Research And Simulation Of Secure Federated Learning For The Internet Of Vehicles

The Internet of Vehicles(IoV)is a communication technology that can connect and exchange data among vehicles,roads,people,transportation infrastructure,environment,and cloud platforms.It is also a key technology to achieve intelligent transportation and unmanned driving.However,while helping vehicles share data and obtain more information and services,the Internet of Vehicles also brings privacy and security issues to vehicle users.In order to obtain more information by using the data collected by each vehicle and at the same time protect the privacy of users,federated learning technology has been introduced into the Internet of Vehicles scenario.However,the attacking behavior of malicious vehicles in the open environment of the Internet of Vehicles poses a security threat to federated learning,and the training quality of federated learning cannot be guaranteed.In addition,the heterogeneous training quality and data sets,computing,and communication resources of different vehicles make designing an optimal federated learning node selection strategy a challenge.This paper focuses on the research on the security issues of federated learning in the Internet of Vehicles environment.The main contents are as follows:(1)Aiming at the malicious node poisoning attack in federated learning,a blockchain based security protection and reputation management mechanism for federated learning in the Internet of Vehicles is proposed,considering the characteristics of hierarchical federated learning in the Internet of Vehicles.This mechanism includes two parts:model detection and reputation value update.In the model detection process,RSU calculates the global model when the local model uploaded by vehicles and the global model issued by BS,considering the size information of the vehicle dataset,to calculate the global model where the local model does not participate in global aggregation.By combining accuracy and classification recall indicators,the impact of the local model on the global model is evaluated.In the process of updating the reputation value,the vehicle reputation value is evaluated based on the impact of the local model on the global model.The cost of malicious behavior of vehicles is improved by comprehensively considering the historical behavior and interaction timeliness of the vehicle.Blockchain is used as a traceable storage platform to record the vehicle reputation value and local model information,providing a mechanism for post validation.The experimental results indicate that the mechanism proposed in this thesis can resist various security attacks in federated learning of the Internet of Vehicles.(2)Aiming at the problems of irresistible attacks,low efficiency of federated learning,and low energy efficiency caused by the superposition of heterogeneous training capabilities of vehicle nodes and federated learning security attacks,this thesis proposes a trusted vehicle node optimization selection algorithm based on deep reinforcement learning.The algorithm uses vehicle computing capability,communication capability,data set size,reputation value,etc.as information features,and takes global model loss and average energy consumption as optimization goals,and selects appropriate vehicle nodes for model training before each round of global iteration.The simulation results show that the algorithm can resist federated learning attacks in the complex environment of the Internet of Vehicles,reduce the global convergence rounds of federated learning,and improve the efficiency of federated learning.Compared with other node selection schemes,this algorithm can reduce vehicle invalid training,thereby reducing Energy consumption.(3)Aiming at the security attack problem of the federated learning of the Internet of Vehicles,combined with the characteristics of the Internet of Vehicles environment,this thesis simulates and implements a secure federated learning system for the Internet of Vehicles.The system implements functional modules including federated learning task management,vehicle resource information processing,vehicle node selection algorithm,federated learning model training,local model detection and reputation value update,blockchain network management,and verifies the aforementioned mechanisms and algorithms on the system.Test results show that the designed system can meet the functional requirements of federated learning for Internet of Vehicles safety.