G-DriverAUT: A Growable Driver Authentication Scheme Based on Incremental Learning

As an important part of intelligent and connected vehicles (ICVs) safety, driver authentication is a hot research direction in recent years. The existing driver authentication schemes in the industry are based on scret key, biometric characteristics or image characteristics, and these schemes always have efficiency and security problems in real-time driver authentication. In order to make up for these shortcomings, researchers propose driver authentication schemes based on driver behaviors and conduct in-depth explorations. However, the existing behavior-based schemes rely on a fixed number of drivers, when a new driver is added as an authorized user, existing schemes need to retrain the whole model to authenticate all drivers, cost many computing resources and can not meet the low latency requirements of ICVs. To solve the above deficiencies, we propose a growable driver authentication scheme based on incremental learning for the first time to improve authentication efficiency in dynamic scenarios where the number of drivers is constantly increasing. We extract driver behavior characteristic data from two real vehicles and design a growable model based on incremental learning for driver authentication. Besides these, we also use the support vector machine (SVM), convolutional neural network (CNN) and some other algorithms to make comparative experiments. The experimental results indicate that our scheme not only has high authentication accuracy, but also greatly reduces the time of model retraining.