Deepfake Detection

This project is a deep learning pipeline for detecting deepfake videos using PyTorch. It includes preprocessing, training, and evaluation scripts, and leverages a ResNeXt backbone with LSTM for temporal modeling.
Dataset Link
Download the .pth file for the Model4 from here
Project Documentation

Features

• Face extraction and super-resolution preprocessing
• Custom PyTorch dataset and dataloader
• ResNeXt50 backbone with LSTM for sequence modeling
• Training with early stopping and backbone fine-tuning
• Evaluation with ROC, F1, and confusion matrix visualization

Directory Structure

DeepfakeModelv4/
├──ERSGAN/(git clone)
├──deepfake_dataset
├──data
   ├──faces
      ├──test
      ├──train
├── best_model.pth         # Trained model weights
├── dataset.py             # Custom dataset class
├── eval.py                # Evaluation script
├── model.py               # Model definition
├── preprocess.py          # Preprocessing (face extraction, super-resolution)
├── train.py               # Training script          
└── __init__.py            # Package marker

Requirements

• Python 3.7+
• PyTorch
• torchvision
• scikit-learn
• tqdm
• numpy
• opencv-python
• pillow
• facenet-pytorch
• seaborn
• matplotlib

You may install dependencies with:

pip install torch torchvision scikit-learn tqdm numpy opencv-python pillow facenet-pytorch seaborn matplotlib

Data Preparation

1. Place your raw videos in deepfake_dataset/train/real, deepfake_dataset/train/fake, deepfake_dataset/test/real, and deepfake_dataset/test/fake.
2. Run the preprocessing script to extract faces and apply super-resolution:

   python preprocess.py

   This will create processed frames in data/faces/train and data/faces/test.

Training

Train the model using:

python train.py

The best model will be saved as best_model.pth.

Evaluation

Evaluate the trained model:

python eval.py

This will print metrics and show confusion matrix and ROC curve plots.

These are the comparison with other model architectures trained on the same dataset

Notes

• The preprocessing script uses ESRGAN for super-resolution. Make sure the ESRGAN model weights and code are available as referenced in preprocess.py.
• Adjust paths in scripts as needed for your environment.

License

MIT License