Attacks in Future
The once the computer is infected with the “WannaCry” attacks, it is possible to
spread to other computers in the same network using “wrom” method. Dealing with
such attacks could be extremely complex with two basic possibilities [29]. The first
one is to take responsive procedures after the computer is infected, and the second
one is to take preventive procedures before infection (ibid). In general, the post-
attack procedures are very limited especially if the ransomware is unknown and the
data cannot be decrypted. Indeed, this was the case that happened with the “Wan-
naCry” recently (Symantec Security Response 2017). While security experts have
released many tools like “WannaKey” and “WannaKiwi” to help users fight against
it, and these tools have very limited abilities and may not be able to recover files in
most scenarios (Roettgers 2017). Indeed, either paying the ransom or relying on the
backups are the only choices users may have (ibid). According to Levy [9], even if
the users pay the ransom, there is no guarantee that the files will be recovered. Given
this scenario, the best method to deal with such attacks would be to take preattack
preventative security procedures that can prevent such attacks. These procedures
could be taken at two levels, organizational and individual. At an organizational
level, the steps that can be taken include the following:
768 S. Algarni
1. Backup important data frequently and effectively in order to decrease the negative
impact if an attack occurs [9].
2. Ensure that all Windows systems are patched with the latest security and operating
system updates, especially for SMB or the Server Message Block [1].
3. Block all vulnerable SMB ports from access by external hosts, especially port
139 and port 445, because these are the main ports that are open to exploits [30].
4. Update the antivirus and antimalware signatures regularly on systems and
network endpoints [9].
5. Use the sandboxes and virtual machines as these make it easier to contain
malwares from affecting the operating system [31, 32].
6. At an individual level, attacks from “WannaCry” can be prevented by exer-
cising caution when opening emails from unknown senders especially if they
contain attachments, because these are the most common methods of infection
[9]. Similar to organizational systems, personal computers can also be protected
from attacks by ensuring that security and operating system patches, antivirus
and antimalware signatures are up to date [9, 1]. Finally, users can back up their
files in order to protect their data (ibid).
The above-mentioned procedures could help to protect individuals and organiza-
tions from future attacks.
3 Conclusion
Based on the “WannaCry” ransomware attack and other similar attacks in the recent
years, it can be deduced that perpetuators of ransomwares are becoming increasingly
sophisticated and ambitious in their approach. The emergence of the anonymous
cryptocurrencies like bitcoins makes it very difficult for experts and governments to
bring the perpetuators to justice. The “WannaCry” attack was especially vicious due
to the combination of a ransomware and worm-like self-propagation capabilities. It
is evident that the success of “WannaCry” was largely due to the lack of preven-
tative measures by both users and organizations. Indeed, since there are currently
no mechanisms to decrypt the files encrypted by this ransomware, the only way to
minimize the damage is to take precautionary measures by both users and orga-
nizations to protect their computers and data. Three years on, ransomware attacks
are still a concern that threaten millions of people worldwide and further study and
investigation are recommended in this regard.
References
1. S. Mohurle, M. Patil, A brief study of wannacry threat: ransomware attack 2017. Int. J. Adv.
Res. Comput. Sci. 8(5), 1938–1940 (2017)
2. L.Y. Connolly, D.S. Wall, The rise of crypto-ransomware in a changing cybercrime landscape:
taxonomising countermeasures. Comput. Secur. 87, 101568 (2019)
3. Internet Security Threat Report: 2017 https://digitalhubshare.symantec.com/content/dam/
Atlantis/campaigns-and-launches/FY17/Threat%20Protection/ISTR22_Main-FINAL-JUN8.
pdf?aid=elq_
4. S. Mansfield-Devine, Ransomware: taking businesses hostage. Netw. Secur. 2016(10), 8–17
(2016)
5. G. Ledin, The Growing Harm of Not Teaching Malware. Commun. ACM 54(2), 32–34 (2011)
6. R. Brewer, Ransomware attacks: detection, prevention and cure. Netw. Secur. 2016(9), 5–9
(2016)
7. A. Liska, T. Gallo, Ransomware: Defending Against Digital Extortion (O’Reilly Media Inc.,
New York, 2016)
8. J.Greene, Ransomware (Salem Press Encyclopedia of Science, New Jersey, 2017)
9. A. Levy, Avoiding the Ransom: Cybersecurity for Business Owners and Managers (Magnet
Solutions Group Press, Austin, 2016)
10. J. Allen, Surviving ransomware. Am. J. Fam. Law 31(2), 65–68 (2017)
11. M. Wecksten, J. Frick, A. Sjostrom, E. Jarpe, A novel method for recovery from Crypto
Ransomware infections. In 2016 the 2nd IEEE International Conference On Computer And
Communications (ICCC), (2016 October)
12. I. Bica, R. Reyhanitabar, Innovative security solutions for information technology and
communications. In: 2016 the 9th International Conference (SECITC) (2016, June)
13. C. Rozas, H. Khosravi, D.K. Sundar, Y. Bulygin, Enhanced Detection of Malware. Intel
Technology Journal 13(2), 6–15 (2009)
14. J.M. Ehrenfeld, WannaCry, cybersecurity and health information technology: a time to act. J.
Med. Syst. 41(7), 104 (2017)
15. J. Stevenson, The WannaCry ransomware attack (2017). Strategic Comments 23(4), vii–ix
(2017)
16. W. Ashford, Ransomware attack could be wake-up call: a failure by many organisations to take
cyber security seriously has long been blamed on the lack of a single significant event to shake
things up. Does WannaCry fit the bill? Comput. Wkly. 4–5 (2017)
17. R. Hackett, Why a global cyber crisis stalled-this time. Time 189(20), 7–9 (2017)
18. A.L. Young, M. Yung, Privacy and security cryptovirology: the birth, neglect, and explosion
of ransomware. Commun. ACM 60(7), 24–26 (2017)
19. B. Brian, WannaCry Exposes Defense Supply Chain Vulnerabilities. National Defence 20–21
(2017)
20. A. Azad, Ransomware: a research and a personal case study of dealing with this Nasty Malware.
Informing Sci. Inf. Technol. 14, 087–099 (2017)
21. R. Upadhyaya, A. Jain, Cyber ethics and cyber crime: a deep dwelved study into legality,
ransomware, underground web and bitcoin wallet. In: 2016 International Conference on
Computing, Communication and Automation (ICCCA) (2016, April)
22. J. Thornton, Don’t click bait from cyber criminals. Public Finance. 11, 43 (2015)
23. D. Clemente, Cyber Security and Global Interdependence: What Is Critical? (2013) https://
www.chathamhouse.org/sites/files/chathamhouse/public/Research/International%20Security/
0213pr_cyber.pdf
24. D. Cliff, Failures in Ultra Large Scale Complex Systems (The Chartered Insurance Institute,
London, 2012)
25. P. Shakarian, J. Shakarian, A. Ruef, Introduction to Cyber-Warfare: A Multidisciplinary
Approach (Elsevier, Waltham, 2013)
26. C. Han, R. Dongre, What motivates cyber-attackers? Technol. Innov. Manag. Rev. 4(10), 40–42
(2014)
27. D. Russell, G.T. Gangemi, Computer Security Basics (O’Reilly & Associates, California, 1993)
28. S.W. Brenner, Cybercrime: Criminal Threats from Cyberspace (Praeger, Santa Barbara, 2010)
29. R. Richardson, North, M: Ransomware: Evolution, Mitigation and Prevention. International.
Manag. Rev. 13(1), 10–21 (2017)
770 S. Algarni
30. Microsoft Website: How to enable and disable SMBv1, SMBv2, and SMBv3 in Windows
and Windows Server. https://support.microsoft.com/en-us/help/2696547/how-to-enable-and-
disable-smbv1-smbv2-and-smbv3-in-windows-and-windows (2017)
31. K. Kendall, Practical Malware Analysis. In: Black Hat Conference, USA (2007)
32. G. Radhamani, G.S. Rao, Web Services Security and e-business (Idea Group Inc., PA, 2007)
33. T. Craig, Internet: Technology, People, Process (Hodder Wayland, London, 2002)
34. T. Brewster, WannaCry Hackers Are Using This Swiss Company To Launder $142,000 Bitcoin
Ransoms. (2017). https://www.forbes.com/sites/thomasbrewster/2017/08/03/wannacry-hac
kers-use-shapeshift-to-launder-bitcoin/#4638139f3d0d
35. G. Slabodkin, WannaCry had wide impact: recent ransomware attack hit much harder than
initial estimates; NHS facilities, two U.S. healthcare systems are among its victims. Health
Data Manag. 25(3), 6 (2017)
Enhancing Trust and Immutability
in Cloud Forensics
Pranay Chauhan and Pratosh Bansal
Abstract We are living in a digital era where most of us have an online presence with
all their data on cloud. “Easy access of data from anywhere” is the biggest advantage
of cloud but this same point becomes the cause of cloud-based crimes. There has
been a drastic increase in cloud-based crime in the recent past. The immoral from any
geolocation can target the victim of any geolocation, even cross-borders. Moreover,
hunting down the immorals/suspects in these cross-border attacks is more difficult
as suspects are easily able to coat themselves. In the cloud, data is stored at multiple
geolocations, and this cross-border cloud data storage is the biggest challenge for law
enforcement agencies. When the crime is committed from a different country this
geographically sparse data becomes a major hindrance for the forensics investigation.
LEA had to request the cloud service providers of all the involved locations (including
the ones which are outside the country) to provide data for fetching out the right pieces
of evidence. In the current state, LEA performs these forensic investigation requests
through web portals or through email conversations. These requests take days or
weeks to process which in turn delays the investigation process, and sometimes
victims are left out without justice. Also, these web portals/applications through
which the LEA demands the pieces of evidence are also based on cloud and thus are
highly insecure. These not so secure portals raise a big question on the authenticity of
the data and evidence received on them. Evidence could be easily tailored or removed
(partially or fully) from the cloud storage.
Keywords Cloud security · Digital evidence · Cross-border · Data localization ·

Smart contract
P. Chauhan (B) · P. Bansal

Department of Information Technology, Institute of Engineering and Technology, Devi Ahilya
Vishwavidyalaya, Indore, MP, India
e-mail: pranaychauhan1985@gmail.com
772 P. Chauhan and P. Bansal
1 Introduction
With the advancement of technology, computing services have also evolved drasti-
cally. Thanks to the “Cloud Computing” technology now even the robust software
could be executed on low-performance devices. Cloud in simpler words refers to “the
computers and storage devices that are kept over a remote location and are allowed to
be accessed by people via the Internet are termed as cloud.” Since these are located
and could be accessed from distant locations, they allow users to run applications
without installing onto their devices [1]. These computing machines kept over a
remote location not only helps people to run heavy softwares on low-end devices but
also allows them to store, access, and process unlimited data. Cloud computing has
achieved a compound annual growth rate of 19% from 2015 to 2020. Grown from
$67B to 162B. Another interesting fact to add here, Vormetric stated that 85% of
the total enterprise organizations keep their sensitive information on the cloud [2].
Each coin has two faces; sadly cloud computing crimes are also increasing at the
alarming rate. RedLock stated in 2018 that 49% of cloud databases are not encrypted
and they have predicted that by 2022, 95% of the cloud security issues would be
caused because of consumers’ fault [2, 3]. Though, there are numerous reasons to
believe cloud computing, but all these security issues become its bottleneck. These
barriers not only question its performance but may compromise the trust of cloud
users on all service providers. Security is the major worry of cloud consumers and
cloud service providers, since risk and trust cannot be considered as weak activity.
Table 1 above depicts various types of attacks and threats in cloud environments.
Apart from these listed conventional attacks, there are other security concerns such as
compliance, availability, data localisation [4]. While massive data storage at multiple
geolocations/distributed across borders, dependencies on other networks, and data
centers are major privacy concerns. Conservative security techniques for integrity,
confidentiality, and authentication could not be effectively integrated on various
levels of a cloud environment. Subsequently, cross-border data storage and public
networks require additional concern. When the cloud infrastructure is distributed
across different nations and not limited to the same nation (within the border) then
the local law enforcement agencies (LEA) faces extra challenges while dealing with
the cross-border storage and security laws. Travel of data across national borders is
another major concern for cross-border jurisdictional [3]. LEA often requires access
to the infrastructure resources, forensic evidence, data, system test details, logs,
contact information, etc. Cloud service providers usually take time for sharing this
data and information, even if they agree to share the information, the process takes
too long, and the CSP’s keep delaying the requests [5].
Enhancing Trust And Immutability In Cloud Forensics 773
Table 1 Shows various cloud threats and attacks

Cloud threats Attacks on cloud
Insider threats Malware can be used
• XSS attacks
• SQL i attacks
Data leakage Multiple side channels methods can be used for data
leakage
Example: Have i Been pwned, firefox monitor, breach
alarm and other tools can be used for breach data
identification.
User access and physical disruption Man in the cloud attacks
External attackers and threats DDOS attacks
Data segregation Various authentication attacks such as
• Phishing attacks
• Replay attacks
• Key loggers
• Brute force attacks
Mis configured server Improper server configuration can lead to big attacks
Crypto jacking Malicious scripts are being used for mining of crypto
currencies
Example: Process explorer can be used to identify the
malicious process running in network and cloud
2 Existing System for Investigation
In the current state, LEA performs these forensic investigation requests through CSP’s
web portals or through email conversations and through mutual legal assistance treaty
(MLAT). These requirements take extensive time to process. This delays the forensics
investigation process, and at times victims go away without justice served. These web
portals and other web applications through which the LEAs demand the pieces of
evidence are also based on the cloud and thus are highly insecure. Due to the lack of
security, the evidence can also be tailored or removed from the cloud storage. These
not so secure portals raise a big question on the authenticity of the data and evidence
received on them [6]. Evidence could be easily tailored or removed (partially or fully)
from the cloud storage.
3 Evidence Collections Methods
Email communications is one of the most common ways for exchanging critical infor-
mation about the evidence [7]. The investigation agencies directly demand informa-
tion from the cloud service providers (CSP’s) or concerned authority. These author-
ities might take weeks to respond to these emails and sometimes might even ignore
them. Most of the email services are insecure and are easily hackable [4]. Even if
the CSP responds with the correct information, the evidence could be malformed by
the suspect without knowledge of the victim or the investigation agencies. Phishing,
spear phishing, spoofing, and zero days attack are few common and easy ways to
hijack email accounts. According to Hostingbrunal, 1.5 million phishing web pages
are formed every month, and 92% of the malwares are delivered through email only
[8].
4 Cloud Service Providers’ Web Portals
To deal with the messiness of the emails, the CSPs or concerned authorities released
their own web portals through which LEAs/investigation agencies can request for
specific evidence. Many CSPs host web portals on their own platform so anyone
could easily authenticate themselves and could request for the critical information
of the suspect. These portals usually ask the user (LEA/Investigator) to fill and
submit the form along with the supporting documents such as court orders, search
warrants, and other necessary documents too. [9]. The CSP or concerned authorities
then analyses the received form and checks for the authenticity of the submitted
documents, and thereafter, they provide the requested information [5]. The CSP’s
or concerned authorities take an extensive amount of time to respond to such portal
requests since these requests are overhead for them. In some cases, they might not
be actively monitoring these portals, and so requests on these portals might lay
unattended for weeks. Thus, the victim has to wait for a long time and would end up
losing hope for justice. Also, these insecure web portals are hackable too. This again
questions the authenticity of the evidence received [10].
5 Mutual Legal Assistance Treaty (MLAT)
MLAT is an agreement among two or more countries which helps them to exchange
critical information, data, and evidence while dealing with the cross-border cyber-
crimes. This treaty establishes a single point of contact between different countries
through which these nations can exchange the pieces of evidence. Each involved
country has to follow a three-level internal communication process before sending
the request draft across the border. This process ensures that all the protocols are
being followed while requesting the information [11, 12]. Below are the steps which
are followed:
• The local city authorities first send the request draft to the state-level authorities.
• State-level authorities forward these request drafts to the central authority of the
country.
• The central authority checks the request draft for all the protocols and some
additional grounds.
• Post request draft examination central authority then forward this draft to law
enforcement agencies.
• LEAs or the investigators then collect all the evidence and revert back to the
central authorities in admissible form.
• The central authority then reverts these pieces of evidence to the state authority
in admissible form.
• The state authority reverts it back to the local city level authorities in admissible
form.
One can easily see how tedious and lengthy the aforementioned process is.
Evidence can be easily tempered before reaching the right hands, and this hierar-
chical flow increases the processing time as well. “Time” is an essence of cross-border
crimes; delay in request processing gives unfair advantage to the suspect [12].
6 Data Localization and Its Advantages
Data localization/Data residency is a law that states that the people’s data must
be stored and processed within the same country. Data must be sent out to other
countries only after it meets all the data protection laws and local data privacy laws.
According to this, the CSPs will have to establish their storage setups within the
country and collect, store, and process within the same country. Data localization
also helps in maintaining the privacy law for the citizens. This includes the protection
of critical information including political judgments and other personal interests of
an individual. With data localization, LEAs and/or investigators could easily get the
case information and could save themselves from a lot of obstacles while dealing with
the cross-border crimes. Data localization provides security for the citizens’ data and
helps in protecting individual’s data from foreign surveillance. Also, it equips LEAs
to have better monitoring over the data and its usage. The LEA and investigation
agencies can get the evidence almost immediately without extended formalities and
without relying on MLAT.
a. Issues with Data Localization
The most prominent issue with data localization is the CSPs’ denial to its adaptation.
Many of the CSPs are opposing this privacy draft because this would require huge
efforts from their side, and there are no tangible benefits for them. Storing data inside
the country requires a lot of investment in terms of time, money, and resources. On
top of that, some countries do not have well-developed facilities and infrastructure
to maintain the security of the data. Without strict monitoring there is no guarantee
that the CSPs will not store data outside the country [2, 13, 14]. They may continue
to carry out such practices for their benefits.
7 Proposed Solution
Blockchain is the modern technology that could help in dealing with the cross-border
crime issues by aiding in exchange of cross-border evidence. It is a data structure that
allows the creation of a distributed digital ledger for storing and recording transac-
tions (events/records) shared by all participating parties over a distributed network of
computers [15]. Blockchain technology ensures data integrity and security and could
overcome all the problems associated with time taking, not so secure conventional
data sharing techniques such as emails and web portals. Blockchain maintains the
hashes of the electronic data which are derived from the hashes of the previously
inserted blocks. Using its consensus mechanism, it allows the insertions of new blocks
only after the approval of all the present nodes in the network. Since blockchain is a
distributed ledger, no single authority has the complete right or control over the data
in the network [16, 17].
Law enforcement can employ blockchain-based solutions through which all the
existing issues of evidence exchanging, evidence tampering, generating trust, main-
taining logs, and metadata could be easily solved. Chain of all the evidence along with
their metadata could be tracked down with this technology. Government agencies,
LEA, and CSPs can use smart contracts, i.e., Blockchain applications for exchanging
evidence instead of conventional procedures. Smart contracts could ensure trust,
prevent malicious activities, and provide easy ways for data sharing. These are
cryptographically sealed and are updated near real time.
Digital evidence could be converted into lines of codes using these smart contracts.
This code could be deployed on a decentralized and distributed blockchain network.
Code would control the execution of evidence and control read/write transactions
in a traceable, immutable, and irreversible manner by managing the permissions.
Code is law, and it could be deployed either on permissionless or permissioned
blockchain platform. Permissioned blockchain platforms have more advantages over
permissionless blockchain platforms [18]. Once government bodies, authorities, or
concerned agencies deploy code/evidence on a permissioned blockchain platform
and they could grant permissions to concerned parties for examination of evidence.
Concerned parties here could be law enforcement agencies, prosecution, defense,
forensic expert, investigation agencies, court of law, central authority, and any other
involved agencies. All these concerned agencies could access the information using
cryptography concept of public and private security keys. Since evidence is stored
on the blockchain, concerned authorities could access these evidence globally too.
Blockchain stores metadata of the transactions along with the timestamps, this could
help in tracking the overall case discussions with the involved individuals and groups.
Blockchain could solve all the issues of conventional data sharing systems/solutions
and provide below list of benefits:
a. Data Integrity—Each block of the blockchain is immutable ensuring the
integrity of digital evidence as data could not be tampered in between by anyone.
b. Data Transparency—Each node of blockchain tracks the timestamp and meta-
data of every single change or transaction done. This could help in maintaining
the transparency among the LEA and concerned authorities as all involved parties
could see changes done by other involved parties.
c. Decentralization—Blockchain is a distributed ledger, and the data is stored in
the form of hashes, and so data is distributed across the multiple nodes present
on the network. In this way, no individual has the complete authority over any
block or data present in the network. This will eventually enhance the trust in the
legal system.
d. Consensus mechanism—Blockchain uses consensus mechanism which states
that every node has an equal right over the digital evidence/data in the network.
Hence, there would be no monopoly among the users of the system, i.e., among
concerned authorities.
e. Data authorization and authenticity—Permissioned blockchain allows you to
control the data authorization of involved parties by setting appropriate read/write
permissions. This will ensure only allowed parties could view the evidence/data.
Blockchain uses public and private keys for allowing users/agencies access to the
nodes. This secure key mechanism ensures authenticity and keeps the information
secure from the outer world by preventing unauthorized data access.
f. Time efficiency—This smart contract way is time efficient and faster compared
to all conventional data sharing systems —Emails and Web Portals
g. No third party dependency—Blockchain technology provides direct access of
the required information to all the authorized personnel and thus eliminates the
dependency on third parties for evidence requests and approvals.
h. Trust in the legal system—Blockchain would provide correct untampered infor-
mation/evidence to the right hands without unnecessary processes and delay
which will speed up the entire investigation and would eventually enhance the
trust in the legal system [19].
8 Conclusion
Conventional data/evidence sharing techniques, such as emails, web portals, etc.,

