G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems

2025-02-16

AI Summary

LLM-driven multi-agent systems (MAS) exhibit critical security vulnerabilities—including susceptibility to adversarial attacks, misinformation propagation, and unintended behaviors—hindering their deployment in safety-critical applications. Method: This paper proposes a topology-aware proactive defense framework that models inter-agent dialogues as directed graphs, integrating graph neural networks with MAS interaction topology to enable interpretable anomaly detection and structured intervention. It introduces a novel topology-guided intervention mechanism and prompt-injection robustness enhancement technique, designed for cross-architecture compatibility and plug-and-play integration. Contribution/Results: Extensive experiments demonstrate that the framework achieves over 40% average performance recovery across diverse attack types. It maintains full compatibility with mainstream LLMs and large-scale MAS deployments, significantly improving both system security assurance and operational deployability without compromising functional integrity.

Technology Category

• Multiagent Systems: Adversarial Agents
• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Large Language Model (LLM)-based Multi-agent Systems (MAS) have demonstrated remarkable capabilities in various complex tasks, ranging from collaborative problem-solving to autonomous decision-making. However, as these systems become increasingly integrated into critical applications, their vulnerability to adversarial attacks, misinformation propagation, and unintended behaviors have raised significant concerns. To address this challenge, we introduce G-Safeguard, a topology-guided security lens and treatment for robust LLM-MAS, which leverages graph neural networks to detect anomalies on the multi-agent utterance graph and employ topological intervention for attack remediation. Extensive experiments demonstrate that G-Safeguard: (I) exhibits significant effectiveness under various attack strategies, recovering over 40% of the performance for prompt injection; (II) is highly adaptable to diverse LLM backbones and large-scale MAS; (III) can seamlessly combine with mainstream MAS with security guarantees. The code is available at https://github.com/wslong20/G-safeguard.

Problem

• Enhance security in LLM-based multi-agent systems
• Detect and remediate adversarial attacks effectively
• Ensure system adaptability across diverse LLM backbones

Innovation

• Graph neural networks for anomaly detection
• Topological intervention for attack remediation
• Adaptable to diverse LLM backbones