dy-TACFL: Dynamic Temporal Adaptive Clustered Federated Learning for Heterogeneous Clients

Abstract

Federated learning is a potential solution for training secure machine learning models on a decentralized network of clients, with an emphasis on privacy. However, the management of system/data heterogeneity and the handling of time-varying client interests still pose challenges to traditional federated learning (FL) approaches. Therefore, we propose the concept of dynamic temporal adaptive clustered federated learning (dy-TACFL) to tackle the issue of of client heterogeneity in time-varying environments. By continuously analyzing and assigning appropriate clusters to the clients with similar behavior, the proposed federated clustering approach increases both prediction accuracy and clustering efficiency. First, a silhouette coefficient-based threshold is used in the temporal adaptive clustering federated learning (TACFL) algorithm to evaluate cluster stability in each round of federated training. Then, an affinity propagation-based dynamic clustering (APD-CFL) algorithm is proposed to adaptively organize clients into an appropriate number of clusters, taking into account the complex underlying pattern. The experimental findings indicate that the proposed time-based adaptive clustered federated learning algorithms can significantly improve prediction accuracy compared to the existing clustered federated learning algorithms.