Research On Task Offloading Decision For Internet Of Vehicles Based On Data Privacy Protection

With the rapid development of automobile industry,a large number of new in-vehicle applications have emerged.These in-vehicle applications not only provide drivers with a lot of traffic and safety information to avoid traffic accidents,but also provide drivers and passengers with rich services to enrich people’s travel experience.However,the emergence of in-vehicle applications brings convenience to people’s lives,but also brings a lot of challenges.At present,the emerging in-vehicle applications have greater computing requirements than ordinary applications.The vehicle’s own computing power and data storage capacity can’t calculate the computing tasks generated by such applications.Traditional cloud computing can provide sufficient computing resources for such applications,but the excessive transmission delay often cannot meet the needs of low delay services.Mobile edge computing(MEC)enables vehicle users to flexibly offload computing tasks to roadside units(RSU)by sinking computing resources to RSU.The vehicle can use the MEC edge server configured by RSU to calculate,which can not only make up for the shortage of computing resources at the vehicle user side,but also effectively reduce the transmission delay.However,the general task unloading decision-making algorithm can’t protect the privacy of the user side,and it can’t adapt to the dynamic changes of vehicle networking and the high-speed movement of vehicles.To solve the above problems,this paper designs a task unloading algorithm that can protect data privacy for the Internet of Vehicles,mainly including the following contents:Firstly,this paper designs the network architecture of vehicle networking based on MEC.This architecture is more suitable for high-speed moving vehicles and dynamic networking of vehicles.Based on the vehicle networking network architecture proposed in this paper,a cooperation scheme between MEC nodes is proposed.The vehicle can make the task unloading decision and resource allocation decision based on the current remaining resource information of RSU.The vehicle will upload all input data and decision information to the RSU where it is currently located.The RSU can distribute the input data to the RSU that decides to unload the vehicle through wired connection,and send the final result back to the vehicle.The vehicle application is modeled as a set of dependent subtasks,and its structure is represented as a directed acyclic graph(DAG).In this paper,the communication model and calculation model are also modeled to simulate the dynamic channel conditions and calculate the time delay in the vehicle networking.Finally,the total unloading delay of the computing task in the unloading process is obtained.Based on the above model,this paper takes minimizing the total unloading delay as the optimization objective,and the task unloading decision problem is set up as a mathematical optimization problemSecondly,the optimization problem proposed in this paper is an integer programming problem.This problem is NP-hard,and the subtasks divided by applications have strong dependencies.Therefore,reinforcement learning is used to solve the proposed optimization problem.In this paper,the problem is transformed into Markov Decision Process(MDP).This paper also establishes the state space,action space and reward function.To reduce the complexity of the algorithm and improve the convergence speed,Double-DQN(Deep Q-Network)algorithm is used to solve this problem.Through experiments and simulations,it is proved that the algorithm can not only achieve convergence quickly,but also improve the task completion rate and resource utilization compared with other algorithms.Finally,reinforcement learning is prone to user data privacy leakage in the training process.In this paper,the framework of federated learning(FL)is added to the designed reinforcement learning algorithm.The user’s private data is kept locally by transmitting model parameters,so as to ensure the privacy of the user’s data.Through experiments and simulations,it is proved that the algorithm with federated learning still has good convergence and environmental adaptability.By comparing the algorithm with other algorithms,it is proved that the algorithm can improve the completion rate of tasks and resource utilization.