REAL: Rectified Adversarial Sample via Max-Min Entropy for Test-Time Defense

Lina Ma, Lei Zhang

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: deeplearning, adversarial defense

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Adversarial attacks expose the vulnerability of neural networks. But it is difficult for existing defense methods to defend against all attacks, which leads to the lack of generalization in adversarial robustness. Inspired by test-time adaptation which leverages model’s prediction entropy to generalize naturally distributed samples during testing, we try to rationally utilize adversarial samples’ entropy for sample rectification, and then achieve test-time defense. In this article, we investigate the entropy properties of adversarial samples and obtain two observations: 1) adversarial samples are often confidently misclassified despite having low prediction entropy and 2) samples with higher attack strength typically show lower prediction entropy. Therefore, we believe directly minimizing the entropy of adversarial samples is not reasonable and propose a two-stage self-adversarial rectification approach: \underline{Re}ctified \underline{A}dversaria\underline{l} Sample via Max-Min Entropy for Test-Time Defense (REAL), consisting of a max-min entropy optimization scheme and an attack-aware weighting mechanism, which can be embedded in the existing models as a plugged-played block. Experiments on several datasets show that REAL can greatly improve the performance of existing sample rectification model.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7541