Efficient foundation model service provision over airborne networks: Collaborative tuning and inference

Deploying foundation models across distributed airborne networks offers a promising solution for delivering flexible, high-coverage, and on-demand generative AI services. However, the deployment and tuning of foundation models present critical challenges on airborne platforms such as Unmanned Aerial Vehicles (UAVs), due to the intensive computational requirements, substantial memory footprint, and high communication overhead, particularly given these platforms’ limited power and memory capacity as well as the limited communication connections. In view of these, a collaborative fine-tuning and inference framework for deploying foundation models over UAV networks is proposed, which employs a split model deployment strategy to distribute computational loads across multiple UAVs. The framework also incorporates a multi-stage fine-tuning approach utilizing a large vision model-based knowledge distillation and personalized local tuning to further enhance performance while maintaining system stability despite UAV mobility. The proposed framework could achieve foundation model fine-tuning in a memory- and computation-efficient manner. To further improve the communication and computation efficiency, two variants of the framework are proposed via leveraging over-the-air computations and parameter-efficient fine-tuning techniques in communication and local computation. Extensive experimental evaluation demonstrates the superior and stable performance of the proposed framework compared to baselines in terms of generalization, communication efficiency, memory efficiency, and scalability.