Data-Driven Misbehavior Generation and Detection Methodologies for Intelligent Transportation Systems
Abstract:
Intelligent Transportation Systems (ITSs) is a major application field for Cyber-Physical Systems. Future public transportation systems will feature different modes of transportation where different driving telematics data and other data sources (e.g., traffic counts, traffic signal phases, and road network map) will be utilized to increase roadway safety and traffic efficiency. Therefore, several new applications will be offered, such as Usage-based Insurance (UBI) programs and Connected Autonomous Vehicles (CAV) systems. However, such driving telematics data usages will bring new possibilities of misbehavior scenarios in terms of data-driven privacy and security aspects. From an attacker's perspective, a location privacy attack uses the driving telematics data to determine the complete roundabouts of a victim vehicle. An attacker can generate an adversarial attack towards a CAV vehicle. To prevent attacks, vehicles can utilize a misbehavior detection system utilizing the driving telematics data along with the current traffic flow and road features.
Motivated by the above scenarios, we design a location privacy attack, a black-box adversarial attack, and a misbehavior detection system utilizing the driving telematics data in this dissertation. First, we design a path inference method using the driving telematics data, which first classifies the brake signal of a trip into four different driving maneuvers. The proposed path inference method predicts a pseudo graph using the classified driving maneuvers. It then utilizes a path search algorithm with the help of the edge score function to select a list of possible paths traversed by the vehicle. From the real driving datasets, we find that the proposed path inference method can predict around 89% of the original routes regardless of drivers and vehicles. Second, we design a black-box adversarial attack in the CAV scenarios, which achieves 71.49% success rates with 27.31 queries to the deep learning model and 4.35% adversarial perturbations on average and an adversarial attack that can avoid being predicted. The offline perturbation generation procedure in the proposed attack reduces the query counts, and the online perturbation generation procedure helps increase the attack success rates. Third, we design and implement a misbehavior detection system in the CAV scenarios to detect and identify the false information received through vehicular communication technologies. The proposed misbehavior detection system utilizes the acceleration patterns, current traffic flow, and maneuver-specific vehicle movements to design a rule-based detection mechanism and identify false information in various traffic scenarios. The proposed system achieves higher detection accuracy, i.e., 91.75% and 87.72% precision and recall evaluated using a realistic traffic simulator with the real driving datasets.
Committee:
- Andrew Grimshaw, Committee Chair (Department of Computer Science)
- Haiying Shen, Advisor (Department of Computer Science)
- Lu Feng (Department of Computer Science)
- Yuan Tian (Department of Computer Science)
- Laura Barnes (Department of Engineering Systems & Environment)