Algoritmo híbrido de otimização por enxame de partículas para o aprendizado federado de redes neurais artificiais

Abstract

The algorithms for managing Federated Learning (FL) training commonly 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, to address these aspects. In contrast to classical FL training techniques, the proposed method receives client errors instead of model parameters, while the PSO-SGD algorithm performs training on the client side. The algorithms were evaluated on classification problems using UC Irvine datasets for PSO-SGD, and the CIFAR-10 dataset was also used with both PSO-SGD and FLPSO-SGD across different convolutional neural network architectures. The results highlight the promising performance of the PSO-SGD algorithm. Moreover, the federated methodology FLPSO-SGD demonstrated better performance in training global models compared to the FedAvg and FedPSO techniques. These results are particularly notable in applications involving data that are not independent and identically distributed (non-IID data). The findings suggest that FLPSO-SGD is an effective alternative for FL training, especially in applications with communication networks that have restricted bandwidth on the client side.