关键词: 车辆轨迹预测, 对抗性攻击, 智能驾驶车辆, 鲁棒性

Abstract:

Considering the lack of extreme traffic scenarios in conventional vehicle trajectory prediction datasets， a novel adversarial attack framework to simulate such scenarios is proposed in this paper. Firstly， a threshold determination method is proposed to judge the effectiveness of adversarial attacks under different scenarios. Then， two adversarial trajectory generation algorithms are designed for different attack objectives， which generate more adversarial samples under physical and concealment constraints. In addition， three new evaluation metrics are proposed to comprehensively assess attack effect. Finally， different defense strategies are explored to mitigate adversarial attacks. Experiments results show that the Perturbation Threshold for Fast Attack （PTFA） algorithm and the Attack Algorithm Based on Dynamic Learning Rate Adjustment （DLRA） achieve shorter attack time and better perturbation effect compared to existing algorithms on the NGSIM dataset， discovering model vulnerabilities more efficiently. By simulating extreme cases， this research enriches trajectory samples， evaluates model robustness in-depth， and lays a foundation for further optimization.

Key words: vehicle trajectory prediction, adversarial attacks, intelligent driving vehicles, robustness