https://github.com/vita-group/sparseadv_homotopy

[ICML 2021 Long Talk] "Sparse and Imperceptible Adversarial Attack via a Homotopy Algorithm" by Mingkang Zhu, Tianlong Chen, Zhangyang Wang
https://github.com/vita-group/sparseadv_homotopy

adversarial-attacks adversarial-robustness homotopy-algorithm homotopy-type-theory sparse-adversarial

Last synced: about 1 month ago
JSON representation

[ICML 2021 Long Talk] "Sparse and Imperceptible Adversarial Attack via a Homotopy Algorithm" by Mingkang Zhu, Tianlong Chen, Zhangyang Wang

• Host: GitHub
• URL: https://github.com/vita-group/sparseadv_homotopy
• Owner: VITA-Group
• License: mit
• Created: 2021-06-06T19:08:18.000Z (almost 4 years ago)
• Default Branch: main
• Last Pushed: 2021-12-30T07:42:30.000Z (over 3 years ago)
• Last Synced: 2025-03-29T09:51:12.709Z (about 2 months ago)
• Topics: adversarial-attacks, adversarial-robustness, homotopy-algorithm, homotopy-type-theory, sparse-adversarial
• Language: Python
• Homepage:
• Size: 604 KB
• Stars: 9
• Watchers: 8
• Forks: 3
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Sparse and Imperceptible Adversarial Attack via a Homotopy Algorithm

[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT)

[ICML 2021 Long Talk] [Sparse and Imperceptible Adversarial Attack via a Homotopy Algorithm.](https://arxiv.org/abs/2106.06027)

Mingkang Zhu, Tianlong Chen, Zhangyang Wang

## Overview

We proposed Homotopy Attack, a homotopy algorithm for sparse adversarial attack that aim to find sparse adversarial perturbations with relatively small magnitudes. The sparsity of the perturbation is regularized by the L_0 norm and the magnitude is constrained by the L_inf norm. The proposed algorithm can achieve much sparser results under the same L_inf constraint when compared to the SOTA.





## Prerequisites

```

pytorch

torchvision

scipy 1.2.1

pillow

Download the pretrained Inception-v3 model from https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth and put it into the pretrain folder

```

## Usage

```

python demo_attack.py --img_dir path/to/imgs

```

## Parameter Setting

```

--imgdir             path to input images

--target_class       target class

--dec_factor         decreasing factor for lambda per iteration

--val_c              constant multiplied with searched intial lambda

--val_w1             weight for post attack stage

--val_w2             weight for post attack stage

--val_gamma          threshold of triggering the post attack stage

--max_update         max L_0 update per outer iteration

--max_epsilon        L_inf constraint      

--maxiter            max inner iteration to run nmAPG per outer iteration of homotopy

```

Hint:

For lower `--max_epsilon`, `--val_c` should be lower and `--max_update` should be higher.

For higher `--max_epsilon`, `--val_c` should be higher and `--max_update` should be lower.

Please adjust the parameters accordingly so that the average perturbation per entry is close to the L_inf constraint `--max_epsilon`, which can lead to sparser results.

For detailed parameter setting, please refer to the Appendix.

## Citation

```

@misc{zhu2021sparse,

    title={Sparse and Imperceptible Adversarial Attack via a Homotopy Algorithm},

    author={Mingkang Zhu and Tianlong Chen and Zhangyang Wang},

    year={2021},

    eprint={2106.06027},

    archivePrefix={arXiv},

    primaryClass={cs.LG}

}

```