A Secure Hierarchical Federated Learning Architecture for UAV-Enabled Edge Computing Systems

Abstract

Unmanned aerial vehicles (UAVs) are emerging as mobile edge nodes for temporary coverage, aerial sensing, disaster response, public-safety monitoring, intelligent transportation, and smart-agriculture services. Although federated learning (FL) enables distributed model training without transferring raw data, conventional flat FL is not well aligned with UAV-enabled edge systems because ground clients may experience mobility-dependent availability, unstable wireless links, and costly longrange synchronization with a remote coordinator. Hierarchical federated learning (HFL) mitigates these limitations by introducing UAV edge aggregators between ground clients and the global coordinator. This article provides a simulation-based assessment of a secure HFL-UAV architecture implemented in Python. The simulator captures non-independent and identically distributed (non-IID) client data, mobility-aware client-UAV association, two-level model aggregation, malicious client behavior, trust-weighted robust aggregation, optional differential privacy, secure-aggregation overhead, communication cost, latency, and UAV energy consumption. Under a model-replacement attack, the secure HFL-UAV scheme maintains the learning performance of the considered synthetic classification task while substantially reducing the aggregation trust assigned to malicious clients. The results also expose important design tradeoffs: security mechanisms increase local communication overhead, UAV-assisted aggregation introduces energy cost, and trust filtering must be calibrated carefully under non-IID data. The proposed simulation framework therefore provides a reproducible basis for evaluating secure aerial edge learning systems.