Adversarial attacks in consensus-based multi-agent reinforcement learning

TL;DR

This paper investigates how adversarial attacks can manipulate consensus-based multi-agent reinforcement learning systems, revealing their vulnerability to malicious agents influencing collective policies.

Contribution

It demonstrates that adversarial agents can successfully persuade all agents in a consensus-based MARL network to adopt harmful policies, exposing a critical security flaw.

Findings

Adversarial agents can influence entire networks to adopt malicious policies

Consensus-based MARL algorithms are vulnerable to adversarial manipulation

Standard algorithms lack robustness against targeted attacks

Abstract

Recently, many cooperative distributed multi-agent reinforcement learning (MARL) algorithms have been proposed in the literature. In this work, we study the effect of adversarial attacks on a network that employs a consensus-based MARL algorithm. We show that an adversarial agent can persuade all the other agents in the network to implement policies that optimize an objective that it desires. In this sense, the standard consensus-based MARL algorithms are fragile to attacks.