Professional Documents
Culture Documents
Joint Optimization of Offloading Utility and Privacy For Edge Computing Enabled Iot
Joint Optimization of Offloading Utility and Privacy For Edge Computing Enabled Iot
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 1
Abstract—Currently, edge computing, emerging as a burgeon- to cloud data center for supporting massive data analytics and
ing paradigm, is powerful to handle real-time resource provision application execution [3].
for the IoT applications. However, due to the spatial distribution However, with the continuous explosion and expansion of
of the geographically sparse IoT devices and the resource limita-
tions of the edge computing units (ECUs), the resource utilization IoT devices, a huge volume of raw data is generated [4].
of corresponding edge servers is relatively insufficient, and the According to a recent survey by Gartner, the total number
execution performance is ineffective to some extent. Meanwhile, of connected devices in 2020 will reach 20.8 billion [5].
privacy leakage including personal information, location and This proliferation of devices and associated data streams has
media data, etc., during the transmission process from IoTs put a huge burden on IoT infrastructures [6]. Consequently,
devices to edge servers severely restricts the application of ECUs
in IoT. To address these challenges, a two-phase offloading it is inadequate to process computation tasks efficiently for
optimization strategy is put forward for joint optimization of cloud computing on account of the ever-growing bandwidth
offloading utility and privacy in edge computing enabled IoT. pressure resulted from data transmission [7]. On the other
Technically, a utility-aware task offloading method, named UTO, side, the cloud is always placed in the center of the network
is devised first to obtain the goal of maximizing the resource and remote from the IoT devices, which causes high latency
utilization of ECUs and minimizing the implementation time cost.
Then a joint optimization method, named JOM, for utility and and transmission energy consumption [8]. In particular, cloud
privacy trade-offs is designed to balance the privacy preservation computing is unsuitable for the IoT applications with real-time
and execution performance. Eventually, experimental evaluations restrictions, e.g., data processing in the Internet of Vehicles.
are designed to illustrate the efficiency and reliability of UTO and Even if there are no stringent timing constraints for the IoT
JOM. applications, uploading all the colossal data to remote cloud
Index Terms—Privacy, Utility, IoT, Edge Computing, Offload- directly also induce vast liability on the Internet [9].
ing A promising philosophy that emerges recently, i.e., edge
computing (EC), is powerful to solve the current problems
I. I NTRODUCTION where resource provisioning are pushed from the cloud infras-
Internet of Things (IoT) is expected to efficiently improve tructure to the edge of the Internet [10]. In the EC philosophy,
the performance of software services through supporting mul- the smart IoT devices at the edge of the Internet are not
tiple intelligence fields like modern medical, digital home only the data producer but also service provider. Benefiting
and smart transportation [1]. Generally, embedded IoT devices from EC, an extensive scope of IoT devices resoundingly
are insufficient to hold the computation-intensive tasks, e.g., gain the requested resources in real-time by migrating their
tanglesome data mining and high definition video decoding. computation-intensive tasks to ECUs for execution, known as
To satisfy the computation and storage requests of the tasks computation offloading [11]. Computation offloading settles
generated by IoT devices, cloud computing has become the the problem of latency and bandwidth efficiently, and improve
most fashionable computation philosophy for dealing with the quality and availability of the services provided by luxu-
these assignments [2]. A convenient implementation manner rious IoT devices [12].
for the process of these tasks is transmitting all generated data Nevertheless, since the data generated by the IoT devices
becomes more and more complex, ECUs need to withstand
X. Xu, C. He and Z. Xu are with the School of Computer and Software, more pressure inevitably. At the same time, the disclosure of
Nanjing University of Information Science and Technology, Nanjing, China. sensitive information is a great calamity, particularly in the
X. Xu is with Jiangsu Engineering Center of Network Monitoring, Nanjing
University of Information Science and Technology, Nanjing, China. running of IoT [13]. During the transmission process from IoT
Email: njuxlxu@gmail.com,cxunhey@gmail.com,zhanyang xu@nuist.edu.cn devices to ECUs, a series of unexpected leakages of sensitive
L. Qi is with the School of Information Science and Engineering, Qufu data, e.g., user location, telephone number, and personal
Normal University, China.
