Reproduce “Characterizing Attacks on Deep Reinforcement Learning” 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

Xinlei Pan, Chaowei Xiao, Warren He, Shuang Yang, Jian Peng, Mingjie Sun, Jinfeng Yi, Zijiang Yang and 3 more

Pan et al. developed efficient black-box and online sequential adversarial attacks against Deep Reinforcement Learning agents, along with novel environment dynamics attacks. Their research demonstrated the first successful physical adversarial attacks on a real DRL-controlled robot, leading to suboptimal behavior and collisions.

Key insights

  • Black-box adversarial attacks on DRL can be significantly optimized for query efficiency through adaptive sampling while retaining high effectiveness.
  • Targeting states with high Q-value variance provides a more potent and efficient strategy for adversarial attacks in sequential decision-making environments.
  • Deep Reinforcement Learning systems are susceptible to adversarial modifications of environment dynamics, not just sensory inputs, and these vulnerabilities extend to physical DRL systems.

Summary by alphaXiv.

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