基于动态分组和贡献感知的联邦客户端选择算法OA

Data heterogeneity(Non-IID)adversely affects the accuracy and convergence speed of the global model in federated learning.In order to solve this problem,this paper proposed a federated client selection algorithm based on dynamic grouping and contribution awareness,named FedGCCS.This algo-rithm employs a label-distribution-driven dynamic grouping mechanism to cluster the clients with similar data distribution.A multi-dimensional contribution awareness framework is designed to quantify client contributions dynamically by combining heterogeneous sensitive indicators,such as model similarity,test accuracy and training loss.Based on this frame work,an adaptive client selection strategy using Thompson sampling is implemented to balance the continuous utilization of high-contribution clients and the exploration opportunities of potential clients.Experimental results show that FedGCCS performs well on MNIST and Fashion-MNIST datasets.In highly heterogeneous scenarios,it improves the accuracy by 19.7%,17.3%and 7.1%over FedAvg,FedProx and FedCor,respectively,and achieves a faster conver-gence speed.These findings validate its effectiveness in solving the Non-IID problem and improving the overall performance.