Federated and Collaborative Deep Learning Optimization in Cloud–Fog Computing

Abstract

The growing pace of the rise in the proportion of data-driven intelligent services in smart cities, industrial systems, and cyber-physical infrastructures have exacerbated the demand of an effective distributed deep-learning over cloud-fog computing activities. Although the cloud- fog architectures offer proximity and scalability requirements, effective learning is difficult because of the decentralization of data, heterogeneity of the system, and constraints of the communication. In this paper, a federated and collaborative deep learning optimization framework is introduced, which allows hierarchical learning of training models on the fog nodes under cloud coordination. It uses adaptive local training, weighted collaborative aggregation, and proximal regularization: it removes non-IID effects caused by data, and enables the stabilization of convergence. The fog nodes do decentralized optimization and the cloud dynamically assembles the models according to the data volume and resource conditions. The experimental findings prove that the given approach has a smaller final global loss of 0.043 than 0.061 and 0.058 of standard federated learning and hierarchical baselines. Moreover, communication cost is also lowered to 29 MB round, which is an improvement of about 3040. These findings justify the usefulness of the suggested framework in scalable and resilient deep learning on cloud-fog systems.