FGFL: Fine-Grained Federated Learning Based on Neural Architecture Search for Heterogeneous Clients

Federated learning (FL) has gained tremendous attention across different machine learning tasks. In large-scale deployments, client heterogeneity is a fact and imposes constraints on model design, training performance and accuracy. This paper introduces a fine-grained federated learning (FGFL) method to tackle resource heterogeneity. FGFL utilizes a configurable architecture search space in order to offer abundant architectures for various devices. FGFL first employs a greedy coarse-grained architecture selection method and a local training optimization strategy to enable most architectures to be readily deployable. Additionally, it conducts a fine-grained multi-objective evolutionary search to automatically identify the optimal architectures for heterogeneous devices. Experimental results demonstrate that FGFL achieves the superior performance while reducing computational costs.