Research On Adversarial Attack On Steganography Based On Dual-discriminator Generative Adversarial Network

In recent years,deep neural networks have developed rapidly.As a transformative technology,it has brought huge economic benefits and social benefits to people,but it has also caused security problems in artificial intelligence.A large number of studies have proved that deep neural networks are very vulnerable to adversarial examples.The perturbations in the adversarial examples are so slight that the human visual system cannot detect them,but they will cause the deep neural network to output erroneous results,causing a great potential threat to safety.With the increasing attention to security issues,we will study the security of steganography in this paper,that is,the adversarial nature of steganography.This article proposes firstly to use adversarial examples to attack the steganographic network: on the basis of detecting the secret information,the secret information is tuned to enable the communication receiver to receive the wrong secret information and complete the interception and reverse investigation of the illegal information transmitted by the illegal elements,and prove the insecurity of the steganographic network to promote the development of steganography.In this paper,a novel method is proposed to autonomously generate adversarial examples,which can attack steganography to change the secret image and make the adversarial image have a high perceived quality.The design of the attack model proposed is inspired by the study of using generative adversarial network to generate adversarial examples.But compared to the ordinary method of generating adversarial examples,the model designed in this paper is more challenging because there is a certain opposition between the quality of the adversarial example(the adversarial steganographic image)and the quality of the attack result(the adversarial secret image),and it is necessary to learn the balance point in the antagonistic relationship while satisfying the requirement that both the adversarial example image and the attack result image are high-perception quality images.Therefore,this article proposes a generative adversarial model with a dual discriminator structure to generate adversarial examples to attack steganography.Themodel is named Steganography-Attack-Dual-Discriminator-GAN,referred to as SA-DDGAN.The model generates perturbations that can successfully attack steganography that the perturbations can enable the steganography to synthesize high-perceived quality adversarial steganographic images and to decode high-perceived quality adversarial secret images.One of the discriminators enables the network to learn to generate perturbations that have the least visual impact on the hidden images,and the other discriminator enables the network to learn to generate perturbations that promote the conversion of real secret images to target images.In order to improve the training effect of the network,this paper uses the LSGAN function as the target function of training GAN,and introduces the structural similarity index function as the standard to measure the image quality in the overall target function.SA-DDGAN is the first adversarial attack proposed for steganography.This article tested the attack effect of the model in two data sets.According to the experimental results,it can be seen that the adversarial examples generated by the model designed in this article can successfully attack steganography.Compared with the traditional algorithm for generating adversarial examples,the examples generated by the method proposed in this paper have better visual effects.