Communication-Efficient Decentralized Federated Learning for Generalization and Personalization Over Wireless Networks

초록

In this letter, we present Communication-Efficient Proxy-Based Federated Learning (CE-ProxyFL), a novel decentralized federated learning (DFL) framework for wireless networks. CE-ProxyFL aims to balance generalization and personalization in a communication-efficient manner. Unlike conventional DFL approaches that rely on full model exchange, CE-ProxyFL employs a proxy-based architecture to reduce communication overhead and improve model performance. Each mobile device (MD) model is divided into three parts: a private component, a private classifier, and a proxy component. The private classifier is trained locally to support personalization, while the proxy component promotes generalization by exchanging distilled knowledge via a public dataset, rather than sharing full models. The private component is trained in a balanced manner to support both personalization and generalization, thereby providing useful output features to both private classifier and proxy component. This hybrid approach enables effective adaptation to both seen and unseen data. During inference, the MDs dynamically select between the generalized and personalized models based on confidence levels. Experiments on the FMNIST and CIFAR-10 demonstrate that CE-ProxyFL outperforms existing methods in both latency and accuracy.

키워드