Research On Dynamic Driver Identification Technology Based On Incremental Learning

In the past few decades,the development of artificial intelligence,5th generation mobile communication technology,edge computing and other technologies has greatly enriched the functions of intelligent and connected vehicles(ICVs)and brought people comfortable travel experiences.However,these diversified functions are inseparable from a large number of access sensor devices and frequent data exchanging process.While ICVs is more closely connected with external devices,attackers also have more ways to invade which can remotely control or steal ICVs.Therefore,researchers propose the identification scheme based on driver behavior to ensure that ICVs are driven by the designated driver,which has great significance in protecting driving and property safety of vehicle owners.At present,the researches on driver identification schemes are mainly divided into two parts:One part studies driver behavior feature data extraction,which improves the accuracy and practicability of driver identification results from data source.The other part studies the construction of driver identification model,which uses more efficient algorithms to improve the accuracy of identification results.However,most of these existing studies are based on the static scenarios with a fixed number of drivers,and only focus on the accuracy results of driver identification schemes.In the dynamic scenarios where drivers continually join the authentication model,these schemes not only need additional space to store the behavior data of existing drivers in model,but also need to retrain the whole identification model when the new driver is added to the model,which consumes a lot of space and time resources.Based on the limitations of existing driver identification schemes,this paper studies the driver identification works in dynamic scenarios,and proposes a dynamic driver identification scheme based on incremental learning for the first time.The main research contents of this paper are summarized as follows:(1)In this paper,we propose a reliable driver behavior recognition and behavior feature extraction method,design experiments to verify the results.Most of the existing schemes obtain driver behavior data from vehicle simulators,external sensors or documents provided by vehicle manufacturers.Different from these works,we obtain CAN bus data from two real vehicles,analyze CAN frames through reverse engineering,recognize driver behaviors,and construct a reliable driver behavior feature data group.(2)Based on the analysis of driver identification works in dynamic scenarios,this paper proposes a dynamic driver identification scheme based on incremental learning: D-Driver ID,which can verify and identify the authorized drivers.D-Driver ID does not need to store the data of drivers already in the identification model.When new drivers join the model,they can dynamically adjust the model structure and efficiently complete the model retraining process.(3)In order to verify the advantages of D-Driver ID in dynamic scenarios,this paper uses two classical driver identification schemes and two dynamic driver identification schemes to do compare experiments with D-Driver ID.The experimental results show that D-Driver ID is significantly better than the existing driver identification schemes in space and time utilization.Compared with similar dynamic identification schemes,D-Driver ID can better combat the problem of catastrophic forgetting and is more suitable for application in real dynamic scenarios.