Task offloading and computing in vehicular edge networks: A joint of dynamic game and federated learning

Federated Learning (FL) is well-suited for the Internet of Vehicles (IoV) due to its privacy-preserving and distributed nature. However, its application in Vehicular Edge Computing (VEC) faces key challenges: frequent task timeouts caused by high vehicle mobility, performance degradation from heterogeneous data, and the disjointed optimization of offloading and computation processes. To address these issues, this paper introduces a Edge Computing Service Provider (ECSP) to manage task allocation. We then decouple the VEC process into task offloading and computation. For the offloading process, we model it as a two-layer Stackelberg game and propose a novel Joint Dynamic Game and Federated Learning (JDGFL) algorithm. JDGFL enables dynamic strategy adaptation for the ECSP and vehicles, and integrates this with a three-tier "Vehicle-ECSP-RSU" FL architecture for the computation process. Extensive simulations demonstrate that our JDGFL strategy achieves significant performance improvements, reducing latency by 15 %, improving energy efficiency by 20 %, and increasing model accuracy by 5 % on average compared to static game-theoretic and standard FL baselines.