Connected autonomous vehicles (AVs) may offer new mobility options to millions of people. Integration of connectivity features into modern vehicles is a main driving force behind the ever-expanding attack surface of connected AVs, rendering them vulnerable to hacking and data theft. Key vulnerabilities arise from the increased coupling of unsecured automotive control networks with multimedia networks and the integration of wireless interfaces such as Bluetooth and Wi-Fi networks. As such, developing robust and reliable solutions to identify, localize, and mitigate cybersecurity threats to connected AVs is of societal importance. Existing solutions, however, are limited in their ability and scope as they are unable to reliably link the received data to the transmitting devices. The goal of this project is to safeguard AVs against growing attack surfaces and vectors by developing a holistic solution called the Linking2Source framework through three seamlessly integrated layers of defense, with each layer aiming to mitigate a specific set of attacks. The project also has a significant educational component, consisting of a set of inquisitive hands-on activities involving vehicle data acquisition, decoding, and data analytics, network packet injection, and intrusion detection aimed at outreach and broadening participation in STEM disciplines, including automotive cybersecurity, cyber-physical system security, statistical data analysis and digital forensics.
The first layer of the proposed Linking2Source framework aims to protect in-vehicle networks by developing real-time message authentication, intrusion detection, and localization tools based on unclonable signal attributes for physical fingerprinting of electronic control units (ECUs). The approach exploits uniqueness in physical signal attributes, leverages statistical signal processing and parameter modeling techniques for physical fingerprint estimation, and uses statistical machine learning methods for transmitting ECU identification and localization. The second layer aims to protect in-vehicle networks against firmware/software-level attacks using ECU behavioral fingerprinting through data-driven statistical graph analytics. The approach targeted by the research team here is the transformation of sequential in-vehicle network data into a directed-graph to leverage statistical graph analytics for ECU behavior modeling and intrusion detection. The third layer of defense aims to protect AVs against attacks at the sensing and actuation layer by using dynamical observers that rely on vehicle-physics-based modeling for fault detection and isolation. The faulty signals such as incorrect steering angle commands that are issued by the rogue ECUs and are not in agreement with the vehicle physics could cause unsafe maneuvers such as excessive yaw motions. The project exploits the physics-based vehicle model for verifying the correctness of the issued ECU signals over the in-vehicle network bus. By leveraging the Dempster-Shafer evidence theory, the decisions from these layers of defense are optimally fused to integrate the three defense solutions in the Linking2Source framework. A key component of this project is to use in-vehicle network data both at the physical and datalink layers for modeling physical, behavioral, and vehicle-state fingerprints and using them for attack detection and localization and mitigation of the impact of malicious ECUs using a proactive cancellation policy. The research team will prototype the proposed solutions and evaluate them on the University of Michigan-Dearborn shuttle, on the University of Michigan MCity Test Facility, and on commercial tools, in addition to collecting large-scale data from a network testbed and from a real vehicle driving and sharing it with the research community.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