have their own flaws. Data integrity and security are few of the major concerns with
these traditional ways of evidence collection. Security concerns with these methods
raises doubts over the integrity of evidence and also weakens trust in the cloud
forensics. Using blockchain and converting all digital evidence to smart contracts
would not only solve all these security problems but it also has its other set of
advantages like decentralization, data transparency, and time efficiency. All these
advantages of blockchain will strengthen and build the trust in the overall cloud
forensics system.
References
1. L. Jianf, et al., An IoT—oriented data storage framework in cloud computing platform. IEEE
Trans. Ind. Inf. 10(2)
2. Cloud computing market and security issues available at https://www.gartner.com/en/inform
ation-technology/glossary/cagr-compound-annual-growth-rate. Accessed on 16 Apr 2020
3. Cloud computing security and privacy issues available at https://www.nist.gov/system/files/doc
uments/itl/cloud/NIST_SP-500-291_Version-2_2013_June18_FINAL.pdf. Accessed on Apr
2020
4. S.O. Kuyoro, F. Ibikunle, O. Awodele, Cloud computing security issues and challenges. Int. J.
Comput. Netw. (IJCN) 3(5), 247–255 (2011)
5. A. Alenerzi, N.H. Zulkipli, H.F. Atlam, R.J. Walters, G.B. Wills, The impact of cloud Forensics
readiness on security, in Proceedings of closer 2017, pp. 511–517
6. C. Modi, D. Patel, B. Borisaniya, A. Patel, M. Rajarajan, A survey on security issues and
solutions at different layers of Cloud computing. J. Supercomput. 63(2), 561–592 (2012)
7. M.E. Alex, R. Kishore, Forensics framework for cloud computing. J. Comput. Electr. Eng. 60,
193–205 (2017)
8. Cloud based email attacks available at https://www.cloudsecuretech.com/types-of-email-att
acks-and-the-damage-they-can-cause/. Accessed on 23 Apr 2020
9. State of cloud adoption and security. (2017). https://www.forbes.com/sites/louiscolumbus/
2017/04/23/2017-state-of-cloud-adoption-mand-security/. Retrieved 25 May 2017
10. N.L.S., Forsenca, Boutaba, Cloud services networking and management in cloud services
(2015), pp. 153–159
11. Mutual assessment legal treaty available at https://mea.gov.in/mlatcriminal.htm. Accessed on
19 Mar 2020
12. Mutual assessment legal treaty format description cross boarder available http://www.cbi.gov.
in/interpol/mlats.php. Accessed on 26 Mar 2020
13. Cross boarder data storage issues available at https://gcn.com/articles/2019/04/09/cloud-act-
international-data.aspx. Accessed on 2 Apr 2020
14. Cross boarder data flows available at https://itif.org/publications/2017/05/01/cross-border-
data-flows-where-are-barriers-and-what-do-they-cost. Accessed on 4 Apr 2020
15. Guardtime cloud assurance with Blockchain (2017). Online available at https://www.guradt
ime.com/solutions/cloud
16. S. Nakamoto, Bitcoin: a peer to peer electronic cash system (2008). Online available at http://
bitcoin.org/bitcin.pdf
17. Distributed ledger technology: beyond Blockchain (2016). Online available at www.gov.uk/
government. Accessded on Mar 2020
18. N. Szabo, The idea of smart contract (1997). Online available at www.fon.hum.uva.nl/pdf.
Accssed on Apr 2020
19. Ethereum platform online available at www.ethereum.org. Accessed on Apr 2020
A Fuzzy Time Series Prediction Model
of the COVID-19 Epidemic
Mohammad Minhazul Alam, S. M. Shahadat Hossain,

Md. Romman Riyadh Shishir, Sadman Hasan, Eumna Huda,
Sabrina Yeasmin, Abdul Motaleb, and Rashedur M. Rahman
Abstract In our study, a number of time series analysis techniques are applied to the
COVID-19 epidemic data. The result of the analysis is used to develop three single
variable fuzzy time series models from the dataset of day to day newly confirmed
cases of the virus in certain countries. Those developed models are later applied
on infected cases of the countries that are not used before in model development.
This helps in appropriate model selection for prediction of cases for unseen coun-
tries. The forecasted results are then compared and analyzed in detail with different
performance metric. The time series analysis process described in this article can
be used to better understand the epidemic trend of a particular country and help the
government plan better intervention policies.
M. M. Alam (B) · S. M. S. Hossain · Md. R. R. Shishir · S. Hasan · E. Huda · S. Yeasmin ·

R. M. Rahman
North South University, 1229 Dhaka, Bangladesh
e-mail: mohammad.alam05@northsouth.edu
S. M. S. Hossain
e-mail: shahadat.hossain10@northsouth.edu
Md. R. R. Shishir
e-mail: romman.shishir@northsouth.edu
S. Hasan
e-mail: sadman.hasan10@northsouth.edu
E. Huda
e-mail: eumna.huda@northsouth.edu
S. Yeasmin
e-mail: sabrina.yeasmin@northsouth.edu
R. M. Rahman
e-mail: rashedur.rahman@northsouth.edu
A. Motaleb
University of Information Technology and Sciences, 1212 Dhaka, Bangladesh
e-mail: abdul.motaleb@uits.edu.bd
780 M. M. Alam et al.
Keywords COVID-19 epidemic · Fuzzy time series analysis · Fuzzy prediction

model · HOFTS · WHOFTS · PWFTS · Intervention planning
1 Introduction
Time series data is a sequence of change of event that spreads over a specific length
of time. In time series analysis, a dataset is analyzed and learned with some math-
ematical or machine learning models to understand the trend and make necessary
prediction. Some of the statistical analysis tools used for the time series analysis
are Simple Moving Average, Exponential Smoothing, Seasonal Autoregressive Inte-
grated Moving Average (SARIMA), etc. Machine learning algorithms like LSTM
which is a variant of recurrent neural network is also used for analysis and prediction
of time series data.
1.1 Time Series Analysis Using Fuzzy Logic
The fuzzy logic approach for time series analysis starts by dividing the universe of
discourse into some discrete fuzzy sets. Each point of the time series is then assigned
to a fuzzy set depending on the membership function value. Afterward, “if-then”
fuzzy inference rules are extracted from the time series pattern and recorded as
transmission from one fuzzy set to another. These rule sets are later used to compute
new values for prediction. The key variables that affect the accuracy of a Fuzzy
Time Series (FTS) model are: number of partitions (fuzzy sets), partition type (grid,
entropy, or cluster), choice of membership function (triangular, trapezoidal, Gaus-
sian, etc.), and order of the function being learned (lag). As the trained model records
the time series as a transition from one set to another, use of too few partition numbers
will not learn the training data well. On the other hand, use of too many partitions
will overfit the training data and will not generalize well. The fuzzy algorithms used
to learn a time series can be weighted (gives different weights to the past values in the
series) or non-weighted. And again, a time series prediction can be made simply from
the past values of the target variable (endogenous variable), or it can be predicted
from a combination of multiple variables (exogenous variables).
1.2 Using FTS Models for the Prediction of COVID-19
The new corona virus, first detected in December 2019 in Wuhan, China is a fast
spreading strain which has almost shut down all the international flights all over the
world. As of writing this article, many countries have implemented strict home lock
down in their major cities which are seriously affecting the economic condition of
A Fuzzy Time Series Prediction Model of the COVID-19 Epidemic 781
the overall population. Apart from the home lock down, governments around the
world are taking new measures to decrease the number of casualties. In this regard,
policymakers need to analyze the region wise infection spread time series data to
understand the effectiveness of the measures and to prepare better for the future. In
this study, we use univariate models of three FTS algorithms, namely—High Order
Fuzzy Time Series (HOFTS), Weighted HOFTS, and Probabilistic Fuzzy Time Series
(PWOFTS) to analyze and model the COVID-19 epidemic data [1]. An FTS model
learns the value of the time series data E(t) at time t, based on a number of previous
values (lag) of the time series (Eq. 1).
E(t) = f (E(t − 1), E(t − 2), E(t − 3), . . . , E(t − n)), lag = n (1)
The weighted version of HOFTS assigns weights to the fuzzy inference rules
with the same precedent to improve the prediction accuracy. The PWOFTS is a
probabilistic model, which assigns weight not only to the precedent of the function,
but also to the consequent. As a result, PWOFTS would work better in complex
environments with large number of training data.
In this work, we have analyzed the dataset of day to day COVID-19 infection
spread along with a mobility trend dataset. Several statistical tools are applied to
the datasets to find correlations and patterns. We develop FTS prediction models
using the infection dataset and describe a model selection procedure for prediction
of the unseen data. Finally, we have also proposed how this study can be extended by
developing a prediction model of the maximum infection size of a particular region
which along with our current models may make better and more effective prediction.
2 Related Works
T. Ganyani et al. estimated the generation time of the COVID-19 disease based on
the onset data of its symptoms [2]. Their study conducted on the data of Tianjin,
China, and Singapore found out that the COVID-19 has a mean incubation period of
5.2 days, and its standard deviation is 2.8 days. The pre-symptomatic transmission
from one host to another was found to be 48% for Singapore and 62% for Tianjin,
China. This is a very important study on the incubation nature of COVID-19 which
explains the high susceptibility of dense population areas from the virus.
In another study, Wei Xia et al. described that the longer than usual incubation
period of the new COVID-19 virus may lead to a quarantine nightmare [3]. They
collected data of 124 confirmed cases from Wuhan, China and analyzed their demog-
raphy, exposure history, and symptom onset data. Their study finds that the mean
incubation period is 4.9 days with 73% of the second-generation patients getting
infected before the symptoms were present in the first-generation cases.
These studies show the reason of COVID-19 being so infectious comparative to
other diseases. From these works, we get a primary estimation that the spread of
the infection at some particular day may depend on the number of infections about
three to seven days back when most of the presently confirmed patients were moving
around in their community or family.
Xia Jiang et al. developed a dynamic transmission model of COVID-19 by dividing
the population into six chambers—un-infected, infected without symptom, infected
and highly infectious but not quarantined, diagnosed and quarantined, potential
victims, and cured [4]. They tested the model with the actual data of COVID-19
infection in Hubei province, China. The work was done at the beginning of the
epidemic and could utilize only a limited amount of data available then. However, at
the early stage of an epidemic, when less than enough data is available, such models
can be useful for short-term prediction and intervention planning.
In another study, Fong et al. described a model for forecasting the COVID-19
with a group of statistical and machine learning algorithms utilizing a small dataset
[5]. The dataset they used comprised of the first fourteen days infection data of
Wuhan, China. They used statistical tools like ARIMA, SARIMA, Holt-Winters
Addictive; along machine learning algorithms like Linear Regression, Decision Tree
learning, SVM, and Neural Network-based algorithm Polynomial Neural Network
with Corrective Feedback (PNN+cf). Their result showed that PNN+cf achieves the
best result with RMSE value of 136.547.
Al-qaness et al. [6] used a modified flower pollination algorithm that uses Salp
Swarm Algorithm (SSA) to improve the performance of an Adaptive Neuro Fuzzy
Inference System (ANFIS)-based COVID-19 prediction model. As another early
research in the domain, the work had little data to work with, but the novel approach
they proposed outperformed several other neural network algorithms including stand-
alone ANFIS implementations. Apart from the COVID-19 outbreak data, the algo-
rithm is also applied to two dataset of weekly influenza cases in US and China and
was shown to be performed well.
Kraemer et al. analyzed the human mobility and government measures, and those
factors may affect the spread of COVID-19 outbreak in Wuhan, China using real-
time mobility data and travel history of specific cases [7]. In their study, they found
out that the initial human mobility data helped spread the virus spatially. However,
after the control measures were taken, the correlation between mobility and virus
spread dropped significantly. At this point, the virus spreads rather locally than over
spatially separated areas. Their report suggests that zoning and controlling mobility
of a large area of dense population during the early outbreak of the virus is the most
effective measure the governments can adopt. Once the virus is spread into different
areas, it will find ways to transmit locally and will pose a formidable task to control
it.
Sunhwa Choi et al. attempted to estimate the possible outbreak size of COVID-19
using Stochastic Epidemic SEIHR model for two regions of Korea [8]. The model
was developed using the outbreak data of daily new confirmed cases in Daegu and
NGP, which were the key outbreak regions in Korea. However, when modeling
the values, they faced challenges like, varying delays between symptom onset to
diagnosis reasoned by changing government measures and individual preferences.
In addition to that they did not take into considerations the number of possible infected
patients incoming from abroad which, according to their statement, could seriously
Fig. 1 COVID-19 infection spread data—Italy (January 22nd–June 1st, 2020)
affect the model. They suggested that in order to reduce the epidemic quick testing
and isolation of possible patients is the key factor.
3 Dataset Overview
Two datasets that are explored in our study are as follows.
3.1 Johns Hopkins University COVID-19 Dataset
In our study, we have used the time series data of COVID-19 infection made avail-
able by the John Hopkins University Center for Systems Science and Engineering
(JHU CSSE) on GitHub [9]. The dataset contains day-wise cumulative number of
confirmed cases, as well as the number of recovered cases and casualty for different
countries and regions (or states). The dataset is updated daily from sources of World
Health Organization, European Center for Disease Prevention and Control (ECDC),
US CDC, and many other government sources of the corresponding countries. At the
time of writing this article, the dataset contained information starting from January
22, 2020 to June 1, 2020. As an example, the normalized data of cumulative confirmed
cases and newly confirmed cases per day is shown in Fig. 1 from the dataset of Italy.
The number of newly confirmed cases at the ith day is determined by subtracting the
number of total confirmed cases of day i − 1 from the total confirmed cases at day
i. The data shows that the highest number of infections per day for Italy was March
11, 2020 after which it has seen a steadily declining trend.
3.2 Apple Mobility Reports
In an ongoing effort to help the researchers fighting COVID-19, Apple Inc. has
published its mobility trend report on the web [10]. The dataset reflects the aggregated
Fig. 2 Day wise Apple mobility trend for Brazil (January 13th–June 2nd, 2020)
number of requests for direction in the Apple Map application by the people of a
specific region/country. From this dataset, we can estimate the mobility of people of
a particular region on a specific day. The dataset is updated daily and contains day
to day data starting from January 13, 2020. At the time of writing this article, the
dataset contains data up to June 2020 with the exception of 11–12 of May. We have
used linear interpolation for estimating the value of these two days. In Fig. 2, we
have shown a sample of this mobility dataset. The country data we presented here is
of Brazil, where the COVID-19 infection spread is one of the highest in the world.
4 Methodology
4.1 Development Environment and Libraries Used
In our experiment with the FTS models, we use the PyFTS library of the MINDS
research lab which is a set of fuzzy logic libraries for Python [1].
The plots of this study are drawn using the Matplotlib library of Python. All
the codes are compiled and executed in Anaconda Python which is a free software
suitable for this work.
4.2 Dataset Exploration and Parameter Values Estimation
Determining Lag from the Infection Time Series Data. In order to determine lag in
the infection spread data, we have used the Autocorrelation Function (ACF) which is
a standard statistical method for this purpose. As several previously discussed studies
on COVID-19 suggests that the virus has an incubation period of two to eight days
(95% Confidence Interval), and 73% of the second-generation patients gets infected
before any onset of symptoms in the first-generation patients it is assumed that the
lag will be around three to seven [1, 2].
We have used the daily new number of confirmed cases for three different countries
(Brazil, Italy, and India) and plot the ACF values for up to fourteen days in Fig. 3.
From figure, we deduce that an estimated lag of eight to ten days should be
enough for learning the models to predict a reasonably good result. A lag value too
high usually makes the learning process slow. Hence, in our experiment, we have
used a lag value of eight to generate the FTS models.
Optimal Partitioning Determination. For determining the preferable number of
partitioning, we make experiment with the day to day newly confirmed infection
number of Brazil, India, and Italy and train the data using HOFTS to check the
RMSE output of the training error. However, as discussed earlier that the total number
of confirmed infections varies from country to country and depends on extraneous
situations; as such, we first normalize the data of each country to one to find the
best result. This way, we are focusing on the trend of the data rather than the actual
number of infected cases. Figure 4 shows the RMSE value for different partition
Fig. 3 Autocorrelation value of the daily new confirmed cases for Brazil, India, and Bangladesh
(January 22nd–June 1st, 2020)
Fig. 4 The effect of the number of partitioning on the training accuracy (RMSE) using HOFTS
numbers when the model is trained using HOFTS with order value of 5. As too many
partitions result in over-fitting of the training data, we chose the number of partitions
to be 35.
Choice of Membership Function. We have experimented with three different
membership functions—Triangular, Trapezoidal, and Gaussian and obtained training
error result for the data of several countries. As the results for these three different
membership functions are found to be similar, we used the simple one—Triangular
membership function, for the experimental models.
Correlation Analysis with the Mobility Trends Data. We have calculated the
Pearson correlation value of the Apple mobility data with the number of new cases
per day for an incubation period of 2–8 days and found negative correlation value for
all the three countries tested (Brazil, Italy, and India). The mobility dataset presents
the daily trend of location search queries made through the Apple maps application
interface by the users of a particular region. This trend reflects how much a particular
population desires to travel to some distant unknown locations. The analysis of this
dataset signifies that, as the infection spreads over time, people become less inclined
to travel to far and unknown locations. However, it is unlikely that people will use
a map application to visit their known business areas, job places, or markets. As a
result, even though the mobility trend data shows significant decrease following the
COVID-19 outbreak, the virus may still spread among people through their necessary
daily interactions. And the perceived change in mobility trend would not affect the
overall aggregated value of infection spread. Figure 5 shows the normalized value of
mobility and confirmed number of infections change with eight days lag for Brazil.
Finding the negative correlation value between mobility trends and infection spread,
we choose not to use the mobility trend as an exogenous variable and opt for the
mono-variable fuzzy time series models instead.
Fig. 5 Day-wise mobility trend versus newly infected cases with seven days lag for Brazil. Infection
data—January 22nd–June 1st, 2020. Mobility data—January 14th–May 24th, 2020
4.3 Development and Selection of Fuzzy Time Series Models
As we draw the day-wise number of confirmed cases for four countries Brazil, India,
Belgium, and Italy, we find the graph of Fig. 6. It can be seen that Italy and Belgium
have successfully checked the infection with number of new infections decreasing
significantly, which means, their data has a full cycle which we can use for training
and testing purpose of similar other countries. On the other hand, countries like India
and Brazil are still in the upward exponential slope. So the challenge now is finding
the most suitable training model for the prediction of a particular country. As such,
we distributed the dataset of 188 total countries randomly into a training set of 151
countries (80%) and testing set of 37 countries (20%). After that we have used the
three FTS algorithms to train with the training set and generate models for each of
the countries separately.
Because different countries have different environment variables and have
different infection spread rate, the challenge is to find the time series model that
can predict the data of a particular country best. The COVID-19 dataset contains day
to day infection spread data of countries from January 22nd to June 1st. So, firstly, we
used all the trained models of all the 151 countries to generate prediction of the trend
of May 19th–May 25th (seven days) of a test country. The model that can predict
the value of these seven day period best is picked and then used for the prediction
of the trend (three days window) of the next seven days (May 26th–June 1st). This
is a reasonable procedure because if we had not known the value of these last seven
days (May 26th–June 1st), we would have picked the model that can best predict
the data of the previous seven days that we know of (May 19th–May 25th). We have
repeated this procedure for all the countries in the test set for all three algorithms
and calculated the Mean, Standard Deviation, and Percentiles of RMSE. The result
of our analysis is presented in the following section.
Fig. 6 Trend of newly confirmed cases—Brazil, Italy, India, Belgium (January 22nd–June 1st,
2020)
Table 1 Training and test accuracy of the prediction models

