Client selection optimization method based on dynamic programming for federated learning

Federated learning is a distributed machine learning technique which enables clients with limited resources to collaboratively train models without sharing private data, effectively protecting the data privacy of the clients. Classic federated learning systems lack a strict mechanism for selecting clients, but generally use an average strategy to aggregate the local model parameters, which may lead to the inclusion poor-quality clients in the training process of the federated learning, consequently affecting the overall performance of the final models. To address the above issues, a federated learning client selection method based on dynamic programming is proposed (FedWeight). The proposed method uses the Shapley Value method to measure the contribution of each client in different communication rounds, addressing the inherent difficulty of evaluating clients’ data quality directly. Using the Shapley Value as an important measurement to dynamically select high-quality clients by the server, and then the server improves the overall performance of model by aggregating these high quality clients. To construct different federated learning scenarios, MNIST, CIFAR-10, Fashion-MNIST, EMNIST and KMNIST datasets are used in our experiments. The experimental results demonstrate that the proposed method can effectively identify high-quality clients, and the performance of the obtained final federated model is almost unaffected by poor-quality clients. Furthermore, our method exhibits significant advantages in terms of convergence speed and model stability. © 2024 Systems Engineering Society of China. All rights reserved.