FedLabCluster: A Clustered Federated Learning Algorithm Based on Data Sample Label

Federated learning is a method to train machine learning models in large-scale distributed devices without uploading local data to the central server. However, in the actual federated learning application environment, the Non-Independent and Identically Distributed(Non-IID) data will reduce the prediction accuracy and performance of federated learning. In this paper, we propose a novel clustered federated learning algorithm, FedLabCluster. This algorithm uploads the data sample labels of all clients participating in the training. The central server forms the sample label matrix and uses the clustering algorithm to cluster the clients. We evaluated the FedLabCluster algorithm based on five datasets and compared the FedLabCluster algorithm with the baseline algorithm. Compared with the FedAvg algorithm in the MNIST dataset, the accuracy of the FedLabCluster algorithm was improved by 6.5% and 6.9% on the CNN and MLP models. In the FEMNIST dataset, accuracy was improved by 15% and 28.6% on CNN and MLP models. Therefore, the FedLabCluster algorithm is feasible in Non-IID data and has batter federated learning performance and robustness.