Towards Efficient and Privacy-Preserving Hierarchical Federated Learning for Distributed Edge Network

Federated learning is a promising paradigm that utilizes widely distributed devices to jointly train a machine learning model while maintaining privacy. However, when oriented to distributed resource-constrained edge devices, existing federated learning schemes still suffer from heterogeneity challenge. Hierarchical federated learning divides devices into clusters based on their resources to enhance training efficiency. In addition, there are also security concerns such as the risk of privacy leakage during the interactions intra and cross the layers. In this paper, we propose PHFL, a Privacy-Preserving Hierarchical Federated Learning Scheme, which includes a personalized task assignment protocol and a heterogeneous-friendly aggregation scheme, so as to achieve an efficient and low-cost training for resource-constrained edge devices. To address the heterogeneity challenge, we combine synchronous and asynchronous aggregation, and design different aggregation weights for device differences. On top of that, a privacy-preserving scheme based on threshold homomorphic encryption is designed to accommodate the privacy requirements of the proposed hierarchical federated learning scenario. The security analysis and performance evaluation demonstrate that this approach improves model generalizability and learning efficiency with acceptable overhead, while also providing privacy protection.