Algorithm Mean RMSE (in %) STD of RMSE (in %) 25th and 75th percentile (RMSE)
HOFTS 11.25 12.85 2.5–12.64
WHOFTS 21.71 23.46 2.4–38.70
PWFTS 25.67 26.91 2.61–47.06
Fig. 7 Day-wise new infection case prediction result of Algeria for HOFTS, WHOFTS, and PWFTS
(May 25th–June 1st)
5 Results
The test accuracies of HOFTS, WHOFTS, and PWFTS for our test scenario are
presented in Table 1. It is seen that HOFTS performed better than both WHOFTS
and PWFTS. As the number of training data is one cycle only and the data differs
from country to country, we see that the more complex models overfit the training
data and fail to predict the test data satisfactorily. As a result, simpler algorithms like
HOFTS generalize well and perform better in the test environment scenario.
The original day-wise trend for one country (Algeria) along with the predicted
value of the three aforementioned algorithms for May 25th–June 1 are shown in
Fig. 7.
6 Conclusion
In this study, we have demonstrated the use of different statistical and fuzzy time
series techniques to develop a simple yet powerful FTS prediction model selection
process for the ongoing COVID-19 epidemic. Three FTS algorithms used in this
study are tested, and results are compared with the original values for a seven day
look ahead prediction accuracy. The study will hopefully help the policymakers to
better prepare for the pandemic and find more suitable prevention measures using
the outcome.
As the pandemic is still going on and there is no seasonality in the data yet,
more complex algorithms like WHOFTS and PWFTS could not perform better than
HOFTS in this study. Moreover, predicting the infection size is a challenging topic
still now as it depends on many outside variables, like swiftness of testing and isola-
tion, government measures, individual precaution, etc. A suitable prediction model
may be developed to predict the maximum infection size (number of infected people)
of a particular country or region based on the aforementioned variables. We can then
use that model to first predict the maximum infection of a country that has not reached
infection peek yet. Afterward, we have to normalize that particular country’s current
infection trend with the peak value and then predict its infection trend with the trained
model of normalized dataset from another country that has seen a full cycle of the
infection. Used in this way, the fuzzy prediction models may be able to make long
term prediction with better accuracy.
References
1. P.C.L. Silva et al., pyFTS: Fuzzy Time Series for Python (Belo Horizonte, 2018). http://doi.org/
10.5281/zenodo.597359
2. T. Ganyani, C. Kremer, D. Chen, C. Torneri, C. Faes, J. Wallinga, N. Hens, Estimating the gener-
ation interval for COVID-19 based on symptom onset data. medRxiv 2020.03.05.20031815
[Preprint]. March 08, 2020. https://doi.org/10.1101/2020.03.05.20031815. Cited 15 June 2020
3. W. Xia, J. Liao, C. Li et al., Transmission of corona virus disease 2019 during the incubation
period may lead to a quarantine loophole. medRxiv 2020.03.06.20031955 [Preprint]. March
08, 2020. https://doi.org/10.1101/2020.03.06.20031955. Cited 15 June 2020
4. X. Jiang, B. Zhao, J. Cao, Statistical analysis on COVID-19. Biomed. J. Sci. Technol. Res.
https://doi.org/10.26717/bjstr.2020.26.004310
5. R.G. Crespo, E. Herrera-Viedma, N. Dey, S.J. Fong, G. Li, Finding an accurate early forecasting
model from small dataset: a case of 2019—nCoV novel coronavirus outbreak. IJIMAI. https://
doi.org/10.9781/ijimai.2020.02.002
6. M.A.A. Al-qaness, A.A. Ewees, H. Fan, M. Abd El Aziz, Optimization method for forecasting
confirmed cases of COVID-19 in China. J. Clin. Med. 9, 674 (2020)
7. M.U.G. Kraemer, C.H. Yang et al., The effect of human mobility and control measures on the
COVID-19 epidemic in China. Science 493–497 (2020)
8. S. Choi, M. Ki, Estimating the reproductive number and the outbreak size of COVID-19 in
Korea. Epidemiol. Health 42 (2020). https://doi.org/10.4178/epih.e2020011
9. COVID-19, Data Repository by the Center for Systems Science and Engineering (CSSE)
at Johns Hopkins University. https://github.com/CSSEGISandData/COVID-19. Accessed 22
June 2020
10. Apple Mobility Reports. https://www.apple.com/covid19/mobility. Accessed 22 June 2020
Modelling of Generation and Utilization
Control for Isolated Hybrid Microgrid
for Rural Electrification
Kuldip Singh, Satyasis Mishra, Davinder Singh Rathee,

Demissie Jobir Gemecha, and R. C. Mohanty
Abstract For ameliorating the villages, electrification Indian regime introduced

many programmes under rural electrification scheme. Under these programmes’,
regime is providing the electrical connection to each village in ON-Grid or OFF-
Grid mode. The rural sector is expected to play a key role for microgrid framework,
because in future rural electrical demand will probably exceed more than 30% of
the total yearly consumption in India. The implementation of microgrid for rural
electrification differs on the type of dispersed generation facilities, accessibility of
sources, distribution control, stability and reliability of system, economic analysis,
and operating policy. Along with there are few issues in isolated microgrid authentic
time electrical power management is the major challenges for isolated hybrid micro-
grid for rural electrification. In this study, we are discussing the basic modelling of
generation, controlling for rural electrification with renewable sources, stability issue
for isolated hybrid microgrid.
Keywords Village electrification isolated hybrid microgrid · Generation ·

Utilization · Stability and reliability
K. Singh
Department of EEE, Centurion University of Technology and Management, R.Sitapur, Odisha,
India
e-mail: ksmann3@gmail.com
S. Mishra (B) · D. S. Rathee · D. J. Gemecha
Department of ECE, Signal and Image Processing SIG, Adama Science and Technology
University, Adama, Ethiopia
e-mail: satyasis.mishra@astu.edu.et
D. S. Rathee
e-mail: jerryrathee@gmail.com
D. J. Gemecha
e-mail: gelmechad@gmail.com
R. C. Mohanty
Department of Mechanical Engineering, Centurion University of Technology and Management,
R.Sitapur, Odisha, India
e-mail: rcmohanty@cutm.ac.in
792 K. Singh et al.
1 Introduction
Powertransmission and distribution is a challenging and arduous task in rural compo-

nents of India. As per the Indian regime statistics, nearly about 30% of the villages
are still in lack of electricity as there is no opportune communication for conveyance.
In India, more than 73% population are predicated on villages in which the major
population in villages depend on irrigation and diminutive business. In today’s life,
electricity plays major role in human life for amending the socio-economic condi-
tions, connectivity with outer world through TV/radio/phone, income amendment
with incipient opportunities of minute household business, and it is additionally
playing the vital role for edification amendment [1]. The electricity use reducing
the traditional way of cooking and lighting, which will have impact on health condi-
tions. As the traditional grid is utilized for supplying the electrical energy to rural and
urban consumers that provide the supply through central high voltage transmission
network with conventional sources predicated central generation units. As conven-
tional sources are not reliable due to economic and environmental impact in India
for rural electrification, so the scenario is transmuting from conventional sources to
non-conventional energy sources like solar, biomass, wind energy, etc [2, 3].
Active efforts are afoot to analyses and providing the solution for village elec-
trification through isolated hybrid microgrid with renewable sources. Albeit most
of the work in the microgrid areas along with astute grid has been done throughout
the world for rural electrification in ON-Grid mode. The literature surveyed unfolds
the rural electrification and utilization scenario in India, different challenges for
village electrification, implementation microgrid, demand side management, power
quality and scope of isolated hybrid microgrid along with power management in
local distribution and generation units. Wimmler et al. the renewable energy is used
for islanding electrification, as renewable energy sources are natural sources. The
smart grid technology is used to shift the loads from source to storage during off
generation time. The demand side management is the major component of smart
grid for communication between electrical load demand and electrical generation
[4]. Goy et al. the study considers the large-scale building load management for
thermal system for improved and optimized the energy consumption, control energy
wastage, peak load demand and integration of renewable energy sources [5]. Alham
et al. the wind power generation unit with ramping and mismatch between genera-
tion and demand is major issue in power system. Arefifar et al. gave the emphasis on
the systematic and optimized approach implementation for evaluating the reliability
indices and improving the supply security of the microgrid [6]. Ma et al. focused
on operation of micro-energy grid with optimal dispatch strategies from multiple
generation sources. The aim of the study is minimizing the daily operation cost with
day ahead dynamic optimal generator model with mixed integer linear programming
optimization based on demand response [7]. Chekried et al. described on energy flow
management system for domestic loads with sola PV system. The solar PV system
with stored system is implemented in the study to fulfil the daily load demand for
domestic uses in each home. The energy management system is the communication
Modelling of Generation and Utilization Control … 793
and control system between the generation sources integration and load demand with
different mode of operation [8]. These factors considered for system planning are
load forecasting, inadequate correlation or power balancing, faster ramp rate and
power quality. The demand side management is implemented for dynamic economic
dispatch for load balance along with dynamic operation and control of power system.
The demand side management system uses for “peak clipping” and load shifting at
utilization to attain a load fluctuating objective by using time of use for changing the
tariff during ultimate period [9]. The study considers the scope and methodology for
implementation control for isolated hybrid microgrid in Sect. 2, Sect. 3 is planning
and load forecasting for rural electrification, Sect. 4 is generation model for isolated
hybrid microgrid, Sect. 5 is the results based on real-time study, and finally the paper
is concluded.
2 The Scope and Methodology for Control Implementation
The isolated hybrid microgrid has gained more attention for improvement of the
stability and reliability of electrical system in rural area with “renewable energy”
sources. In islanded hybrid microgrid, all the sources are natural resources to produce
electrical generation. The generation is depending on the climate and natural condi-
tion and availability of sources. The load demand in rural area does not steady all the
time, during peak demand is higher than off peak hours, due to sudden load change
the reliability and quality of power generation is exceptionally low. At some time, the
generation from solar PV power plant is excess compare to maximum load demand.
The excess generating power is flowing to biomass generation plant and creating the
transient stability problem, which interrupted the inverter generation from solar PV
plant and biomass power plant. To overcome this problem, electrical power manage-
ment system is required. It is the real-time power management system which acts
as a bridge between the utility and generation sources, and it is also the planning,
implementation and monitoring of electrical utility at each house in the village. As a
result, it changes the time of utility based on priority and time pattern for control the
peak load demand. The main objective of real-time power management system is to
control the electricity utilization based on the generation during peak demand and
off-peak demand in each household in the villages in rural area [10–12]. In large-
scale power system, the energy mismatch between electrical load demand and power
generation is controlled by slack bus. However, in isolated hybrid microgrid, such
power reserve is often absent. Thus, it is more interesting for development of new
mechanism for isolated hybrid microgrid to enable the generation control, distribu-
tion control and load control with optimization control game theory. In this study, the
methodology for isolated hybrid microgrid is implemented based on Time, priority
of load, generation and integration with central control unit along with balance distri-
bution system. The central control unit RPMS is implemented on non-cooperative
game theory approach based on mixed Nash equilibrium optimization techniques.
794 K. Singh et al.
In isolated hybrid microgrid, for village electrification, the first challenge is plan-
ning and load forecasting in each household with local radial electrical distribution
system. In the rural electrification, with renewable energy, stability and reliability of
electrical distribution system are major challenges due to uncontrol load utilization
and power generation from solar PV plant and biomass power plant. The study is
carried out for 125 households in the village.
3 Planning and Load Forecasting for Rural Electrification
In electrical generation and distribution system, the planning is major concern for
assuring that the increasing demand for electricity in the village can be fulfilled by
isolated hybrid microgrid in additions to technical adequate with economic feasi-
bility. Even though considerable work has been done for technical adequate for the
sources of generation in microgrid with systematic approach to electrical power
generation and local distribution, its application of real-time distribution load and
generation control for isolated hybrid microgrid planning is considered in the study,
which has been unfortunately somewhat neglected. For future electrification, it needs
a fast and economic development tool to evaluate the consequence of different
proposed alternatives for village electrification and their impact on the lively hood,
economical, reliable and safe electric demand for domestic system [13].
3.1 Load Demand Module
A set of Pd = {1, 2, . . . , n} houses that share the load on solar power plant and
biomass power plant. For each house η, let L ηh represent the total load demand
with respect to appliance utilization based on time t ∈ τ = {1, 2, . . . , T }. For the
integration based on load demand, the time slots are taken based on hourly total time
T = 24.
The total daily load demand for user n is

L n L 1n , . . . , L nT (1)
The total load demand for all users during a time t is

L dt L tn (2)
n∈N
Daily peak load demand
L dPeak = max (3)

t∈τ L dt
Average Load demand
1
L Davg = L dt (4)
T t∈τ
Peak to average ratio of load demand is
T max L dt
Pdr ms = t∈τ (5)
t∈τ L dt
We define τ be the set of operating time slot of user n ∈ N , and the user is
restricting the minimum and maximum power consumption in different operating
time slots.
0 ≤ Pdn
t
≤ Pdn
max
,∀t ∈ τ (6)
Let Pd M = predetermine total daily load demand of connected load from each
home n ∈ N .
The energy balance equation is

t
Pdn = Pd M (7)
t∈τ
In isolated hybrid microgrid, loads and sources are in distribution nature. We

pretended to be equivalent injected power from microgrid or load flow analysis. Here
(k)spec (k)spec
Pn and Q n represent to specified active power and reactive power injection
in the bus. The Q (k) (k)
n and Pn represent to equivalent active power and reactive power
(k)cal
injection in bus. The Pn and Q (k)cal
n are calculated values of injected active power
and reactive power in the bus [13].
The resultant active power
Pn(k) = Pn(k)spec − Pn(k)cal (8)
The reactive power
Q (k) (k)spec
n = Qn − Q (k)cal
n (9)
Here

n
(k) (k)
Pn(k)cal = V V |Yni | cos θni + δ (k) − δ (k) (10)
n i i o
i=1

n
(k) (k)
qn(k)cal = − V V |Yni | cos θni + δ (k) − δ (k) (11)
n i i o
i=1
796 K. Singh et al.
The maximum demand in the rural electrification is changing with respect to load
scheduling and time. In this study isolated hybrid microgrid, the maximum demand
is occurring during priority four (P4 ) [14].
In summer during priority, (P4 ) maximum connected load/home L Dmix P4 =
2.145 kW/home
In winter, during priority (P4 ) maximum connected load/home L Dmix P4 =
2.005 kW/home
The diversified power demand of the composite group of loads or unrelated loads
over a specified period in each home.
Diversified Demand/Home L DDiv = 0.750 kW.
The non-coincident demand is the sum of group of loads without any restrictions
over the interval of time.
Non-coincident demand/home L DNC = 0.2 kW
The connected loads the sum or maximum load consuming apparatus in each
home
Maximum connected load/home L maxCL = 2.245 kW
The demand factor is the ratio of maximum demand of a local distribution system
in microgrid.
Maximum Demand 268.125

L DF = = 0.953 (12)
Total Conneected Demand 281.25
The utilization factor is the ratio of maximum demand of a connected load to the
rated capacity.
Maximum Demand 268.125

L UF = = 0.893 (13)
Rated System Capacity 300
The load factor is the ratio of the average load over a designated period of time to
the peak load. In village, load factor can be calculated based on priority scheduling
or time scheduling.
Average Load L AVHome 1.495

L LF = = = 0.665 (14)
Peak Load L PLHome 2.245
The diversity factor is the sum of the maximum demand of individual home in
the local distribution system to maximum load demand on isolated microgrid.
n
L HD1 + L HD2 + L HD3 + . . . + L HDn LH 48.68
L DIVF = i=1 Di = = 0.961
L CON D max L CON D max 50.62
(15)
The load diversity is the difference between the sum of peak of more than one
individual loads as the peak of the combine load.

n
L LD L HDi − L maxCL = 50.62 − 48.68 = 1.94 kW (16)
i=1
The coincidence factor is the ratio of maximum coincident total demand on to the
sum of maximum power demand from each home at microgrid.
L CD max 50.62
L CONF n = =1 (17)
L
i=1 H Di 48.68
The contribution factor is define Ci “the contribution factor in isolated hybrid

microgrid is depends on the priority scheduling of load w.r.t. maximum demand of
number of houses in village.
n
Ci × L h i
L CF n
i=1 (18)
i=1 L h i
4 Generation and Controlling of Isolated Hybrid Microgrid
The generation module of isolated hybrid microgrid is based on the solar power
plant and biomass power plant. The generation is controlled by generation control
unit with respect to time and priority of Loads. The Generation solar power plant is
based on AC power flow solution for distribution system. The generation of solar PV
plant depends on time and light intensity.
Let
t
PPV = generation from solar power plant ∀ t ∈ τ
max = maximum generation from solar power plant ∀ t ∈ τ
t
PPV
min = minimum generation from solar power plant ∀ t ∈ τ
t
PPV
The generation of solar power plant with respect to light intensity λ

PPV gen = Pλ1 . . . Pλt . . . Pλτ (19)
where Pλt = maximum power generation.

