Reproduce “Universal and Transferable Attacks on Pathology Foundation Models” 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

Yuntian Wang, Xilin Yang, Che-Yung Shen, Nir Pillar, Aydogan Ozcan

Researchers at UCLA developed Universal and Transferable Adversarial Perturbations (UTAP) that generate a single, fixed, visually imperceptible noise pattern capable of compromising state-of-the-art pathology foundation models. This perturbation consistently degrades feature representations and reduces classification accuracy by over 80% on both seen and black-box models, including out-of-distribution data.

Key insights

  • Pathology foundation models exhibit fundamental vulnerabilities to universal, transferable attacks that corrupt upstream feature representations rather than merely altering decision boundaries.
  • Crafting adversarial perturbations by targeting feature space similarity effectively enables broad transferability across diverse black-box models and out-of-distribution data.
  • Strategic regularization during attack generation, such as random masking and attention dropping, is critical for developing perturbations that generalize effectively.

Summary by alphaXiv.

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