Deepfake Photo Detector

A web application that detects deepfake content in videos by extracting frames and analyzing them with a CNN model.

Features

• Upload short video files through a user-friendly interface
• Process videos by extracting 10 representative frames
• Compress frames efficiently without quality loss
• Analyze frames using a CNN model to detect deepfakes
• Display results with confidence level
• Optimized for performance with no database or permanent storage requirements

Getting Started

Prerequisites

• Python 3.8 or higher
• pip (Python package manager)

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/DeepFake-Photos-Genrator.git
   cd DeepFake-Photos-Genrator
   

2. Install dependencies:

   pip install -r requirements.txt
   

Running the Application

Start the application with:

python app.py

The application will be available at http://localhost:5000

How It Works

1. Video Upload: Users upload a short video file through the web interface
2. Frame Extraction: The system extracts 10 evenly spaced frames from the video
3. Image Compression: Frames are compressed to reduce size while maintaining quality
4. Model Analysis:
   • Frames are normalized and fed into a CNN model
   • Each frame receives a prediction score
   • Scores are averaged to determine final classification
5. Result Display: The system shows whether the video is authentic or manipulated with a confidence score

Technical Implementation

• Framework: Flask web server
• Video Processing: OpenCV for frame extraction
• Image Processing: PIL/Pillow for efficient compression
• Model: TensorFlow CNN model for deepfake detection
• In-memory processing: No database or storage requirements

Performance Optimizations

• Only 10 frames are extracted to minimize processing time
• Images are resized to 224x224 pixels for the model
• Efficient JPEG compression reduces memory usage without losing quality
• Temporary files are cleaned up after processing
• All processing happens in memory for better performance
• Model is loaded once at startup to save time during analysis

Customization

Replace the dummy model in app.py with your trained CNN model:

# Replace with your actual model loading code
model = tf.keras.models.load_model('path_to_your_model')

License

MIT License