| With the development of technologies such as big data and the Internet of Vehicles,vehicle-road collaboration has become an important research direction in the field of intelligent transportation.Through the data fusion and information interaction between vehicles and intelligent roadside equipment,it can provide more accurate road traffic status and more convenient services for intelligent transportation systems and driving vehicles,but there are also a series of problems in practical application.Due to the serious risk of user privacy leakage caused by data sharing,and the heterogeneity of data collected by a large number of devices,it is difficult for existing machine learning schemes to integrate and use massive data training to achieve the complete perception of traffic state.Due to the potential malicious attackers in the data providers,the traffic model faces the security threat of poisoning attack,and the existing defense methods are based on specific data sets and have high computational overhead,so they have low applicability in complex practical application scenarios.In order to solve the above problems,the main work of this dissertation includes:(1)Aiming at the problem of limited environmental perception caused by user privacy and data heterogeneity,this dissertation proposes a vehicle-road collaborative learning method combining hierarchical federated learning and deep reinforcement learning,which realizes vehicle-road collaborative perception and system centralized decision-making.The integrated use of data collected by vehicles and roadside equipment enhances the environmental perception ability,protects the security of vehicle data privacy through federated learning technology,and makes globally optimized management decisions according to real-time traffic conditions through deep reinforcement learning technology.(2)Aiming at the problems existing in the traffic signal light control scene,the method of effective cooperation among multiple devices is considered in the practical application scene.Based on the proposed vehicle-road collaborative learning method,a vehicle-road collaborative intelligent signal light control scheme is designed.Under the premise of protecting the privacy of the original data of the vehicles,the driving data collected by the vehicles is used to analyze the degree of congestion around them,and the road traffic state data collected by the roadside equipment is combined to expand the dimension of the traffic state data and optimize the signal light control strategy,thereby further improve road traffic efficiency and ease traffic congestion.(3)Aiming at the problem that the defense scheme of poisoning attack in federated learning is not applicable in traffic scenarios,this dissertation proposes a corresponding hierarchical defense method based on the hierarchical federated learning method in the Internet of Vehicles scenario.The bottom layer performs suspicious client detection based on the similarity between local models in a distributed manner on the roadside servers,reducing the impact of malicious attackers on the global model;the upper layer on the central server further adjusts the deviation of the global model according to the verification data set extracted from the historical traffic data,and reduces the computational overhead while improving the security of the model.The work of this dissertation is based on federated learning technology,while ensuring the security of user data privacy and model,it realizes vehicle-road collaboration to perceive the real-time traffic status,and provides it to the intelligent transportation system to make control decisions.The vehicle-road collaborative learning method and the hierarchical defense method proposed in this dissertation are experimentally analyzed in the signal light control scenario to verify the feasibility of this scheme and its privacy and performance advantages compared with the existing schemes,proving that this scheme can meet the requirements of data privacy,computing efficiency and model security in the intelligent traffic control scenarios. |
