Fixed evaluation of models with random defenses

[Buntender]

Thank you for your outstanding contributions.

@LYMDLUT and I put forward this PR to improve the evaluation of models with random defenses.

We've noticed that AutoAttack's current strategy for selecting the final output (clean/APGD etc) based on one time evaluation, regardless of whether the target models implement random defenses or not. This overlooks the variability of outputs in models with random defenses.

Relying on a single evaluation to filter samples for subsequent attacks leads to inflated success rate and hinders the exploration of attack methods that could potentially yield superior outcomes.

To address this, we propose to perform multiple time evaluations for models with random defenses and chose the adversarial example with the highest robustness as final output.

[LYMDLUT]

@fra31 Could you please review this pr？

[ScarlettChan]

您好，您的邮件已收到!

[LYMDLUT]

In Appendix L of our paper, we provide a detailed report on our fix for AutoAttack and its impact. We encourage future research to adopt this updated version when evaluating models with randomness, as it effectively reduces the risk of overestimating robustness.

If you find our work useful for your research, please consider citing it:

@Article{liu2024towards,
title={Towards Better Adversarial Purification via Adversarial Denoising Diffusion Training},
author={Liu, Yiming and Liu, Kezhao and Xiao, Yao and Dong, Ziyi and Xu, Xiaogang and Wei, Pengxu and Lin, Liang},
journal={arXiv preprint arXiv:2404.14309},
year={2024}
}

[ScarlettChan]

您好，您的邮件已收到!