README.md

metadata

title: Unmask
emoji: 🕵️‍♂️
colorFrom: purple
colorTo: indigo
sdk: docker
sdk_version: '1.0'
app_file: app.py
pinned: false

UnMask - Enhanced Deepfake Detection

This project integrates multiple deepfake detection models with a modern web interface.

🚀 Quick Start

Backend Setup

Updated the model to tf_efficientnet_b7_ns_0_40

1. Navigate to backend directory:

   cd backend
   

2. Install Python dependencies:

   pip install opencv-python numpy flask flask-cors torch torchvision dlib
   

3. Start the server:

   Option A: Using the batch file (Windows):

   start_server.bat
   

   Option B: Using Python directly:

   python simple_server.py
   

   Option C: Using the Flask API (if dependencies available):

   python app.py
   

The backend will start on http://localhost:8000 (simple server) or http://localhost:5000 (Flask API).

Frontend Setup

1. Navigate to frontend directory:

   cd frontend
   

2. Install dependencies:

   npm install
   

3. Start the development server:

   npm run dev
   

The frontend will be available at http://localhost:3000.

🎯 Features

Detection Models

• MesoNet (Meso4 & MesoInception4): Lightweight CNN architectures
• Enhanced Xception: Deep learning model optimized for deepfake detection
• Ensemble Voting: Combines multiple models for better accuracy

Web Interface

• Drag & Drop Upload: Easy file upload interface
• Real-time Processing: Live feedback during analysis
• Detailed Results: Comprehensive detection metrics
• Multi-model Analysis: Shows individual model predictions

📊 API Endpoints

Flask API (app.py)

• GET / - Server status and information
• POST /api/detect - Upload video for deepfake detection
• GET /api/models - Get information about loaded models

Simple Server (simple_server.py)

• GET / - API status (JSON)
• GET /detect - Upload interface (HTML)
• POST /upload - Process uploaded video

🔧 Technical Details

Supported Formats

• Video: MP4, AVI, MOV, MKV, WebM
• Maximum Size: 100MB
• Processing: Automatic face detection and analysis

Models Integration

The project integrates the HongguLiu/Deepfake-Detection repository, providing:

1. MesoNet Models: Specialized for deepfake detection
2. Transfer Learning: Pre-trained Xception networks
3. Face Detection: Automatic face extraction from videos
4. Statistical Analysis: Advanced frame-level analysis

Architecture

frontend/          # Next.js React application
├── src/
│   ├── app/       # Main application pages
│   └── Components/ # Reusable UI components
│
backend/           # Python detection server
├── Deepfake-Detection/  # Cloned detection models
├── app.py         # Flask API server
├── simple_server.py     # Standalone HTTP server
└── requirements.txt     # Python dependencies

🛠️ Troubleshooting

Backend Issues

1. Python dependencies missing: Install required packages:

   pip install -r requirements.txt
   

2. Port conflicts: Change port in server files if needed

3. Model loading errors: Ensure the Deepfake-Detection repository is properly cloned

Frontend Issues

1. CORS errors: Ensure backend server is running and accessible

2. Upload failures: Check file format and size limits

3. Connection refused: Verify backend server is running on correct port

📈 Performance

Processing Speed

• Face Detection: ~1-2 seconds per video
• Model Inference: ~0.5-1 second per model
• Total Analysis: ~3-5 seconds average

Accuracy

• Ensemble Model: Combines multiple approaches for better reliability
• False Positive Rate: Reduced through multi-model voting
• Confidence Scoring: Provides reliability metrics

🔮 Future Enhancements

1. Additional Models: Integration of more detection algorithms
2. Batch Processing: Multiple file uploads
3. Real-time Streaming: Live video analysis
4. Model Training: Custom model training interface
5. Advanced Analytics: Detailed forensic analysis

📝 Notes

• This is a research/educational tool - results should be verified
• Processing time depends on video length and system performance
• GPU acceleration available if PyTorch CUDA is installed
• Regular model updates recommended for best performance