Deep Learning Anomaly Classification Using Multi-Attention Residual Blocks for Industrial Control Systems

This paper proposes a novel method monitoring network packets to classify anomalies in industrial control systems (ICSs). The proposed method combines different mechanisms. It is flow-based as it obtains new features through aggregating packets of the same flow. It then builds a deep neural network (DNN) with multi-attention blocks for spotting core features, and with residual blocks for avoiding the gradient vanishing problem. The DNN is trained with the Ranger (RAdam + Lookahead) optimizer to prevent the training from being stuck in local minima, and with the focal loss to address the data imbalance problem. The Electra Modbus dataset is used to evaluate the performance impacts of different mechanisms on the proposed method. The proposed method is compared with related methods in terms of the precision, recall, and F1-score to show its superiority.