Recent years, researchers focused on personalized federated learning (pFL) to address the inconsistent requirements of clients causing by data heterogeneity in federated learning (FL). However, existing pFL methods typically assume that local data distribution remains unchanged during FL training, the changing data distribution in actual heterogeneous data scenarios can affect model convergence rate and reduce model performance. In this paper, we focus on solving the pFL problem under the situation where data flows through each client like a flowing stream which called Flowing Data Heterogeneity under Restricted Storage, and shift the training goal to the comprehensive performance of the model throughout the FL training process. Therefore, based on the idea of category decoupling, we design a local data distribution reconstruction scheme and a related generator architecture to reduce the error of the controllable replayed data distribution, then propose our pFL framework, pFedGRP, to achieve knowledge transfer and personalized aggregation. Comprehensive experiments on five datasets with multiple settings show the superiority of pFedGRP over eight baseline methods.

翻译：近年来，研究者们聚焦于个性化联邦学习（pFL）以解决联邦学习（FL）中因数据异质性导致的客户端需求不一致问题。然而，现有的pFL方法通常假设本地数据分布在FL训练期间保持不变，而实际异质数据场景中变化的数据分布会影响模型收敛速度并降低模型性能。本文致力于解决一种特定情境下的pFL问题：数据像流动的溪流一样流经每个客户端，即受限存储下的流动数据异质性，并将训练目标转向模型在整个FL训练过程中的综合性能。为此，基于类别解耦的思想，我们设计了一种本地数据分布重构方案及相关的生成器架构，以减少可控回放数据分布的误差，进而提出了我们的pFL框架——pFedGRP，以实现知识迁移与个性化聚合。在五个数据集上进行的多设置综合实验表明，pFedGRP优于八种基线方法。