Reward Poisoning Attacks on Offline Multi-Agent Reinforcement Learning

In offline multi-agent reinforcement learning (MARL), agents estimate policies from a given dataset. We study reward-poisoning attacks in this setting where an exogenous attacker modifies the rewards dataset before see The wants to guide each agent into nefarious target policy while minimizing Lp norm of reward modification. Unlike on single-agent RL, we show that can install as Markov Perfect Dominant Strategy Equilibrium (MPDSE), which rational are guaranteed follow. This attack be significantly cheaper than separate attacks. works various MARL including uncertainty-aware learners, and exhibit linear programs efficiently solve problem. also relationship between structure datasets minimal cost. Our work paves way for studying defense MARL.