Deepfake Detection with Swin Transformer

This repository contains code for deepfake detection using the Swin Transformer architecture. The project leverages PyTorch and PyTorch Lightning for training and evaluation.

Dataset Structure

The dataset should be organized as follows:

/data/
    train/
        Fake/
            fake_0.jpg
            fake_1.jpg
            ...
        Real/
            real_0.jpg
            real_1.jpg
            ...
    val/
        Fake/
            ...
        Real/
            ...
    test/
        Fake/
            ...
        Real/
            ...

Requirements

• Python 3.9+
• PyTorch
• PyTorch Lightning
• timm (PyTorch Image Models)
• albumentations
• torchmetrics
• Other dependencies listed in requirements.txt

Install the required packages using:

pip install -r requirements.txt

Configuration

The project's settings are defined in configs/config.yaml. You can modify parameters such as:

• data_root: Path to the dataset root directory
• model_name: Swin Transformer model name from timm
• batch_size, num_epochs, learning_rate, etc.
• train_augmentations and val_augmentations: Albumentations transformations

Training

To train the model, run:

python3 train.py

Training progress and metrics will be logged to TensorBoard. Checkpoints will be saved in the checkpoints directory.

Evaluation

To evaluate a trained model on the validation set, use:

python3 eval.py

Evaluation metrics will be printed to the console.

Key Files

• train.py: Main training script.
• eval.py: Evaluation script.
• swin_base.py: Defines the Swin Transformer model.
• dataloader.py: Handles dataset loading and augmentations.
• augmentations.py: Defines image transformations.
• metrics.py: Calculates evaluation metrics.
• configs/config.yaml: Project configuration file.

Acknowledgements

• The Swin Transformer implementation is based on the timm library.
• Albumentations is used for image augmentations.
• PyTorch Lightning simplifies training and experimentation.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

License

This project is licensed under the MIT License.