Awesome

Anti-Forgery

An example of Anti-Forgery: Towards a Stealthy and Robust DeepFake Disruption Attack via Adversarial Perceptual-aware Perturbations (to be presented at the IJCAI-ECAI 2022). This repository contains code for crafting perceptual-aware perturbation in the Lab color space to attack an image-to-image translation network.

Preparation

CelebA Dataset

bash download.sh celeba

StarGAN Model

bash download.sh pretrained-celeba-256x256

More information about the CelebA dataset can be found here.

Attack Testing

Here is a simple example of testing our method to attack StarGAN on the CelebA dataset.

# Test
python main.py --mode test --image_size 256 --c_dim 5 --selected_attrs Black_Hair Blond_Hair Brown_Hair Male Young --model_save_dir='stargan_celeba_256/models' --result_dir='./results' --test_iters 200000 --attack_iters 100 --batch_size 1

Related Work

We use some code from the original Disrupting-Deepfakes, which does a good work.

Citation

If you find this work useful, please cite our paper:

@article{wang2022anti,
  title={Anti-Forgery: Towards a Stealthy and Robust DeepFake Disruption Attack via Adversarial Perceptual-aware Perturbations},
  author={Wang, Run and Huang, Ziheng and Chen, Zhikai and Liu, Li and Chen, Jing and Wang, Lina},
  journal={arXiv preprint arXiv:2206.00477},
  year={2022}
}

The IJCAI camera-ready version (pdf) is available here