PrivHFL: A privacy-preserving scheme for hierarchical federated learning

In the Internet of Things (IoT) field, where interconnected devices generate sensitive data, ensuring privacy is a major challenge. Federated Learning (FL) addresses this by allowing devices to collaboratively train a model without sharing their local data, improving privacy in IoT systems. While the traditional two-layer FL framework is commonly used, adopting a hierarchical client-edge-cloud architecture can significantly accelerate model training, especially in resource-constrained IoT networks. Hierarchical Federated Learning (HFL) offers significant advantages, yet concerns persist regarding potential privacy breaches from analyzing client or edge server data. To address these concerns, we propose PrivHFL, a privacy-preserving solution for HFL that leverages threshold homomorphic encryption. Security and performance analyses demonstrate that the proposed scheme is scalable, supporting larger FL scenarios, including diverse IoT environments, while ensuring data privacy. PrivHFL is resilient to collusion among nearly half of the clients and effectively handles client dropouts. Our approach achieves high accuracy in IID and non-IID scenarios, as demonstrated using the MNIST, CIFAR-10, and CIFAR-100 datasets. Additionally, we show that the added encryption overhead is reasonable, making our solution feasible for real-world IoT applications.