Abstract
Hierarchical federated learning (HFL) enhances the scalability of federated learning by deploying edge servers closer to the clients, facilitating intermediate aggregation before sending the aggregated models to the cloud server. To improve the privacy and integrity of HFL, we propose a lightweight hash-based verification technique for the edge and cloud model aggregations. Our Verifiable HFL (VHFL) technique utilises immutably stored client model hashes to verify the aggregated models and their weights without the need to retrieve individual client models preserving their privacy. We leverage homomorphic hash function to avoid complicated key exchange protocols to guarantee the models' integrity. Simulation-based experiments highlight that the proposed VHFL achieves significantly better model accuracy than HFL without verification when the HFL system is under attacks. Moreover, VHFL has low training time overheads and can successfully recover the cloud model under different edge server attacks.