Reproduce “Don't Trust Stubborn Neighbors: A Security Framework for Agentic Networks” on OpenResearch

About this paper

Samira Abedini, Sina Mavali, Lea Schönherr, Martin Pawelczyk, Rebekka Burkholz

Researchers at CISPA Helmholtz Center for Information Security and the University of Vienna established a security framework for LLM-based Multi-Agent Systems by adapting the Friedkin Hendash Johnsen opinion dynamics model. This framework effectively describes and predicts agent behavior, demonstrating how network topology influences adversarial success and leading to the development of trust-adaptive defense mechanisms that mitigate cascade attacks, including those from adaptive adversaries.

Key insights

• The Friedkin Hendash Johnsen opinion dynamics model accurately describes and predicts the complex belief formation and propagation in LLM-based Multi-Agent Systems across various LLM families and network topologies.
• A single stubborn and influential agent can dominate system dynamics, with attack success significantly dependent on the attacker's structural position (e.g., hub in a star network) and the agreeableness of benign agents.
• Dynamic and continuous trust management is crucial for mitigating cascade attacks, particularly against sophisticated adaptive adversaries who can initially mask their malicious intent.

Summary by alphaXiv.