Professional Documents
Culture Documents
Sanjay DRC
Sanjay DRC
ABSTRACT
PROBLEM STATEMNT
LITERATURE SURVEY
PROPOSED SYSTEM
SYSTEM REQUIREMENTS
SYSTEM ARCHITECTURE
FLOW CHART
ALGORITHM STEPS
SAMPLE CODE
INPUT
OUTPUT
CONCLUSION
ABSTRACT
As artificial intelligence (AI) continues to advance, the generation of highly realistic and deceptive
images using deep learning techniques poses a significant challenge to online platforms and content
moderation systems. The proliferation of AI-generated images, commonly referred to as deepfakes,
raises concerns about misinformation, identity theft, and the erosion of trust in digital media. The
problem at hand is the urgent need for robust and real-time methods to detect AI-generated images
within online content. Current detection mechanisms often fall short in keeping pace with evolving
AI techniques, leading to the widespread dissemination of manipulated images that can deceive,
manipulate, or harm individuals and organizations.
LITERATURE SURVEY
Authors: Aakanksha Chordia, Anirudh Goyal, Shubham Jha, Shweta Khare, Aniruddha Saha, Ayush Verma
Title: A Survey on the Possibilities & Impossibilities of AI-generated Text Detection
Authors: Muhammad Bilal Zafar, Muhammad Usman Qureshi, Muhammad Umair Khan, Shahzad Rashid
Title: AI vs. AI: Can AI Detect AI-Generated Images?
Date: ct October 2023
Source: MDPI Sensors
Summary: This survey explores the use of AI models to detect AI-generated images. It proposes a framework
using Convolutional Neural Networks and Class Activation Maps for accurate image classification. The authors
achieve promising results and emphasize the importance of using detection tools as an evaluation tool for image
generation models.
PROPOSED SYSTEM
This model simulates the continuous release of new generative models by creating batches of real
and generated images in a specific order. The order mimics how new generators might be released in
practice, ensuring the training data reflects the dynamic nature of the problem. A CNN classifier is
then trained on these batches, progressively adapting to detect features indicative of different
generative models. Once trained, the classifier can predict the probability of each pixel in an image
being real or generated. This allows for identifying both fully AI-generated images and those
containing mixed real and generated content (hybrid images).The model demonstrates strong
performance on this task, suggesting this can effectively distinguish between different image origins.
SOFTWARE REQUIREMENTS
EfficientNet Model :
1. Create Baseline : Develop EfficientNet-B0 through neural architecture search for optimal
efficiency.
2. Scale Dimensions : Systematically scale network width, depth, and resolution with fixed
coefficients.
3. Produce Variants : Generate B1-B7 models, scaling up from B0 for varying computational
capacities.
4. Optimize Efficiency : Fine-tune for a balance between model size and accuracy, surpassing
conventional architectures.
5. Adapt Application : Deploy scalable EfficientNet models across diverse platforms, from mobile
devices to cloud servers.
SAMPLE CODE
INPUT
OUTPUT
CONCLUSION
This model simulates the continuous release of new generative models by creating batches of real
and generated images in a specific order. The order mimics how new generators might be released in
practice, ensuring the training data reflects the dynamic nature of the problem. A CNN classifier is
then trained on these batches, progressively adapting to detect features indicative of different
generative models. Once trained, the classifier can predict the probability of each pixel in an image
being real or generated. This allows for identifying both fully AI-generated images and those
containing mixed real and generated content (hybrid images).The model demonstrates strong
performance on this task, suggesting this can effectively distinguish between different image origins.
THANK YOU !
PROJECT DRC-1 PRESENTATION ON
Dr. N.JAYALAKSHMI
Associate professor
The project is to solve the difficulty of telling apart photos and artwork created by humans from
those generated by AI due to the frequent advancements in AI-generated images. Unlike previous
methods that focus on one AI generator at a time, we study the generalization across multiple
generators released over time. Images today often combine real and AI-generated elements, and we
extend our analysis to predict pixels, demonstrating strong performance. Additionally, we explore if
pixel detectors can be trained exclusively on synthetic images when commercial models aren't
accessible for data generation. my project contributes to improving the detection of AI-generated
images online by considering the continuous release of new generators and addressing the
complexity of images with both real and AI-generated parts.