Reproduce “Content-based Unrestricted Adversarial Attack” 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

Zhaoyu Chen, Bo Li, Shuang Wu, Kaixun Jiang, Shouhong Ding, Wenqiang Zhang

Researchers from Fudan University and Tencent's Youtu Lab developed Adversarial Content Attack (ACA), a new unrestricted adversarial attack method that leverages the latent space of Stable Diffusion to generate photorealistic adversarial examples. ACA achieves 13.3% to 50.4% higher transferability against normally trained models and consistently bypasses state-of-the-art defenses, setting new benchmarks for attack efficacy and realism.

Key insights

  • Leveraging the latent space of powerful generative models like Stable Diffusion allows for the creation of adversarial examples that are intrinsically photorealistic and semantically consistent, capable of diverse content modifications.
  • Efficiently computing gradients through complex diffusion processes can be achieved via 'skip gradient' approximations, significantly reducing computational load and memory requirements.
  • Existing adversarial defenses, including input pre-processing and adversarial training, exhibit a 'Dunning-Kruger effect,' performing poorly against novel, unrestricted attacks and underscoring a need for more generalizable robustness.

Summary by alphaXiv.

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