Adaptive Hierarchical Parameter Transmission for Federated Learning in MEC Environments With Dynamic Network Awareness

Federated learning (FL) in mobile edge computing (MEC) operates over wireless links with time-varying bandwidth, latency, and packet loss, which makes conventional full-model synchronization inefficient and vulnerable to stragglers. To address this issue, we propose Adaptive Hierarchical Parameter Transmission for Federated Learning (AHPT-FL), a network-aware communication-control framework that combines real-time link sensing with three adaptive transmission behaviors: full upload, distilled–incremental upload, and cached synchronization. AHPT-FL further incorporates reliability-aware aggregation to preserve training continuity under fluctuating MEC conditions. Experimental results under time-varying MEC traces show that, compared with FedAvg, AHPT-FL reduces communication cost by 25.3% and normalized uplink energy consumption by 22.8%, while improving convergence speed by 31.6%, accuracy stability by 18.7%, and successful update delivery by 9.3%. Additional ablation, sensitivity, and multi-seed results confirm that these gains remain stable across the tested threshold settings and network conditions.