Data-Driven Battery Attack Detection and Control Behavior Determination for Vehicle Driving Safety

Pure electric vehicles (EVs) have become popular in current transportation systems because of their zero air pollution emissions. The battery management system (BMS) in an EV monitors battery information (current, voltage and temperature) in real time to prevent batteries from overcharging or overheating and also shares these battery information to outside-vehicle environments (e.g., smartphone apps) to enrich vehicle usage experiences. Many researches have studied various aspects regarding vehicle driving safety, few works have comprehensively studied battery security and its effects on driving safety of an EV. Besides, autonomous vehicles (AVs) have been adopted to reduce traffic congestion and multiple AVs will drive on the same road with the AV population growth. Therefore, It is critical to make optimal control decisions of multiple AVs in real time to ensure driving safety.

Motivated by the above scenarios, we focus on three areas to improve vehicle driving safety: (1) battery authentication system for detecting battery attacks (i.e., malicious AC-turn-on requests or battery-charge-stop requests) in electric vehicles; (2) control policy based driving safety system for an individual AV; and (3) multi-AV control decision making system for multiple AVs to ensure their driving safety. First, we propose the first battery attack, which can turn on air condition and stop battery charging process by sending requests through a smartphone without being noticed by users, and design a Battery authentication method (Bauth) to detect such battery attacks. Bauth describes a user’s habits in turning on air condition and stopping battery charging using a data-driven behavior model. It then applies the behavior model into a reinforcement learning model to judge whether an AC-turn-on or a battery-charge-stop request is from a real user. From real-life daily driving experiments, we find that Bauth can prevent EV batteries from being attacked accurately and its accuracy reaches as high as 95.6%. Second, we propose a control policy based driving safety system (Polsa) to help improve driving safety of a given AV. For a given AV, Polsa extracts its control policies and determines the safest control behavior among multiple control behaviors for each given trigger condition. Accordingly, Polsa has a control policy extraction method using dynamic time warping and k-means clustering technologies to cluster historical driving data with the same control behavior type together and then analyzes positions and driving speeds in each cluster to extract control policies of a target AV. It then develops an optimal control policy determination method to determine the safest control behavior for each given trigger condition by considering time-varying driving state of its nearby vehicle. We use an industry-standard AV platform (Baidu Apollo) to evaluate optimal control policy success rate of Polsa and find that Polsa can extract control policies with as much as 83% accuracy, and improve optimal control policy success rate by 28% compared with existing methods. Third, we propose a multi-AV control decision making system (MADM), which considers multi-AV coexistence driving situations. MADM builds a policy formation method to form policies to learn driving behaviors of an expert based on the expert driving trajectory data. It then builds a multi-AV control decision making method, which adjusts the formed policies through a multi-agent reinforcement learning and forms safety driving state of each AV, to make multiple control decisions with safety guarantee. We used a real-world traffic dataset to evaluate optimal control decision making performance of MADM and experimental results show that MADM reduces its emergency rate by as high as 51% compared with existing methods.