OpenEvoShield: Dual Non-Stationary Continual Defense for Open-World Multi-Agent System Attacks

Short summary

OpenEvoShield is a co-evolutionary continual defense framework for LLM-based multi-agent systems that addresses doubly dynamic threats where adversaries refine injection strategies while normal-agent behavior drifts. It combines an asymmetric rate controller, dynamic behavioral boundary updater, EWC-regularized policy ensemble, and energy-based multi-granularity detector. Across 100 deployment rounds and five benchmarks, it outperforms static and continual baselines, detecting most unseen attacks while maintaining low false positive rates.

• •Co-evolutionary defense framework for LLM multi-agent systems against dynamic adversarial instruction injection
• •Four-component architecture handles dual drift: fast attack-side adaptation and slow normal-side boundary updates
• •Outperforms static and continual baselines across 100 deployment rounds on five benchmarks and four MAS topologies

Generated with AI, which can make mistakes.