E-mail: lianyongqi@gmail.com (corresponding author) picture, makes the computation offloading more sophisticated
S. Wan is with School of Information and Safety Engineering, Zhongnan [14]. Although the applying of EC intellectualizes the use of
University of Economics and Law, Wuhan, China. IoT devices, privacy information leakage still demands to be
E-mail: shaohua.wan@ieee.org
M. Z. A. Bhuiyan is with the Department of Computer and Information considered as a critical problem [15]. With these observations,
Sciences, Fordham University, New York, NY 10458 USA. it remains challenging to improve the implementation utility,
E-mail: zakirulalam@gmail.com including the resource utilization and time consumption of
Copyright (c) 20xx IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be ECUs, and achieve privacy preservation simultaneously. Given
obtained from the IEEE by sending a request to pubs-permissions@ieee.org. this challenge, a two-phase offloading optimization strategy is
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 2
• Conduct experimental evaluation and comparison analysis - Edge Server - Data Offloading - Edge Computing Unit
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 3
where gw (X) is used to mark the occupation status of cw , length, the amount of employed VMs, and the processing per-
defined by formance of each VM instances. The processing performance
of each VM is denoted as p and the instruction length for tn
1, cw is occupied,
gw (X) = (5) is denoted as mipsn, then the time cost of process is obtained
0, otherwise.
by
With the resource utilization of each ECU by (2) and the N
X mipsn
total occupied amount of ECUs by (4), the average resource Kexe (X) = . (12)
λn · p
utilization of all the ECUs is acquired by n=1
PW After task processing, the corresponding result data will be
uw (X)
U (X) = w=1 . (6) generated, which should be transmitted to the IoT devices and
γ(X)
the data center. The feedback time of the results for tn from
C. Time Cost Model of Tasks in IoT the edge server to the IoT devices also consists of three parts
which are similar to the task transmission time. Denote πn as
Time cost for implementing the tasks is the main concern for the data size of the feedback results, then the three parts (i.e.,
the IoT customers [17] [18]. With shorter time consumption, the data transmission time from edge server to AP, the data
comes better QoS. The time consumption mainly comprises transmission time among APs, and the data transmission time
four parts, i.e., the migration cost through APs, the process from AP to the IoT device) of feedback time are calculated
time in the ECU, the feedback time for returning the result by
data, and the offloading time from edge servers to the cloud πn
ZnEA (xn ) = , (13)
data center. ωA2E
The migration time for the tasks to the target ECUs is
divided into three parts, i.e., the transmission time from IoT πn · (σn (xn ) + 1)
ZnAP (xn ) = (14)
device to AP, the transmission time among APs, and the ωA2A
transmission time from AP to edge server.
and
The transmission time of tn from the IoT device to AP is πn · (σn (xn ) + 1)
defined by ZnAP (xn ) = . (15)
θn ωA2A
KnDA (xn ) = , (7) Then the total data feedback time is determined by
ωD2A
where θn is the image size of tn needs migration and ωD2A Znreturn (xn ) = ZnDA (xn ) + ZnAP (xn ) + Ztrans
AE
(xn ). (16)
is the data transmission rate between the IoT device and AP.
The task image should be transferred across multiple APs The execution results also demand to be transmitted to the
from the original edge server to the destination one. Suppose cloud for storage, therefore the time of data transmission from
the amount of the passing APs is denoted as σn (xn ), then the the edge server to the cloud platform is calculated by
transmission time for tn to pass through APs is calculated by πn πn
GCloud
n (xn ) = + . (17)
θn · (σn (xn ) + 1) ωA2E ωA2C
KnAP (xn ) = , (8)
ωA2A As the feedback process from the edge servers to the IoT
where ωD2A is the data transmission rate between two con- devices is concurrent with the uploading process from the edge
nected APs. servers to the cloud, the overall data offloading time for tn is
Then the task image should be transmitted from the desti- calculated by
nation AP to the goal edge server. The relevant transmission Knof f (xn ) = max{Znreturn (xn ), GCloud (xn )}. (18)
n
time for tn is calculated by
θn Correspondingly, the total data offloading time for all the
AE
Ktrans (xn ) = , (9) tasks is defined by
ωA2E
N
where ωD2A is the data transmission rate from the edge server X
to AP. K of f (X) = Knof f (xn ). (19)
From (7), (8) and (9), the total time cost of transmission for n=1
tn from the IoT device to the target edge server is calculated In summary, the total amount of time consumption is
by calculated by
Kntrans (xn ) = KnDA (xn ) + KnAP (xn ) + Ktrans
AE
(xn ). (10) K(X) = K trans (X) + K of f (X). (20)
Then the overall time cost of transmission is calculated by
N IV. A U TILITY-AWARE TASK O FFLOADING M ETHOD
X
trans
K (X) = Kntrans (xn ). (11) In this section, a NSGA-III based utility-aware task offload-
n=1 ing method for EC enabled IoT, named UTO, is devised first
When the task image is transferred, the task starts execution to obtain the goal of maximizing the resource utilization of
immediately. The processing time is decided by the instruction ECUs and minimizing the implementation time cost.
