针对车载环境感知系统的对抗样本生成方法

Adversarial Example Generation Method for Vehicle Environment Perception System

摘要:

针对车载环境感知场景中的目标检测系统，提出了一种针对目标检测器的对抗样本生成方法。该方法能够实现对目标检测器的白盒对抗攻击，包括目标隐身攻击和目标定向攻击。在Rail数据集和Cityscapes数据集中进行测试，测试结果验证了所提方法对YOLO目标检测器对抗攻击的有效性。

Abstract:

A method of generating adversarial examples against object detectors was proposed for object detection system in vehicle environment perception scenarios. The method achieves white-box adversarial attacks on object detectors， i.e.， object invisible attacks and object targeted mis-detectable attacks. On the Rail dataset and Cityscapes dataset， experimental results indicate that the method has good performance on the object invisible attacks and the object targeted mis-detectable attacks in the process of YOLO object detection.