Reproduce “When Flatness Does (Not) Guarantee Adversarial Robustness” 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

Nils Philipp Walter, Linara Adilova, Jilles Vreeken, Michael Kamp

This work by Walter et al. (2025) investigates the link between neural network loss landscape flatness and adversarial robustness, demonstrating that while flatness offers local guarantees, adversarial examples can paradoxically inhabit 'Uncanny Valley' regions of high confidence and low sharpness despite misclassification. It provides theoretical bounds showing an inverse relationship between sharpness and the certified adversarial radius, alongside empirical confirmation of this phenomenon.

Key insights

  • Flatness in parameter space guarantees local adversarial robustness, directly increasing the certified perturbation radius within which loss increase remains bounded.
  • Relative sharpness is inversely related to model confidence, implying that high confidence in a prediction, even an incorrect one, can locally flatten the loss surface.
  • Adversarial examples can reside in 'Uncanny Valley' regions, which are exceptionally flat and high-confidence areas despite leading to model misclassification.

Summary by alphaXiv.

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