Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PeerGuard: Defending Multi-Agent Systems Against Backdoor Attacks Through Mutual Reasoning

Falong Fan, Xi Li

TL;DR

The paper addresses safety gaps in multi-agent systems by defending against backdoor attacks through mutual reasoning and inconsistency detection during agent debates. It introduces PeerGuard, a defense integrated into two-agent (and extendable to multi-agent) MAS, where agents generate reasoning steps and evaluate each other’s justifications to identify malicious manipulation. Empirical results on GPT-4o and Llama3 show PeerGuard achieves high detection performance with low false positives and generalizes to other MAS frameworks like AutoGen and CAMEL, without compromising benign task performance. This work advancing robustness and trustworthiness in inter-agent AI interactions, with practical plug-and-play applicability in API-based MAS deployments.

Abstract

Multi-agent systems leverage advanced AI models as autonomous agents that interact, cooperate, or compete to complete complex tasks across applications such as robotics and traffic management. Despite their growing importance, safety in multi-agent systems remains largely underexplored, with most research focusing on single AI models rather than interacting agents. This work investigates backdoor vulnerabilities in multi-agent systems and proposes a defense mechanism based on agent interactions. By leveraging reasoning abilities, each agent evaluates responses from others to detect illogical reasoning processes, which indicate poisoned agents. Experiments on LLM-based multi-agent systems, including ChatGPT series and Llama 3, demonstrate the effectiveness of the proposed method, achieving high accuracy in identifying poisoned agents while minimizing false positives on clean agents. We believe this work provides insights into multi-agent system safety and contributes to the development of robust, trustworthy AI interactions.

PeerGuard: Defending Multi-Agent Systems Against Backdoor Attacks Through Mutual Reasoning

TL;DR

The paper addresses safety gaps in multi-agent systems by defending against backdoor attacks through mutual reasoning and inconsistency detection during agent debates. It introduces PeerGuard, a defense integrated into two-agent (and extendable to multi-agent) MAS, where agents generate reasoning steps and evaluate each other’s justifications to identify malicious manipulation. Empirical results on GPT-4o and Llama3 show PeerGuard achieves high detection performance with low false positives and generalizes to other MAS frameworks like AutoGen and CAMEL, without compromising benign task performance. This work advancing robustness and trustworthiness in inter-agent AI interactions, with practical plug-and-play applicability in API-based MAS deployments.

Abstract

Multi-agent systems leverage advanced AI models as autonomous agents that interact, cooperate, or compete to complete complex tasks across applications such as robotics and traffic management. Despite their growing importance, safety in multi-agent systems remains largely underexplored, with most research focusing on single AI models rather than interacting agents. This work investigates backdoor vulnerabilities in multi-agent systems and proposes a defense mechanism based on agent interactions. By leveraging reasoning abilities, each agent evaluates responses from others to detect illogical reasoning processes, which indicate poisoned agents. Experiments on LLM-based multi-agent systems, including ChatGPT series and Llama 3, demonstrate the effectiveness of the proposed method, achieving high accuracy in identifying poisoned agents while minimizing false positives on clean agents. We believe this work provides insights into multi-agent system safety and contributes to the development of robust, trustworthy AI interactions.
Paper Structure (12 sections, 2 figures, 8 tables)

This paper contains 12 sections, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Threat Model and Defender's Assumptions
  5. Classic Two-Agent Systems
  6. PeerGuard in Two-Agent Systems
  7. Generalization to Multi-Agent Systems
  8. Experiment
  9. Experimental Setup
  10. Experimental Results
  11. Limitation and Future Work
  12. Conclusion

Figures (2)

  • Figure 1: Illustration of a backdoor attack on an LLM-based agent (left) and an overview of PeerGuard defense strategy in a two-agent system (right).
  • Figure 2: TPR of the proposed method in various multi-agent frameworks in S1 setting.