Preprocessing defense method aiming at target detection confrontation attack

Abstract

The invention provides a preprocessing defense method for target detection countermeasure attacks. The method comprises the following steps: step 1, adding adversarial disturbance to an image in an original image data set to generate a target detection adversarial sample; 2, performing bilateral filtering and data normalization processing on the target detection confrontation sample to obtain a processed target detection confrontation sample; 3, constructing a noise reduction auto-encoder model, taking the processed target detection confrontation sample as input, taking the corresponding original image as a label, and carrying out noise reduction training on the noise reduction auto-encoder model; 4, acquiring an image to be subjected to target detection, and performing bilateral filtering and data normalization processing on the image to be subjected to target detection to obtain a processed image to be subjected to target detection; and 5, inputting the processed image to be subjected to target detection into the trained noise reduction auto-encoder model to obtain a denoised image to be subjected to target detection, and taking the denoised image to be subjected to target detection as the input of the target detection model to realize target detection.

Description

technical field [0001] The invention relates to the technical field of target detection and confrontation attack defense, in particular to a preprocessing defense method for target detection and confrontation attack. Background technique [0002] Object detection is one of the most basic tasks in computer vision. With the rapid development of object detection applications, it plays an important role in medical images, automatic driving, and pedestrian detection, and its safety is particularly important. However, the object detection based on deep learning inherits the shortcomings of deep networks while having excellent performance, and the field of object detection has been proved to be extremely vulnerable to the attack of adversarial examples. Many studies have introduced the idea of ​​image classification into target detection. Many adversarial attacks of image classification have migrated to the field of target detection. Adversarial methods based on pixel perturbation ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problems of time-consuming, labor-intensive and poor defense effect in the existing way of using confrontational training to defend against target detection against attacks, the present invention provides a preprocessing defense method for target detection against attacks. The time cost requirements are low, and it has robust defense capabilities against most attack methods

Images