Code for paper "Detecting Adversarial Data Using Perturbation Forgery".

Get Started

Our codebase accesses the datasets from ./data/ and checkpoints from ./results/checkpoints/ by default.

├── ...
├── data
│   
├── checkpoints
│   
├── results
│   
├── main.py
├── ...

Dataset

We use official CIFAR10 and ImageNet100 datasets for training. All of the adversarial data are generated using torchattacks. The training and testing datasets are available here. Download these datasets and unzip them in the /data folder.

Generate noise by gen_noise.py and extract distribution by gen_dist.py. You can directly download the distribution from Baidu Disk or Google Disk. Put distributions under 'data/dist/'.

Checkpoints

We offer checkpoints for testing. Download these checkpoints in the /checkpoints folder.

Dataset	Attack	Checkpoints
CIFAR-10	Linf	ckpt
ImageNet100	Linf	ckpt

Train Detector

Train detector against gradient-based adversarial attacks on CIFAR-10

python main.py
--config configs/datasets/general/DIS_CIFAR10.yml
configs/pipelines/train/DIS_train_CIFAR10.yml
--force_merge True
--preprocessor.name base

Train detector against gradient-based adversarial attacks on ImageNet100

python main.py
--config configs/datasets/general/DIS_ImageNet100.yml
configs/pipelines/train/DIS_train_ImageNet100.yml
--force_merge True
--preprocessor.name ImageNet

Test

Test detector against gradient-based Linf adversarial attacks on CIFAR-10

python main.py
--config configs/datasets/DIS_CIFAR10.yml
configs/pipelines/test/DIS_test_CIFAR10.yml
--force_merge True
--preprocessor.name base

Test detector against gradient-based Linf adversarial attacks on ImageNet100

python main.py
--config configs/datasets/DIS_ImageNet100.yml
configs/pipelines/test/DIS_test_ImageNet100.yml
--force_merge True
--preprocessor.name ImageNet

## Dependencies
python 3.8.8, PyTorch = 1.10.0, cudatoolkit = 11.7, torchvision, tqdm, scikit-learn, mmcv, numpy, opencv-python, dlib, Pillow

Citation

@article{wang2023detecting,
  title={Detecting adversarial faces using only real face self-perturbations},
  author={Wang, Qian and Xian, Yongqin and Ling, Hefei and Zhang, Jinyuan and Lin, Xiaorui and Li, Ping and Chen, Jiazhong and Yu, Ning},
  journal={arXiv preprint arXiv:2304.11359},
  year={2023}
}