Adversarial patch generation method

Abstract

The invention discloses an adversarial patch generation method, which comprises the following steps of: collecting a picture with a camouflage target from a data set, and randomly initializing a patch; defining a target real frame to which the patch is pasted and other real frames to which the patch is not pasted; according to targets of different sizes, scaling the patch according to a certain proportion; adding the patch to the camouflage target according to an anchor frame constructed by the target real frame; and inputting the picture pasted with the adversarial patch into a detection network, calculating loss, and iteratively updating the adversarial patch until the camouflage target is not detected by the detection network. Important ground object targets are camouflaged, and the whole target cannot be detected as long as the adversarial patch covers a small part of the directional target; an adversarial patch is placed at the central position of a directional target, and decision features of the directional target in a target detector are hidden, so that the confidence coefficient of a prediction frame in the detector is lower than a threshold value, and a detection result of the detector is misguided.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for generating an anti-patch. Background technique [0002] In recent years, in the target detection task of remote sensing images, the target detector technology based on deep learning has played an important role. However, although the target detector technology based on deep learning has good performance in target detection tasks, However, some studies have found that these target detectors are sensitive to the input with a small perturbation, and are easily disturbed by these perturbations, resulting in erroneous outputs. These samples with specific perturbations are called adversarial samples. [0003] The phenomenon of adversarial samples exists in the target detector. Although it brings a certain threat to the performance of the detector, this phenomenon also provides us with a new way to disguise some important ground objects and escape the d...

AI Technical Summary

Problems solved by technology

However, DPatch does not limit the pixel value of the generated patch, and it is possible that the pixel value exceeds the valid pixel value range of the image.

Images