Method for generating adversarial image

Abstract

The invention provides a method for generating an adversarial image. The method comprises: on the basis of a gradient algorithm, a loss value of a first deep neural network model is obtained accordingto an image obtained by last-turn iteration, and a momentum of current-turn iteration is generated based on the loss value; and on the basis of the momentum of current-turn iteration, an image obtained by current-turn iteration is generated based on the image obtained by the last-turn iteration until iteration reaches a preset iteration turn number, and the image obtained by the last-turn iteration is used as an adversarial image. According to the method provided by the invention, iteration of an original image is carried out by using the momentum and the adversarial image being capable of attach the deep neural network model is obtained, so that coupling between a white-box attack success rate and a migration performance is reduced effectively. The high attack success rates of the whitebox model and the black box model are realized. The method can be used for adversarial training, thereby improving the correction rate of image classification by using the deep neural network model and also can be used for attacking the deep neural network model.

Description

technical field [0001] The present invention relates to the technical field of machine learning, and more specifically, to a method for generating confrontational images. Background technique [0002] As a type of machine learning method, deep neural network has gained widespread attention in recent years due to its remarkable results in many fields such as speech recognition, image classification, and object detection. However, deep neural network models that can achieve high accuracy on many tasks are vulnerable to attacks in adversarial environments. In the adversarial environment, the deep neural network will be input with some maliciously constructed adversarial samples based on normal samples, such as image or voice information. These adversarial examples are easily misclassified by deep learning models, but it is difficult for human observers to detect the difference between adversarial examples and normal examples. Since adversarial examples can measure the robustn...

AI Technical Summary

Problems solved by technology

Existing methods for generating adversarial images for deep neural network models, it is difficult for the generated images to have a high attack success rate against both white-box and black-box models, so that the deep neural network cannot achieve high performance under adversarial conditions. accuracy rate

Images