A Cybersecurity-Enhanced Connected and Automated Vehicle System Achieved up to 96 Percent Cyberattack Detection Accuracy Using Real-World Roundabout Data in Michigan.
Study Evaluated Cybersecurity Approaches for CAV Using Collaborative Perception Models and Infrastructure-Based Anomaly Detection.
Ann Arbor, Michigan, United States
Hardening the CAV Ecosystem to Reduce Cybersecurity Risks – Year One
Summary Information
Connected and Automated Vehicle (CAV) systems face growing cybersecurity threats across sensors, communication links, cloud infrastructure, and artificial intelligence (AI) models. This study developed a holistic cybersecurity framework to strengthen and secure the CAV ecosystem, covering four major areas: (i) threat analysis and risk assessment (TARA), (ii) shared-state cybersecurity defense, (iii) collaborative perception under lossy networks, and (iv) infrastructure-based anomaly detection.
METHODOLOGY
The study evaluated the four interconnected cybersecurity components via data analysis, field experiments, and simulation. Data sources included three multi-agent perception benchmark datasets for collaborative perception evaluation and a real-world roundabout trajectory dataset comprising 278 vehicle trajectory pairs collected at a signalized intersection in Ann Arbor, Michigan. Benefits were evaluated using several performance metrics across the four components.
- For the TARA, evaluation was based on identifying attack feasibility, likelihood, and potential impacts on safety and system operations.
- For the shared-state teleoperation security approach, assessment focused on anomaly detection capability and communication overhead.
- For the perception framework, performance was measured using object detection Average Precision (AP) under varying network packet loss, localization errors, synchronization delays, and combined noise conditions.
- For the infrastructure-based anomaly detection framework, performance metrics included accuracy, precision, recall, F1-score (a single metric that balances precision and recall).
FINDINGS
- The robust collaborative perception framework improved collaborative perception performance by 5.9 to 13.2 percentage in AP across multiple synthetic and real-world collaborative 3D object detection datasets under both standard and noisy scenarios. It maintained higher detection accuracy under severe network packet loss (i.e., an 80 percent loss rate), outperforming competing approaches by up to 7.5 percent.
- Under synchronization delays, the robust collaborative perception framework improved perception accuracy by 1.1 percent to 11.6 percent.
- The infrastructure-based anomaly detection achieved 100 percent accuracy in offline detection and 96 percent accuracy in online detection for time-interference cyberattacks, with an average 1.4-second early detection lead time and achieved 92.86 percent online detection accuracy for V2X communication attack scenarios.
