Reproduce “Delving into adversarial attacks on deep policies” on OpenResearch

About this paper

Jernej Kos, Dawn Song

Research from the National University of Singapore and UC Berkeley demonstrates that deep reinforcement learning policies are highly susceptible to small, imperceptible adversarial perturbations. It introduces a value function-guided method for efficient attack injection and explores re-training as a defense, finding that while resilience can be improved, the policy's decision boundaries often remain fragmented.

Key insights

• Adversarial perturbations are orders of magnitude more effective at degrading DRL agent performance compared to random noise, even at very low magnitudes.
• Utilizing the agent's internal value function to guide perturbation injection allows for highly efficient attacks, significantly degrading performance with infrequent intervention.
• Re-training DRL agents with adversarial examples or noise can confer resilience that generalizes across perturbation magnitudes, though this does not fundamentally smooth the policy's highly fragmented decision boundaries.

Summary by alphaXiv.