CAGCN: Centrality-Aware Graph Convolution Network for Anomaly Detection in Industrial Control Systems

In industrial control systems, the utilization of deep-learning-based methods achieves improvements for anomaly detection. However, most current methods ignore the association of inner components in industrial control systems. In industrial control systems, an anomaly component may affect the neighboring components therefore the connective relationship can help us to detect anomalies effectively. In this paper, we propose a centrality-aware graph convolution network (CAGCN) for anomaly detection in industrial control systems. Unlike the traditional graph convolution network model (GCN), we utilize the concept of centrality to enhance the ability of graph convolution networks to deal with the inner relationship in industrial control systems. Our experiments show that compared with GCN, our CAGCN has a better ability to utilize this relationship between components in industrial control systems. The performances of the model are evaluated on the Secure Water Treatment (SWaT) dataset and the Water Distribution (WADI) dataset, the two most common industrial control systems datasets in the field of industrial anomaly detection. The experimental results show that our CAGCN achieves better results on precision, recall, and F1 score than the state-of-the-art methods.