Consensus-Guided Redundant Memory Isolation for Poisoning Mitigation in Collaborative Agent Frameworks

Le résumé fourni par la source

Memory poisoning can destabilize collaborative decision-making when agents rely on shared state synchronization. This study introduces a consensus-guided redundant memory isolation mechanism to prevent cross-agent contamination. Agents maintain parallel memory replicas and compute consensus confidence scores using weighted majority voting across trusted peers. Memory segments with confidence below 0.65 are temporarily isolated and revalidated before reintegration. Evaluation was conducted on a distributed task allocation benchmark with 60 agents under varying poisoning rates (5%–25%). The isolation mechanism reduced collective performance degradation by 42.1% and limited contamination spread to fewer than 18% of agents, compared with 61% under standard synchronization. Communication overhead increased by 12.7%, remaining within acceptable system bounds. The approach strengthens robustness while maintaining scalability in collaborative environments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.