Energy-Efficient Model Decoupling for Personalized Federated Learning on Cloud-Edge Computing Networks

Federated Learning (FL) has emerged as a key distributed learning approach for privacy-preserving data scenarios. However, with the demonstrated effectiveness of scaling laws by large language models, the increasing parameter size of neural networks has led to substantial communication overhead, posing significant challenges for distributed learning systems. To address these issues, we propose a novel energy-efficient personalized federated learning framework called FedEMD, which utilizes model decoupling to divide deep neural networks into a body, consisting of the early layers of the network, and a personalized head, comprising the layers beyond the body. During training, the personalized head does not need to be uploaded to the central server for aggregation, thereby saving significant bandwidth resources. Additionally, we propose a performance-resource balancing mechanism that adaptively adjusts the number of body layers uploaded based on the available resource of the client. Finally, we conducted experiments on six datasets, comparing our method with five state-of-the-art model decoupling approaches. Our method was able to save about 10.7% in bandwidth consumption while providing comparable performance.