Energy-Efficient Over-the-Air Computation for Federated Generative Model Fine-Tuning in Unmanned Vehicle-Assisted Disaster Relief

Abstract

<p>The utilization of emergent generative artificial intelligence (GAI) within the realm of unmanned vehicles (UVs) can boost edge intelligence and potentially enhance the efficiency and effectiveness of rescue operations. However, to facilitate specialized generative edge intelligence in dynamic UV networks, GAI models need to tap into local data and conduct online fine-tuning, and the dispersed distribution of UVs makes distributed fine-tuning paradigms crucial. Although the federated generative model fine-tuning brings a potential solution, the large number of parameter transmissions involved in its fine-tuning process are often constrained by communication bottlenecks, which are more pronounced for resource-constrained UV networks. To cope with these challenges, we introduce an energy-efficient GAI model fine-tuning framework for the hierarchical UV networks, which employs over-the-air technology to save computational resource costs for unmanned aerial vehicle (UAV) during federated aggregation within a federated learning paradigm. Thereafter, we study the resource allocation problem in the fine-tuning process of the generative model and formulate a joint optimization problem for the bandwidth allocation, power control, computation resource allocation, and denoising factor control, aiming to minimize the system energy consumption. Then, we decouple the original problem into four tractable subproblems, and propose a block coordinate descent algorithm to solve them iteratively. Particularly, we derive a closed-form solution for the denoising factor to minimize the local model uploading transmission energy consumption under a specific AirComp communication error threshold. Simulation based on the state-of-the-art generative models shows that our AirComp-assisted federated generative model fine-tuning scheme can achieve satisfactory customized field-of-view image generation capability compared with the traditional fine-tuning scheme in an energy-efficient way.</p>