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 4
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 5
Algorithm 1 Generating candidate computation offloading Then, the correlation of tn,l and pn,l is demonstrated by
strategies by UTO
tn tn,1 tn,2 · · · tn,l
Require: R = . (26)
Pn pn,1 pn,2 · · · pn,l
Ensure: X
1: Obtain computing tasks from IoT devices The average implementation time of l-th type of privacy
2: for i = 1 to I do data is calculated by
3: r=1 N L
while r <= R do 1 X X θn,l
4: K 0 (t) = pn,l (27)
5: Crossover operations to generate the offspring N n=1 ω
l=1
6: Mutation operations to ensure the offspring
7: for Individuals in population do where θn,l represent the size of l-th type data in computing
8: Calculate resource utilization by (1) to (6) task tn and ω is the data process rate from IoT devices to
9: Calculate total time consumption by (7) to (20) ECUs,
10: end for Invite Y = {y(t1 ), y(t1 ), · · · , y(tn )} to represent the value
11: Environmental selection operation to ensure the of entropy for the computing tasks collection in T . The privacy
offspring entropy value of tn is calculated as
12: r=r+1 L
X
13: end while y(tn ) = − pn,l · log2 pn,l . (28)
14: Evaluate utility value of the individuals with SAW and l=1
MCDM Therefore, the correlation of the optimal strategy X and the
15: Select the optimal solution entropy value Y is demonstrated by
16: end for
17: return X X x1 x2 · · · xn
= . (29)
Y y(t1 ) y(t2 ) · · · y(tn )
Eventually, the summation of privacy entropy value of all
the computing tasks, which are waiting to be transmitted, are
transmitted data generated by the computing tasks placed in
divided into several parts. By this way, even if some parts
xn is obtained as
of the computing tasks are intercepted, the criminals could
obtain the whole information in the computing tasks. Entropy N
1 X
is actually a measure of disorder, with more types of data y(t) = y(tn ). (30)
N n=1
are cut, comes higher level of randomness. Higher the level
of randomness is, safer the computing computing tasks are.
However, the high entropy results in some problems such as B. Privacy Formulation
how to deploy the tasks in the same breath to improve the In this section, our purpose is to achieve the goal of
average resource utilization and transmission time. Therefore, maximizing the privacy entropy and minimizing the average
how to optimize the trade-off between the privacy entropy and transmission time of privacy data as possible as we can. The
the deployment strategy of computing tasks is resolved in this objective problem of formalization is defined as
section.
max(Y (t)), min(K 0 (t)), (31)
A. Privacy Entropy Model And the concomitant constraint is defined as
As described in section II , there are N IoT devices L
and computing tasks denoted as D and T respectively. The
X
s.t. pn,l = 1. (32)
relationship between the optimal strategy X and the placed l=1
computing tasks T is demonstrated by
X x1 x2 · · · xn VI. C OMPARISON AND A NALYSIS OF E XPERIMENTAL
= . (24)
T t1 t2 · · · tn RESULTS
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 6
TABLE I: Key notations of employed VMs in each ECU and the number of used
ECUs. Fig. 3 describes the average resource utilization based
Notations Description Notations value
The amount of IoT devices 100,200,300,400,500
on Benchmark, NF, FF and UTO in different scales of tasks.
The amount of VMs on each server 7 It is clear that UTO performs well while NF method performs
The transmission rate between APs 450 Mbit/s worst in all scales of tasks. It is concluded that UTO max-
The transmission rate between AP and edge server 500 Mbit/s
The transmission rate between AP and data Center 550 Mbit/s
imizes the resource utilization of ECU, and the computing
The transmission data of VM [0.5GB,0.8GB] resources are saved.
