新疆大学学报（自然科学版中英文）2025，Vol.42Issue(4)：425-433,9.DOI:10.13568/j.cnki.651094.651316.2024.12.30.0001
非独立同分布与长尾分布下的联邦学习优化方法
Federated Learning Optimization Method in Non-IID and Long-Tail Distributions
摘要
Abstract
To address the challenges of non-independent and identically distributed data and long-tail distribu-tions in federated learning,a novel federated learning method is proposed by integrating contrastive learning with a two-stage learning strategy.The approach employs contrastive learning to align feature representations between client models and the global model,thereby reducing feature discrepancies across clients.Simultaneously,it aggre-gates and uploads client model gradients,enabling retraining of the classifier through virtual features on the server side to enhance the global model's learning capability for minority class data.Experimental results demonstrate that the proposed method achieves maximum accuracy improvements of 0.36％on the Fashion-MNIST dataset and 1.64％on the CIFAR-10 dataset.关键词
联邦学习/非独立同分布/长尾分布/对比学习/两阶段学习Key words
federated learning/non-independent and identically distributed/long-tail distribution/contrastive learning/two-stage learning分类
信息技术与安全科学引用本文复制引用
冯文举,杨焱青,贾振红,张琳琳..非独立同分布与长尾分布下的联邦学习优化方法[J].新疆大学学报（自然科学版中英文）,2025,42(4):425-433,9.基金项目
新疆维吾尔自治区天山英才科技创新团队"面向公共安全的信号检测与处理技术研究"(2023TSYCTD0012) （2023TSYCTD0012）
新疆维吾尔自治区教育厅高校科研计划"基于深度生成模型的工控网络攻击检测技术研究"(XJEDU2021Y003) （XJEDU2021Y003）
新疆维吾尔自治区重大科技专项"能源数据标准化体系研究"(2022A01007-4). （2022A01007-4）
