Resource Management Of Federated Learning In Vehicular Edge Computing

Vehicular Edge Computing(VEC)is the integration of emerging Mobile Edge Computing(MEC)with traditional vehicle networks.The MEC server at the edge of the network provides communication,computing,storage and data resources for nearby vehicles to meet the demand of some new application services for low latency and high reliability.In order to further reduce communication costs and protect the privacy of vehicle users,Federated Learning(FL)is introduced into VEC.In VEC,designing an efficient resource management scheme,which can meet the needs of FL is crucial to improve the quality of service and user experience of real-time applications.Resource management mainly focuses on resource allocation and resource efficiency optimization.Most of the existing resource allocation studies lack full consideration of the high mobility of vehicles and data offloading assisted FL,and cannot meet the highly dynamic scenario requirements of VEC.In resource-constrained VEC,existing resource efficiency optimization studies allocate the same local training epochs for all vehicles,ignoring the system heterogeneity and data imbalance,resulting in a long waiting time and low resource utilization.Therefore,this paper focuses on the resource management of FL in VEC,and conducts research from two aspects:resource allocation and resource efficiency optimization.The main research contents include:(1)A dynamic resource allocation method based on vehicle mobility and data offloading is proposed to minimize the delay and model training loss.This method combines the high mobility of vehicles and considers the heterogeneity and limitation of computing and communication resources,proposes a vehicle selection algorithm based on scoring rules.The resource allocation problem is abstracted as a multi-objective optimization problem,and proposes a genetic algorithm with constraints to solve it.Simulation results show that the proposed algorithm can adapt well to VEC scenarios and has higher model training performance compared with the existing schemes.(2)In resource-constrained VEC,an adaptive FL local training method for resource efficiency optimization is proposed.The method fully considers the impact of system heterogeneity and data imbalance on computation and communication overhead.The method adaptively determines different local training epochs and global aggregation weights according to the computing power and the size of local data of each vehicle to minimize the FL training delay and accelerate the model training speed,and then improve the resource utilization efficiency of the whole system.Simulation results show that the proposed algorithm can achieve fast convergence of FL model under limited resource conditions and improve model training performance and resource utilization.