Research On Autonomous Driving Strategy For Internet Of Vehicle Based On Federated Learning

With the rapid development of Internet of Vehicles and artificial intelligence technology,federated learning,as an emerging machine learning paradigm,has unique advantages in protecting privacy and reducing communication overhead.In the existing intelligence connected vehicles scenarios based on federated learning,insufficient data volume and privacy issues seriously limit the optimization of vehicle driving strategies,lead to insufficient model accuracy of vehicles,and even affect the safety of vehicles.At the same time,in the process of federated learning,malicious nodes attack the global model,resulting in model poisoning and accuracy degradation.In order to ensure the reliability of the model,accelerate the convergence speed and reduce the communication overhead,identifying malicious nodes is a major challenge in federated learning.This paper studies the optimization of intelligent connected vehicles driving strategy based on federated learning in the Internet of Vehicles scenario,and the research of client selection in federated learning as follows:(1)In this paper,we propose a research on driving strategy optimization of intelligence connected vehicles based on federated learning.Under complex and variable road conditions,conditional imitation learning is used to obtain single-vehicle intelligent driving strategies,and the driving strategies of multiple vehicles are aggregated based on simultaneous federated learning,and the aggregation stage assigns the same weight to all vehicles to ensure the fairness of aggregation.It breaks the barrier of insufficient data and optimizes the driving strategies of intelligence connected vehicles.Vehicle models are dynamically validated on the Carla platform,and the experimental results show that the proposed federated learning-based intelligence connected vehicles model significantly improves the success rate of turning task by 15% and the success rate of straight driving by 21% compared with the single vehicle intelligent model,in addition to the accuracy of the model and the convergence speed are better.(2)In this paper,we propose a reputation federated learning research for energy saving client selection.In the Internet of Vehicles scenario,the energy consumption model of the vehicle is formed together with the original on-board energy based on the amount of natural energy drawn by the vehicle through the green energy absorption device.Considering that malicious devices will attack the global model during the federated learning process,a reputation mechanism is designed to calculate the reputation value of a vehicle through three parameters,namely honesty,model accuracy,and interaction timeliness,and set a variable threshold to evaluate multiple vehicles in this way and screen out malicious vehicles.In order to minimize vehicle energy consumption while maximizing the number of vehicles with high reputation that participate in federated learning,the energy consumption model and the reputation mechanism are jointly optimized and modeled as a Markov Decision Process(MDP).The vehicles with low energy consumption and high reputation are selected to participate in federated learning.To solve this MDP problem,the Twin Delayed Deep Deterministic policy gradient algorithm is used in this paper,and the simulation is based on a Python simulator.The experimental results show that the scheme designed in this paper successfully selects vehicles with low energy consumption and high credibility,and achieves the expected results.