This repository is an PyTorch implementation of the paper Image Super-Resolution as a Defense Against Adversarial Attacks

We use wavelet denoising and image super resolution as pre-processing steps to defend images against adversarial attacks. If you find our work useful in your research or publication, please cite our work:

We provide scripts for reproducing all the results from our paper. You can check the efficacy of our defense on your own adversarial images.

• Python 3.6
• PyTorch >= 0.4.0
• imageio
• tqdm

Clone this repository into any place you want.

git clone https://github.com/aamir-mustafa/super-resolution-adversarial-defense
cd super-resolution-adversarial-defense

You can test our wavelet denoising + super-resolution algorithm on your own adversarial images and their corresponding ground truth labels.

Wavelet_Denoising.py -- (for image wavelet denoising).

• The denoised images will be saved in test folder.

Place your denoised images in test folder. (like test/<your_image(s)>) We support jpg files.

Run the script in src folder.

cd src       # You are now in */super-resolution-adversarial-defense-master/src
sh super_resolution.sh

• You can find the result images from experiment/test/results-Demo folder.

Accuracy.py (Evaluate the performace of our method by comparing accuracies on adversarial and recovered images).

@article{mustafa2019image,
  title={Image Super-Resolution as a Defense Against Adversarial Attacks},
  author={Mustafa, Aamir and Khan, Salman H and Hayat, Munawar and Shen, Jianbing and Shao, Ling},
  journal={arXiv preprint arXiv:1901.01677},
  year={2019}
}