Abstract
In industrial control systems (ICSs), safeguarding system security is vital, and intrusion detection is an effective protective measure. In real-world ICS environments, however, the training data for intrusion-detection models are dominated by normal traffic, with very few abnormal samples. This scarcity prevents the model from fully learning abnormal characteristics, thereby limiting anomaly-detection performance. ICS traffic is also noisy and often captured incompletely, which is especially problematic because it possesses strong temporal dependencies. Moreover, ICS operations tend to be highly regular and follow a strict logical order, causing data to cluster densely in certain regions while edge regions remain sparsely populated—an imbalance that easily leads to misclassification. To tackle data imbalance, missing temporal features, and edge-sample scarcity, we propose the DualCGAN-AE approach. It employs two conditional generative adversarial networks to perform dynamic learning on global and edge features and to enrich samples with temporal information; a multimodel discriminator is incorporated to enhance training stability. We validated the method on two datasets, assessing both the quality of the generated samples and their impact on intrusion detection performance. Experimental results show that the samples produced by DualCGAN-AE significantly boost detection accuracy.
Funding Information
  • National Natural Science Foundation of China (U22B2026)

This publication has 28 references indexed in Scilit: