LLM-Driven Hierarchical Federated Orchestration for Privacy-Preserving 6G TN-NTN Networks

The integration of terrestrial and non-terrestrial networks (TN-NTN), spanning LEO satellites, high-altitude platform stations (HAPS), unmanned aerial vehicles (UAVs), and ground infrastructure, is a cornerstone of 6G connectivity. However, the resulting system is characterized by highly dynamic topology, heterogeneous resources, and privacy-sensitive distributed data, making scalable and privacy-preserving resource orchestration a significant challenge. To address these issues, we propose LHPFO, an LLM-guided hierarchical federated orchestration framework that enables intent-aware and privacy-preserving resource management across the TN-NTN hierarchy. The key idea is to leverage a domain-adapted large language model (LLM) as a semantic decision engine that transforms structured network context into orchestration directives and a semantic prior, which guides distributed policy updates in a hierarchical federated learning (HFL) architecture. This design improves coordination across tiers while avoiding centralized data collection. We further develop a composite privacy framework integrating (ϵ, δ)-differential privacy, secure aggregation, and geo-indistinguishability, supported by an adaptive privacy controller. The overall system is formulated as a multi-objective optimization problem that jointly balances QoS utility, latency, energy consumption, and privacy leakage. Extensive simulations demonstrate that LHPFO consistently outperforms strong baselines, achieving lower latency, higher throughput, improved energy efficiency, and better QoS fairness, while maintaining strict privacy guarantees. These results highlight the effectiveness of semantic-guided federated orchestration for next-generation TN-NTN systems.