H-FedSN: Personalized Sparse Networks for Efficient and Accurate Hierarchical Federated Learning for IoT Applications

Abstract

H-FedSN pushes the boundaries of IoT with a unique approach that uses structured masking to train a personalized sparse network, enhancing personalization through client-based transfer learning and Bayesian aggregation. Applied to non-IID IoT datasets, it achieves high accuracy while reducing communication cost by up to 238x compared to traditional federated learning approaches.

Yuangang Li

PhD Student at UCI