Staircase Sign Method for Boosting Adversarial Attacks

This is the Tensorflow code for our paper "Staircase Sign Method for Boosting Adversarial Attacks". Pytorch version can be found here.

In our paper, we rethink the limitation of Sign method (SM), e.g., I-FGSM, and propose a novel Staircase Sign Method (SSM) to alleviate this issue, thus boosting both targeted and non-targeted transfer-based attacks. Comparing with state-of-the-art targeted attacks, we significantly improve the transferability (i.e. on average, 5.1% for normally trained models and 12.8% for adversarially trained defenses).

Implementation

• Requirement

  • Python 3.7
  • Tensorflow 1.14.0
  • pandas 1.1.3
  • gast 0.2.2
  • matplotlib 3.3.4
  • tqdm 4.43.0

• Download the models

  • Normlly trained models
  • Ensemble adversarial trained models

• Then put these models into "models/"

• Run the code

  • The vanilla I-FGSSM attack method

  python attack_iter_SSM_NT.py  # If the victim's model is in normally trained models
  

  • The more powerful P-T-DI2++-FGSSM

  python attack_iter_SSM_EAT.py  # If the victim's model is in ensemble adversarially trained models  
  

• The output images are in "output/"

Visualization

Experimental Results

Citing this work

If you find this work is useful in your research, please consider citing:

@article{Zhang2021SSM,
    title={Staircase Sign Method for Boosting Adversarial Attacks},
    author={Zhang, Qilong and Zhu, Xiaosu and Song, Jingkuan and Gao, Lianli and Shen, Hengtao},
    journal   = {CoRR},
    volume    = {abs/2104.09722},
    year      = {2021}
}