mena-ml-deepfake-detection

MENA ML Challenge - Deepfake Detection Solution

🎯 Overview

This repository contains our solution for the MENA ML Challenge focusing on deepfake detection and image manipulation identification.

🏗️ Architecture

Our solution combines three key modules:

1. Vision Encoder (CLIP): Extracts semantic features from images
2. Forensic Analysis: Handcrafted features detecting manipulation artifacts
3. Vision-Language Model (BLIP-2): Provides semantic reasoning about image authenticity

📦 Installation

pip install -r requirements.txt

🚀 Usage

python predict.py --input_dir /path/to/test_images --output_file predictions.json

Input

• Directory containing .jpg or .png images

Output

JSON file with predictions in the format:

{
  "image_name": "000001.jpg",
  "authenticity_score": 0.91,
  "manipulation_type": "inpainting",
  "vlm_reasoning": "The window reflection is inconsistent. Shadow direction does not match the light source."
}

📊 Model Details

• Vision Encoder: OpenAI CLIP ViT-B/32
• VLM: Salesforce BLIP-2 OPT-2.7B
• Custom Components:
  • Forensic MLP (32 → 512 dimensions)
  • Fusion Head (768 + 512 → 1)
  • VLM Decision Head (512 + 4096 → 1)

🎓 Technical Report

See TECHNICAL_REPORT.pdf for detailed architecture and methodology.

📝 Manipulation Types Detected

• authentic: No manipulation detected
• inpainting: Object removal/addition
• full_synthesis: Fully AI-generated image
• filter: Color grading/enhancement
• deepfake: Face manipulation
• splicing: Image composition

👥 Team

[Your Team Name]

📧 Contact

For questions, contact: [[email protected]]

📄 License

[Specify your license]