The generation biomass power plant is based on AC power flow solution for
distribution system. The generation of biomass power plant depends on time, load
demand and availability of biomass.
Let
798 K. Singh et al.
t
PBM = generation from biomass power plant ∀ t ∈ τ
max = maximum generation from biomass power plant ∀t ∈ τ
t
PBM
min = minimum generation from biomass power plant ∀t ∈ τ
t
PBM
The generation of biomass power plant with respect to total biomass availability
Bmt to fulfil the demand for 24 h.

PBM gen = PB1mt . . . PBt mt . . . PBτmt (20)
where PBt mt = maximum power generation.

At each time slot t,
gen + PBM gen ≥ 0

t t
PPV (21)
The total demand in isolated microgrid is shared by three different sources

t
Pdη = PPV gen + PBM gen (22)
t∈τ t∈τ
4.1 Non-cooperative Game Theory
The load balancing for isolated hybrid microgrid with real-time optimization is set
by threshold s to achieve the power load balancing. To measure the load demand in
different houses, we define the electric load ratio in home n at time t as [15]
E nt (s) + Bnt
rnt (s) = (23)
Cm
where E nt is the energy consumption of home n and it depends on the energy threshold
s for all homes Bnt = background power load and Cm = capacity of source m.
Let qit = Cm − Bnt maximum supply available to the load i at time t. Usually in
each household use the electricity in inflexible patterns and the demand on microgrid
can be inelastic. Base on the load ratio rit . The electric load index (ELI) across all
locations and time slots as
2
ELI rit (s) Ci (24)
t∈τ i∈N
where ELI is motivated by the index measurement techniques used for feeder load
balancing is isolated hybrid microgrid distribution system. ELI not only measures the
overall load ratio but also gives different weights based on the capacities in different
locations. The RPMS aims at balancing the load ratio rit (s) at all locations and all
time slots by minimizing the ELI.
Optimal response time: Any source has maximum response time at most twice
the optimal response time. Let PV and biomass has maximum load PL max .
Let load PL be assigned to PV and biomass
r1 (L 1 ) ≤ rs L s + d j ∀s ∈ S (25)
By Nash equilibrium, the response function is linear

We have ri (L i ) = L i
L1 ≤ Ls + d j ∀ s ∈ S (26)

m L1 ≤ Ls + d j (27)
s∈S
1
L1 ≤ Ls + dj (28)
m s∈S
At optimal allocation: Load PL is assigned to solar power plant and biomass power
plant. So, opt ≥ d j .

Ms = dL (29)
s∈S ρ L ∈PL
Maximum loaded on solar power plant and biomass power plant at least average
is
1 1
= dL = Ms (30)
m ρ ∈P m s∈S
L L
Total time or average time is

= Ms rs (L s ) (31)
s∈S
5 Simulation Results
The study is carried out with and without real-time power management (RPMS)
control for isolated hybrid microgrid with solar power plant and biomass power plant
DG Set. Figure 1 shows the load demand in 24 h without distribution control, and the
demand on distribution system is unstable. Figure 2 shows the load demand control
800 K. Singh et al.
Fig. 1 Total load demand in microgrid without control
Fig. 2 Total load demand with RPMS control
with RPMS system with respect to time, which will improve the stability of isolated
hybrid microgrid with load controlling on distribution system with distributed control
in each home with respect to time and priority.
Figure 3 shows the integration of biomass and solar power plant with respect
to load demand without generation control and the excess reverse power flow to
biomass DG set which changes the frequency and speed of DG set. The transient
instability condition is occurring, which interrupts the solar power plant generation
and damages the installation of biomass DG set.
To overcome the transient stability problem generation control implemented with
RPMS control unit for isolated hybrid microgrid, which integrate the solar power
plant and biomass power plant w.r.t. load demand and control the reverse power flow
by scaling the solar generation as shown in Fig. 4.
Fig. 3 Reverse power flow without control
Fig. 4 Reverse power flow control with RPMS system
Figure 5 shows the frequency of distribution system with and without RPMS
control unit, which is the major cause for poor power quality in local distribution
system. Figure 6 shows the biomass DG set speed with and without RPMS generation
control unit, which is controlling the voltage and frequency instability issue.
The planning of distribution generation system for rural electrification has the
important role for improvement of stability and reliability. The paper describes the
different characteristics of load and distribution system planning in different sections.
In the characteristics, the different factors for load planning, which are affecting
the stability, reliability and quality of hybrid microgrid, have been considered. The
802 K. Singh et al.
Fig. 5 System frequency with and without RPMS control
Fig. 6 Speed of biomass alternator with and without RPMS control
characteristics of load planning describe the different parameters of load, maximum

and minimum load demand on microgrid, planning of load based on the time, priority
and session load demand. The RPMS control improves the stability of isolated hybrid
microgrid by controlling the excess power generation as shown in Fig. 6. The real-
time power management system is the effective solution for zero feeding in biomass
generation unit for stabile operation for village electrification with renewable energy
sources. The focus of study is planning the stabile, reliable and cost optimize system
for village electrification system with renewable sources.
6 Conclusion
The optimize scheduling of generation from renewable energy sources with respect
to time and priority to fulfil the electrical load demand. The major challenge for
isolated microgrid for village electrification is load scheduling in each home with
respect to time and generation, which is directly related to stability and reliability of
distribution generation system. In this study, real-time power management system
(RPMS) is proposed for optimize generation scheduling by integration of different
renewable energy sources with respect to time and load demand and distribution
control for balancing the load distribution in radial distributed system.
References
1. D. Palit, A. Garimella, M. Shardul, S. Chaudbury, Analysis of the electrification programme

in India using the ‘Energy plus’ framework and the key lessons. Final report 2015 GNESD,
Teri.2015, pp. 1–83
2. A. Jain, S. Ray, K. Ganesh, M. Aklin, C. Cheng, J. Urpelainen, Access to clean cooking Energy
and Electricity Survey of States. CEEW Report—2015, pp. 1–65
3. R. Bilolikar, R. Deshmukh, Rural Electrification in India—An Overview (National Power
Training Institute, 2014), pp. 1–12
4. C. Wimmler, G. Hejazi, E. de Oliveira Fernandes, C. Moreira, S. Connors, Impact of load
shifting on renewable energy integration. Elsevier Energ. Procedia 107, 248–252 (2017)
5. S. Goy, A. Ashouri, F. Marechal, D. Finn, Estimating the potential for thermal load management
in buildings at a large scale: overcoming challenges towards a replicable methodology. Elsevier
Energ. Procedia 111, 740–749 (2017)
6. S.A. Arefifar, Y.A.-R.I. Mohamed, T.H.M. EL-Fouly, Optimum microgrid design for enhancing
reliability and supply security. IEEE Trans. Smart Grid 4(3), 1567–1575 (2013)
7. T. Ma, J. Wu, L. Hao, Energy flow modelling and optimal operation analysis of the micro
energy grid based on energy hub. Elsevier Energ. Convers. Manag. 133, 292–306 (2017)
8. F. Chekired, Z. Smara, A. Mahrane, M. Chikh, S. Berkane, An Energy flow management
algorithm for a photovoltaic solar home. Elsevier Energ. Procedia 111, 934–943 (2017)
9. M.H. Alham, M. Elshahed, D.K. Ibrahim, E.E.D.A. EI Zahab, Optimal operation of power
system incorporating wind energy with demand side management. Elsevier J. Ain Shams Eng.
J. 8, 1–7 (2017)
10. D. Salomonsson, A. Sannino, Low-voltage dc distribution system for commercial power
systems with sensitive electronic loads. IEEE Trans. Pow. Del. 22(3), 1620–1627 (2007)
11. W. Lu, B.T. Ooi, Optimal acquisition and aggregation of offshore wind power by multiterminal
voltage-source HVDC. IEEE Trans. Pow. Del. 18(1), 201–206 (2003)
12. A. Barbato, A. Capone, Optimization models and methods for demand side management of
residential users: a survey. MDPI J. Energ. 7, 5787–5824 (2014)
13. K. Singh, M.N. Kumar, S. Mishra, Load flow study of isolated hybrid microgrid for village
electrification. Int. J. Eng. Technol. 7(2), 232–234 (2018). ISSN 2227-524X. http://dx.doi.org/
10.14419/ijet.v7i2.23.11925
804 K. Singh et al.
14. T. Gonen, Electrical Power Distribution Enegineering, 3rd edn. CRC Press, pp. 1–88. ISBN
978-1-1381-9802-9
15. C. Miao, Y. Zhang, Fully mixed nash equilibria for the load balancing game on uniform parallel
machines. Int. Math. 33, 1607–1612 (2010)
Toyama de Walking: Practice
of Location-Based Data Collection Using
Augmented Reality Applications
Yuya Ieiri, Hung-Ya Tsai, and Reiko Hishiyama
Abstract Location-based data is used in a wide variety of applications and can be

collected using various methods. Besides these methods, it is expected that a method
to collect location-based data by walking applications using augmented reality (AR)
will be common in the future. AR has the nature that encourages users to collabo-
rate with the real world, and it is supposed to help collect location-based data about
the atmosphere and quality of the city. However, the features of the location-based
data collected by walking applications using AR have not yet been revealed. Clari-
fying the features is important to grasp the position of the collection method using
AR. Therefore, we developed an AR walking application “Toyama de Walking” for
collecting location-based data and conducted a demonstration experiment using the
application. As a result, the features of the location-based data collected by walking
applications using AR were revealed.
Keywords Location-based data · Augmented reality · User behavior analysis
1 Introduction
Location-based data can be used in various situations. For example, there has been a
study on the analysis of users’ behavior in outdoor activities using location informa-
tion [1]. In addition, there has been an attempt to represent commercial and social
activities in the city by using location-based data [2], and an analysis of a crowd of
Y. Ieiri (B) · H.-Y. Tsai · R. Hishiyama

Faculty of Science and Engineering, Waseda University, Tokyo, Japan
e-mail: ieyuharu@ruri.waseda.jp
H.-Y. Tsai
e-mail: hungyatsai@akane.waseda.jp; hungyatsai@osensetech.com
R. Hishiyama
e-mail: reiko@waseda.jp
H.-Y. Tsai
Osense Technology Co., Ltd., Taipei City, Taiwan
806 Y. Ieiri et al.
people in the city [3]. One of the reasons for the widespread use of location-based
data is the use of devices, such as smart phones that are equipped with global posi-
tioning system (GPS) function. Such devices enable the collection of various types
of location-based data such as that transmitted using social media websites such as
Twitter,1 Facebook,2 and Foursquare3 via smart devices, and a huge volume of such
data can be collected at present.
Various researches have been conducted on the location-based data collected by
the above methods. In particular, Twitter data with location information has been
shown to be useful in investigating human movement patterns [4] and there are
numerous ways to use it [3, 5]. In addition, as a part of the location-based social net-
works (LBSN), various studies have been conducted on location-based data collected
using Foursquare [2, 6].
Besides these methods, it is expected that a method to collect location-based
data by walking applications using augmented reality (AR) will be common in the
future. AR has matured rapidly in recent years, showing great potential for prac-
tical applications [7]. In addition, AR has been used for various purposes [8], and
walking applications are typical example of using AR. Walking applications using
AR encourage users to collaborate with the real world, and the game aspects of the
applications represented by “Pokemon Go4 ” encourage users to continue using it.
In particular, the nature that encourages users to collaborate with the real world is
supposed to help collect location-based data about the atmosphere and quality of
the city that each person feels. Such location-based data is useful for clarifying the
image of the city. Clarifying the image of the city is important in city management
[9].
Tursi et al. [1] attempted to collect users’ location data by walking application
using AR. However, the features of the location-based data collected by walking
applications using AR have not yet been revealed. Clarifying the features is important
to grasp the position of the collection method using AR in the future. The positions
of the collection methods are necessary to judge what kind of method is useful in a
situation.
Therefore, the purpose of our study is to clarify the features of the location-
based data collected by walking applications using AR. In order to achieve this goal,
we developed a walking application, called “Toyama de Walking”, using AR. In
addition, we conducted an experiment near Waseda University, Japan, and analyzed
the results of the experiment to reveal the features of the location-based data collected
by walking applications using AR.
The remainder of this paper is structured as follows: Sect. 2 contains the details
of our “Toyama de Walking” application. Details of the conducted experiment and
the results of the experiment are given in Sect. 3. Conclusions from this work and
ideas for future work have been discussed in Sect. 4.
1 https://twitter.com.
2 https://www.facebook.com.
3 https://foursquare.com.
4 https://www.pokemongo.com/en-us/.
Toyama de Walking: Practice of Location-Based Data … 807
2 Toyama de Walking
“Toyama de Walking” is a native application using AR and the users of the application
walk around using mobile devices such as smart phones. In this application, the
users can send the location-based data, and the location-based data contains location
information and text data, similar to Twitter. In addition, in order to investigate where
the location-based data is transmitted, data on the characteristics of the transmission
point can also be collected.
Firstly, in “Toyama de Walking,” the user registration screen is displayed. On this
screen, the user enters their user name and the password to register on the application.
Following this, a destination selection screen, similar to the one shown on the left
side of Fig. 1, is displayed. When the user selects a destination on this screen, the
screen transitions to the walking screen similar to the one shown in the center in
Fig. 1. In the walking screen, the destination is displayed as an AR-Point of Interest
in the augmented reality space and the direction of the destination is displayed by
a compass at the bottom of the screen. Ambiguous navigation using a compass is
suitable for walks while feeling the surrounding atmosphere [10], and is compatible
with walking applications using AR, which encourages the user to collaborate with
the real world.
The user can send out location-based data by pressing the tweet button at the
bottom left of the walking screen if he or she gets any thoughts or feelings. When
the user presses the tweet button, it transitions to the tweet screen, similar to the
one shown on the right side of Fig. 1. On the tweet screen, the user’s thoughts and
feelings are transmitted as text data along with the user’s location information. At
the same time, the characteristics of the area are evaluated by a rating between one
to four stars in three aspects: “How good is the atmosphere?”, “How lively is the
Fig. 1 User interface of the “Toyama de Walking” application

808 Y. Ieiri et al.
Fig. 2 System architecture of the “Toyama de Walking” application
area?”, and “How surprised are the users are?” The number of stars indicates the
degree, with four stars indicating the highest degree.
When the user approaches a destination, AR-Point of Interest changes into a
present. By tapping on this present, the user can earn points. In addition, the user can
also earn points by sending out location-based data on the tweet screen.
The system architecture of “Toyama de Walking” is shown in Fig. 2. In this
application, mobile clients retrieve content data, such as destination data and other
tweet data, from the backend system and display it as an AR object. When the
user uses the application, the mobile client uploads the user’s username, location
information, and data about the tweet to the database. In addition, the backend system
stores the updated content data in the database and can check all the data in the
database by querying.
3 Experiment and Results
A demonstration experiment using the “Toyama de Walking” application was con-

ducted in the neighborhood of Waseda University in order to clarify the features of
the data that can be collected by the method to collect location-based data using
walking applications based on AR. The target area of this experiment is shown in
Fig. 3. The black circles labeled 1, 2, and 3 in the figure indicate the location of the
destinations in this experiment.
The experiment was conducted as an event. The event experiment was held for the
Waseda University students and they could participate in a lottery by walking around
the neighborhood and accumulate certain points using our application. The star in
Fig. 3 Target area of the

experiment
Fig. 3 indicates the starting point, at which the instructions of the application and the
lottery were held. Therefore, it can be said that the experiment participants used the
application freely to explore the neighboring areas and transmitted location-based
data as one likes.
In the experiment, 14 participants transmitted location-based data, and the total
of 51 location-based data were collected. Then, we divide these 51 data into three
groups based on the characteristics of the text part of the collected data. 8 out of 51
data (Group 1) are data about the application features such as “I’m going to reach
60 points,” and 26 out of 51 data (Group 2) are data about what users felt when they
were walking such as “I’m starving! I want to go for lunch.” In addition, 17 out of 51
data (Group 3) are data about the atmosphere and quality of the city such as “There
was a beautiful cherry blossom tree,” and “The curry is 470 yen.” Checking the data
of Group 3, it was found that Street X in Fig. 3 is a street with cherry blossoms and
has a good atmosphere, and Area Y is an area where many stores are crowded. As
described above, the collected location-based data showed the characteristics of the
target city.
Figure 4 shows the 51 data points on the map corresponding to group shown
above. Each pin in Fig. 4 indicates the location where the data for each group was
transmitted. We discuss the surroundings of the three destinations 1, 2, and 3 in Fig. 4.
It can be seen that a large amount of data was transmitted from around destinations
1 and 2, while that transmitted from around destination 3 was small. There were
landmarks around destination 1 such as the starting point where the instruction and
lottery for this event was held. There were many restaurants and supermarkets in the
vicinity of destination 2. On the other hand, the area around destination 3 was a quiet
area, the number of the stores in this area was small, and there were few people on
this street. Therefore, it was speculated that a considerable amount of location-based
data would be transmitted when there are objects of interest for the users around the
destination.
810 Y. Ieiri et al.
Fig. 4 Location-based data obtained from the experiment shown on the map corresponding to each
group
Table 1 Experimental results focusing on the four-level evaluation

