Self-supervised Learning of Adversarial Example: Towards Good Generalizations for Deepfake Detection
This repository contains PyTorch implementation of the CVPR oral presentation paper:
Self-supervised Learning of Adversarial Example: Towards Good Generalizations for Deepfake Detection.
Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, Jue Wang
The proposed method uses adversarial self-supervised training to improve the generability of current deepfake detectors. The pipeline is illustrated in the following figure:
Please refer to the requirements.txt for details.
Download Xception pretrained weights and dlib landmark predictor and put them in the weights folder.
We use the FaceForensicsDataset (FF++) for training. Please go to their project page for downloading. For every video in FF++ dataset, we extract 270 frames for training, and 100 each for evaluation and testing rigously following their data splitting strategy. The data structure is like:
SLADD project
|---README.md
|---...
|---data
|---FF
|---image
|---FF-DF
|---071_054
|---0001.png
|---...
|---...
|---FF-F2F
|---FF-FS
|---FF-NT
|---real
|---mask
|---FF-DF
|---071_054
|---0001_mask.png
|---...
|---...
|---FF-F2F
|---FF-FS
|---FF-NT
|---config
|---train.json
|---test.json
|---eval.json
We use the DFDC, CelebDF, and DF1.0 for testing. These datasets are organized similar to FF++. Please go to their sites for downloading.
python train.py --resolution 256 --dataname none --dset FF-DF --meta FF-DF -n 1 -g 8 -nr 0 -mp 5555
If you find this code useful for your research, please cite:
@inproceedings{chen2022self,
author = {Liang Chen and Yong Zhang and Yibing Song and Lingqiao Liu and Jue Wang},
title = {Self-supervised Learning of Adversarial Examples: Towards Good Generalizations for DeepFake Detections},
booktitle = {CVPR},
year = {2022}
}
Please open an issue or contact Liang Chen ([email protected]) if you have any questions or any feedback.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent