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

This dataset contains adversarial debate trajectories and agent interaction sequences used to train SentinelNet, a decentralized framework for detecting malicious behaviors in multi-agent systems. It includes labeled examples of benign and malicious messages along with features for contrastive learning-based credit evaluation. The data enables reproduction of SentinelNet experiments and supports research on securing Web-based collaborative multi-agent environments.