PerCFed: An Effective Personalized Clustered Federated Learning Mechanism to Handle non-IID Challenges for Industry 4.0

Federated learning (FL) has emerged as an effective solution to build a collaborative intrusion detection model for Industry 4.0 in a privacy-preserved way. However, due to the use of real-time sensors and different Industrial Internet of Things (IIoT) devices, it results in heterogeneous or non-Independent and Identically Distributed (non-IID) data. This non-IID nature of data affects the FL-based intrusion detection models performance and degrades the accuracy. Moreover, FL also suffers from model poisoning and data poisoning i.e., even with a few poor local models can affect the performance of the overall global model. Therefore, this work proposes a Personalized Clustered Federated learning (PerCFed) mechanism for Industry 4.0 to achieve collaboration between similar clients, thereby reducing the effect of non-IID and poisoning attack challenges. The proposed PerCFed also balances the models training between learning from the previous training and collaboratively learning for intrusion detection and reduces the impact of the above challenges on the FL-based training. A customized model built using PerCFed could thus achieve increased intrusion detection accuracy and more effectively fit the majority of various industry data types. Experimental results shows that our proposed mechanism maintains high accuracy for intrusion detection compared to other state-of-the-art techniques with non-IID data and poisoning attacks.