Efficient Communication Method for Hierarchical Federated Learning Based on 1-Bit Compressed Sensing

Federated learning is a distributed learning framework that allows multiple computing nodes to collaboratively train deep learning models without accessing each other's data. Under this framework, computing nodes transmit and aggregate model uploading with a common cloud server, efficiently completing deep learning training tasks. However, a major challenge currently faced by federated learning is the significant communication overhead generated by frequent model aggregation between computing nodes and the central server. Hierarchical federated learning is a novel federated learning framework that incorporates edge learning, introducing edge servers to enable model aggregation at the edge, thus avoiding the substantial communication cost incurred during aggregation between computing nodes and the central server. Several approaches have been proposed to reduce communication overhead in federated learning, including random sparsification and model quantization. In this work, they presented a 1-bit compressed sensing-based hierarchical federated learning algorithm for compressing communication models in hierarchical federated learning. The proposed algorithm significantly reduces the size of transmitted models during federated communication by performing sparsification, dimensionality reduction, and quantization on model updates. Experimental results on the MNIST and FashionMNIST dataset demonstrate that the proposed algorithm can significantly reduce communication overhead in hierarchical federated learning while ensuring virtually no performance degradation.