Efficient Vehicle Selection and Resource Allocation for Knowledge Distillation-Based Federated Learning in UAV-Assisted VEC
Chunlin Li, Yong Zhang, Yu Long, Mengjie Yang
Abstract
In Vehicular Edge Computing (VEC), the high mobility of vehicles and periodic of traffic flow present challenges to the effectiveness of roadside units. Unmanned Aerial Vehicles (UAVs) can serve as aerial base stations to address this issue. Federated Learning (FL) is employed to reduce backhaul load. However, the limited battery and bandwidth of UAVs constrain long-term training capabilities. We propose a collaborative deployment of multiple UAVs to maximize communication coverage, utilizing a Particle Swarm Optimization (PSO) algorithm for optimal deployment decisions. We take into account the mobility of vehicles during vehicle selection to prevent network interruptions. Furthermore, knowledge distillation is used to compress the local model without sacrificing accuracy, thereby reducing transmission overhead and accelerating model convergence. Finally, the Deep Deterministic Policy Gradient - Double Dueling Deep Q-Network (DDPG-D3QN) algorithm addresses optimal vehicle selection and resource allocation in dynamic scenarios. Experimental results demonstrate that our approach effectively meets communication needs in urban areas while enhancing training efficiency and accuracy.