DHFL: Decentralized Hierarchical Federated Learning With Dynamic Global Aggregation for Privacy-Aware Fog Computing
Abstract
Federated Learning represents a promising paradigm for distributed machine learning while guaranteeing privacy. However, traditional FL frameworks face challenges like poor communication latencies, dependency on the availability of the centralized servers, and inefficient handling of Non-Independent and Identically Distributed (non-IID) data. This paper introduces a new Distributed Hierarchical Federated Learning (DHFL) framework that shifts global aggregation from the cloud to the fog layer, dynamically selecting the best fog node for aggregation based on metrics such as data quality, workload, communication cost, and availability. In this framework, an enhanced multi-criteria particle swarm optimization algorithm (MCPSO) is employed to optimize the selection of aggregation nodes in order to minimize latency and enhance resource utilization. Extensive experiments using a real-world testbed on DigitalOcean demonstrate that DHFL achieves superior accuracy with faster convergence and reduced communication delays as compared to traditional hierarchical FL frameworks. The proposed system will be particularly suitable in those latency-sensitive applications, such as healthcare monitoring and industrial automation, for which fast model training is crucial.