Federated learning client selection algorithm based on gradient similarity

Federated Learning (FL) is an innovative machine learning method where terminal devices collaboratively train a global model coordinated by a central server, addressing data privacy and data silo issues without transferring data to the central server. However, in federated learning, the heterogeneity of client data significantly impacts FL performance. To address issues of low model accuracy and slow convergence speed, a client selection for federated learning based on gradient similarity (FedGSCS) algorithm was proposed. This algorithm selected clients by comparing the similarity between client gradients and the average gradient, prioritizing clients that could accelerate model convergence for model aggregation. Experiments on the MNIST, FEMNIST, and Shakespeare datasets showed that, compared to the Federated Averaging (FedAvg) algorithm, the Power-of-Choice algorithm selecting clients with the highest loss, and the federated averaging with diverse client selection (DivFL) algorithm, FedGSCS reduced the number of communication rounds by up to 80% and improved accuracy by up to 16.38%.