Local Anomaly Detection with Partial Observation in Multi-agent Systems as a Data Matching Game

Local anomaly detection in a multi-agent system is a pervasive but challenging problem. The challenge entails how agents with heterogeneous objectives and partial data collection train local anomaly detectors for heterogeneous domain-specific tasks. This paper proposes a distributed training method to address this question. Our approach involves a game-theoretic framework to address agents' heterogeneous objectives and a transformer-based model to handle partial data observation. Our game, conditionally proven as a potential game, guides agents under the same local objectives into a data-sharing group for local training. Compared to other topperforming SOTAs, our evaluation outcomes empirically reflect the efficiency and robustness of our method in multi-agent scenarios.