| In the Internet of Vehicles,vehicles can share information such as vehicle trajectory,status,and accidents through data sharing,thereby assisting humans to make correct decisions,which has very important application value in the fields of unmanned driving and traffic control.With the rapid development of on-board sensor systems,vehicles can sense massive data related to vehicle control and vehicle communication from sensors.Large-scale data transmission and processing require the Internet of Vehicles to provide new and more efficient data exchange and storage platforms.With the continuous development of big data,artificial intelligence and other technologies,it has become possible to use machine learning methods to process and analyze massive data,which will further promote data sharing in intelligent connected vehicles.Traditional centralized data processing methods can cause data leakage and loss of privacy,which will bring security risks to vehicles and occupants,and this method is not suiTab.for the Internet of Vehicles.Federated learning is a distributed data processing technology widely used in machine learning,which can complete the training of distributed models under the premise of protecting user privacy.However,federated learning often faces data security problems caused by poisoning attacks and inference attacks,and does not fully protect the driving safety and personnel safety of the Internet of Vehicles.Blockchain technology guarantees the privacy and security of data through cryptography,verification mechanisms,etc.,and model parameters can be stored in the blockchain.Therefore,integrated blockchain and federated learning techniques can not only ensure the privacy and security of data,but also ensure the integrity of the training model.Aiming at the data security and user data privacy issues in the process of Internet of Vehicles data sharing,a data sharing method with the ability of local model update legitimacy check is proposed,and the blockchain and federated learning technology are effectively integrated.Firstly,the proposed method updates and shares the model among multiple parties,so that the vehicle can update the model parameters locally and collaboratively complete the model exchange and verification on a global scale.Then,in view of the fact that the participating users are not all trustworthy,the malicious nodes in the network are identified and the disturbed models are filtered through the model verification function,so as to resist the influence of erroneous model parameters on the global model.After that,the grouping method is adopted,and the association list is set when the network is initialized,which reduces the number of communication links between nodes and solves the communication overhead problem.The simulation results show that the global model accuracy of DBL is improved by 6.88%and 8.61%compared with FL chain,respectively,under the two conditions of trusted users and malicious users,indicating that DBL protects the privacy of vehicle data and resists poisoning attacks to a certain extent while obtaining accurate ML models.Aiming at the problem of large waste of computing resources in PoW algorithm,this paper proposes a consensus mechanism of joint proof-of-work and proof-of-stake PoDaS to adapt to the public environment of data sharing.Firstly,in the process of model generation-verificationfiltering aggregation,the rewards corresponding to the nodes are designed,and each node can update its own reward value after receiving data processing rewards,verification rewards,and aggregation rewards according to its own contributions.Then,miner nodes compete to obtain the accounting rights of the blockchain according to the rewards they have,thereby reducing the difficulty of mining,further shortening the block time and improving the efficiency of the system.After that,malicious members will not be rewarded due to evil behavior,and their rewards will be less than normal users,which can better motivate participating nodes and ensure fairness.The experimental results show that the average block generation time of PoDaS method decreases by 5.02%and 4.51%compared with PoW under different datasets,respectively.Under different proportions of malicious users,the model learning accuracy of PoDaS algorithm is improved by 0.55%,0.98%,3.95%and 4.81%compared with traditional PoW,respectively,and the results show that the data sharing method based on PoDaS can reduce the block generation time and reduce the impact of error parameters on the global model accuracy. |
