计算机技术与发展2026，Vol.36Issue(8)：69-77,9.DOI:10.20165/j.cnki.ISSN1673-629X.2026.0030
结构感知异构特征对齐的个性化联邦学习方法
Structure-aware Heterogeneous Feature Alignment for Personalized Federated Learning
摘要
Abstract
Personalized federated learning(PFL)customizes models for client-specific data,addressing challenges of non-IID distributions.Existing methods suffer from intra-client feature inconsistency,causing dispersed intra-class representations and blurred boundaries,and inter-client misalignment,where heterogeneous feature scales hinder knowledge transfer and hyperparameter tuning.To overcome these issues,we propose SPAlign,a structure-aware heterogeneous feature alignment framework.It employs intra-client structure rectification via exponential moving average and intra-class standardization to enhance feature compactness,and inter-client representation alignment through unified statistical mapping to mitigate cross-client scale differences and reduce distillation temperature sensitivity.Guided by historical personalized models,SPAlign balances global knowledge with local features.Experiments on CIFAR-10,CIFAR-100,FMNIST,and TinyImageNet show consistent improvements in accuracy,robustness,and personalization,especially under high heterogeneity and large-scale client scenarios.关键词
个性化联邦学习/异构特征对齐/结构感知/知识蒸馏/数据异质性Key words
personalized federated learning/heterogeneous feature alignment/structure-aware/knowledge distillation/data heterogeneity分类
信息技术与安全科学引用本文复制引用
吴晶,陈姚节,袁鑫,刘金鑫..结构感知异构特征对齐的个性化联邦学习方法[J].计算机技术与发展,2026,36(8):69-77,9.基金项目
湖北省自然科学基金青年B类(原青年项目)(2025AFB056) （原青年项目）
湖北省教育厅科学研究计划重点项目(D20241103) （D20241103）
智能信息处理与实时工业系统湖北省重点实验室开放基金项目(ZNXX2023QNO3) （ZNXX2023QNO3）
