Reproduce “Generative Adversarial Perturbations” 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

Omid Poursaeed, Isay Katsman, Bicheng Gao, Serge Belongie

This paper introduces Generative Adversarial Perturbations (GAP), a framework utilizing deep generative models to produce adversarial examples at significantly faster speeds than previous iterative methods. The work demonstrates effective universal and image-dependent attacks for both classification and semantic segmentation, including novel targeted universal perturbations on ImageNet.

Key insights

  • Generative models enable real-time adversarial attacks, fundamentally altering the feasibility and threat landscape compared to slow iterative methods.
  • A single, trainable generator architecture can effectively create diverse attack types (targeted, non-targeted, universal, image-dependent) and generalize across different computer vision tasks like classification and semantic segmentation.
  • The success of targeted universal perturbations on large-scale datasets like ImageNet indicates that models exhibit consistent exploitable biases across varied inputs.

Summary by alphaXiv.

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