Sparse-Adversarial-Attack

This repository provides a simple implementation of our recent work "Sparse Adversarial Attack via Perturbation Factorization", ECCV 2020.

Dependencies

• Python 3.6
• Pytorch 0.4.0 (other versions may be also OK, but we didn't verify it)
• Other Python packages: numpy, time, PIL, skimage, json

Demo

The following demo can generate sparse adversarial perturbations by attacking a CNN model trained on CIFAR-10, using the proposed attack method.

python main.py --attacked_model cifar_best.pth --img_file img0.png --target 1 --k 200   

• Inputs: attacked_model indicates the checkpoint to be attacked; img_file denotes the benign image; target is the target attack class; k represents the number of perturbed pixels.
• Outputs: The generated perturbation (saved as .npy) and the adversarial image (saved as .png file) will be saved in ./results.

Citations

@inproceedings{sapf-ECCV2020,
  title={Sparse Adversarial Attack via Perturbation Factorization},
  author={Fan, Yanbo and Wu, Baoyuan and Li, Tuanhui and Zhang, Yong and Li, Mingyang and Li, Zhifeng and Yang, Yujiu},
  booktitle={European conference on computer vision},
  year={2020}
}