A clustering federated learning algorithm based on three-layer structure

Federated learning is a new distributed machine learning method. Decentralized clients can train data locally, and multi-party machine learning can be implemented efficiently by aggregating clients to update models through a central server. Therefore, federated learning has become an effective tool to solve the problem of data privacy leakage and data silos. However, there are huge differences in the distribution of data between clients in the real world. These data are usually Non-IID. Using such data can lead to problems such as reduced model accuracy, performance degradation, and slow convergence. To solve the problems, we propose a clustering federated learning algorithm based on three-layer structure, TCFL. In TCFL, the clients in each cluster are trained in a pre-arranged order. The central server collects the in-cluster model uploaded by the local server to obtain the optimal target. The experimental results show that the test accuracy of TCFL is improved by 3.42% ∼ 9.49% compared with the federal average algorithm on Non-IID data classification training of different degrees. Compared with the baseline algorithm, TCFL has achieved excellent results on tests of time rate, accuracy and training loss, and reduced the communication overhead between the central server and the client. Robustness in solving Non-IID data had been verified.