基于DBSCAN聚類的聯(lián)邦學(xué)習(xí)算法優(yōu)化

打開文本圖片集

引文格式：，.基于DBSCAN聚類的聯(lián)邦學(xué)習(xí)算法優(yōu)化[J].理工大學(xué)學(xué)報（自然科學(xué)版），2025，45（2）：92-99.DOI：10. 3969/j. issn.1672-1098. 2025.02.012

中圖分類號：TP391 文獻標志碼：A 文章編號：1672-1098（2025）02-0092-08

Federated Learning Algorithm Optimization Based on DBSCAN Clustering WANG Guoming，LI Miaomiao

（School ofComputerScienceandEngineering，Anhui UniversityofScienceand Technology，HuainanAnhui 232Ool，China） Abstract：Objective In federated learning，each node generates its own local data independently，resulting in heterogeneity between the data.During the training，heterogeneity can cause a gradient drift in the local model generated by federated learning，whichdecreases the accuracy and convergence speed of the federated learning model. Methods To address this isse，a federated learning optimization algorithm were introduced based on DBSCAN clustering，denoted as FLDC，which utilized the DBSCAN algorithm for clustering and grouping local clients.The client data distribution within the layer was similar.Additionally，it utilized a hierarchical sampling method to select clients from each layer and created a subset of clients，improving the diversity of training dataand ensuring thatthe clientdata samples in the subset reflected the global data distribution characteristics.Results Experiments on MNIST and CIFAR-1O showed that FLDC achieved 0.29% to 8.38% higheraccuracyand faster convergencecompared to the benchmark algorithm.Conclusion FLDC efectively reduces the impact of heterogeneous data on model performance in heterogeneous scenarios.

Key Words ： machine learning; federated learning; data heterogeneity； clustering

為了應(yīng)對集中式機器學(xué)習(xí)中必須要求用戶數(shù)據(jù)共享的挑戰(zhàn)，Google提出了聯(lián)邦學(xué)習(xí)（FederatedLearning，F(xiàn)L）[1]。（剩余10472字）