Algoritmo Híbrido de Otimização por Enxame de Partículas para o Aprendizado Federado de Redes Neurais Artificiais

Abstract

The aggregation algorithms of Federated Learning (FL) often overlook the balance between performance and communication cost. Therefore, in this study, we propose a new federated training method called FLPSO-SGD, based on the hybrid Particle Swarm Optimization-Stochastic Gradient Descent (PSO-SGD) algorithm, aiming to address these aspects. In contrast to classical FL training techniques, our method collects errors from the clients to the server, instead of model parameters, while the PSO-SGD algorithm conducts training on the client side. The algorithms were evaluated on classification problems using the UC Irvine datasets for the PSO-SGD, and the CIFAR-10 dataset for the FLPSO-SGD. The results highlight the promising performance of the PSO-SGD algorithm. Furthermore, the FLPSO-SGD algorithm demonstrated superior accuracies in global training compared to the FedAvg and FedPSO techniques. The results suggest that FLPSO-SGD is an effective alternative for FL training, particularly in applications with restricted bandwidth communication networks on clients.