Privacy-Preserving Federated Distillation GAN for CIDSs in Industrial CPSs

Intrusion Detection System (IDS) is an effective way to detect both internal and external abnormal behaviors, which has been widely deployed in industrial Cyber-Physical Systems (CPSs). However, due to the data island problem caused by the imperativeness of confidentiality of sensitive information, most existing IDSs are limited to be trained and evaluated in isolated CPSs, resulting in the cyber systems vulnerable to various newly-emerging attacks. In this article, a secure and collaborative IDS solution, called PFD-GAN, is proposed. Specifically, we firstly develop a novel semi-supervised IDS model by improving External Classifier (EC)-Generative Adversarial Network (GAN) with Wasserstein distance and label condition, to strengthen the classification performance through the use of synthetic data. Furthermore, Local Differential Privacy (LDP) is adopted to prevent against adversaries learning sensitive information in collaboration. Moreover, a Decentralized Federated Distillation (DFD)-based collaboration is designed, allowing multiple industrial CPSs to collectively build a comprehensive IDS to recognize the threats under the entire cyber systems without sharing a uniform template model. Experimental evaluation and theory analysis demonstrate that the proposed PFD-GAN is secure from the threats of privacy leaking and highly effective in detecting various types of attacks on industrial CPSs.