Privacy preserving federated fine tuning of large language models for Autonomous Aerial Vehicles command generation

Fine tuning large language models for Autonomous Aerial Vehicles command generation exposes flight logs, patrol routes, and Global Positioning System (GPS) coordinates during collaborative training. This paper presents a privacy preserving federated fine tuning framework that maps operator instructions to structured JavaScript Object Notation (JSON) drone commands while defending against privacy leakage, server visibility, and poisoned clients. The framework builds on Low Rank Adaptation (LoRA) based Federated Averaging (FedAvg). Differential privacy (DP) is enforced through per step Differentially Private Stochastic Gradient Descent (DP-SGD) with Rényi Differential Privacy (RDP) accounting, which bounds per client leakage. Elliptic Curve Cryptography (ECC) based pairwise secure aggregation over the National Institute of Standards and Technology (NIST) P-256 curve with hash based key derivation hides individual updates. Krum and coordinate wise trimmed mean filter Byzantine updates. We evaluate 14,446 command pairs across six mission categories under non independent and identically distributed (non IID) Dirichlet partitions. With five simulated clients, TinyLlama-1.1B, ɛ = 8, and Krum, the full framework reaches a Bilingual Evaluation Understudy (BLEU) score of 0.8608 and 100% JSON validity under one sign flip attacker, compared with 0.8618 BLEU for the centralized baseline. LoRA subsampling makes per step DP noise negligible (∼10⁻⁶ per parameter), so data heterogeneity dominates utility. Without Byzantine resilient aggregation, FedAvg falls to 76.6% JSON validity.