Adversarial sample generation method and system, storage medium and device

Abstract

The invention discloses an adversarial sample generation method and system, a storage medium and a device. The method comprises the following steps: training a data set to obtain an initial model; attacking the initial model by using an attack algorithm, and calculating to obtain an initial adversarial sample; putting the initial adversarial sample into a first DCGAN network for training to obtaina first generator corresponding to the initial model, and generating an unprotected attack sample by the first generator; performing AND operation on the initial adversarial sample and the data set to obtain a protection model with protection adversarial capability; attacking the protection model by using an attack algorithm, and calculating to obtain a protection confrontation sample; and putting the protection confrontation sample into a second DCGAN network for training to obtain a second generator corresponding to the protection model, and generating a protection attack sample by the second generator. According to whether an original model has protection or not, a new confrontation sample generation method is provided by utilizing the GAN, and malicious samples with new features can be produced in batches.

Description

technical field [0001] The present invention relates to the field of image recognition, in particular to a method, system, storage medium and device for generating an adversarial example. Background technique [0002] In the modern society with the rapid development of information technology, deep learning technology is gradually being recognized and accepted. In many fields, deep learning can complete preset tasks with accuracy close to or even surpassing that of humans. However, while deep learning technology is widely used, the importance of its security is self-evident. Existing neural networks are vulnerable, the researchers found. Especially in the field of image recognition, only a small amount of transformation of the image will cause the classifier to produce wrong classification results. Among other things, the total number of variables required may be so small that it is imperceptible to humans. As attack algorithms become more and more advanced, adversarial s...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art, provide a method, system, storage med

Images