Avaliação de métodos de agregação em ambientes federados: aplicações em classificação de imagens com redes neurais convolucionais

Abstract

This study presents a detailed technical analysis of the FedAvg, FedProx, and WeightedFedAvg aggregation methods in the context of federated learning, applied to convolutional neural networks for classification tasks on three datasets: MNIST, FashionMNIST, and CIFAR-10. Federated learning, a decentralized model training approach, facilitates collaboration across multiple devices while maintaining data locally, which is crucial for preserving privacy in distributed scenarios. The experiments were conducted in a non-IID (non-independently and identically distributed) data environment, replicating real-world conditions where data distribution among clients is heterogeneous. For each aggregation method, local training was performed for 5 epochs on each client, followed by central aggregation of the weight updates, with this cycle repeated for 10 rounds. Performance evaluation was based on traditional metrics such as accuracy and loss, as well as more sophisticated analyses using ROC/AUC curves and confusion matrices, providing a granular assessment of the models’ discriminative capabilities and specific error patterns. The results reveal that the WeightedFedAvg method consistently outperforms the others in environments characterized by high heterogeneity and data imbalance, particularly on CIFAR10, achieving superior accuracy and lower loss, which reinforces its suitability for challenging federated learning scenarios. FedProx demonstrated stability in settings with significant variability among client data, mitigating divergence in local updates and preserving the cohesion of the global model. In contrast, while FedAvg is efficient in more homogeneous data distribution environments, it showed significant limitations in more complex scenarios, aligning with the challenges discussed in the literature. The research concludes that, although FedAvg remains an effective solution for homogeneous federated environments, the increasing complexity of real-world scenarios necessitates the adoption of more advanced aggregation methods, such as FedProx and WeightedFedAvg, which exhibit greater resilience and accuracy in contexts of high data variability. These findings indicate the need for future investigations to develop new aggregation algorithms capable of further balancing the stability and accuracy of models in complex and challenging federated environments.