Score 1 2 3 4
(a) How good is the atmosphere? 0 2 22 27
(b) How lively is the area? 1 9 25 16
(c) How surprised are the users? 2 6 26 17
Total 3 17 73 60
In the experiment, location-based data not only in the form of text data, but also
in the form of the four-level evaluation in terms of three aspects, namely, “How
good is the atmosphere?”, “How lively is the area?”, and “How surprised are the
users?”, was collected using the “Toyama de Walking” application. The results of
the four-level evaluation of the 51 data points are given in Table 1. Out of 51 data
inputs received, the data corresponding to the scores of 1 or 2 are classified as red,
while those corresponding to scores of 3 or 4 are classified as blue. The results of the
classification corresponding to the three aspects are shown by the red and blue pins
on the map in Fig. 5. From Table 1 and Fig. 5, it can be seen that the location-based
data was frequently transmitted at a pleasant, lively and surprising area.
Furthermore, we analyzed whether the characteristics of the collected data (Group
1, Group 2, Group 3) differ depending on the difference in the evaluation from these
three aspects. The differences in the evaluation for each of the three aspects and the
characteristics of the collected data are summarized in Table 2. The first column of
these tables shows the evaluation value from each aspects, and the first row of these
tables shows the characteristics of the data. Each value in the table indicates the
number of data.
Pearson’s Chi-square test was performed on each contingency table in order to
confirm whether the characteristics of the collected data differ depending on the
difference in the evaluation level from the three viewpoints. The results are shown in
Fig. 5 Location-based data shown on the map focusing on the four-level evaluation. The pins in red
indicate the location-based data corresponding to scores of 1 or 2 while the pins in blue correspond
to scores of 3 or 4
Table 2 Contingency tables showing the relationship between the evaluation results of three aspects
and the characteristics of the data
Evaluation Group 1 Group 2 (feeling) Group 3 Total
(application (atmosphere and
features) quality)
(a) How good is the atmosphere?
1 and 2 0 2 0 2
3 and 4 8 24 17 49
Total 8 26 17 51
(b) How lively is the area?
1 and 2 0 9 1 10
3 and 4 8 17 16 41
Total 8 26 17 51
(c) How surprised are the users?
1 and 2 0 6 2 8
3 and 4 8 20 15 43
Total 8 26 17 51
Table 3 The results of Pearson’s Chi-square test

Aspect Chi-square value Degree of freedom p-value
(a) How good is the 2.002 2 0.368
atmosphere?
(b) How lively is the 7.698 2 0.021*
area?
(c) How surprised are 2.760 2 0.252
the users?
∗ p < 0.05
812 Y. Ieiri et al.
Table 3. From Table 3, it was confirmed that the difference in the evaluation level of
“How lively is the area?” makes a difference in the characteristics of the collected data
( p < 0.05). The result shows that the lively of the city influences the characteristics
of the collected data. Specifically, it was speculated that a lot of data about feelings
would be sent out in calm places, and a lot of data about the atmosphere and quality
of the city would be transmitted in lively places.
4 Conclusion
We collected location-based data using “Toyama de Walking.” Analysis of the col-

lected data revealed the following features of the location-based data collected by
walking applications using AR: (i) Location-based data about the atmosphere and
quality of the city can be collected during the walk. (ii) A lot of location-based data
can be collected around attractive destinations. (iii) Location-based data was fre-
quently transmitted at a pleasant, lively and surprising area, and the lively of the area
influences the characteristics of the collected data.
A future work is to try a city management by conducting a large-scale experiment
using “Toyama de Walking” and collecting a large amount of location-based data.
Furthermore, it is necessary to clarify the features by comparing the collection method
using AR with other collection methods using Twitter, Facebook and Foursquare.
Acknowledgements This research was supported by the Waseda University Grant for Special
Research Projects and a Grant-in-Aid for JSPS Research Fellow (19J22365).
References
1. Tursi G, Deplano M, Ruffo G (2014) AIRCacher: virtual geocaching powered with augmented
reality. In: Proceedings of 25th ACM conference on hypertext and social media, pp 330–332
2. T. Agryzkov, P. Martí, L. Tortosa, J.F. Vicent, Measuring urban activities using Foursquare
data and network analysis: a case study of Murcia (Spain). Int J Geogr Inf Sci 31(1), 100–121
(2017)
3. Khalifa MB, Díaz Redondo RP, Vilas AF, Rodríguez SS (2017) Identifying urban crowds
using geo-located social media data: a Twitter experiment in New York City. J Intell Inf Syst
48(2):287–308
4. Al-Jeri M (2019) Towards human mobility detection scheme for location-based social network.
In: 2019 IEEE symposium on computers and communications (ISCC), pp 1–7
5. M.M. Mostafa, Mining and mapping halal food consumers: a geo-located Twitter opinion
polarity analysis. J Food Prod Market 24(7), 858–879 (2018)
6. M. Li, R. Westerholt, A. Zipf, Do people communicate about their whereabouts? Investigating
the relation between user-generated text messages and Foursquare check-in places. Geo-spatial
Inf Sci 21(3), 159–172 (2018)
7. Diaz C, Walker M, Szafir DA, Szafir D (2017) Designing for depth perceptions in augmented
reality. In: 2017 IEEE international symposium on mixed and augmented reality (ISMAR), pp
111–122
8. Azuma RT (1997) A survey of augmented reality. Presence: Teleoper Virt Environ 6(4):355–
385
9. Lynch K (1960) The image of the city. MIT Press, p 11
10. Ieiri Y, Mizukami T, Nakajima Y, Ayaki R, Hishiyama R. (2017) Effect of first impressions in
tourism by using walk rally application. In: Proceedings of the 2017 international conference
on culture and computing (culture and computing), pp 1–6
Graphene-Based Wireless Power
Transmission: Charge as You Drive
P. C. Nissimagoudar, H. M. Gireesha, R. M. Shet, Nalini C. Iyer,

Ajit Bijapur, and H. R. Aishwarya
Abstract Many automotive industries aim to limit their impact on the environment
and transform the mobility of automobiles into a sustainable mode of transporta-
tion. Electric vehicles seem to be key milestones in this regard. This paper intends to
introduce a “Charge as you drive”—a newly evolved concept that can be used to over-
come limitations of existing electric vehicles and aid in better battery management.
Wireless power involving a high-efficiency charging system that uses magnetic reso-
nance coupling to generate electricity could revolutionize the highway transportation
system. With this strategy, when using a fairly small battery pack, a vehicle effec-
tively has a limitless electrical range, which can significantly reduce the consumption
of oil. These coils are powered by graphene-based solar panels. As solar power is
renewable, infinite, eco-friendly and also considered to be a form of energy that
is much safer to use. Thus, by introducing the above-proposed solution in electric
vehicles, the obstacles of charging range, time, and cost can be easily mitigated.
Keywords Magnetic resonance coupling · Graphene-based solar panel · Wireless

power transmission
1 Introduction
According to the environmental protection agency, 75% of carbon monoxide comes

from automobiles. These automotive vehicles emit a wide variety of gases and solid
matter, causing changes in air and global climate. Burning large quantities of fossil
fuels, such as gasoline and diesel used in motor cars, have caused global temperatures
to increase by 0.6 °C or 1 °F. The pollutant emissions from these vehicles contribute
to unhealthy, immediate, and long-term effects on the environment as well as human
health. This has its effect on wildlife, sea levels, and natural landscapes as well.
P. C. Nissimagoudar (B) · H. M. Gireesha · R. M. Shet · N. C. Iyer · A. Bijapur · H. R. Aishwarya

School of Electronics and Communication Engineering, KLE Technological University, Hubli,
Karnataka 580031, India
e-mail: pcnissimagoudar@gmail.com
816 P. C. Nissimagoudar et al.
With the world becoming more eco-conscious, automobile industries aim at

limiting their impact on the fossil fuel-driven internal combustion engines, by intro-
ducing the vehicles powered by electric motor. Electric vehicles recognized as a
future of mobility is an attempt to help reduce gas emissions and improve air quality.
They offer a key path achieving in greater than 80% reduction in global warming, oil
dependency, and smog formation. More the electric vehicles on the road, the lower
the production of harmful gases and hence lower the pollution rate.
Unlike the conventional fuel-driven automobiles, electric vehicles can be powered
using batteries which acts like a “Gas tank.” These vehicles are fitted with onboard
batteries which stores and uses the energy needed to power a set of electric motors,
which ultimately gives the car the acceleration. The batteries used powers all the
electronic devices in the vehicle. It can be either recycled or used elsewhere after
a battery has worn out, as it will still have nearly 70–80% of its original energy
capacity. The controller controls the amount of power the batteries get, so that the
engine does not burnout [1, 2]. Because electric vehicles use an electric motor, the
driver can take advantage of the momentum of the motor when applying pressure on
the brakes. The forward momentum of the motor can be used to recharge the battery
as well. When the vehicle is stationary, there is no electrical current being processed,
so no energy is being used up. One of the electric vehicle’s main feature is that they
can be plugged into off-board charging power sources. These vehicles are powered
from the electricity by regenerative braking, as well.
However, charging of electric vehicle battery through charger and wire is incon-
venient, hazardous, and expensive. Although, these plug-in electric vehicles are
burdened by the need for cable and plug charger, with this system user has to phys-
ically plug-in for charging purposes. In addition to these, batteries seem to be the
limiting factors as recharging these batteries takes time. On a single charge, these
vehicles can travel about only 100 miles on an average [3, 4].
In this paper, we propose an innovative idea of wireless power transfer using the
principle of magnetic resonance coupling to overcome the limitations of existing elec-
tric vehicles. Dynamic charging will allow the transportation sector to supply a very
large fraction of the energy and greatly reduce petroleum consumption. Transmitting
coils embedded in the highway are powered by graphene-based solar panels. Solar
cells require materials that are conductive and allow light to get through; graphene is
used here due to its superb conductivity and optical transparency. A layer of graphene
has an optical transparency of 97.7%.
The proposed model would be onetime investment, revolutionizing the highway
transportation system. This is a wonderful idea which can be implemented, or the
government can undertake this as a huge project and completely change the Indian
approach toward electrification for transportation
The organization of the rest of the paper is as follows. Section 1.1 describes
the basic principle of magnetic resonance coupling, Sect. 2 explains the proposed
methodology, and experimental results for the proposed model are briefed in Sect. 3,
and conclusions in Sect. 4.
Graphene-Based Wireless Power Transmission: Charge as You Drive 817
1.1 Magnetic Resonance Coupling
The inductively coupled power system consists of two coils called L 1 and L 2 , respec-
tively, the transmitter coil and the receiver coil, all of which come from a magnetically
coupled inductor system. A magnetic field is created in the receiver coil by an alter-
nating current in the transmitter coil that induces a voltage. For charging purposes,
this voltage can be used [5].
The efficiency of the power transfer depends on the relation between the inductors
and their quality. The link between the inducers (z) and D2 /D ratios is determined.
Also defining the relation between the structure and the angle of the coils.
2 Methodology
Electrification for transport was a major concern for energy, resources, the environ-
ment, and many other factors. Wireless power transfer technology plays an important
role in this connection by eliminating all the problem with charging. A key difference
is that a transformer/receiver is replaced by loosely coupled coil systems between
a wireless power transmission and conventional gas-fueled vehicles. The series of
coils connected to an electrical current would be embedded onto the highway, and
the receiving coils would be attached to the vehicle bottom. On the road would
be built up the series of coils connected to an electric current, and the receiving
coils would be attached to the base of the vehicle. This echoes with the pace of the
vehicle. Power can be transferred to the battery without significant energy losses,
if the operating frequency of the battery is between 50 and 95% of the frequency.
Figure 1 shows magnetic fields that continuously transmit the electricity to charge
the battery. The solar panels are powered by graphene-based transmission coils. In
order for the transmission coil to be powered by a primary compensation network,
the DC power produced from solar panel is converted to high-frequency graphene,
the thinnest material known to people with good conducting, optical clarity, and very
Fig. 1 Magnetically
coupled inductor
high strength, and is used to improve the efficiency of sun cells. These solar graphic
cells also can be used to extract electricity from raindrops. Unfair weather like rain
and snow therefore has no effect on the load capacity [6] (Fig. 2).
Figure 3 shows the vertical magnetic flux of the transmitter and receiver coil. From
this magnetic flux path, where loops are generated and the receiving coil receives
the vertical magnetic flux that is fitted at the bottom of the electric vehicle, a high-
frequency power source is distributed through transmitting coil. The transmit and
receive coil are 60 cm wide. The transmitting coil and the surface of the road have
an air gap of 20 cm between them. To transfer 100 kw of power, five 20 kw pickup
modular is to be installed in the electric vehicles.
Fig. 2 Wireless power transmission
Fig. 3 Design of power transmitter

2.1 Design Process
The magnetic induction coupling compresses of two stages which are as discussed
below.
2.1.1 Wireless Power Transmission Between the Coils
A resonant wireless transmission network comprises magnetically coupled transmit-

ting, receiving coils, and various electronic control circuits. The resistance of the coil
essentially increases as the frequency changes. It is therefore important to design a
low loss belt that helps improve the quality factor. The constant current source shown
in figure shows the wireless power transfer system. (a) The simple passive elements
shown in Fig. 4. (b) The transmitting and receiving coil inductances, where the LS
and the LR are, may be designed. When a transmitter coil is driven by the source
(M is a mutual inductance and k a magnetic coupling coefficient), the transmitting
and receiving coils shall be magnetically coupled. For maximum power transference,
the condensers CS and CR are connected in series with each coil in the model. The
Fig. 4 a WPT system configuration. b The analogous circuit model, load transfer power and
efficiency of the actual transmitted power; the coil resonance frequencies are set at 20 kHz with the
transmission power being 50 W at the resonant frequency
internal resistances of each coil are RS and RR, and the load resistant is RL. The
model can be used to calculate the loop impedance of the ZS transmitter and the ZR
receiver loop impedance
1
Z s = Rs + + jωL s (1)
jωCs
1
Z R = R R + RL + + jωL R (2)
jωC R
where the source current is IS and IR, respectively, the load current. The PL
transmitted power can be calculated with the real-power transfer efficiency η
ω2 M 2 . R L . Is . Is∗
PL = Re VL . I R∗ = (3)
Z R . Z ∗R

Re VL . I R∗
n= (4)
Re Vs . Is∗
ω2 M 2 . R L
n= (5)
Re Z s Z R Z ∗R + ω2 M 2 . Z ∗R
Figure 4b demonstrates both the efficiency of power transfer and the transmission
of power from the illustration case. The power transfer efficiency is dependent on
the parameters of the passive circuit and operating frequency, and the performance
of the power transmission is typically maximized at the resonant frequency.
2.1.2 Generation of Power from Graphene-Based Solar Panels
The solar panel, based on graphene, converts sunlight into DC electricity, acting as
a source for powering the connecting coils. Solar panels may be wired in series or
in parallel, respectively, to increase voltage or current. A 12 V solar panel’s rated
terminal voltage is typically around 17.0 V, but this voltage can be reduced to around
13–15 V as required for charging by the use of a solar regulator. Solar regulators or
charge controllers have the purpose of regulating the current from the solar panels
and ensuring the correct supply of the power to the inverter without damaging it. In
order to drive the transmitter through an inverter network, the dc voltage is converted
to a high-frequency alternative current. Reflected sunlight and different conditions
of temperature raise the output current by 25% above a solar panel’s average output.
The solar regulator must be sized in such a way that 125% of the rated short-circuit
current of the solar panel can be handled. The solar panel based on graphene can also
be used to obtain power from the impact of the raindrops. The gout is not pure water.
They contain salts which dissociate themselves into ions. The positively charged ions
will bind to the surface of graphene and get enriched in delocalized electrons. The
Fig. 5 Transmitter circuit
consequence of this process is a double layer of electrons and positively charged

ions. This difference in associated potential is sufficient to generate a voltage and the
current needed to power coupling coils. So solar panels will operate in all climatic
conditions with the aid of graphene [7].
Transmitter circuit
The module of the transmitter is a direct power supply connected to an oscillator a
coil. The transmitting coil generates an oscillating magnetic field from the AC current
during this operation, which transforms the electrical supply into a high-frequency
alternating current (Fig. 5).
Receiver circuits
This receiving module consists of a voltage regulator to provide a constant voltage
and connecting voltage to the battery mounted in an electric vehicle [7]. It receives
all its power in the magnetic field produced by the transmission coil of the receiver
version of a wireless power transfer device (Fig. 6).
3 Experimental Results
From Table 1 analysis, we get to know that the transformer output will give 11 V
but expected is 12 V same way rectifier will give 11.5 V, and the oscillator will give
10 V.
Fig. 6 Receiver circuit
Table 1 Output voltage

Parameter Voltage type Theoretical Actual value
measurement of the
value (V) (V)
transmitter module
Transformer AC 12 11
output
Rectifier DC 11.9 11
output
Oscillator AC 11.9 11.5
output
3.1 Critical Pointing
Coupling between two coils increases with decreasing distances between two
resonators this trench is roughly proportional to increases in device efficiency with
shorter distances (Fig. 7).
Fig. 7 Matlab distance simulation versus frequency analysis and spiral to spiral
Table 2 Transfer efficiency

