Robust Memory Update Mechanisms Against Poisoning Attacks in Multi-Agent Reinforcement Learning

Abstract

In multi-agent reinforcement learning (MARL), shared replay buffers and inter-agent memory exchange introduce vulnerability to memory poisoning attacks. This work proposes a confidence-weighted memory validation mechanism integrated into the experience sharing pipeline. Each memory entry is assigned a trust score derived from temporal consistency and reward deviation metrics. A Bayesian filtering process excludes anomalous transitions before propagation to peer agents. Experiments were conducted on cooperative navigation and resource allocation benchmarks with 12–24 agents. Under a 20% poisoning injection rate, baseline MARL performance dropped by 37.8%, whereas the proposed defense limited degradation to 11.4%. Convergence time improved by 23.6% compared to anomaly-blind training.The method effectively mitigates adversarial memory contamination in collaborative reinforcement learning environments.