基于LightFL平台的联邦学习算法对比研究OA

This paper addresses the lack of systematic evaluation of federated learning algorithms in practical applications.Using the lightweight platform LightFL,it compares the performance of 10 mainstream algorithms on the MNIST dataset under two data partitioning methods:IID and Dirichlet Non-IID.In the IID scenario,most algorithms perform well with minimal differences.However,in the highly heterogeneous Dirichlet Non-IID scenario,the robustness of each algorithm is challenged.SCAFFOLD,through its unique client-side bias correction mechanism,effectively mitigates the performance degradation caused by data heterogeneity,demonstrating good robustness.The research shows that the performance of federated learning algorithms is highly dependent on the characteristics of the data distribution.When the data distribution is relatively uniform,simple and efficient algorithms such as FedAvg can be given priority;however,in highly heterogeneous scenarios,algorithms with strong robustness,such as SCAFFOLD,are more reliable choices.These findings provide empirical evidence for the selection of federated learning algorithms in different scenarios and have reference value for their promotion and application.