Driver identification and authentication with active behavior modeling

Abstract

The legitimate driver of a vehicle traditionally gains authorization to access their vehicle via tokens such as ignition keys, some modern versions of which feature RFID tags. However, this token-based approach is not capable of detecting all instances of vehicle misuse. Technology trends have allowed for affordable and efficient collection of various sensor data in real time from the vehicle, its surroundings, and devices carried by the driver, such as smartphones. In this paper, we propose to use this sensory data to actively identify and authenticate the driver of a vehicle by determining characteristics which uniquely categorize individuals' driving behavior. Our approach is capable of continuously authenticating a driver throughout a driving session, as opposed to alternative approaches which are either performed offline or as a session starts. This means our modeling approach can be used to detect mid-session driving attacks, such as carjacking, which are beyond the scope of alternative driver authentication solutions. A simulated driving environment was used to collect sensory data of driver habits including steering wheel position and pedal pressure. These features are classified using a Support Vector Machine (SVM) learning algorithm. Our pilot study with 10 human subjects shows that we can use various aspects of how a vehicle is operated to successfully identify a driver under 2.5 minutes with a 95% confidence interval and with at most one false positive per driving day.