An Optimized Global Disturbance Adversarial Attack Method for Infrared Object Detection

Deep neural network (DNN) has various applications in various fields. It realizes the function of real-time object detection used in transportation, medical treatment, and industry. It is closely related to our daily life. During the epidemic, infrared object detection has also been widely used, but the vulnerability and security of neural networks should be paid attention. To understand the security of object detection networks and empha-size the urgent need to develop robust systems, a Faster-RCNN object detection attack method for infrared images is proposed in this paper. By controlling the gradient update direction and loss function optimization, an adversarial disturbance is customized for each input image so that the object detection network (Faster-RCNN) marks the target error or creates a new wrong target, thus deceiving the object detection network and makes it detect errors and invalidate the task. In this paper, we show how to generate negative infrared images by adding a tiny disturbance to the infrared image to deceive the object detection network at a significant level, making it visually imperceptible and impossible to detect in the network. Experiments are performed on FLIR dataset, and our methods are compared and verified.