Fed-BSE: Federated Learning Based on Bayesian Selective Ensemble

• Yan Li,
• Guihua Wen,
• Bo Liu,

Vol. 19, No. 3, March 31, 2025
10.3837/tiis.2025.03.002, Download Paper (Free):

• Federated learning
• Non-independent identical distribution
• Bayesian Selective Ensemble
• knowledge distillation
• Stochastic weight average

Abstract

Federated learning (FL) is a crucial method for addressing data security challenges. Notably, the key problem in improving FL's model aggregation lies in handling Non-independent identically distributed (Non-IID) data. This paper introduces a novel federated learning method called Federated Bayesian Selective Ensemble (Fed-BSE), which can achieve better performance by increasing the diversity of distribution and combining the ensemble learning model. Firstly, the client trains the local model using the private database and, uploads the trained model to the public server. Then, four common distributions, Gaussian, Dirichlet, Random and Poisson, are used to construct the model together. Moreover, the global model is obtained by knowledge distillation and multiple models, which is constructed ensembled models and sent to the client together. Finally, the prediction results at the client are improved by relative majority voting ensemble learning and selective ensemble learning. Experiments demonstrate that the proposed Fed-BSE algorithm significantly outperforms other methods.

Statistics

Cite this article