This repository is the official implementation of Real Face Foundation Representation Learning for Generalized Deepfake Detection.

In this study, we propose Real Face Foundation Representation Learning (RFFR), which aims to learn a general representation from large-scale real face datasets and detect potential artifacts outside the distribution of RFFR.

• To install requirements:

pip install -r requirements.txt

• Download datasets at FaceForensics++, ForgeryNet, Celeb-DF, etc.

• Run face detection. With cropped face images, generate json files in the following format:

[
    {
        "path": "path/to/image_0.png",
        "label": 0
    },

    {
        "path": "path/to/image_1.png",
        "label": 1
    },

    ...
]

• Link the json files in the config files to load the data. You can also just modify the dataset files to load the images in any other ways you want.

Make sure to specify paths to labels, pretrained models, as well as hyperparameters in rffr_classifier/configs/config.py and rffr_generative/configs/config.py.

To train the generative model that produces residuals:

cd rffr_generative
python train.py

To train the classifier:

cd rffr_classifier
python train.py

Specify models to test in the aforementioned config files.

View demos of the generative model's output by running:

cd rffr_generative
python evaluate.py

Test the deepfake detector by running:

cd rffr_classifier
python test.py

This repository partially borrows from SSDG (for code structure) and MAE (for implementation of the generative model). Thank you!

Consider citing our paper if you find it helpful:

@article{shi2023real,
  title={Real Face Foundation Representation Learning for Generalized Deepfake Detection},
  author={Shi, Liang and Zhang, Jie and Shan, Shiguang},
  journal={arXiv preprint arXiv:2303.08439},
  year={2023}
}

Feel free to contact us with any comments or feedback.