Attack image generation method, electronic equipment and readable storage medium

Abstract

The invention discloses an attack image generation method, electronic equipment and a readable storage medium. The attack image generation method comprises the following steps: determining initial noise corresponding to an original image; determining a visual sensitivity weight map corresponding to the original image through a visual attention model; determining a weighted noise map corresponding to the original image according to the visual sensitivity weight map and the initial noise; determining a target noise map according to the weighted noise map; and generating a first attack image according to the target noise map and the original image. The visual sensitivity weight map corresponding to the original image is determined through the visual attention model, the weighted noise map is determined according to the visual sensitivity weight map, so that the target noise map is determined, and finally the first attack image is generated according to the target noise map and the original image. The problem that attack images have obvious visual differences is solved, and the attack concealment is poor.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an attack image generation method, electronic equipment and a readable storage medium. Background technique [0002] The deep convolutional neural network has excellent performance in computer vision tasks, but it also has a defect, that is, vulnerability. When the input image is not too much interference noise, that is, it is not visually obvious, the model will do If a wrong judgment is made, the attack is successful. The current attack method is to maximize the success rate of the attack under the constraint that the power or maximum value of the attack noise is guaranteed. However, the attack pictures generated under the condition of limited noise power or maximum value have obvious visual differences, so that the human eye can easily perceive the attack, and the attack concealment is poor. [0003] The above content is only used to assist in understanding the tech...

AI Technical Summary

Problems solved by technology

[0004] The main purpose of the present invention is to provide an attack image generation method, which aims to determine the visual sensitivity weight map corresponding to the original image through the visual attention model, reduce the visual difference of the attack image, and solve the obvious visual difference of the attack image. The problem of poor attack concealment

Images