Error Correcting by Agreement Checking for Adversarial Robustness against Black-box Attacks

Ashutosh Kumar Nirala, Jin Tian, Olukorede Fakorede, Modeste Atsague

Keywords: adversarial defense; AT; black-box; SQA

TL;DR: A biologically inspired error correction mechanism, ECAC, that enhances model robustness by exploiting the distinct classification features of naturally and adversarially trained models against a spectrum of black-box attacks.

Abstract:

Drawing inspiration from the vulnerability of the initial feed-forward phase of biological perception in humans and primates to adversarial attacks, we propose a novel defense strategy named Error Correcting by Agreement Checking (ECAC). This strategy is designed to mitigate realistic \emph{black-box} threats where attackers don't have full access to the model. We exploit the fact that natural and adversarially trained models rely on distinct feature sets for classification.
Notably, naturally trained models retain commendable accuracy against adversarial examples generated using adversarially trained models. Leveraging this disparity, ECAC moves the input toward the prediction of the naturally trained model unless it leads to disagreement in prediction between the two models, before making the prediction.
This simple error correction mechanism is highly effective against leading SQA (Score-based Query Attacks) black-box attacks as well as decision-based and transfer-based black-box attacks.
We also verify that, unlike other black-box defense, ECAC maintains significant robustness even when adversary has full access to the model. We demonstrate its effectiveness through comprehensive experiments across various datasets (CIFAR and ImageNet) and architectures (ResNet as well as ViT).

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 10477