Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SentinelNet: Safeguarding Multi-Agent Collaboration Through Credit-Based Dynamic Threat Detection

Yang Feng, Xudong Pan

TL;DR

SentinelNet tackles the vulnerability of large-language-model-powered multi-agent systems to malicious participants by introducing a decentralized defense built into sentinel nodes. The approach combines adversarial trajectory generation with contrastive learning to produce a credit-based detector, and employs bottom-k elimination for autonomous isolation during debates. Empirical results across six benchmarks show near-perfect detection within two rounds and up to 95% recovery of system accuracy from compromised baselines, with strong cross-domain generalization. This work offers a practical, scalable path toward robust, distributed security in collaborative MAS, enabling safe deployment in high-stakes settings like medicine and finance.

Abstract

Malicious agents pose significant threats to the reliability and decision-making capabilities of Multi-Agent Systems (MAS) powered by Large Language Models (LLMs). Existing defenses often fall short due to reactive designs or centralized architectures which may introduce single points of failure. To address these challenges, we propose SentinelNet, the first decentralized framework for proactively detecting and mitigating malicious behaviors in multi-agent collaboration. SentinelNet equips each agent with a credit-based detector trained via contrastive learning on augmented adversarial debate trajectories, enabling autonomous evaluation of message credibility and dynamic neighbor ranking via bottom-k elimination to suppress malicious communications. To overcome the scarcity of attack data, it generates adversarial trajectories simulating diverse threats, ensuring robust training. Experiments on MAS benchmarks show SentinelNet achieves near-perfect detection of malicious agents, close to 100% within two debate rounds, and recovers 95% of system accuracy from compromised baselines. By exhibiting strong generalizability across domains and attack patterns, SentinelNet establishes a novel paradigm for safeguarding collaborative MAS.

SentinelNet: Safeguarding Multi-Agent Collaboration Through Credit-Based Dynamic Threat Detection

TL;DR

SentinelNet tackles the vulnerability of large-language-model-powered multi-agent systems to malicious participants by introducing a decentralized defense built into sentinel nodes. The approach combines adversarial trajectory generation with contrastive learning to produce a credit-based detector, and employs bottom-k elimination for autonomous isolation during debates. Empirical results across six benchmarks show near-perfect detection within two rounds and up to 95% recovery of system accuracy from compromised baselines, with strong cross-domain generalization. This work offers a practical, scalable path toward robust, distributed security in collaborative MAS, enabling safe deployment in high-stakes settings like medicine and finance.

Abstract

Malicious agents pose significant threats to the reliability and decision-making capabilities of Multi-Agent Systems (MAS) powered by Large Language Models (LLMs). Existing defenses often fall short due to reactive designs or centralized architectures which may introduce single points of failure. To address these challenges, we propose SentinelNet, the first decentralized framework for proactively detecting and mitigating malicious behaviors in multi-agent collaboration. SentinelNet equips each agent with a credit-based detector trained via contrastive learning on augmented adversarial debate trajectories, enabling autonomous evaluation of message credibility and dynamic neighbor ranking via bottom-k elimination to suppress malicious communications. To overcome the scarcity of attack data, it generates adversarial trajectories simulating diverse threats, ensuring robust training. Experiments on MAS benchmarks show SentinelNet achieves near-perfect detection of malicious agents, close to 100% within two debate rounds, and recovers 95% of system accuracy from compromised baselines. By exhibiting strong generalizability across domains and attack patterns, SentinelNet establishes a novel paradigm for safeguarding collaborative MAS.
Paper Structure (35 sections, 12 equations, 4 figures, 1 table, 1 algorithm)

This paper contains 35 sections, 12 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminary
  4. Task Formulation
  5. Security Settings
  6. Adversary Capabilities and Goals.
  7. Defender Assumptions
  8. Framework of SentinelNet
  9. Adversarial Trajectory Generation
  10. Contrastive Training of SentinelNet
  11. Constructing the Training Dataset for Contrastive Learning
  12. Loss Function Design
  13. Defense Integration
  14. Runtime Credit Scoring
  15. Adaptive Isolation Mechanism
  16. ...and 20 more sections

Figures (4)

  • Figure 1: Comparison of (a) a traditional MAS without defenses and (b) a SentinelNet-enhanced MAS with sentinel nodes for detecting malicious behavior and improving accuracy and recovery.
  • Figure 2: Overview of the SentinelNet framework, which transforms agents into sentinel nodes for proactive threat detection through three stages: adversarial trajectory generation, contrastive learning-based training, and dynamic ranking with bottom-k elimination. It overcomes centralized defense limitations, providing scalable protection across diverse attack scenarios.
  • Figure 3: Task accuracy trends across three debate rounds under six different attack methods on six benchmark datasets.
  • Figure 4: Comparison of SentinelNet with the baselines across six multi-agent debate benchmarks, where SentinelNet consistently outperforms the baselines in terms of Detection Accuracy, False Positive Rate (FPR), and False Negative Rate (FNR).