Intrusion Detection for Industrial Control Systems Based on Improved Contrastive Learning SimCLR
Chengcheng Li, Fei Li, Liyan Zhang, Aimin Yang, Zhibin Hu, Ming He
Abstract
Since supervised learning intrusion detection models rely on manually labeled data, the process often requires a lot of time and effort. To make full use of unlabeled network traffic data and improve intrusion detection, this paper proposes an intrusion detection method for industrial control systems based on improved comparative learning SimCLR. Firstly, a feature extraction network is trained on SimCLR using unlabeled data; a linear classification layer is added to the trained feature extraction network model; and a small amount of labeled data is used for supervised training and fine-tuning of the model parameters. The trained model is simulated on the Secure Water Treatment (SWaT) dataset and the publicly available industrial control dataset from Mississippi State University, and the results show that the method has better results in all evaluation metrics compared with the deep learning algorithm using supervised learning directly, and the comparative learning has research value in industrial control system intrusion detection.