DomainForensics: Exposing Face Forgery across Domains via Bi-directional Adaptation

The official PyTorch implementation for the following paper:

DomainForensics: Exposing Face Forgery Across Domains via Bi-Directional Adaptation

Qingxuan Lv; Yuezun Li*; Junyu Dong*; Sheng Chen; Hui Yu; Huiyu Zhou; Shu Zhang

IEEE Transactions on Information Forensics and Security

Recomended Development Environment

Basic dependencies

opencv-python
numpy
easydict
pytorch-lightning==1.7.5
rich
matplotlib

Other dependencies

RetinaFace-pytorch

install retinaface-pytorch.

Jpeg-turbo

install Libjpeg-Turbo.

Jpeg2dct

install jpeg2dct

Setup

Datasets

FF++

1. Download FF++ from ff++ github
2. Extract to /path/to/FF++
3. Crop faces by using utils/crop_retinaface_ff.py e.g. python crop_retinaface_ff.py -d Deepfakes -n 8

Celeb-DF

1. Download Celeb-DF v2 from celeb-df github
2. Extract to /path/to/Celeb-DF
3. Crop faces by using utils/crop_retinaface_ff.py e.g. python crop_retinaface_ff.py -d Celeb-real -n 8 -s train

Training

We train the model on two RTX2080Ti with 11Gx2 GPU memory.

Change the config within train.py for different adaptation setting. e.g.

    # from Deepfakes to Face2Face
    cfg.DATAS.SOURCE = ["Deepfakes"]
    cfg.DATAS.TARGET = ['Face2Face']

run training by executing python train.py

Testing

1. Change the cfg.TESTING.MODEL_WEIGHT to the pretrained weight
2. run testing by executing python testing.py

Pretrained Models

Task	Dataset	AUC	Model	Size
FaceSwap -> Face2Face	FF++	99.13	google drive	855MB
NeuralTextures -> FaceSwap	FF++	97.67	google drive	855MB

Citation

@ARTICLE{10601589,
  author={Lv, Qingxuan and Li, Yuezun and Dong, Junyu and Chen, Sheng and Yu, Hui and Zhou, Huiyu and Zhang, Shu},
  journal={IEEE Transactions on Information Forensics and Security}, 
  title={DomainForensics: Exposing Face Forgery Across Domains via Bi-Directional Adaptation}, 
  year={2024},
  volume={19},
  number={},
  pages={7275-7289},
  keywords={Forgery;Deepfakes;Detectors;Feature extraction;Faces;Training;Bidirectional control;Digital forensics;DeepFake detection;DomainForensics},
  doi={10.1109/TIFS.2024.3426317}}