面向资源异构的通信高效去中心化联邦学习

打开文本图片集

中图分类号： TP393 文献标志码：A DOI：10.12305/j.issn.1001-506X.2025.04.34

Abstract： In order to alleviate the negative impact caused bydata heterogeneity of different terminal nodes in decentralized federation learning，and to enhance asynchronous compatibility while reducing the overall communication overhead，a decentralized federated learning algorithm based on mask location graph is proposed. Specifically，an asymmetric mask updating scheme is designed，which binds the mask norm to the training degree by gradually increasing the sparsity. It can effectively utilize the sparse gradient and guarantee system security while using trusted sparse federated aggregation. Secondly，a dynamic mask community segmentation algorithm is designed to combine gradient mask with community segmentation，which can efectively utilize the similarity between gradients across the entire network，while actively selecting similar aggregation targets and improving model performance.Furthermore，separating the model layer from the mask layer in the network structurecanreduce the impact ofarithmetic heterogeneity on the systemscalability.Finally，a single-threaded based experimental scheme is developed to simulate data heterogeneity，computility heterogeneity and terminal node asynchrony simultaneously. Experimental results show that compared with existing relevant methods，the proposed algorithm maintains high accuracyin both two datasets and strict asynchronous condition setings，and reduces communication overhead by 14%～21% ：

Keywords：edge computing； federated learning；decentralized system；sparse training

0 引言

随着信息技术的不断发展，终端设备的通信和计算能力按照吉尔德定律与摩尔定律飞速提升，催生了基于大规模终端设备的物联网技术，大量终端节点运行时产生海量的数据。（剩余17943字）