关键词: 联邦学习, 数据异构, 个性化联邦学习, 模型预分配, 自蒸馏

Abstract: Federated Learning (FL) is a distributed machine learning method that utilizes distributed data for model training while ensuring data privacy. However, it performs poorly in scenarios with highly heterogeneous data distributions. Personalized Federated Learning (PFL) addresses this challenge by providing personalized models for each client. However, many existing PFL algorithms primarily focus on optimizing client-side local models, while largely overlooking optimization of server-side global models. Consequently, server-side computational resources remain underutilized. To overcome these limitations, FedPASD, a personalized federated learning method based on model pre-assignment and self-distillation, was proposed. FedPASD was operated by addressing both server-side and client-side aspects. On server-side, client models for the next round were strategically identified and pre-assigned, which not only enhanced model personalization but also optimized server computational resources. On client-side, models were trained hierarchically and fine-tuned using self-distillation to better align with local data distribution. Experiments were conducted on three datasets, CIFAR-10, Fashion-MNIST, and CIFAR-100, comparing FedPASD with representative algorithms such as FedCP, FedPAC, and FedALA. Experimental results demonstrate that FedPASD consistently achieves higher test accuracy than that of benchmark algorithms such as FedALA across various data heterogeneity settings. Specifically, FedPASD achieves 30 percentage points higher accuracy than traditional FL algorithms on CIFAR-100 dataset and outperforms existing PFL algorithms by 2.42 to 5.92 percentage points. On CIFAR-10 dataset, FedPASD achieves a maximum accuracy of 88.54%.

Key words: Federated Learning &#40, FL&#41, data heterogeneity, Personalized Federated Learning &#40