Balancing Efficiency and Personalization in Federated Learning via Blockwise Knowledge Distillation

Abstract

Dear Editor, Federated learning (FL) has emerged as a pivotal approach in distributed machine learning, allowing models to be trained across decentralized data sources while maintaining privacy [1], [2]. However, FL faces significant challenges, particularly in balancing personalization, privacy, and computational efficiency, especially when deployed in heterogeneous environments with varied client capabilities [3]. To address these challenges, we introduce FedBW, a novel framework that integrates FL with blockwise knowledge distillation.