This project implements a deep learning pipeline for detecting facial video forgeries, specifically focusing on DeepFake and Face2Face manipulation techniques. It utilizes the lightweight and efficient MesoNet architecture to classify whether a given facial image is real or forged.

The system can operate as a complete pipeline: from video frame extraction, through face detection, to classification using pretrained models.

.
├── test_images/           # Contains test images for evaluation
├── weights/               # Pretrained MesoNet weights for DeepFake (_DF) and Face2Face (_F2F)
├── LICENSE                # Project license (Apache-2.0)
├── README.md              # This README file
├── classifiers.py         # Defines MesoNet models and classification logic
├── example.py             # Example script for running classification
├── pipeline.py            # End-to-end pipeline with face detection and classification
├── test.py                # Script for testing model performance

Python 3.5 is recommended for compatibility with original Keras versions.

Install dependencies:

pip install numpy==1.14.2 keras==2.1.5 imageio face_recognition

Ensure FFMPEG is installed on your system for video processing.

• DeepFake model: Meso4_DF
• Face2Face model: Meso4_F2F

python example.py --model weights/Meso4_DF.h5 --input test_images/

Or use the full pipeline with face detection:

python pipeline.py --video input_video.mp4 --model weights/Meso4_F2F.h5

Note: You can switch between models by replacing the .h5 file accordingly.

The project is trained on aligned face datasets, split into:

• Training set size: ~150 MB
• Validation set size: ~50 MB

📩 Dataset Access:
The dataset is not publicly hosted but can be made available upon request for academic and research purposes. Please contact the project maintainer or raise an issue in the repository to request access.

You can find pretrained models inside the weights/ folder:

• Meso4_DF.h5 – trained for DeepFake detection
• Meso4_F2F.h5 – trained for Face2Face detection

These models are ready to use and achieve high classification accuracy on aligned face images.

This project is licensed under the MIT License.