Match-based model offloading for edge federated learning
Aiming at problems such as the"straggler effect"caused by resource heterogeneity in federated learning in edge computing environments,this paper proposed a match-based model offloading for edge federated learning(Fed-MBMO).This method collected performance analysis results of edge devices,divided devices into strong and weak clients,and considered the time proportion of the four phases of model training,weak clients saved the time of backpropagation on the feature layers by freezing part of the model,and offload the model to the strong client for additional training,finally,the strong clients'feature layers were then reconstructed with the weak clients'fully connected layers.In order to improve the efficiency of model off-loading,the offloading cost matrix is constructed by comprehensively considering the similarity of model feature layers and task completion time,and transform the problem into an iterative solution of the optimal matching problem based on bipartite graph,the proposed approach used a KM-based model offloading algorithm and further analyzed the time complexity of the Fed-MBMO algorithm.Experimental results show that in the case of extremely heterogeneous resources and datasets,this method can accelerate model convergence,and the model training time can be reduced by an average of 46.65 percent,12.66 percent and 38.07 percent compared to FedAvg,FedUE and Aergia,respectively.The experimental results show that the Fed-MBMO algorithm can effectively solve the"straggler effect"problem and significantly improve the efficiency of federated learning.
federated learningstragglers'effectmodel offloadingstrong and weak matchingresource heterogeneitymodel reconstructionedge computing
万方数据
维普
