Abstract:
The data of industrial cyber physical systems (ICPS) are imbalanced. As such, the detection rate of rare attacks based on deep learning is low. In this study, an enhanced attack detection model for industrial CPS is proposed to address this problem. The model selects rare attack samples from the original data set. It then uses the gradient penalty Wasserstein distance generative adversarial network to expand the rare attack samples. Afterward, it mixes the expanded samples with the original data set to form a new data set for training multilayer perception and consequently detect an attack. A real data set with Modbus as the communication protocol is used for testing to verify the performance of the model. Experimental results show that the proposed model can significantly improve the ability to detect an attack compared with that of widely used data augmentation methods.