Reproduce “On Adaptive Attacks to Adversarial Example Defenses” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Florian Tramer, Nicholas Carlini, Wieland Brendel, Aleksander Madry

The authors show that adaptive evaluation must target each defense’s failure modes, rather than rely on a familiar attack. They investigate 13 defenses from ICLR, ICML, and NeurIPS, using defense-specific objectives and attack methods. On CIFAR-10, a score-based attack reduces k-WTA’s reported 50.0% accuracy to 0.16% under an ℓ∞ bound of 8/255; the result covers selected defenses, not all defenses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.