Adaptive attack recognition method based on probability model for autonomous vehicle

Abstract

The perception system is essential for autonomous vehicle safety and stability. However, on-board sensors are vulnerable to external attacks, compromising driving strategies and security. Traditional attack detection methods, relying on static anomaly thresholds from data distributions, falter in complex driving scenarios. Recent studies have sought more nuanced detection techniques, including deep learning and inter-vehicle communication, but these approaches face limitations related to data security and environmental dependency. This letter introduces an innovative, adaptive method for sensor attack detection in autonomous vehicles. By modelling the perception system's functions and employing the Gaussian process for probabilistic modelling, we generate dynamic uncertainty estimates that serve as adaptable anomaly boundaries. Our method's efficacy and resilience are validated through quantitative analysis of the real-world KITTI dataset, demonstrating superior adaptive detection capabilities.