Federated Learning Communication-Efficiency Framework via Corset Construction.

This article focuses on Corset-Based Federated Learning (CBFL), a novel framework designed to improve communication and computation efficiency in federated learning (FL) by addressing data redundancy. CBFL introduces a FedCorset construction algorithm that extracts smaller, representative subsets ("corsets") from clients’ local datasets, considering the non-independent and identically distributed (non-IID) nature of FL data, thereby reducing redundant information. To fit these corsets, CBFL employs a distributed network evolution mechanism that dynamically adapts a smaller sparse neural network by removing less important connections and adding significant ones during training, ensuring model accuracy is maintained. Theoretical convergence of CBFL is proven, and experiments on popular datasets demonstrate that CBFL reduces communication bits by about 87% and computation time by approximately 56%, with only a minor (around 2%) decrease in model accuracy compared to standard FL methods. The framework currently targets horizontal FL scenarios, with future work aimed at extending it to vertical FL settings.