The Certification Paradox: Certifications Admit Better Evasion Attacks
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: certified robustness, adversarial attacks, risk, randomised smoothing
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Certified defences can be exploited by attackers to produce better adversarial attacks.
Abstract: In guaranteeing the absence of adversarial examples in bounded spaces, certification mechanisms play an important role in demonstrating neural network robustness. Within this work we ask if certifications themselves can potentially compromise the very models they help to protect? By demonstrating a new attack surface that exploits certified guarantees to construct norm minimising evasion attacks, we demonstrate the heretofore unexplored risks inherent in releasing certifications. Our new *Certification Aware Attack* produces smaller, more difficult to detect adversarial examples more than $74$% of the time than comparable attacks, while reducing the median perturbation norm by more than $10$%. That this is achievable in significantly less computational time highlights an apparent paradox---that releasing certifications can reduce security.
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: 7024
Loading