Research On Image Steganography Based On Generative Adversarial Networks

In information security,information steganography is an important information protection method.Traditional steganography,whether adaptive or non-adaptive algorithms,requires human experience to be involved in designing distortion algorithms,which more or less produce traces of modification on the image.With the progress of deep learning,the processing of complex data by deep learning networks and their ability to extract image feature information are widely used in the field of steganography,in which the basic idea of mutual opposition between generators and discriminators in generative adversarial networks has common features with the opposition between steganography and steganalysis,so this thesis investigates the image based on generative adversarial networks based on the relative characteristics of steganography methods.The main work in this thesis is as follows:(1)Embedded image steganography scheme based on improved SteGAN.Based on the SteGAN model,this thesis proposes an improved SteGAN model,designs a generator and discriminator based on Dense Net,and transmits the features extracted from each layer of the network to the next layer by using the dense connection blocks unique to Dense Net.,retain the carrier image features to the maximum extent,strengthen feature reuse,and combine the attention mechanism to further enhance the ability to extract deep-level features of carrier images,and generate high-quality dense images with embedded information.The experimental results on the 10 data set show that compared with the original model’s maximum steganographic capacity of 0.4bpp,the improved model’s steganographic capacity can reach 0.6bpp,and the imperceptibility of the secret image is better than the original model.Further use of steganalysis the tool Step Expose detects the secret image,and the results show that the secret image of 0.6bpp can effectively resist the detection of steganalysis tools.(2)A secure steganography scheme based on generative adversarial networks.In the process of image steganography,considering the possible impact when the secret image is attacked,this thesis adds a noise layer between the model embedding information and the decoding information from this perspective to train the secret image to make it resistant to The ability of noise attack,when designing the noise layer,combined with common noise attack types,seven different noises and mixed noises that combine these noises are designed,each noise is trained separately to make it resistant to a specific noise,and finally mix the noise training to make it have anti-noise ability to all designed noise attacks,and in order to stabilize the training of the discriminator,the original batch normalization is replaced with spectral normalization to ease the discriminator convergence speed,It enables continuous training between networks and continuously enhances the confrontation between steganography and steganalysis.Finally,the model is trained on the complex and diverse COCO dataset.The experimental results show that when the steganographic capacity of the secret image is 0.6bpp,the secret image obtained by training with the noise layer performs well in imperceptibility.Except for JPEG compression,the rest the decoding accuracy of the encrypted images trained by noise is above 90%,and can effectively resist the detection of steganalysis tools.