Reproduce “GroupGuard: A Framework for Modeling and Defending Collusive Attacks in Multi-Agent Systems” on OpenResearch

About this paper

Yiling Tao, Xinran Zheng, Shuo Yang, Meiling Tao, Xingjun Wang

GroupGuard introduces a framework for modeling and defending against novel group collusive attacks in multi-agent LLM systems, which leverage sociological strategies like false consensus and targeted suppression to achieve higher attack success rates than individual attacks. The framework employs a three-layer defense that monitors interactions, probes suspicious agents, and prunes malicious connections, achieving improved detection accuracy and significant performance recovery across diverse settings.

Key insights

• Group collusive attacks are more potent and covert than individual attacks, leveraging social dynamics to amplify malicious influence and undermine collective decision-making in LLM-based MAS.
• A multi-layered defense combining real-time graph-based interaction analysis, active probing with optimized honeypot questions, and structural isolation is effective against complex, coordinated threats.
• The framework's training-free nature and generalizability across diverse LLM architectures, communication topologies, and tasks provide a practical and robust security solution for dynamic multi-agent environments.

Summary by alphaXiv.