A Novel Unsupervised Anomaly Detection Method on Adversarial Attacks for Autonomous Vehicles Trajectory Prediction
Jiping Fan, Zhenpo Wang, Guoqiang Li · 2024
Current trajectory prediction methods for autonomous vehicles commonly rely on deep neural networks, which are vulnerable to adversarial attacks. To enhance the security of trajectory prediction, this paper proposes an anomaly detection method based on generative adversarial networks. Firstly, a novel unsupervised anomaly detection model is proposed, taking into account both temporal and spatial features of trajectories with Long Short-Term Memory. The networks are trained using max-min game theory between the generator and the discriminator to capture the normal driving feature distribution. Furthermore, trajectory data is mapped to the latent space, and the generator reconstructs data from the latent space to compute reconstruction loss, while the discriminator detects trajectory data to calculate discrimination loss. Finally, anomalies are detected using an anomaly score that represents the extent to which the data point deviates from normal behavior and determines whether the trajectory of this segment is anomalous within the time window. We evaluate the method on three public datasets, and experimental results demonstrate its excellent performance under adversarial attacks.