Enabling Privacy-Preserving and Drop-Out Resilient Federated LLM Fine-Tuning for the Low-Altitude UAV Swarm Networks

Advancement of unmanned aerial vehicle (UAV) swarm networks presents transformative opportunities for low-altitude surveillance, disaster response, and distributed sensing, where federated large language models (LLMs) enable collaborative learning while preserving data privacy, enhance swarm-level situational awareness through decentralized knowledge fusion, and support adaptive decision-making across dynamic low-altitude operational environments. However, federated LLM fine-tuning for UAV swarm networks operating in low-altitude settings faces three unresolved security and practical issues: (1) Lack of efficient methods to protect parameter security during uplink/downlink transmission under low-altitude communication constraints; (2) Absence of effective mechanisms to handle frequent UAV dropouts caused by low-altitude dynamics that may compromise the robustness of federated LLM systems; and (3) Constraints in UAVs’ computing, storage and communication resources under typical low-altitude mission profiles. To address these challenges, this paper proposes a Secure and privacy-preserving federated fine-tuning (SPFF) scheme for low-altitude UAV swarms that enables: efficient and privacy-preserving one-to-many distribution of global parameters for downlink federated fine-tuning; secure and efficient uplink local parameter uploading adapted to low-altitude network conditions; and encrypted-parameter-based global model fine-tuning. The scheme also incorporates an efficient supervised key update mechanism to address UAV dropout issues common in low-altitude operations. Moreover, we design a delegable extensional SPFF (DE-SPFF) scheme that employs proxy re-encryption to allow UAVs to delegate tasks to other drones before exiting the federated fine-tuning process in volatile low-altitude environments, while providing public verifiability for re-encryption operations performed by semi-trusted edge nodes. Formal security proofs demonstrate the security of the proposed schemes under low-altitude threat models. Theoretical analysis and experimental results confirm their superiority and practicality for low-altitude UAV swarm applications.