FedEdgent：基于DNN分割的端边云协同联邦学习加速框架

打开文本图片集

中图分类号：TN919.5；TP181 文献标志码：A文章编号：1003-3114（2025）06-1306-11

Abstract：ThedeepintegrationofAIandedgecomputingemergesintheIndustrialInteretofTings （Io）scenarios.Howeve devices haveliitedcommunicationndomputatioalresouces，anddatafromthsedevicsareusuallhighlyhtergeneousndpiate sohowtoaccelerateraiingndifereceilsuringatascuitisstlloproblderatedLeang（L），aseatdt gesandpotentialsinpriacyprotetionddtascurityhileastesearchpoblemihtefintcomputatioaldocationdue totetlenckofresoucconstraieddevices.TispaperproposesFedEdgentfrmeork，ahybridevicedgloudsgy framework for accelerating FL based on the Deep Neural Network （DNN） partitioning approach，and implements a dynamic FL task offloading strategythroughDepReinforcementLeang（DRL）toiprovethetrainingefciencythroughdevice-edge-cloudcollborationCombined withmodelcompresionFeEdgentselectsDayersihigercontribtistotieteloalFodelhichdcesh nicationtrafsndproestoicatiofecyExprintalsultssowthatpaditttalizedFdg duces the training time by about 60% and the amount of uploaded parameters by 25% on average while remaining comparable accuracy.

Keywords：FL; computing offloading; DRL; neural network partitioning

0 引言

随着AI的迅猛发展，其与边缘计算的融合日益紧密，在IIoT场景中催生了边缘智能的概念。（剩余17297字）