New research precisely characterizes reward poisoning vulnerabilities in reinforcement learning

This paper provides a precise characterization of when reinforcement learning agents are vulnerable to reward poisoning attacks. The research focuses on linear Markov Decision Processes (MDPs) and establishes necessary and sufficient conditions for an adversary to successfully manipulate an agent's policy within a constrained budget. The framework extends to deep reinforcement learning by approximating environments as linear MDPs, demonstrating its theoretical and practical significance in identifying and exploiting attackable agents. AI

IMPACT Provides theoretical groundwork for understanding and mitigating adversarial attacks in reinforcement learning systems.

RANK_REASON The cluster contains an academic paper detailing theoretical research on reinforcement learning vulnerabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• end-to-end reinforcement learning
• Gotit.pub
• Haoyang Hong
• Hugging Face
• IArxiv
• Linear MDPs
• reinforcement learning
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →