On the Importance of Contextual Information for Building Reliable Automated Driver Identification Systems.

Recent studies on machine learning based driver identification have shown that leveraging deep neural networks to learn latent features from vehicle sensor data boosts the performance of the models to high levels of accuracies. However, models produced by deep neural networks are difficult to explain and the interpretability of their results is limited. Consequently, the reliability of the learned models heavily depends on the amount, quality and diversity of the training dataset as well as on the validation scenarios. In this work, we evaluate state-of-the-art deep learning networks for driver identification using a very rich dataset of more than 395,000 kilometres of different driving scenarios and environmental conditions (e.g. route, vehicle, traffic, weather) collected over two years. It turns out that the neural networks achieve high accuracy levels when training and testing on the same type of driving conditions. However, accuracy drops and importance of individual signals varies when testing on different driving conditions, although all best practices like stratification and cross validation have been applied. Our findings suggest that relying solely on the vehicle sensor data without taking the contextual information about driving conditions into account is not a practical solution.