IEEE Expert Lecture Notes

Introduction

• Welcome to the IEEE expert session.
• Presentation topic: Deep Fake Face Detection using Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM).
• Emphasis on the importance of this project for society and future generations.

Background and Context

• Recent viral incident with Rashmiya Mandana's deep fake face video.
• AI technology is advancing rapidly, but poses risks such as deep fake videos.
• Deep fakes can lead to severe societal issues, including mental health crises among affected individuals.
• Need for proactive measures to combat the negative implications of AI.

Project Overview

• Project Title: Deep Fake Face Detection.
• Base Papers: Three recent IEEE papers from 2023 on deep fake detection.
• Objective: Use AI to detect fake videos and prevent misuse.

Technical Details

Definitions

• Deep Fake Videos: Videos generated using AI that manipulate faces, audio, or video to mislead viewers.
• Existing System: Based on CNN with drawbacks identified.

Limitations of CNN

1. Limited Temporal Understanding: CNN struggles with understanding features in fake vs. original videos.
2. Large Computational Requirements: Needs high-end hardware (e.g., graphics processors).
3. Vulnerability to Adversarial Attacks: Can be hacked, leading to inaccuracies.
4. Training Data Imbalance: Insufficient data can lead to poor accuracy.

Proposed Solution

• New Approach: Implement LSTM to improve detection capabilities.
• Advantages of LSTM:
  • Robust detection of deep fake incidents.
  • Efficient handling of video data in real-time.
  • Achieves high accuracy (up to 98%) for real vs. fake detection.

Project Methodology

1. Data Collection: Use datasets from Kaggle, focusing on video formats.
2. Pre-Processing Module:
   • Extract frames from videos.
   • Detect and crop facial features.
3. Training Module:
   • Utilize LSTM for training on extracted frames.
4. Testing Module:
   • Evaluate new video uploads against trained models to classify as real or fake.

Software and Hardware Requirements

• Minimum requirements:
  • Processor: i3
  • RAM: 4GB
  • OS: Windows or Mac
• Programming Language: Python
• Web Development: HTML, CSS, JavaScript
• Framework: Flask

Demonstration

• Project demonstration using Anaconda Navigator and TensorFlow.
• Users can upload videos to check for authenticity:
  • Results provided as confidence percentages for real vs. fake classifications.
• Example results:
  • Real video: 91% confidence.
  • Fake video: 99.9% confidence.

Conclusion

• The use of LSTM enhances detection capabilities over traditional CNN methods.
• Overall accuracy achieved: 98% in testing scenarios.
• For more information or to access this project, contact iwxpert.com.