Cohort-based kernel principal component analysis with Multi-path Service Routing in Federated Learning

Abstract

Federated Learning (FL) is a machine learning (ML) strategy that is performed in a decentralized environment. The training is performed locally by the client on the global model shared by the server. Federated learning has recently been used as a service (FLaaS) to provide a collaborative training environment to independent third-party applications. However, the widespread adoption in distributed settings of FL has opened venues for a number of security attacks. A number of studies have been performed to prevent multiple FL attacks. However, sophisticated attacks, such as label-flipping attacks, have received little or no attention. From the said perspective, this research is focused on providing a defense mechanism for the aforesaid attack. The proposed approach is based on Type-based Cohorts (TC) with Kernel Principal Component Analysis (KPCA) to detect and defend against label-flipping attacks. Moreover, to improve the performance of the network, we will deploy Multi-path Service Routing (MSR) for edge nodes to work effectively. The KPCA will be used to secure the network from attacks. The proposed mechanism will provide an effective and secure FL system. The proposed approach is evaluated with respect to the following measures: execution time, memory consumption, information loss, accuracy, service request violations, and the request’s waiting time.