基于选举策略的低空物联网稳定联邦学习方法  

Stable federated learning method for low-altitude IoT networks based on election strategy

作　　者：申凌峰 王光辉 白天水 朱政宇[3] 张千坤 SHEN Lingfeng;WANG Guanghui;BAI Tianshui;ZHU Zhengyu;ZHANG Qiankun(School of Software,Henan University,Kaifeng 475004,China;Henan International Joint Laboratory of Intelligent Network Theory and Key Technology,Kaifeng 475004,China;School of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450001,China;China Information Technology Designing&Consulting Institute Co.,Ltd.,Beijing 100048,China)

机构地区：[1]河南大学软件学院,河南开封475004 [2]河南省智能网络理论与关键技术国际联合实验室,河南开封475004 [3]郑州大学电气与信息工程学院,河南郑州450001 [4]中迅邮电咨询设计院有限公司,北京100048

出　　处：《物联网学报》2024年第3期55-65,共11页Chinese Journal on Internet of Things

基　　金：国家自然科学基金重大研究计划(No.92367302);河南省科技攻关项目(No.242102210139,No.242102211097);河南省高等学校重点科研项目(No.25A510015,No.25A520008);河南省交通运输厅科技项目(No.2023-3-2)。

摘　　要：随着无人机(UAV,unmanned aerial vehicle)与物联网(IoT,Internet of things)技术的深度融合,低空物联网中传输了大量包含敏感信息的数据,存在严重的隐私泄露风险。联邦学习(FL,federated learning)允许多个参与者共同训练模型而无须共享敏感数据,为低空物联网安全应用提供了隐私保护的方案。但是,随着应用场景越来越丰富,节点异构性、网络动态性等特点导致低空物联网下的联邦学习非常不稳定。提出了一种结合Raft选举算法和权重计算的新型联邦学习方法(FedPRE-W,federated fearning based on proxy Raft election and weight calculation),提高了联邦学习的稳定性和效率。针对遮挡、网络动态变化以及节点能量耗尽等导致的代理设备中断问题,通过Raft选举算法选举新的代理设备,保障联邦学习的稳定性。结合节点异构性,通过计算节点权重,选举性能强的节点当选代理,提升了联邦学习的效率。最后,在公开数据集上对所提方法进行验证,结果显示,FedPRE-W算法在减少通信轮数、加速模型收敛以及提高系统稳定性等方面有显著优势。该方法为低空物联网进行安全、稳定、高效的联邦学习提供了一种可行的解决方案。The deep integration of UAV and Internet of things(IoT)transmits a large amount of sensitive data in the air-to-ground intelligent network,posing a serious risk of privacy leakage.The proposal of federated learning(FL)provides a privacy-preserving solution for low-altitude IoT applications,allowing multiple participants to jointly train models with‐out sharing sensitive data.However,the federated learning performance is unstable because of various application sce‐narios,heterogeneous nodes and dynamic environments.An federated fearning based on proxy Raft election and weight calculation(FedREP-W)method was proposed,which combined classical Raft election and weight calculation,signifi‐cantly improving the stability and efficiency of federated training.To be more specific,the use of Raft to choose new agent devices keeped federated learning stable.By incorporating the concept of weight elections,the effectiveness of fed‐erated learning could be enhenced by designating the most powerful node as an agent.The experimental results publicly available datasets show that the proposed strategy and algorithm perform well in lowering the number of communication rounds,speeding up model convergence,and making the system stable.This provides a feasible solution for efficient,se‐cure,and stable federated learning in low-altitude IoT networks.

关 键 词：低空物联网 联邦学习 设备选举策略 稳定性 训练效率 

分 类 号：TN915.08[电子电信—通信与信息系统]