[TOC]
- Title: Reward Poisoning Attacks on Offline Multi Agent Reinforcement Learning
- Author: Young Wu et. al.
- Publish Year: 1 Dec 2022
- Review Date: Tue, Dec 27, 2022
Summary of paper
Motivation
Contribution
- unlike attacks on single-agent RL, we show that the attacker can install the target poilcy as a Markov Perfect Dominant Strategy Equilibrium (MPDSE), which rational agents are guaranteed to follow.
- This attack can be significantly cheaper than separate single-agent attacks.
Limitation
- the goal of the attacker is still wanting the agents to learn a target policy $\pi^\dagger$ rather than slowing down the learning speed.
Some key terms
susceptible
- while the above empirical success is encouraging, MARL algorithms are susceptible to data poisoning attacks: the agent can reach the wrong equilibria if an exogenous attacker manipulate the feedback to agents.