Distance (cm) Transfer efficiency (%)
test result
0 96
6 80
11 75
16 68
21 55
26 48
Power transfer system the full wireless power transfer system consists of two
components: the transmitter (Tx) and the receiver (Rx). To optimize the transmission
capacity, choose identical resonators for both transmitter and receiver. The wireless
power transfer device is here developed as a linear series.
From Table 2, we can get to know that as distance increases and frequency
decreases and transfer efficiency decreases.
4 Conclusion
Transportation based on fuel has caused enormous pollution problems. In the present
case, the use of electric vehicles offers an eco-friendly alternative but suffers from
a major downside when it comes to charging. So our proposed model—“Ride as
U-Drive” is used to provide an beneficial alternative to wireless vehicle charging.
The coupling coils are powered using graphene-based solar cells where graphene
increases the solar cell’s performance. The analysis of radiations produced by the
coupling coils does not affect the human health nor vehicles performance.
For a wireless charging electric vehicle, a safe zone should always be established.
When an automobile is usually made of steel, a very strong defensive material, the
magnetic flow density must be assured to comply with the safety requirements at
usual positions such as standing outside the car or inside the vehicle.
References
1. W. Khan-ngern, H. Zenkner, Wireless charging system for electric vehicle. Date of Conference
3–5 Oct 2016
2. M. Shidujaman, H. Samani, M. Arif, Wireless power transmission trends. Date of Conference
(2014)
3. R. Pudur, V. Hanumante, S. Shukla, K. Kumar, Wireless power transmission: a survey. Date of
Conference 9–11 May 2014
4. Z. Zhang, H. Pang, A. Georgiadis, C. Cecati, Wireless power transmission—overview. Date of
Conference 10 May 2018
5. A. Kurs, A. Karalis, R. Moffatt, J.D. Joannopoulos, P. Fisher, M. Soljacic, Wireless power
transfer via strongly coupled magnetic resonances. Science 317, 83–86 (2007)
6. S.K. Chaganty, S. Siva Prasad, Contactless vehicle charging using magnetic resonance coupling
method. Date of Conference, 20–23 Aug 2011
7. A. Karalis, J.D. Joannopoulos, M. Soljac, Efficient wireless nonradiative mid-range energy
transfer. Ann. Phys. 323, 34–38 (2008)
Vision-Based Driver Authentication
and Alertness Detection Using HOG
Feature Descriptor
P. C. Nissimagoudar, A. V. Nandi, H. M. Gireesha, R. M. Shet,

and Nalini C. Iyer
Abstract Driver’s drowsiness is one of the main causes of the car collision, which
raise fatality and wound count worldwide. This makes it necessary to develop a
system that ensures the safety of the driver as well as co-passengers. Further, driver
authentication plays an important role in preventing car robberies and fraudulent
switching of designated drivers, thus ensuring the security of the vehicle and passen-
gers. In this work, we propose an algorithm for real-time and continuous driver
authentication and drowsiness detection. Driver authentication is implemented using
face recognition with dlib’s face detector which is found to be robust compared to
openCV face detector. The SMS regarding authentication of the driver is sent to
the vehicle owner with details so that he can keep track of drivers. Further, if the
driver is authenticated, then he is monitored continuously using the webcam to detect
the early signs of drowsiness. Behavioral measures of the driver-like eyelid move-
ments and yawning are considered for the detection. First, for face detection, we
will apply facial landmark detection made using Histogram of Oriented Gradients
(HOG) feature and then extract the eye and mouth regions using shape predictor with
68 salient points next to the eye aspect ratio and mouth aspect ratio, i.e., EAR and
MAR, respectively, are computed in order to determine if these ratios indicate that
the driver is drowsy, if so then he is alerted using speech message.
Keywords Driver authentication · Driver drowsiness · Histogram of Oriented

Gradients (HOG) · Behavioral measures
1 Introduction
The driver’s status monitoring system includes the driver’s drowsiness and authenti-
cation of a driver to access the vehicle. In this work, we will discuss how to provide
better security to a vehicle and to detect fatigue and drowsiness of the driver to avoid
P. C. Nissimagoudar (B) · A. V. Nandi · H. M. Gireesha · R. M. Shet · N. C. Iyer

KLE Technological University, Hubli, Karnataka 580031, India
e-mail: pcngoudar@gmail.com
accidents. One of the major reasons for the accident is the driver’s not paying atten-
tion to the road. As of September 9, 2018, Global Road Safety statics showed that
1.47 lakh deaths caused due to road accidents in India, and annually, road accidents
claim more than 1 lakh lives every year. The majority of road accident is due to the
driver’s fault, this includes the driver’s distraction, over speeding, drunken driving.
Road deaths due to the accident wipe out a city per year that is over 1.3 lakh per
year. 60% of deaths are of young people in age of 18–35. Sadly, India loses 3% of
its GDP due to road accidents. The vigilance of driver degrades due to lack of sleep,
long continuous driving, or any other illness like brain disorders, etc. According to
Central Road Research Institute (CRRI) study, 40% of critical accidents are caused
due to fatigue.
Driver drowsiness is a major factor in increasing road accidents. There is a require-
ment for a system to monitor the driver’s actions and reactions, check for fatigue,
and on detection of fatigue or drowsiness. Addressing this issue requires monitor
of driver’s alertness using non-invasive techniques using emerging technologies like
computer vision [1], machine learning [2], and deep learning [3–5]. As this problem is
claiming lives and losses to individuals and sometimes public property, it increased
our interest to design a cost-effective warning system with authentication which
enhances the safety of drivers, thus saves lives. To build an application that monitors
the driver non-intrusively. The main objectives of the contribution are
• To provide the security by authentication of driver and safety by the drowsiness
detection system.
• Design a system for authentication using face recognition techniques and
drowsiness detection using the behavioral measures of the driver.
• To make the system work, in despite of driver in glasses and in different
illuminations.
1.1 Related Work
Many works have been done to implement different algorithms and techniques for
driver authentication and drowsiness detection, some of them are briefed below.
Hajinoroozi et al. [6], this work is the investigation of deep learning algorithms
for the identification of driver’s intellectual states (drowsy and alert) using EEG data.
The proposed CCNN consists of many convolution layers as well fully connected
layer but it uses a 1D filter for convolution along the channels. Back-propagation is
the method used for training CCNN weights.
Derman et al. [7], in this paper, they have proposed an approach based on deep
neural networks for real-time and continuous authentication of vehicle drivers. For
testing in realistic conditions, from 52 different subjects, they have collected 130
in car driving videos. For this scenario, they have first collected facial images of
authorized drivers, trained a classifier based on those features, and have performed
the real-time verification. For feature extraction, they have applied transfer learning
Vision-Based Driver Authentication and Alertness Detection … 827
and fine-tuning on a pre-trained deep neural network model, and for the classification
purpose, the SVM and cosine similarity-based approaches were examined.
Yashwanth et al. [8], this work proposes a framework for the detection of drowsi-
ness and the rate of drowsiness of the driver. For the prediction of a driver as drowsy
and alert, they have concluded the data set of nine subjects that Neural Network Clas-
sifiers (NNC) and decision tree have given better results compared to linear SVM
and LDA. So, attributes such as eye blinks, eye closure are used for making decisions
regarding drowsiness.
The organization of the paper is as follows, Sect. 2 discusses the system design
which provides detailed design. In the next Sect. 3 comprises of implementation
details which describe the structural information of the design. In Sect. 4, the results
and discussion section, we discuss the output and results of the proposed algorithm for
driver authentication and drowsiness detection algorithm. In Sect. 5, the conclusion
and future scope provides the information about the discussion on the outcome of
our work, and further, the improvements and the advancements concerning our work
are discussed in the future scope section.
2 System Design
The proposed solution uses a behavioral measure based on the computer vision
approach. This measure is more flexible to drivers and convenient for the imple-
mentation. In this approach, drowsy behavior is identified by a driver’s behavior like
eye closure, yawning, and eye blinking. It may not work properly in low-lighting
conditions; to overcome this limitation, we can make use of infrared camera or light.
Figure 1 represents the block diagram of our proposed system. It provides a high-
level view of the system representing components involved and their relationship.
Input component, the camera which is focused on driver captures video, then trans-
mits it to a processor for processing to interpret the driver’s information sends the
owner with driver access via SMS, and alerts the driver through voice output in case
of asleep. The Dlib library [9], which has a pre-trained facial detector. Since facial
Fig. 1 Block diagram of vision-based driver authentication and alertness detection system
detection is a kind of object detection, it detects 68 landmarks on the face to identify

an object as a face. As soon driver ignites car, the camera takes the input video stream
a Histogram of Oriented Gradients (HOG)-based detector in Dlib finds the location
of face and encodes it into 128 numerical facial features. Then, it compares with the
encoded database, the result of the comparison is sent to the owner via SMS alert
regarding car usage and driver details, avoiding the fraudulent usage of the car.
3 Implementation Details
In this section, we discuss the implementation details of the proposed solution.

Thereby, meeting the aforesaid objectives specified. Implementation details provide
the information of actual flow and things utilized in the work.
3.1 Algorithm
An algorithm is set of rules or the process that must be followed in the manipulations
or problem-solving operations. The problem can be solved step by step by making
use of an algorithm. Technically, machines use algorithms to carry out an operation
by listing the detailed instructions. The algorithm for our work is shown below:
1. Initialization of camera for an input video stream.

2. Reading the frame from the video stream.
3. Resizing and conversion of RGB image to gray-scale image
4. Detection of the frontal face from the captured frame using HOG feature
descriptor which is trained using SVM.
5. If the face is not detected or there are too many faces in the captured frame and
if the attempt counter is less than six go to step 2.
6. Encoding the content of the captured frame.
7. Comparison of the encoded image with an array of encodings created in step 4.
8. Send the details along with authentication status to the vehicle owner, if the
person is authorized else send the notification to vehicle owner.
9. Initialization of EAR and MAR thresholds
10. Read next frame for detection of drowsiness
11. Detect the frontal face from the frame using the HOG feature descriptor.
12. If the face is not detected, then alert the driver and go back to step 10.
13. Extraction of facial features using shape predictors such as Eye Aspect Ratio
(EAR).
Mouth Aspect Ratio (MAR).
14. If the EAR is less than the threshold or MAR is greater than the threshold,
consecutively for the defined number of frames, then the person is considered
as drowsy. So alert the driver. Display “Drowsiness Alert!!!” on the frame.
15. Go back to step 10.

16. The drowsiness detection process stops when the user turns off the vehicle.
3.2 Mathematical Model
Shape Predictor After detecting a face in the image, the basic idea is to find 68
facial landmarks including inner and outer edge of eyebrows, points of nose, lips,
and eyes. The further center position of the face is obtained through manipulations
like rotation, increase as shown in Fig. 2b.
128D Vector Conversion (Encoding) In our algorithm of face recognition, we are

encoding the contents of image into 128-dimensional vector that quantifies face
in an image. Based on this encoding, we can measure similarities between two face
images for further classification. According to research, it is found that encoding into
a 128D vector gives a higher accuracy compared to other dimensions. Hence, from
this analysis, we choose a 128-dimensional byte vector for compactly representing
each face, and it is optimal for applications like recognition and clustering [8].
Eye Aspect Ratio (EAR) EAR is computed using Eq. (1). The numerator of this
equation calculates the distance amid the vertical eye landmarks while the denomi-
nator calculates the distance amid horizontal eye landmarks, weighting the denomi-
nator appropriately since there is only one group of horizontal points but two groups
of vertical points.
P2 − P6 + P3 − P5
EAR = (1)
2P1 − P4
where P1 , P2 , P3 , P4 , P5 , P6 are 2D facial landmark locations of eye, which is shown

in Fig. 2a.
Eye blink count eye blink rate is calculated by taking average of eye aspect ratios
of both eye as shown below:
Fig. 2 a 2D facial landmark locations of eye. b 68 points of reference to identify object as face.
c 2D facial landmark locations of the mouth
EARleft eye + EARright eye

EYEblink detect = (2)
2
Mouth Aspect Ratio (MAR) MAR is computed as shown below. The numerator
of this equation calculates the distance amid the vertical mouth landmarks while the
denominator calculates the distance amid horizontal mouth landmarks, weighting
the denominator appropriately since there is only one group of horizontal points but
two groups of vertical points.
P2 − P8 + P3 − P7 + P4 − P6

MAR = (3)
2P1 − P5
where P1 , P2 , P3 , P4 , …, P8 are 2D facial landmark locations of the mouth, which

is shown in Fig. 2c.
Euclidean Distance The distance between the vertical landmarks and the horizontal
landmarks of facial features considered for the calculation of EAR and MAR ratios
is the Euclidean distance which is the ordinary straight-line distance between two
points. Suppose, if we consider two points x = (x1, x2, …, x n ) and y = (y1 , y2 , …,
yn ) in the Euclidean space, then Euclidean distance between x and y is given by,

d(x, y) = d(y, x) = (y1 − x1 )2 + (y2 − x2 )2 + . . . + (yn − xn )2

n

= (yn − xn )2 (4)
i=1
where, d(x, y) = d(y, x) = Euclidean distance between 2 points x and y. (x 1 , x 2 , x 3 ,

…, x n ) = Coordinates of the point x. (y1 , y2 , y3 , …, yn ) = Coordinates of the point y.
3.3 System Specifications
Defining the system parameters with the right values plays a vital role in an accurate
execution of the software, which are shown in Table 1. After evaluating the different
test conditions, we have arrived at the mentioned thresholds. We have taken the
images with the resolution of 500 * 500 for processing, and the frame rate of the
video stream is 30 fps.
Table 1 System
S. no. Parameter Threshold
specifications
1 Eye aspect ratio 0.2
2 Mouth aspect ratio 0.58
3 Eye aspect ratio frame count 15
4 Mouth aspect ratio frame count 5
5 Resolution of processing image 500 * 5000 pixels
6 Video frame rate 30 fps
The proposed model for the driver authentication and drowsiness detection is tested
for various test conditions as explained in the below section,
Accuracy is one of the metrics for evaluating classification models. Initially, for the
accuracy analysis of authentication algorithm images of ten people were considered.
Among the predictions obtained, we found that nine were right and one was wrong.
Therefore, from this analysis, it is clear that the accuracy of the proposed model
for driver authentication is 90%. Next, for testing drowsiness detection algorithm,
images of 50 people containing different test cases were considered; among the
results obtained, it was found that 41 predictions were right and the other 9 were
wrong. Hence, the accuracy of the proposed drowsiness detection algorithm is 82%.
Finally, the overall accuracy of our proposed algorithm for the “driver authentication
and drowsiness detection” system can be expressed as the average of these two, i.e.,
85%.
4.1 Unauthorized and Authorized User Entry
The algorithm was tested by considering the unauthorized person as a driver. Five
attempts are made to check if the person is authorized; if his image is not present in
the data set provided, then he is considered as unauthorized. Considering one of the
people from the data set to test the proposed algorithm gives the following output,
which is shown in Fig. 3a, and the real-time captured image of the person is compared
with data set, and he will be considered as the authorized driver since the captured
image matches with one of the images of data set, which is shown in Fig. 3b, then
the algorithm proceeds further for the continuous real-time drowsiness detection of
driver.
Fig. 3 a Result of unauthorized entry. b Result of authorized entry
4.2 Drowsiness Detection
After the authentication of the driver is completed and he is found to be authorized

driver, then the real-time stream of video with the driver’s face is considered for
drowsiness detection. Here, we have considered two factors, i.e., EAR and MAR for
classification of the driver as drowsy or alert. We have considered a few test cases,
and they are described below:
Case 1—Normal Condition: The first case, where the driver is normal condition
with no signs of drowsiness; hence, there is no alert as shown in Fig. 4a.
Case 2—Person is wearing spectacles: There is the possibility that the driver
may wear spectacles, so it is mandatory to consider this case as well and the
proposed algorithm should work for such cases also. Here, is the case is shown
in Fig. 4b.
Fig. 4 Drowsiness not detected while person is a normal, b with spectacles, c low light condition,
d eyes closed, e yawning, f yawning and eyes closed, g obstacle in front of face, h driver distraction
Case 3—Driving in low light: As per our objective, the proposed algorithm
should work fine even if there is low light, e.g., while driving at night. So here
is the case is shown in Fig. 4c, where even the algorithm works fine with the
detection of eyes and facial features.
Case 4—Eyes closed condition: Eyelid movement or eye blinking rate is one of
the important behavioral measures which tells about the drowsiness of a human
being. EAR ratio varies as the eyelid movement changes, so if the ratio is found
to be lower than the predefined threshold continuously for the prescribed number
of frames, which means he has closed his eyes for a longer time; then the driver
is declared as drowsy as shown in Fig. 4d. A speech message is used to alert the
drowsy driver.
Case 5—The person is yawning: Yawning is considered as one of the significant
sign of drowsiness. So once the face recognition and prediction of mouth region
are done, we are computing the MAR ratio, and if it is found to be more than the
predefined threshold, then it indicates that a person is yawning, i.e., he is drowsy
as shown in Fig. 4e, and hence, he is alerted with a voice message.
Case 6—Eyes closed and person is yawning: When the driver is yawning as
well as his eyes are closed, the driver is considered to be drowsy. So in such cases
also the driver will get an alert from the alerting system as represented in Fig. 4f.
Case 7—If there is an obstacle in front of the face: The proposed framework is
robust enough to recognize the face and detect the eyes and mouth; from Fig. 4g,
we can see even when we have a hand in front of the face as an obstacle, eyes,
and mouth are detected, and the driver is alerted as he is drowsy since eyes are
closed as well the person is yawning.
Case 8—Driver distraction: Even though we have considered the cases of
drowsiness to alert the user, there may be a situation where the driver may be
distracted due to some other reasons. Hence, it’s necessary to consider the case
when the driver’s face is tilted which is a sign of losing concentration. There is a
need to alert the driver in such situations as shown in Fig. 4h. The defined algo-
rithm was tested for different test cases and driving conditions, and we found that
our objectives are achieved.
This work provides the software solution for the real-time authentication and drowsi-
ness detection of the driver based on face recognition and feature extraction (shape
predictor) algorithm. This work will help the vehicle owners to keep track of their
drivers, ensuring the security of vehicles as well aids drivers to stay alert while
driving, in turn, helps in reduction in car accidents due to drowsiness as well provide
transportation safety. The proposed framework provides information to owners about
driver authentication through SMS alert. Drowsiness detection is achieved by taking
into consideration the eyelid movement and yawning of the driver. Finally, we
conclude saying that the proposed algorithm for driver authentication and drowsiness
detection offers a non-invasive approach without any annoyance and interference for
low cost, continuous and real-time verification, and detection.
Although the proposed solution for authentication and drowsiness detection
provided accurate results, then there is scope for improvement. A more robust system
can be built which will not only verify the driver and alert the owner, but also will
not allow the driver to turn on the ignition if he is not an authorized person, and
further, algorithm can be improved to send an image of a person attempting to start
the vehicle to the vehicle owner.
References
1. F. Schroff, D. Kalenichenko, J. Philbin, Facet: a unified embedding for face recognition and
clustering, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(2015), pp. 815–823
2. P.C. Nissimagoudar, A.V. Nandi, Precision enhancement of driver assistant system using EEG
based driver consciousness analysis & classification, in Computational Network Application
Tools for Performance Management(2020), pp. 247–257. https://doi.org/10.1007/978-981-32-
9585-8_21
3. P.C. Nissimagoudar, A.V. Nandi, H.M. Gireesha, EEG signal analysis using wavelet transform
for driver status detection, in Innovations in Bio-inspired Computing and Applications, IBICA
2018. Advances in Intelligent Systems and Computing, vol. 939. (Springer, Cham, 2019). https://
doi.org/10.1007/978-3-030-16681-6_6
4. P.C. Nissimagoudar, A.V. Nandi, H.M. Gireesha, EEG based feature extraction and classification
for driver status detection, in Innovations in Bio-inspired Computing and Applications, IBICA
2018. Advances in Intelligent Systems and Computing, vol. 939 (Springer, Cham, 2019). https://
doi.org/10.1007/978-3-030-16681-6_15
5. A.M. Kulkarni, A.V. Nandi, P.C. Nissimagoudar, Driver state analysis for ADAS using EEG
signals, in 2019 2nd International Conference on Signal Processing and Communication
(ICSPC), Coimbatore, India (2019), pp. 26–30. https://doi.org/10.1109/icspc46172.2019.897
6799
6. M. Hajinoroozi, Z. Mao, Y. Huang, Prediction of driver’s drowsy and alert states from EEG
signals with deep learning, in 2015 IEEE 6th International Workshop on Computational
Advances in Multi-sensor Adaptive Processing (CAMSAP) (IEEE, 2015), pp. 493–496
7. E. Derman, A.A. Salah, Continuous real-time vehicle driver authentication using convolu-
tional neural network-based face recognition, in 2018 13th IEEE International Conference on
Automatic Face & Gesture Recognition (FG 2018) (IEEE, 2018), pp. 577–584
8. C. Yashwanth, J.S. Kirar, Driver’s drowsiness detection, in TENCON 2019–2019 IEEE Region
10 Conference (TENCON) (IEEE, 2019), pp. 1622–1625
9. Dlib Library. https://medium.com/@aanilkayy/extract-eyes-in-face-i%CC%87mages-with-
dlib-and-face-landmark-points-c45ef480c1
ICT4D in Channelling Welfare Services
in India: Case Study of Haqdarshak
Komal Ramdey and Hasnain Bokhari
Abstract In 2015 Digital India project was launched to improve governance by

