Self-Supervised Adversarial Training for Robust Face Forgery Detection

Introduction

This is the source code for Self-Supervised Adversarial Training for Robust Face Forgery Detection. Our work has been accepted by BMVC 2023. This is the pipeline of our method.

Quickly get: [Paper Link] [Home Page]

Updates

• [12/2023] Training code released.
• [11/2023] Conference paper released.

Install

The implementation bases on Python 3.6. The packages needed in this project are shown in requirements.txt. You can install these packages by:

pip install -r requirements.txt

Dataset

We train our model using the Faceforencis++ dataset. We utilize the first 270 frames of each training video and the first 110 frames of each validation/testing video. The data structure is like:

|---README.md
|---...
|---data
	|---FF-DF
		|---000
			|---0000.png  
			|---0001.png  
			|---...
		|---...
	|---FF-F2F
	|---FF-FS
	|---FF-NT
	|---real

The organization of DeeperForensics dataset is similar to FF++.

API

Training

Start training by:

python train.py --train_batchSize 32 --resolution 256 --nEpochs 10

There are different settings for your training process, such as batch size, resolution, distributed training, etc. Check train.py for more augments.

Testing

Test the model on FF++.

python test.py --test_batchSize 32 --resolution 256 --pretrained 'the weight of pretrained discriminator'

Result

Robustness to Compressed Images

Methods	RAW	HQ	LQ
Xception	0.989	0.961	0.895
Face X-ray	0.988	0.866	0.631
SBI	0.994	0.950	0.903
Two-branch	0.981	0.958	0.890
Cvit	---	0.937	---
SAT(ours)	0.992	0.988	0.934

Robustness to Perturbed Images

More details about these experiments can be found in Section 4 of our paper.

Citation

If you find this work useful for your research, please kindly cite our paper:

@article{gao2023self,
  title={Self-Supervised Adversarial Training for Robust Face Forgery Detection},
  author={Gao, Yueying and Lin, Weiguo and Xu, Junfeng and Xu, Wanshan and Chen, Peibin},
  year={2023}
}