D2D辅助的双阶段空中计算模型聚合方案  

D2D-assisted two-stage model aggregation scheme based on overthe-air computation

作　　者：张冉强 邓娜 卫海超 邢成文[3] 赵楠[1] Ranqiang ZHANG;Na DENG;Haichao WEI;Chengwen XING;Nan ZHAO(School of Information and Communication Engineering,Dalian University of Technology,Dalian 116024,China;School of Information Science and Technology,Dalian Maritime University,Dalian 116026,China;School of Information and Electronics,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]大连理工大学信息与通信工程学院,大连116024 [2]大连海事大学信息科学技术学院,大连116026 [3]北京理工大学信息与电子学院,北京100081

出　　处：《中国科学：信息科学》2024年第10期2487-2502,共16页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:62201115,62325103,62371086);辽宁省自然科学基金联合基金(批准号:2023-MSBA-015);航空科学基金(批准号:2022Z001063001);中央高校基本科研业务费专项资金(批准号:DUT24MS015,3132024246)资助项目。

摘　　要：在无线联邦学习中,不同位置的设备信道质量差异性使得基于空中计算的联邦学习的模型聚合误差由信道质量最差的设备所主导.为此提出了一种终端直通(device-to-device,D2D)辅助的空中计算联邦学习方案,其中选择信道质量好的辅助设备辅助边缘设备将本地模型更新至服务器.构建了最小化均方误差的优化问题,提出了一种交替优化的算法,以低复杂度优化所有设备和基站的运行参数.为了评估该方案的性能,本文通过理论分析验证了所提方案相比于传统的基于空中计算的模型聚合方案具有优势.同时,设计了两种不同的设备分布场景,并基于神经网络和真实数据集构建了联邦学习实验.结果表明,所提算法能够很快收敛,与传统的空中计算方案以及现有的基于调度和基于中继的空中计算方案相比,该方案能够显著减小模型聚合误差,并提高联邦学习的预测精度.In wireless federated learning,the channel quality heterogeneity of the devices at different locations results in model aggregation errors dominated by the device with the worst channel quality.To address this,a device-to-device(D2D)assisted over-the-air computation(OAC)federated learning scheme is proposed,where the devices with good channel quality are selected to assist edge devices in updating their local models to the server.An optimization problem is formulated to minimize the mean squared error and an alternating optimization algorithm is proposed to optimize the operational parameters of all devices and the base station with low complexity.To assess the performance of the proposed scheme,this paper validates its advantages over traditional OAC-based model aggregation schemes through theoretical analysis.Meanwhile,two different device distribution scenarios are designed to construct federated learning experiments based on neural networks and real datasets.The results show that the proposed algorithm converges quickly.Compared with traditional OAC schemes and existing scheduling or relay-based OAC schemes,the proposed scheme significantly reduces model aggregation errors and improves the prediction accuracy of federated learning.

关 键 词：D2D 联邦学习 空中计算 交替优化算法 

分 类 号：TN929.5[电子电信—通信与信息系统] TP18[电子电信—信息与通信工程]