计算机科学与探索2025，Vol.19Issue(10)：2648-2666,19.DOI:10.3778/j.issn.1673-9418.2501059
去中心化联邦学习综述
Review of Decentralized Federated Learning
摘要
Abstract
With the rapid growth of large-scale heterogeneous data,centralized federated learning faces challenges in data processing and privacy protection.Decentralized federated learning addresses these issues by eliminating reliance on cen-tral servers,enhancing system fault tolerance and adaptability,while distributing communication loads and significantly improving privacy protection.This paper systematically elaborates on the fundamental principles of centralized and decen-tralized federated learning,highlighting their differences through multi-dimensional comparative analysis.Building on this,it delves into the technical advantages and innovative methods of decentralized federated learning in communication optimization,privacy protection mechanisms,and model aggregation strategies.Additionally,it comprehensively analyzes the application prospects and development trends of decentralized federated learning in healthcare,smart manufacturing,and smart cities.Finally,through comparing the performance of prevalent decentralized federated learning frameworks on commonly used datasets,this paper highlights their respective advantages,provides a summary of current mainstream open-source frameworks,and offers perspectives on potential technical challenges and development opportunities for future research.关键词
去中心化联邦学习/通信机制/隐私保护/模型聚合策略/应用场景Key words
decentralized federated learning/communication mechanisms/privacy protection/model aggregation strategy/application scenarios分类
信息技术与安全科学引用本文复制引用
陈丽芳,许恺龙,赵人喆,韩阳,代琪..去中心化联邦学习综述[J].计算机科学与探索,2025,19(10):2648-2666,19.基金项目
国家自然科学基金(52074126).This work was supported by the National Natural Science Foundation of China(52074126). （52074126）
