tools-only / evadingfakedetector

We propose a statistical consistency attack (StatAttack) against diverse DeepFake detectors.

Evadingfakedetector

We propose a statistical consistency attack (StatAttack) against diverse DeepFake detectors.

Requirements

• numpy==1.24.1
• opencv_python==4.8.0.76
• Pillow==10.0.0
• scikit_learn==1.3.0
• scipy==1.11.1
• torch==2.0.1+cu118
• torchattacks==3.4.0
• torchvision==0.15.2+cu118
• umap==0.1.1

DataSet

we conduct a comprehensive evaluation based on 4 generation methods. The generated face dataset includes entire face synthesis images, face identity swap images, and face manipulation images

• StyleGANv2: https://github.com/NVlabs/stylegan2
• StarGAN: https://github.com/yunjey/stargan
• ProGAN: https://github.com/tkarras/progressive_growing_of_gans
• FaceForensics++: https://github.com/ondyari/FaceForensics

Usage

1. Clone this repository and install the required modules as listed in requirements.txt.
2. Place the detection model code in the ./model directory. Follow the resnet50.py example inside the ./model folder to add hooks for obtaining the mmd loss.
3. Run demo.py to generate adversarial examples.

Bibtex

@inproceedings{hou2023evading,
  title={Evading DeepFake Detectors via Adversarial Statistical Consistency},
  author={Hou, Yang and Guo, Qing and Huang, Yihao and Xie, Xiaofei and Ma, Lei and Zhao, Jianjun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12271--12280},
  year={2023}
}