面向联邦学习的6G大规模物联网资源跳跃多址方案

Resource Hopping Multiple Access Scheme for Federated Learning in 6G Massive IoT

随着移动设备的广泛应用和大数据的快速增长,联邦学习作为一种在分散数据环境中进行机器学习的新兴范式,吸引了越来越多的关注。同时,5G/6G均将大规模物联网场景作为其核心场景之一,以通过实现大规模设备连接来完成未来海量分散数据的实时传输。因此,6G大规模物联网可以为联邦学习中海量终端的数据处理提供有力支撑。多址技术是6G大规模物联网实现海量连接的关键,现有研究提出了多种面向大规模物联网的新型多址方案,其中资源跳跃多址方案考虑信道资源的跳跃,通过给不同用户分配不同的资源跳跃图案从而实现海量用户接入。提出了资源跳跃多址与联邦学习的结合方案,将联邦学习客户端的通信信道划分为多个子信道,然后根据其数据特征和计算资源分配资源跳跃图案。结果表明,所提出的结合方案不仅能够提高联邦学习模型的训练速度,而且能够有效保护用户数据的隐私。

With the ubiquitous deployment of mobile devices and the exponential growth of big data,federated learning has emerged as a novel paradigm for machine learning in decentralized data environments,attracting increasing attention.Concurrently,5G/6G networks identify massive Internet of Things(IoT)scenarios as one of their core applications,aiming to facilitate real-time transmission of massive decentralized data through extensive device connectivity.Thus,6G massive IoT provides robust support for data processing from numerous endpoints in federated learning.Multiple access technology is crucial for achieving large-scale connectivity in 6G massive IoT.Current research has proposed various novel multiple-access schemes tailored for massive IoT,among which the resource-hopping multiple-access scheme incorporates channel resource hopping to allocate distinct hopping patterns to different users,thereby enabling massive user access.This paper proposes an integration of resource-hopping multiple access with federated learning,partitioning the communication channels of federated learning clients into sub-channels and allocating resource-hopping patterns based on their data characteristics and computation resources.Results demonstrate that the proposed integration not only enhances the training speed of federated learning models but also effectively protects user data privacy.