The transmission data of VM [0.5GB,0.8GB]
The probability of delivering l-th type data [0, 1] 1.2
Benchmark
NF
1 FF
500 FF
UTO 0.00
100 200 300 400 500
400
The number of tasks
300
200
Fig. 4: Comparison analysis of total time cost with diverse
task ranges by Benchmark, NF, FF and UTO.
100
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 7
10.54, 22.06, 35.86, 42.56 and 54.31 (s), and the migration 180.00
Benchmark
time is the least compared with the other three methods. 160.00
NF
140.00 FF
Privavy Entropy
70.00 120.00 JOM
Benchmark
NF 100.00
60.00
FF
UTO 80.00
50.00
Migration Time (s)
60.00
40.00 40.00
30.00 20.00
0.00
20.00 100 200 300 400 500
The number of tasks
10.00
澳
0.00
100 200 300 400 500 Fig. 7: Comparison analysis of privacy entropy with diverse
The number of tasks task ranges by Benchmark, NF, FF, and JOM.
Fig. 5: Comparison analysis of migration time cost with
diverse task ranges by Benchmark, NF, FF and UTO.
VII. C ONCLUSION AND F UTURE W ORK
5) Feedback Time consumption: The calculation of feed- In this paper, we focus on the investigation of the of-
back time is approximately one-fifth of the migration time. The floading problem considering the implementation utility (i.e.,
feedback time is used to indicate that the tasks are transmitted the resource utilization of ECUs and the time cost for task
back. If the feedback time equals 0, the tasks, which have been execution). To achieve the trade-offs of utility and privacy,
processed, are not transmitted back to the first ECU. With the a two-phase offloading strategy has been designed. In the
growth of the value of the computing tasks, the diversity of the first phase, we selected the utility-aware offloading strategy
four methods has not become bigger. Based on the analyze of by using NSGA-III to achieve the optimization of utilization
Fig. 6, the migration time in our method is 2.11, 4.41, 7.17, and time cost. Then in the second phase, the utility and the
8.51 and 10.27 (s). The feedback time also costs the least privacy entropy have been both taken into consideration as
compared with Benchmark, NF and FF methods. two main metrics to be optimized at same time. For future
work, we will continue to improve and extend UTO and JOM
to be applied in real world environment. Besides, the privacy-
14.00
Benchmark aware offloading problem in edge computing for 5G will be
NF
12.00 FF investigated.
UTO
10.00
Feedback Time 澻瀆澼
8.00
VIII. ACKNOWLEDGMENT
6.00
This research is supported by the National Natural Sci-
ence Foundation of China under grant no. 61702277 and
4.00
no. 61872219, the Priority Academic Program Development
2.00 of Jiangsu Higher Education Institutions (PAPD) fund, and
0.00 Jiangsu Collaborative Innovation Center on Atmospheric En-
100 200 300 400 500 vironment and Equipment Technology (CICAEET).
The number of tasks
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2944007, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL 8
[8] W. Hou, W. Li, L. Guo, Y. Sun, X. Cai, Recycling edge devices in Zhanyang XU received Ph.D. degree from the Nan-
sustainable internet of things networks, IEEE Internet of Things Journal jing University of Posts and Telecommunications,
4 (5) (2017) 1696–1706. China, in 2011. He worked as a research scholar at
[9] X. Xie, T. Yuan, X. Zhou, X. Cheng, Research on trust model in Arizona State University, USA, from May. 2014 to
container-based cloud service, Computers, Materials and Continua 56 (2) June 2015. He is currently an associate professor
(2018) 273–283. at the school of computer and software, Nanjing
[10] J. Zhang, N. Xie, X. Zhang, K. Yue, W. Li, D. Kumar, Machine learning University of Information Science and Technology.
based resource allocation of cloud computing in auction, Computers, He has published more than 15 papers in the in-
Materials & Continua 56 (1) (2018) 123–135. ternational journals and conferences. His research
[11] P. Mach, Z. Becvar, Mobile edge computing: A survey on architecture 澳
interests include Software Engineering, IoT, cloud
and computation offloading, IEEE Communications Surveys & Tutorials computing and big data.
19 (3) (2017) 1628–1656.
[12] L. Qi, Y. Chen, Y. Yuan, S. Fu, X. Zhang, X. Xu, A qos-aware virtual
machine scheduling method for energy conservation in cloud-based
cyber-physical systems, World Wide Web (2019) 1–23.