employing information and communication technologies (ICTs). The ICT-led reform
espouses to transform India into a digitally empowered society where electronic
public services are accessible via web and mobile. Despite outstanding achievements
in the ICT sector there are over 200 million people that live below the poverty line in
India. In past decades several welfare schemes have been initiated to promote equi-
table development. However, due to the prevalent information divide, these schemes
suffer to reach those at the bottom of the pyramid. This paper attempts to contex-
tualise ICT4D in India within the sustainable development perspective. It tries to
examine the potential ICTs offer to the marginalised sections of India. The paper
presents a case study of social enterprise Haqdarshak that uses artificial intelligence
to channelise social welfare schemes.
Keywords Digital india · ICT4D · Haqdarshak · Sustainable development · AI
1 Introduction
In 2015 India launched its highly ambitious eGovernment programme titled “Digital
India” aimed at reforming the government processes by employing modern informa-
tion and communication technologies (ICTs) [1]. ICT-led reform promises to trans-
form India into a digitally empowered society thereby ensuring G2C services via
digital communication channels. Digital India programme by the Indian Ministry of
Electronics and Information Technology (MeitY) stands on nine pillars that pledges
to improve online infrastructure, Internet connectivity, government procedures as
well as access to mobile connectivity [2]. The foundations of Digital India can be
K. Ramdey (B)
Willy Brandt School of Public Policy, University of Erfurt, Erfurt, Germany
e-mail: komal.ramdey@uni-erfurt.de
H. Bokhari
University of Erfurt, Erfurt, Germany
e-mail: hasnain.bokhari@uni-erfurt.de
836 K. Ramdey and H. Bokhari
traced back to 1990 and 2000s when the groundwork for digitisation was interpo-
lated. A noteworthy regulation in this regard was the National Telecom Policy (1999)
that paved the way for structural reform in the telecom sector in India [3]. The policy
allowed companies to provide mobile services on a revenue share basis instead of
fixed fees [4]. This helped in triggering a radical transformation in the ICT sector
in India that led to low-cost internet bandwidth and affordable data plans. These
initiatives have resulted in telephone subscribers to reach a 1.2 billion landmark by
the end of 2019, with 514 million rural subscribers and 669 million in urban India.
Similarly, the internet subscribers in India stand at 637 million, with 410 million
residing in urban areas whereas 227 million come from rural parts of India [5]. The
statistics from 2019 reveal an exponential growth in telephone and internet services
that also assists in bridging the digital divide. Another crucial precursor to the initi-
ation of the Digital India programme is also the National eGovernance Plan (NeGP)
by the MeitY launched in May 2006 [6]. NeGP’s focus was to develop Mission
Mode Projects (MMPs) to establish data centres, wide area networks and service
delivery gateways. Amidst the euphoria about ICTs at the national level, the Indian
government continued to show its commitment towards ICTs at international fora.
MeitY represented Indian perspective at the UN ICT Task Force in 2001 where the
digital divide received a global recognition [7]. Similarly, MeitY remained instru-
mental in the World Summit on Information Society (WSIS) in Geneva (2003) and
Tunis (2005) [8]. MeitY’s presence at these platforms eventually provided India
with an opportunity to host the third UN Internet Governance Forum in Hyderabad
in 2008 [9]. Such a commitment to the ICT sector in and abroad has today resulted
in achievements like unique digital ID (Aadhar card) and/or digital financial solu-
tions to complement the challenge that Digital India programme has undertaken.
At the heart of the Digital India programme lies the development of citizen-centric
eServices. Two of the nine aspects of Digital India are decisive for the promise of
digital empowerment and these include, Information for All and electronic public
services (or e-Kranti) [10]. Both these pillars promise to be highly citizen-focused
and aim to reform bureaucratic procedures to provide eServices over multiple digital
channels. Nevertheless, despite these achievements in the ICT sector, the fact that
270 million people in India still live below the poverty line has been a matter of
concern [11]. Several national and international indicators highlight, for instance,
the challenge of access to proper education for children, adequate healthcare and the
job prospects for youth [12]. In the past seven decades, several attempts have been
made to address these issues through various economic and social welfare schemes.
Considering India’s enormous size and diversity, the inclusion of the wider popula-
tion in these welfare schemes has remained a major hurdle [13]. While the digital
infrastructure in India is going through its prime phase, it’s the social strata that
require a special emphasis particularly when it comes to bridging the information
divide. It remains to be seen if these apprehensions elucidate some of the pertinent
ICT4D challenges such as ‘can Digital India programme help to uplift the population
at the bottom of the pyramid?’ With low (digital) illiteracy rates, how far will Digital
India’s 6th pillar (information for all) be able to bridge the information divide? Here-
inafter, this paper while acknowledging the tremendous growth in the ICT sector,
ICT4D in Channelling Welfare Services in India: Case Study … 837
aims to look at the potential that ICT could offer to the marginalised and poor popu-
lation of India. In this regard, this paper explores the case of social welfare schemes
within the context of Sustainable Development Goals (SDGs). More importantly, this
paper discusses a case study of an Indian social enterprise “Haqdarshak” that intends
to bridge the information divide by employing digital solutions to bring the social
welfare schemes for the marginalised Indian population. This paper also explores
the role of SDG 17 i.e. Global Partnerships for Sustainable Development and how it
can help the objectives of the Digital India programme.
2 Contextualising ICT4D in Indian Context
The past two decades have seen tremendous growth in the ICT sector worldwide.
According to World Digital Competitiveness Ranking 2019, India advances four
places to 44th position with the biggest improvement in the technology sub-factor
level, holding the first position in telecom investment [14]. However, within the
ICT4D context, Indian performance has met a mixed response. In United Nations
eGovernment Survey, India’s position has fluctuated between 87 to 125 in the past
20 years [15]. Limited ICT infrastructure, payment gateways, digital divide and lack
of public policy remained the key concerns. The UN’s Millennium Development
Goals (MDGs) encouraged developing countries to employ ICTs as a parallel medium
of development [16]. Between 2000 and 2010 a number of countries started ICT-led
projects either at international motivation, at times initiated by donor organisations
and sometimes out of national enthusiasm around ICTs. The response remained
both celebratory as well as critical [17]. Celebratory because developing countries
found an opportunity to extend their modern ICTs and digital infrastructure from
urban to rural areas. Critical because several studies blamed the approach towards
ICT4D as too centralised, top-down and disregarded indigenous knowledge [18].
Thus, using ICTs effectively to serve citizens online has remained a struggle because
of the technological complexity, complicated organisational routines and technology
acceptance – a reason that several ICT4D projects initially partially failed [19]. Never-
theless, within the ICT4D context Indian presence has also been acknowledged at
the international scale. For instance, the government of Karnataka’s land registry
system Bhoomi serves as the best practice within the developing country context
[20]. Similarly, Gyandoot project by the State of Madhya Pradesh introduced the
concept of rural information kiosk to provide a variety of G2C services. Gyandoot’s
rural telecentre model has since then been replicated in several low-income countries
to bridge the gap between digital haves and have-nots [21]. Andhra Pradesh’s eSeva
project also opened doors to alternative thinking in eGovernment development [22].
Indian ICT4D initiatives challenged the web-only eService solutions and guided the
international debate towards multi channels electronic service delivery [23]. The
unique aspects of these projects are their focus on the realisation of local realities,
indigenous knowledge and existing communication patterns. Such initiatives have
also received critical acclaim from several academic quarters where Subhash Bhat-
nagar, M.P. Gupta and Rahul De stand as notable mentions. While stocktaking ICT4D
initiatives in India it is imperative, as the case of Gyandoot, eSeva or Bhoomi has
reminded us, to acknowledge the presence of unheard and focus on the voices from
within as we proceed to more citizen-centric electronic public services.
In 2015 UN MDG’s mandate came to an end and new set of targets for human
development have been set out in the form of Sustainable Development Goals
(SDGs). Incidentally, SDGs announcement also coincides with the commencement
of Digital India programme. Both programmes espouse to improve the service sector
to be accessible to the bottom of the pyramid. SDGs represent a holistic agenda
that entails achieving progress across social, economic and environmental pillars.
India with 17% of the world’s population, holds an important place in achieving
the SDGs [24]. India has raised the levels of prosperity for its people in the recent
past. Several programmes and welfare schemes have been launched at the state and
national level in past decades. While targeting economic growth, infrastructure devel-
opment and industrialisation, India’s fight against poverty has become fundamentally
focused on social inclusion and empowerment of the poor [25]. Even though with
such progressive ideas and social welfare schemes, the challenges remain in many
areas including health, nutrition, basic infrastructure, quality of education among
others. Currently, India faces a bigger challenge of information divide along with
digital imbalance. The accessibility of welfare schemes to a wider public remains an
issue [26]. Delivery of social welfare can accelerate the progress to achieve SDGs by
relying on ICTs and eGovernment programmes. SDG 9 (Building resilient infrastruc-
ture) and SDG 17 (Partnerships for the Goals) present a unique opportunity in India to
utilise the presence of civil society organisations and technology firms. They can not
only complement the objectives of SDGs and the Digital India programme but also
India’s commitment to the welfare of its low-income and marginalised groups. The
following section explores the case of Pune based social enterprise called Haqdar-
shak where social innovation meets digital entrepreneurship and looks at such an
ICT4D initiative how it can help meet several SDGs.
3 Case Study: Collaborative Partnerships for Development

“Haqdarshak Empowerment Solutions”
Since 1951, the Indian government struggled hard to promote equitable and economic
development with a focus on disadvantaged and vulnerable groups. For the past few
decades, central and state governments in India introduced social welfare schemes to
provide a safety net to marginalised sections of society. Indian Ministry of Finance
reported over 900 welfare schemes promulgated by the central and state governments.
These schemes are built upon three tiers that provide cash-based, subsidy-oriented
and services-based welfare benefits [27]. The rapid growth of ICTs has motivated
Indian government to revamp its Public Distribution System (PDS) particularly now
that every citizen has a unique ID to rely more on technology-based Direct Benefit
Transfer (DBF). The welfare schemes, however, have been a matter of scrutiny by
media and civil society alike [28]. The critique goes, as also pointed out by High
Court of Kerala, that due to the information divide, many a time a majority at the
bottom of the pyramid has not even heard of the existence of welfare benefits [29].
In 2016, a social impact organisation named Haqdarshak (“Haq” means rights and
“Darshak” means “showing the path”) in Pune was established with an aim to pursue
an ICT-based collaborative effort to bring welfare schemes to people’s doorstep
[30]. The rationale behind Haqdarshak is that information is a fundamental right
and every citizen deserves to know his/her entitlements. Haqdarshak realised that
even though governments’ have launched many social welfare schemes, however,
their outreach still remains a big challenge. Poor publicization, complex paperwork,
illiteracy and above all information and digital divide are attributed as some of the
foremost factors that result in welfare budget to remain unused [31]. The lack of
awareness of welfare schemes inspired Haqdarshak to exploit the potential of social
and digital innovation to bring efficient collaboration between the government and
private enterprise. Haqdarshak launched an artificial intelligence-based mobile and
a web platform for citizen-service delivery platform that aims at dissemination of
information about social welfare schemes particularly those at the bottom of the
pyramid and helps them avail their entitlements.
3.1 Haqdarshak Model
Technology, collaboration and social entrepreneurship are at the heart of Haqdarshak

model. Haqdarshak realised that there is a lack of single repository (a sort of one-
stop-shop) for the welfare schemes in India. Haqdarshak started building a search-
able data bank for welfare schemes that could parse through citizens’ eligibility to
various scheme(s) and also explain its procedure. It follows a two-pronged approach
in which Haqdarshak trains community members in India’s rural hinterlands referred
to as village entrepreneurs (VE) [32]. The VEs are taught to use Haqdarshak app
in their local language. The VEs engage with their fellow community members by
going door to door to inform about the possibility of their entitlements. The VEs’
incentive is a possibility to earn a livelihood in due process. Haqdarshak model relies
on collaborations and its wishes to partner with civil society organisations, multina-
tionals as well state governments. VEs receive their monetary incentive through these
partner organisations or from beneficiaries for which they pay a monthly service fee
to Haqdarshak.
3.2 Technology and Impact
Haqdarshak’s research team regularly builds welfare scheme databank in a searchable

way to provide entitlement results for every Indian citizen. Haqdarshak’s mobile/web
platform relies on artificial intelligence by using the rule-mapping system to screen
citizens. Citizen is asked a number of questions based on personal, occupational and
financial status before showing the eligibility for welfare schemes. Eligible schemes
can be displayed in the local language to reduce the information gap between citizens
and their entitlements. Haqdarshak app does not depend upon internet connectivity
and can work as a stand-alone application but the app does have the possibility
to synchronise with Haqdarshak server once connected to the internet. The VEs
can use their own mobile phones or sometimes the ones provided by the partner
organisation to screen the citizens for their eligibility to various welfare schemes.
Since its beginning, Haqdarshak focussed on establishing partnerships and managed
to form collaborations with organisations such as Tata Trust, Save the Children as
well as Rajasthan government. Initially, its focus was limited to Maharashtra and
Rajasthan but now the Haqdarshak’s network has spread across 20 Indian states
where 6136 VEs have been trained. The initial results look promising as Haqdarshak
has been able to reach out 226,717 (0.23 million) citizens, out of which 111,277(0.11
million) were rural and 8329 were urban. The Haqdarshak app has logged in 173,647
(0.17 million) applications through which 142,646 (0.14million) people received
benefits from the the government’s welfare schemes through Haqdarshak, out of
which 43.51% were male, 56.45% female and 0.04% were from the third gender. In
monetary terms, Haqdarshak has been able to mobilise an amount of Rs. 587,241,806
(590 million) value of benefits that would have otherwise remained underutilised
[33]. Haqdarshak’s socio-technological solution not only assisted the government in
the right allocation of welfare money but it is also helping the Indian government
reach its global commitments towards SDGs. As such Haqdarshak’s impact areas are
aligned with six SDGs namely SDG 1, 3, 4, 5, 6 and 9 that range from gender equality
to quality education, good wellbeing to no poverty and clean water to innovation and
industry.
3.3 Analysis of Issues and Challenges
Haqdarshak is a social enterprise and its business model is for-profit, therefore, it

relies on funding which is one of its major issues faced. Currently their revenue
model primarily relies mainly on private funders. Similarly, several welfare schemes
were designed decades ago, and India has gone through major structural and fiscal
changes. Thus, Haqdarshak requires data that would involve working directly with
the community [34]. Similarly, a major challenge in partnerships domain is working
with the government for scale-up and sustainability, however, government process
tends to suffer from the red tape that hampers Haqdarshak’s prospects of quicker
collaborations. Haqdarshak has been able to adopt the SDGs through its work and is
able to create a strong impact in uplifting the communities that it has worked with.
Even though Haqdarshak has sturdily used the power of partnerships to maximise
the outreach which is also a contribution to SDG17 i.e. Partnerships for Sustainable
Development Goals, there is no detailed evidence of the methodology used in the
impact assessment. The data that represents the number of people who have benefitted
from the government welfare schemes and the value of the benefits through Haqdar-
shak, doesn’t signify the individual sectoral impact or individual welfare scheme’s
impact. Having said that, Haqdarshak has managed to socially mobilise people and
enable the environment to make the social welfare schemes more accessible to the
bottom of the pyramid through different partnerships and keeping people at the centre
of solutions. Haqdarshak has so far been able to take over 25 high profile partners
including Bill and Melinda Gates Foundation, UNDP, Aga Khan Foundation and
Tata Trust. In terms of monetary benefits, Haqdarshak’s compound an annual growth
rate of 114% increased from 2017 to 2019 with annual profit of Rs. 1.95 crore in
profit after taxes [30].
4 Conclusion
The case of Haqdarshak has shown as to how community informatics can be used to
bridge the information divide that can also help reach development goals. Its AI-based
solution targeted at the population at the bottom of the pyramid is a prime example
of the potential that ICTs offer in sustainable human development. Haqdarshak’s
model based on social innovation and collaboration also presents a good example
for the Indian government’s Digital India programme. It promises to promote citi-
zens’ access to ICT to encourage their participation. However, providing access to
these ICT services still seems very limited. Citizens must be enabled to use these
services for government-citizen interaction. In a country like India, where social
welfare schemes are very important to bridge the inequalities and where technology
is still not widespread, relying on ICT4D alone can result in the exclusion of many.
Organisations and CSOs with direct outreach to these communities have a signifi-
cant role to reach those in rural areas. As highlighted in SDG 17, Partnerships for
the Goals, there is a huge potential for exploring partnerships with organisations
like Haqdarshak and other CSOs, who have far-reaching impact in the communities.
They could not only empower these communities with information but also bring the
concerns and voices of the ones which have not been heard by the governments. As a
future recommendation, these partnerships can also open a door for further develop-
ments for making the welfare schemes accessible. Similarly, the concept of open data
can also help improve measurement of policies, better government efficiency, deeper
analytical insights, greater citizen participation, and a boost to the local companies
by way of products and services that use government data.
References
1. Digital India Homepage. https://www.digitalindia.gov.in. Last accessed 26 June 2020

