Adversarial examples detection for object detectors using consistency check framework

Automatic driverless technology has been increasingly deploying object detectors that adopt deep learning networks to perceive running conditions, leading to security risks for object detector systems since recent studies have pointed out the existence of adversarial examples. Existing methods are mostly limited to classifier networks and have limitations in detecting adversarial examples targeting object detection networks, lacking generalization and decision foundations. However, accurate detection of adversarial examples targeting object detection networks is crucial for the field of automatic driving. In this paper, we propose an adversarial example detection method tailored for object detection networks, utilizing image processing techniques and consistency checks. The method first processes an input image and then feeds both the original input and its processed image into object detector systems in parallel. Subsequently, we propose an Intersection over Union (IOU) computation method for adversarial example detection. By calculating the IOU value between two images processed in parallel, we can determine whether an image constitutes an adversarial example. Our method is evaluated on the traffic running environment in the real world, and the detection method reaches high accuracy. Our research will inspire further efforts in detection and defense measures against adversarial examples in other scenarios.