UAV-Assisted Task Offloading and Resource Allocation in Internet of Vehicles: An Integration of Digital Twin and Generative AI

With the increasing deployment of environment-aware services in the Internet of Vehicles (IoV), vehicles are required to execute multiple computational tasks in real time. However, resource allocation and task offloading in unmanned aerial vehicles (UAVs)-assisted IoV systems remain challenging due to the growing number of vehicle terminals (VTs), potential privacy leakage, and resource-constrained edge devices. This paper proposes a digital twin (DT) and generative artificial intelligence (GAI)-powered hierarchical aerial-ground cooperative architecture (DTG-HACA) that achieves dynamic resource optimization through a three-layer framework. The DT layer enables real-time synchronization of vehicle/UAV states and simulated trajectory planning. The high altitude platforms (HAPs) layer provides low-latency offloading channels through stratospheric wide-area coverage and solar-powered endurance, while the physical entity layer executes energy-efficient edge computing via UAV-vehicle-roadside units (RSUs) collaboration. For UAV trajectory optimization, we introduce the multi-agent deep deterministic policy gradient (MADDPG)-improved prioritized experience replay (MADDPG-IPER) algorithm that minimizes communication overhead and energy consumption while integrating DT-simulated trajectory planning. For the joint challenge of edge caching and task offloading under privacy preservation constraints, we develop a federated deep reinforcement learning (FDRL) based generative adversarial network (FDRL-GAN) algorithm. This solution addresses critical challenges in dynamic task offloading and resource allocation for UAV-assisted IoV by leveraging GAI to predict task demands for cache hit rate optimization, while implementing FDRL for distributed privacy-preserving decision-making without raw data sharing, thereby achieving global resource allocation optimality. Extensive simulation experiments confirm that our proposed scheme demonstrates significant advantages over existing benchmark algorithms across five critical performance metrics, including training stability, computational capacity, task offloading efficiency, cache hit rate, and energy consumption.