Reproduce “Diffusion Policy Attacker: Crafting Adversarial Attacks for Diffusion-based Policies” 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

Yipu Chen, Haotian Xue, Yongxin Chen

This paper introduces DP-Attacker, the first comprehensive framework to analyze the adversarial vulnerability of diffusion policies in robotic control. The framework demonstrates that both imperceptible digital perturbations and physical adversarial patches can effectively disrupt robotic task completion, often reducing success rates by over 50%.

Key insights

  • The encoder component, which maps raw observations to latent representations, is the most susceptible part of diffusion policies to adversarial attacks, exhibiting significant feature distortion from small input perturbations.
  • Attacks that target the internal noise prediction mechanism of diffusion models are both more computationally efficient and effective than end-to-end optimization through the entire diffusion process.
  • Offline, pre-computed adversarial perturbations can effectively transfer across different frames and timesteps, demonstrating that real-time, computationally intensive attacks are not always necessary to disrupt policy performance.

Summary by alphaXiv.

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