摘要
Hierarchical Federated Learning (HFL) has emerged as a prominent framework for jointly addressing communication overhead and statistical heterogeneity, which are the core design elements of current Federated Learning systems. However, HFL triggers new challenges in client assignment and the decision to choose different levels of aggregation frequency among clients, edge, and cloud. This paper proposes an Adaptive Hierarchical Federated Learning (AdaptHFL) framework to handle the above issues jointly. First, we advocate the inter-edge similarity metric as our client assignment strategy to minimize edge model divergence, which amplifies HFL's inherent advantages on non-IID data. Then, we design an HFL framework booster by dynamically balancing local training iterations and edge aggregation frequency. Experimental results demonstrate that our AdaptHFL framework has superior model accuracy, 7% higher than other baselines, on various non-independent and identically distributed (non-IID) scenarios. In addition, our proposed dual-level adaptive mechanism can reduce time consumption by 28% and flexibly complement state-of-the-art client assignment approaches to improve the HFL system's convergence speed further. © 2026 IEEE.