FedSC: A federated learning algorithm based on client-side clustering

Abstract

<abstract><p>In traditional centralized machine learning frameworks, the consolidation of all data in a central data center for processing poses significant concerns related to data privacy breaches and data sharing complexities. In contrast, federated learning presents a privacy-preserving paradigm by training models on local devices, thus circumventing the need for data transfer. However, in the case of non-IID (non-independent and identically distributed) data distribution, the performance of federated learning will drop. Addressing this predicament, this study introduces the FedSC algorithm as a remedy. The FedSC algorithm initially partitions clients into clusters based on the distribution of their data types. Within each cluster, clients exhibit comparable local optimal solutions, thus facilitating the aggregation of a superior global model. Moreover, the global model trained by the previous cluster serves as the initial model parameter for subsequent clusters, enabling the incorporation of data contributions from each cluster to foster the development of an enhanced global model. Experimental results corroborate the superiority of the FedSC algorithm over alternative federated learning approaches, particularly in non-IID data distributions, thereby establishing its capacity to achieve heightened accuracy.</p></abstract>