Abstract
Abstract
Power grid anomaly detection requires modeling both network topology and device dependencies. Graph neural networks (GNNs) are suitable for this task. However, few prior works have considered the stealthy camouflage behavior of grid anomalies, where faulty devices or malicious attackers intentionally mask their abnormal signals to resemble normal operational patterns. Such camouflages—specifically feature camouflage (manipulating local measurement data) and relation camouflage (establishing deceptive connections with healthy clusters)—significantly degrade the performance of existing GNN-based detectors. This paper proposes a camouflage-resistant graph neural network framework for power grid anomaly detection, termed CR-PGNN. Specifically, the framework first devises a physics-aware similarity measure to quantify the consistency between neighboring equipment based on electrical laws. It then leverages reinforcement learning (RL) to adaptively select the most informative neighbors and relations, effectively filtering out “camouflaged” connections that might dilute anomaly signals. A relation-aware aggregator fuses information across multiple grid relations. Extensive experiments on IEEE benchmark power systems with simulated PMU measurements and synthetically injected stealthy anomalies demonstrate that CR-PGNN outperforms state-of-the-art GNN-based detectors in terms of both detection performance and robustness against camouflage attacks. Code and data are publicly available at
https://github.com/123gaoyi/CR-PGNN
.