Research On Information Hiding Method For Anti-CNN Stegananalysis

With the development of the Internet,network security has become the focus of current attention,ensuring that the covert transmission of information on the channel becomes more important.Compared to traditional image encryption techniques,digital image steganography can be used to embed secret information into an image in an imperceptible form.However,steganography is not absolutely safe.Using steganalysis can detect whether there is confidential information in the image,which eventually leads to steganography failure.Through the acceleration of the GPU,the deep neural network can be quickly trained.Therefore,the CNN-based steganalysis method can accurately detect whether a large number of images are dense in a short time.There is a huge challenge to steganography.In image classification,the antagonistic sample technique can be subtly perturbed on the image to make it misclassified.Therefore,the transplantation of the antagonistic sample technology to the anti-steganalysis,that is,minor changes in the steganographic image,so that CNN recognizes it as unsteeled,can ensure the security of the steganographic image.In the work of this paper,the development status of digital image steganography and digital image steganography analysis is first studied.Next,the concept of digital image steganography is described in detail,and a framework based on CNN steganalysis is given.It can be seen from the framework that if CNN is used for steganalysis,the image needs to be pre-processed to show the feature map of the steganographic image.Then use CNN for model training.In this paper,an anti-steganalysis method based on local embedding is proposed.The enhanced features of anti-steganalysis are combined with the unmodified perceptual features to increase the recognition error rate of the CNN steganalysis model.The method first selects some pixels of the image,uses the method of generating the antagonistic sample to modify the gradient of the image,and increases the embedding rate to ensure the amount of embedded secret information.The remaining portions of the pixel are not modified,retaining their unobscured perceptual characteristics.This paper proposes a security test platform for anti-CNN steganalysis.The platform includes two sub-modules: Steganography and Anti-Steganography.In order to better port it to the platform,the front-end separation method is used on the overall structure of the platform,and the current popular Vue.js+Django is used for development.In the steganalysis module,the three representative CNN steganalysis models are modularized to realize deep learning engineering.In the anti-steganalysis module,the anti-CNN steganography proposed in this paper is combined with the general framework to integrate it into the test platform.Finally,functional testing and simulation testing were performed on the entire platform.In the simulation test,three kinds of CNN steganalysis methods were first simulated,and their performance achieved the expected results.Then the two local embedding strategies proposed in this paper and the general anti-steganization analysis framework are used to confront the three CNN steganalysis models in the test platform,and the experimental results are compared.The results show that choosing the appropriate local embedding strategy can improve the performance of the original method,and thus protect the security of the steganographic image.