异构联邦学习在无人系统中的研究综述  

Review of research on heterogeneous federated learning in unmanned systems

作　　者：俞浩 范菁[1,2] 孙伊航 Yu Hao;Fan Jing;Sun Yihang(College of Electrical&Information Technology,Yunnan Minzu University,Kunming 650000,China;Yunnan Key Laboratory of Unmanned Autonomous System,Kunming 650500,China)

机构地区：[1]云南民族大学电气信息工程学院,昆明650000 [2]云南省无人自主系统重点实验室,昆明650500

出　　处：《计算机应用研究》2025年第3期641-649,共9页Application Research of Computers

基　　金：国家自然科学基金资助项目(61540063);教育部人文社会科学研究青年基金资助项目(20YJCZH129);云南省吴中海专家工作站项目(202305AF150045);云南省教育厅科学研究基金资助项目(2025Y0670);云南民族大学硕士研究生科研创新基金资助项目(2022SKY004)。

摘　　要：异构联邦学习(heterogeneous federated learning,HFL)是一种用于解决数据和设备异构性问题的分布式机器学习方法,广泛应用于包括无人系统在内的多种场景。随着无人系统(如无人机、自动驾驶车辆)的快速发展,如何有效应对非独立同分布(non-IID)数据及设备计算能力差异,已成为提升联邦学习效率和性能的重要挑战。综述了异构联邦学习在无人系统中的最新研究进展,重点分析了数据、设备及模型异构性带来的主要问题,并总结了现有的解决方案,如分层联邦学习、模型压缩与剪枝技术在此领域的应用。还讨论了这些技术的实际应用场景,评估了它们的优缺点,并提出了未来可能的研究方向,以进一步提升无人系统中的联邦学习性能和数据隐私保护水平。HFL is a distributed machine learning approach designed to address the challenges of data and device heterogeneity,applicable to various domains,including unmanned systems.As unmanned systems(e.g.,drones and autonomous vehicles)continue to evolve,efficiently handling non-independent and identically distributed(non-IID)data and the computatio-nal differences between devices has become a critical challenge for improving the performance and efficiency of federated lear-ning.This paper reviewed recent advances in HFL within unmanned systems,focusing on the challenges posed by data,device,and model heterogeneity,and summarized existing solutions,such as hierarchical federated learning,model compression,and pruning techniques.The paper also discussed practical applications of these techniques,evaluated their strengths and limitations,and proposed future research directions to further enhance the performance of federated learning in unmanned systems and improve data privacy protection.

关 键 词：联邦学习 异构性 无人系统 非独立同分布 数据隐私 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]