Attention Consistency Refined Masked Frequency Forgery Representation for Generalizing Face Forgery Detection
- torch 1.10.0+cu113
- timm 0.4.12
More details about dependencies are shown in requirements.txt.
- We include the dataset loaders for WildDeepfake and Celeb-DF. You can enter the dataset website to download the original data. https://github.com/deepfakeinthewild/deepfake-in-the-wild and https://github.com/yuezunli/celeb-deepfakeforensics
- For WildDeepfake, you should first extract the facial images from the sequences and store them. We use RetinaFace to do this.
To test the model in the Celeb-DF in FAC stage, run the following script in your console. The model will start training and return the AUC at each epoch.
python test-cele.py
To test the model in the WDF in FAC stage, run the following script in your console. The model will start training and return the AUC at each epoch.
python test-wdf.py
You can load the pre-trained model in test files. You can download the pre-trained model from here. https://pan.baidu.com/s/1sULCnA-5WVtGHtfCjc9_PQ key：ozcc or https://drive.google.com/file/d/1T3bNEJwsxLVlU6R6bZKWG7weMW8jvaKl/view?usp=drive_link, https://drive.google.com/file/d/1l1ZRTjT7X4QlprTCBVhUjJJsTtmbk9zI/view?usp=drive_link
@article{liu2023attention,
title={Attention Consistency Refined Masked Frequency Forgery Representation for Generalizing Face Forgery Detection},
author={Liu, Decheng and Chen, Tao and Peng, Chunlei and Wang, Nannan and Hu, Ruimin and Gao, Xinbo},
journal={IEEE Transactions on Information Forensics and Security},
year={2024}
}