[13] P. Gope, B. Sikdar, Lightweight and privacy-preserving two-factor
authentication scheme for iot devices, IEEE Internet of Things Journal Lianyong Qi received his PhD degree in Department
6 (1) (2018) 580–589. of Computer Science and Technology from Nanjing
[14] C. Li, B. Palanisamy, Privacy in internet of things: from principles to University, China, in 2011. Now, he is an asso-
technologies, IEEE Internet of Things Journal 6 (1) (2018) 488–505. ciate professor of the School of Information Science
[15] S.-C. Cha, T.-Y. Hsu, Y. Xiang, K.-H. Yeh, Privacy enhancing technolo- and Engineering, Chinese Academy of Education
gies in the internet of things: Perspectives and challenges, IEEE Internet Big Data, Qufu Normal University, China. He has
of Things Journal 6 (2) (2018) 2159–2187. already published more than 50 papers including
[16] M. Chen, Y. Hao, Task offloading for mobile edge computing in JSAC, TCC, TBD, FGCS, JCSS, CCPE, ICWS and
software defined ultra-dense network, IEEE Journal on Selected Areas ICSOC, etc. His research interests include services
in Communications 36 (3) (2018) 587–597. computing, big data and IoT.
[17] S. Bi, Y. J. Zhang, Computation rate maximization for wireless pow-
ered mobile-edge computing with binary computation offloading, IEEE
Transactions on Wireless Communications 17 (6) (2018) 4177–4190.
[18] Z. Ding, P. Fan, H. V. Poor, Impact of non-orthogonal multiple access
on the offloading of mobile edge computing, IEEE Transactions on
Communications 67 (1) (2018) 375–390.
[19] K. Deb, H. Jain, An evolutionary many-objective optimization algorithm Shaohua Wan received the joint Ph.D. degree from
using reference-point-based nondominated sorting approach, part i: solv- the School of Computer, Wuhan University and the
ing problems with box constraints, IEEE Transactions on Evolutionary Department of Electrical Engineering and Computer
Computation 18 (4) (2013) 577–601. Science, Northwestern University, USA in 2010.
[20] K. R. Sollins, Iot big data security and privacy versus innovation, IEEE Since 2015, he has been holding a post-doctoral
Internet of Things Journal 6 (2) (2019) 1628–1635. position at the State Key Laboratory of Digital Man-
[21] Y. Guo, S. Wang, A. Zhou, J. Xu, J. Yuan, C.-H. Hsu, User allocation- ufacturing Equipment and Technology, Huazhong
aware edge cloud placement in mobile edge computing, Software: University of Science and Technology. From 2016
Practice and Experience (2019). to 2017, he was a visiting professor at with the
[22] S. Wang, A. Zhou, R. Bao, W. Chou, S. S. Yau, Towards green service Department of Electrical and Computer Engineering,
composition approach in the cloud, IEEE Transactions on Services Technical University of Munich, Germany. He is
Computing (2018). currently an Associate Professor with the School of Information and Safety
Engineering, Zhongnan University of Economics and Law. His main research
interests include deep learning for Internet of Things and Cyber-Physical
Xiaolong Xu received Ph.D. degree from the Nan- Systems. He is an author of over 60 peer-reviewed research papers and books.
jing University, China, in 2016. He worked as a He is a senior member of IEEE.
research scholar at Michigan State University, USA,
from Apr. 2017 to May 2018. He is currently an
assistant professor at the school of computer and
software, Nanjing University of Information Science
and Technology. He has published more than 60
peer review papers in the international journals and Md Zakirul Alam Bhuiyan (M’09-SM’17) is cur-
conferences, including TCC, TETCI, TCSS, TBD, rently an Assistant Professor of the Department of
JNCA, SPE, CCPE, FGCS, ICWS, ICSOC, etc. Computer and Information Sciences at the Fordham
His research interests include fog computing, edge University, NY, USA, the Founding Director of
computing, IoT, cloud computing and big data. Fordham Dependable and Secure System Lab (De-
pendSys). Earlier, he worked as an Assistant Profes-
sor at the Temple University. His research focuses on
dependability, cybersecurity, big data, and IoT/CPS
Applications. His work in these areas published in
top-tier venues, including IEEE TC, TPDS, TDSC,
Chengxun He received his B.S. in Computer Sci- 澳
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.