Multiple-user-oriented Model Training And Data Sharing Mechanism For Assistant Driving Businesses

With the development of smart cities,road pressure continues to increase with the increasing number of vehicles,and traffic safety and travel efficiency have become current hot issues.Assisted driving technology can provide guarantees for driving safety and improve people’s driving experience.The provision of assisted driving services includes two stages:model training and data sharing.In terms of model training,the assisted driving scheme based on traditional distributed machine learning has problems such as high network latency,vehicle privacy leakage,and poor model universality caused by differences in slope,height limit,and weight limit in different regions.As a distributed computing architecture that can address issues of data privacy and communication efficiency,Federated Learning(FL)can provide personalized model training services for assisted driving services.However,in the process of FL model training,existing synchronous aggregation and asynchronous aggregation methods have problems with long aggregation time and large convergence fluctuations,respectively,and are not suitable for delay sensitive and high accuracy required auxiliary driving services.In addition,some terminals with low data quality or limited available resources participating in training and uploading low-quality model data will result in poor service quality for the Internet of Vehicles.On the other hand,in the process of data sharing,traditional assisted driving solutions mostly adopt a centralized storage and processing mode,which poses risks such as database failures and tampering of transaction records.Blockchain technology with distributed storage and tamper resistance can effectively ensure the security of data.However,the continuous expansion of the blockchain ledger consumes too much storage space at blockchain nodes,and data between different institutions should be stored separately to prevent data leakage.Therefore,it is necessary to design a reasonable blockchain structure to improve the scalability and data security of the blockchain.In addition,blockchain requires a consensus process to ensure data consistency,and the difficulty for vehicles to handle this process will result in a large number of computationally intensive tasks,leading to high latency.Therefore,the combination of mobile edge computing(MEC)and blockchain system can improve the computing efficiency and accelerate the consensus process through the cooperation between terminals and edge nodes.However,the existing computing task offloading mechanisms lack refined task allocation methods and cannot adapt corresponding offloading decisions based on the dynamic changes in terminal available resources,resulting in low resource utilization.Therefore,in response to issues such as user privacy data leakage and low model training efficiency during the data utilization process,this paper designs an auxiliary driving model training mechanism based on semi asynchronous federated learning.Firstly,a system model including an auxiliary driving service framework,a delay model,and a trust value model was proposed.Then,a training node selection algorithm for regional business was proposed,effectively avoiding the participation of low-quality nodes in model training.Furthermore,a double-layer FL semi asynchronous aggregation mechanism based on directed acyclic graphs was proposed,and a double-layer FL framework consisting of a vehicle terminal aggregation layer and an edge node aggregation layer was constructed.The local model asynchronous aggregation mechanism based on directed acyclic graphs and the region model semi asynchronous aggregation mechanism based on model accuracy were adopted,respectively,to further improve the aggregation efficiency of the FL system.The simulation results show that the proposed mechanism outperforms traditional synchronous/asynchronous FL mechanisms in terms of training delay and model accuracy.In response to the issues of user privacy data leakage and low sharing efficiency in the data sharing process,this article first constructs a 1+n main side blockchain architecture,reducing the number of consensus nodes on each side chain.By sharing index information of driving data on the main chain and sharing driving data and vehicle information on the side chain,the communication cost of the consensus process is reduced,consensus efficiency is improved,and data leakage between different types of vehicles is avoided.Specifically,for the calculation tasks(mining tasks)that some vehicles cannot process in a timely manner during the consensus process,MEC is integrated with the blockchain system to design a double-layer unloading mechanism for blockchain mining tasks based on edge collaboration.This enables edge nodes and adjacent vehicles to form a collaborative mining network and collaborate with accurate unloading rates to process mining tasks.The simulation results show that the mechanism proposed in this paper outperforms binary and single point offloading algorithms in traditional single chain blockchain systems in terms of average profit and consensus delay.