2. Economic Times, Digital India: 15 salient things to know about PM Narendra
Modi’s project. https://economictimes.indiatimes.com/tech/internet/digital-india-15-salient-
things-to-know-about-pm-narendra-modis-project/articleshow/47893380.cms. Last accessed
26 June 2020
3. R. Prasad, The impact of policy and regulatory decisions on telecom growth in
India, https://kingcenter.stanford.edu/publications/impact-policy-and-regulatory-decisions-tel
ecom-growth-india. Last accessed 26 June 2020
4. Department of telecommunication, new telecom policy. https://dot.gov.in/new-telecom-policy-
1999. Last accessed 26 June 2020
5. Ministry of communications, telecom statistics India. https://dot.gov.in/sites/default/files/Tel
ecomStatisticsIndia-2019.pdf. Last accessed 26 June 2020
6. Ministry of Electronics & Information technology, information national e-Governance Plan,
ministry of electronics and information technology, Government of India. https://meity.gov.in/
divisions/national-e-governance-plan. Last accessed 26 June 2020
7. United Nations, ICT Task Force-Membership. https://www.un.org/webcast/ict/members.htm.
Last accessed 26 June 2020
8. International telecommunication union, world summit on the information society. http://www.
itu.int/net/wsis/index.html. Last accessed 26 June 2020
9. Internet Governance Forum, the IGF 2008 Meeting. https://www.intgovforum.org/multiling
ual/content/the-igf-2008-meeting. Last accessed 26 June 2020
10. Digital India, eKranti, Digital India Programme. https://digitalindia.gov.in/content/ekranti.
Last accessed 26 June 2020
11. World Bank, India’s Poverty Profile. https://www.worldbank.org/en/news/infographic/2016/
05/27/india-s-poverty-profile. Last accessed 26 June 2020
12. M. Gwozdz, India: a world of uncertainty—future directions international. http://www.future
directions.org.au/publication/india-world-uncertainty. Last accessed 26 June 2020
13. A. Antony, Almost half govt schemes fail to deliver the goods. https://economictimes.indiat
imes.com/news/economy/indicators/almost-half-govt-schemes-fail-to-deliver-the-goods/art
icleshow/3911028.cms?from=mdr. Last accessed 26 June 2020
14. IMD World competitiveness center, IMD World digital competitiveness ranking 2019. https://
www.imd.org/globalassets/wcc/docs/release-2019/digital/imd-world-digital-competitiveness-
rankings-2019.pdf. Last accessed 26 June 2020
15. United Nations department of social and economic affairs, e-Government surveys. https://pub
licadministration.un.org/en/research/un-e-government-surveys. Last accessed 26 June 2020
16. T. Unwin, ICT4D: Information and communication technology for development (Cambridge
University Press, Cambridge, 2009)
17. D. Kleine, T. Unwin, Technological revolution, evolution and new dependencies: what’s new
about ict4d? Third World Q. 30(5), 1045–1067 (2009)
18. D. Gichoya, Factors affecting the successful implementation of ICT projects in government.
Electr. J. e-Government 3(4), 175–184 (2005)
19. D. Dada, The failure of E-government in developing countries: a literature review. Electr. J.
Inf. Syst. Dev. Countries 26(1), 1–10 (2006)
20. P. Thomas, Bhoomi, Gyan Ganga, e-governance and the right to information: ICTs and
development in India. Telematics Inform. 26(1), 20–31 (2009)
21. World Bank, Gyandoot Project: ICT initiative in the district of Dhar, Madhya Pradesh. https://
documents.worldbank.org/en/publication/documents-reports/documentdetail/621851468267
614396/gyandoot-project-ict-initiative-in-the-district-of-dhar-madhya-pradesh. Last accessed
26 June 2020
22. Government of Andhra Pradesh, e-Seva, https://www.esevaonline.ap.gov.in. Last accessed 26
June 2020
23. S. Bhatnagar, Unlocking E-government potential: concepts, cases and practical insights, Sage
Publication, New Delhi (2009)
24. NITI Aayog, India Voluntary National review report on implementation of sustainable devel-
opment goals. https://sustainabledevelopment.un.org/content/documents/16693India.pdf. Last
accessed 26 June 2020
25. Ministry of Finance, Government of India, economic survey 2012–13. https://www.indiab
udget.gov.in/budget2013-2014/survey.asp. Last accessed 26 June 2020
26. D. Herald, Welfare scheme access challenge for urban poor. https://www.deccanherald.com/
city/top-bengaluru-stories/welfare-scheme-access-challenge-for-urban-poor-report-787468.
html. Last accessed 26 June 2020
27. Ministry of Finance, Government of India, Economic Survey 2017–18 (Volume I and Volume
II). http://mofapp.nic.in:8080/economicsurvey. Last accessed 26 June 2020
28. R. Bansal, India has social schemes for poor in crises like Covid. But it needs a ‘who
to pay’ database. https://theprint.in/opinion/india-needs-a-who-to-pay-database-covid-crisis/
406783. Last accessed 26 June 2020
29. The Hindu, Welfare schemes not reaching tribes: HC. https://www.thehindu.com/news/
national/kerala/welfare-schemes-not-reaching-tribespeople-hc/article22880604.ece. Last
accessed 26 June 2020
30. Haqdarshak—Every Citizen Matters. https://www.haqdarshak.com. Last accessed 26 June
2020
31. M. Karnik, Community-driven innovation—haqdarshak solution. https://solve.mit.edu/challe
nges/community-driven-innovation/solutions/10131. Last accessed 26 June 2020
32. N. Raj, Haqdarshak: helping communities understand their rights and entitlements
(2017). http://blog.impactpreneurs.com/seedsofchange/haqdarshak-helping-communities-und
erstand-rights-entitlements
33. Haqdarshak, impact report. https://haqdarshak.com/assets/reports/Impact_Report_2019_F
inal.pdf. Last accessed 26 June 2020
34. Yourstory, Using a mobile app, Haqdarshak is helping rural India benefit from govt welfare
schemes, 2017. [Accessed: 27-Jun-2020]
Author Index
A C
Aalam, Zunaid, 721 Castanha, Jick, 317
Aditya, S., 39 Chafe, Shreya Shailendra, 245
Adlakha, Sushant, 395 Chakravaram, Venkamaraju, 675
Adnan, Md Nasim, 473 Chandra, Theodore S., 279
Adu-Gyamfi, Samuel, 741 Chatterjee, Punyasha, 535
Aiholli, Nagaraj R., 595 Chauhan, Pranay, 771
Aishwarya, H. R., 815 Chauhan, Sudakar Singh, 29
Alam, Mohammad Minhazul, 779 Chin, Cheng Siong, 185
Chitrao, Pradnya Vishwas, 49, 209
Algarni, Sumaiah, 763
Chopra, Akanksha Bansal, 341
Ananth, V. Vignaraj, 39
Christian, Annan, 19
Apte, Tejaswini, 365
Ashish, Ravi, 439
D
Dalal, Vipul, 177
Deochake, Saurabh, 563
B
Deshmukh, Abhishek, 503
Badrinarayanan, M. K., 1 Deshmukh, Rashmi J., 621
Bandyopadhyay, Saubhik, 535 Dinesan, Abhishek, 629
Bangad, Divya Ashok, 245 Divekar, Rajiv, 49
Bansal, Divya, 553 Dixit, Veer Sain, 341
Bansal, Pratosh, 771 Domb, Menachem, 165
Bhagvandas, Ashni Manish, 629
Bhagwat, Alok, 209
Bhandari, Suchandra, 535 E
Bhatia, Sajal, 463 Edla, Damodar Reddy, 157
Bhaumik, Rahul, 649
Bhavsar, Bansari, 19
Bhoyar, Pravin Kumar, 49 G
Bijapur, Ajit, 815 Ganeshan, M., 383
Biswas, Arghya, 29 Geetha, V., 585
Bokhari, Hasnain, 835 Gemecha, Demissie Jobir, 791
Borwankar, Antara, 29 Ghatkamble, Rajlakshmi, 289
Buyya, Rajkumar, 661 Gireesha, H. M., 815, 825
and Computing 1270, https://doi.org/10.1007/978-981-15-8289-9
846 Author Index
H Kranthi kumar, V., 89

Hai, Do Huu, 707 Kuddus, Khushboo, 685
Hamid, Khaild, 639 Kumar, Amith Kiran, 279
Han, Shengnan, 297 Kumari, Kamlesh, 233
Harsh, 697 Kumar, Navin, 101
Hasan, Sadman, 779 Kumar, Tarun, 649
Hassan, Mohamed, 639
Hatture, Sanjeevakumar M., 573
Hegde, Vinayak, 417 L
Hiremath, P. S., 307 Li, Ping, 741
Hishiyama, Reiko, 805 Lobiyal, D. K., 269
Hossain, S. M. Shahadat, 779 Long, Nguyen Tien, 707
Huda, Eumna, 779
M
I Mahmud, Khan Raqib, 483
Ieiri, Yuya, 805 Mamud, Samiul, 535
Indrawati, 317 Manjrekar, Amrita A., 621
Iyengar, S. S., 661 Marathe, Nilesh, 503
Iyer, Nalini C., 815, 825 Medvedeva, L. N., 733
Meghana, H. N., 417
Mehnaz, Laiba, 357
J Mekruksavanich, Sakorn, 753
Jacob, Rohan Ninan, 11 Mishra, Satyasis, 791
Jadeja, Mahipal, 147 Mistry, Nilay, 19
Jain, Anshul, 223 Mohanty, R. C., 791
Jain, Ketan Kumar, 395 More, Abhilasha, 177
Jain, Nitesh, 223 Motaleb, Abdul, 779
Jaisingh, Kelavath, 77 Mukesh, Rajeswari, 1
Jayakumar, P., 447 Mukhopadhyay, Debajyoti, 563
Jayan, Abhijith, 447
Jegatha, M., 493
Jha, Ravi Shankar, 513 N
Jindal, Rajni, 357 Naik, Amrita, 157
Jitpattanakul, Anuchit, 753 Naik, Sapan, 63
John, Anice, 525 Nandi, A. V., 825
Nasir, Afsana, 473
Nazli, Nursyafiqah Binte, 185
K Nguyen, Duc Dinh, 123
Kabir, Moumita, 483 Nguyen, Long, 123
Kakulapati, V., 89 Nissimagoudar, P. C., 815, 825
Kannadasan, R., 439
Kaur, Amandeep, 431
Kaur, Amritpal, 721 O
Kaur, Saravjit, 721 Obiora, Sandra Chukwudumebi, 741
Kaur, Satnam, 721
Kaushik, Abhinesh, 269
Keerthiga, B., 39 P
Kesavan, S. Ragul, 39 Pai, Shashidhar, 453
Khan, Nafis Mahmud, 685 Pallavi, B., 661
Khara, Satvik, 113 Pallavi, M. S., 417
Khot, Shubhangi Tanaji, 621 Panchami, V., 77
Kofi Frimpong, Adasa Nkrumah, 741 Patil, Kanishk, 405
Kothari, Komal, 113 Patil, Prakashgoud, 307
Author Index 847
Patil, Ratna, 405 Shetgaonkar, Sima, 607

Patnaik, L. M., 661 Shet, R. M., 815, 825
Paul, Bijan, 473, 483 Shishir, Md. Romman Riyadh, 779
Pawar, Manjula K., 307 Singh, Alok, 607
Pente, Yash, 503 Singh, Kuldip, 791
Pillai, Subhash Kizhakanveatil Bhaskaran, Singh, Manwinder, 639
317 Singh, Tanya, 223
Pingili, Madhavi, 373 Singh Tomar, Rudra Pratap, 585
Pokala, Sai Surya Kiran, 77 Sinthia, Fatema Zaman, 473
Popy, Fatima Noor, 483 Solanki, Manishkumar R., 257
Prajapati, Himanshu, 113 Sonune, Harsha, 245
Prajapati, Sunny Prakash, 649 Spandana, R., 417
Prakash Raj, G., 543 Sriram, Padmamala, 629
Prem Kumar, S., 543 Srivastava, Sanjeev Kumar, 133
Purohit, Hinddeep, 199 Sujanani, Anish, 453
R T
Rachh, Rashmi, 595 Tamane, Sharvari, 405
Rahman, Rashedur M., 779 Trung, Ngo Sy, 707
Rajan, E. G., 373 Tsai, Hung-Ya, 805
Rajesh Khanna, M., 543 Tuan Van, Pham, 707
Rajeswari, P., 279
Ramanjinailu, R., 89
Ramdey, Komal, 835 V
Rana, Sanjeev, 233 Vaghela, Ravirajsinh, 199
Rashid, Mohammad Rifat Ahmmad, 473, Varma, Datla Anurag, 439
483 Vats, Prashant, 721
Rastogi, Shubhangi, 553 Venkadeshan, R., 493
Rathee, Davinder Singh, 791 Venkataraman, Reethesh, 629
Ratnakaram, Sunitha, 675 Venkata Sai Sandeep, V., 439
Roiss, O., 733 Venkatesh, B., 439
Roopa, M. S., 661 Venugopal, K. R., 661
Roy, Sohini, 329 Verma, Atul Kumar, 147
Verma, Vijay, 395
Vidyasagar Rao, G., 675
S Vignesh Raaj, N. S., 543
Sahoo, Priti Ranjan, 513 Vigneshwaran, P., 383
Saini, Gurdeep, 585 Vihari, Nitin Simha, 675
Saini, Jatinderkumar R., 365 Viswavardhan Reddy, K., 101
Sai Sandeep, R., 89
Sait, Unais, 649
Salvi, Sanket, 585 W
Samuel, Subi George, 447 Wali, Uday V., 595
Sapraz, Mohamed, 297 Warang, Kshitij, 503
Saxena, Rahul, 147 Wejinya, Gold, 463
Sen, Arunabha, 329
Sethi, Gunjan, 697
Shah, Aagam, 113 Y
Shah, Hinal, 63 Yadav, Naveen, 585
Shankar Lingam, M., 373 Yankati, Pallavi V., 573
Shankar, S. K., 1 Yaratapalli, Nitheesh Chandra, 629
Sharma, Aditya, 395 Yeasmin, Sabrina, 779
Sheik, Reshma, 525 Yedla, Sandeep Kumar, 77

ICT Systems and Sustainability: Milan Tuba Shyam Akashe Amit Joshi Editors

Uploaded by

Copyright:

Available Formats

You might also like

ICT Systems and Sustainability: Milan Tuba Shyam Akashe Amit Joshi Editors

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ICT Systems and Sustainability: Milan Tuba Shyam Akashe Amit Joshi Editors

Uploaded by

Copyright:

Available Formats

Advances in Intelligent Systems and Computing 1270

ICT Systems and

More information about this series at http://www.springer.com/series/11156

ISSN 2194-5357 ISSN 2194-5365 (electronic)

The Fifth International Conference on ICT for Sustainable Development (ICT4SD

Belgrade, Serbia Milan Tuba

Challenges in Adoption of Secure Digitization of Graduate

A Framework for Remote Health Monitoring . . . . . . . . . . . . . . . . . . . . 101

SOLID: A Web System to Restore the Control of Users’

An Experimental Approach Toward Type 2 Diabetes Diagnosis

Automated Dietary Monitoring System: A Novel Framework . . . . . . . . 535

Social Interaction-Enabled Industrial Internet of Things for Predictive

Modelling of Generation and Utilization Control for Isolated

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845

Milan Tuba is the Vice Rector for International Relations, Singidunum

Shyam Akashe is a Professor at ITM University, Gwalior, Madhya Pradesh, India.

omnipresent. He has authored two books entitled Moore’s Law Alive:

Amit Joshi is currently the Director of Global Knowledge Research Foundation,

S. K. Shankar, Rajeswari Mukesh, and M. K. Badrinarayanan

Abstract Blockchain technology finds variety of applications in everyday life. It

Keywords Blockchain · Assessments · Accreditation

S. K. Shankar (B) · R. Mukesh · M. K. Badrinarayanan

1.1 Blockchain Technology

1.2 Digitized Assessments

3.1 Perceived Benefits

3.2 Perceived Challenges

Fig. 1 Perceived benefits of digitized assessments

Fig. 2 Perceived challenges in digitized assessments

Fig. 3 Overall usefulness of digitized assessments

4 Multiple Regression Analysis

Multiple regression analysis is a powerful technique used to predict the value of

Table 2 Multiple regression analysis—significance statistics

Table 3 Multiple regression analysis—challenges

Table 5 Multiple regression analysis—significance benefits statistics

Table 6 Multiple regression analysis—benefits

as the significant variables in influencing the overall perceived usefulness of the

multiple choice questions, online role-plays, scenario-based activities, judged mathe-

4. M. Memon, U.A. Bajwa, A. Ikhlas, Y. Memon, S. Memon, M. Malani, Blockchain beyond

Rohan Ninan Jacob

Keywords NPA · Parameters · XGBoost · Gradient boosting · Feature importance

2 Algorithm and Program Flow

While gradient boosting has an approach of negative gradients, in order to optimize

Fig. 1 Accuracy equation

Fig. 2 AUC equation

Fig. 3 Program flow

Fig. 4 Training using XGBoost classifier

Fig. 5 Results obtained after selecting the best iteration

• With organization seeking help in need of open-source solutions, the program

1. Non-performing loan, https://en.wikipedia.org/wiki/Non-performing_loan. Last accessed on 4

3. XGBoost Algorithm: Long May She Reign, https://www.kdnuggets.com/2019/05/xgboost-alg

Nilay Mistry, Annan Christian, and Bansari Bhavsar

Keywords RAMDisk · Random access memory · RAM · Anti-forensics ·

N. Mistry (B) · A. Christian · B. Bhavsar

2 Different Software Available for RAMDisk

As discussed, a plethora of softwares are available for creating a RAMDisk. However,

2.1 AMD Radeon RAMDisk

Fig. 1 AMD Radeon RAMDisk driver