Research On Client Scheduling And Resource Optimization For Wireless Federated Learning

Wireless Federated Learning(FL)is a distributed learning framework that allows user equipment to jointly train models without uploading private data,which not only reduces the demand for network resources,but also reduces the risk of data leakage.However,with the development of mobile networks and the Internet of Things,the number of user devices is increasing rapidly.Due to the scarcity of wireless resources,it is difficult for large-scale user equipment to participate in the federated learning and training process at the same time,resulting in low efficiency and poor accuracy of model training.In order to tackle this problem,this paper proposes corresponding algorithms to improve the performance of wireless federated learning system from the perspectives of both client scheduling and resource optimization.Firstly,for the problem of scheduling on large-scale mobile clients,this paper proposes a density-based weighted clustering method.Specifically,Density-Based Spatial Clustering of Applications with Noise(DBSCAN)and Low Energy Adaptive Clustering Hierarchy(LEACH)are applied to divide user devices into different initial clusters,and the threshold values are established in each initial cluster.Then,a scoring mechanism considering location,speed,link lasting time,and training delay of user equipment is proposed.According to the relationship between the score and threshold,the cluster head in each cluster is determined.Finally,the simulation results show that the proposed method can reduce the communication overhead and training delay of wireless federation,compared with randomly selected cluster heads and traditional centralized FL.Secondly,to alleviate the contradiction between scarce resources and massive mobile clients,this paper considers both computing energy consumption model and communication energy consumption model of clients,then the allocation methods on both load and bandwidth are proposed.Specifically,in order to reduce the communication load during each training round,the clients participating in the training are selected according to Multi-Armed Bandit(MAB)theory.To further reduce the total energy consumption of the devices,we establish an energy consumption optimization equation and obtains the optimal solution including bandwidth allocation,central processing unit(CPU)and graphics processing unit(GPU)workloads.Finally,simulation results show that the proposed method can reduce the energy consumption generated in the wireless federated learning and training process compared to traditional filtering client schemes and only optimizing load and bandwidth resources.