面向6G的时间敏感异构联邦边缘智能  

Latency-Sensitive Heterogeneous Federated Edge Intelligence for 6G

作　　者：王连 何升涛 蔡浩晖 李莹玉 肖泳 石光明 WANG Lian;HE Shengtao;CAI Haohui;LI Yingyu;XIAO Yong;SHI Guangming(School of Mechanical Engineering and Electronic Information China University of Geosciences,Wuhan 430074,China;School of Electronic Information and Communications,Huazhong University of Science and Technology,Wuhan 430074,China;Peng Cheng Laboratory,Shenzhen 518055,China;School of AI,Xidian University,Xi’an 710071,China;Pazhou Laboratory(Huangpu),Guangzhou 510335,China;Key Laboratory of Electronic Information Countermeasure and Simulation Technology,Ministry of Education,Xi’an 710071,China)

机构地区：[1]中国地质大学(武汉)机械与电子信息学院,湖北武汉430074 [2]华中科技大学电信学院,湖北武汉430074 [3]鹏城实验室,广东深圳518055 [4]西安电子科技大学人工智能学院,陕西西安710071 [5]琶洲实验室(黄埔),广东广州510335 [6]电子信息对抗与仿真技术教育部重点实验室,陕西西安710071

出　　处：《移动通信》2024年第7期66-72,共7页Mobile Communications

基　　金：国家自然科学基金“面向6G群体智能资源共享博弈基础理论研究”(62071193);“面向智能语义理解的计算成像方法研究”(61976169);“语义通信基础理论与方法研究”(62293483);“基于多通道压缩感知的高分辨高动态范围红外成像方法研究”(61871304);鹏城实验室重大攻关项目(PCL2021A12);中央高校基本科研业务费资助,HUST“基于联邦学习的数字孪生网络建模研究”(2023JYCXJJ029);中国地质大学(武汉)“地大学者”人才岗位科研启动经费资助(2021164);湖北省国际科技合作计划项目“面向印尼通信工程复杂场景的云边协同智能运维系统研发及应用”(2023EHA009)。

摘　　要：与人工智能技术深度融合并实现智能内生是6G网络的重要演进趋势之一。联邦边缘智能是一种新兴的分布式协同训练架构,能够在为用户提供智能化服务的同时有效保护本地数据隐私。然而,由于算力网络中不同边缘节点的数据特征、可用的通信与计算资源等差异巨大,呈现出高度异构性,传统的联邦学习架构全局模型收敛速率较低,服务响应时延高。介绍了一种时间敏感异构联邦边缘智能系统架构,该架构采用异步的方式进行模型协同,可有效避免落后者效应对时延的影响,同时在对数据进行贡献度分析的基础上,采用低性能边缘节点数据丢弃与负载转发相结合的策略,进一步降低异步协同机制下陈旧性效应对服务响应时延的影响。基于树莓派的异构联邦边缘智能原型系统中的测试实验表明,所提方案可有效降低服务响应时延。The deep integration of artificial intelligence and the realizaiton of intelligent endogenous capabilities are critical evolutoi nary trends for6G networks.Federated Edge Intelligence(FEI)is an emerging distributed collaborative training architecture that provides intelligentservices to users while effectively protecting local data privacy.However,significant differences in data characteristics and availablecommunication and computing resources among different edge nodes in computing networks result in high heterogeneity.Traditionalfederated learning architectures exhibit poor global model convergence rates and high service response latency under these conditions.This paper proposes a latency-sensitive heterogeneous federated edge intelligence system architecture that employs asynchronousmodel collaboration to mitigate the latency impact of lagging nodes.Additionally,based on data contribution analysis,a strategycombining low-performance edge node data discarding with load forwarding is introduced to further reduce the staleness effect onservice response latency in asynchronous collaboration mechanisms.Experimental tests on a heterogeneous FEI prototype systembased on Raspberry Pi demonstrate that the proposed scheme significantly reduces service response latency.

关 键 词：智能内生 时间敏感网络 联邦学习 异构网络 

分 类 号：TN929.5[电子电信—通信与信息系统]