Digital Image Steganography Based On Deep Adversarial Networks

As an important research direction in the field of information hiding,steganography is an important means of information covert communication in the network age.The traditional digital image steganography method manually designs the hiding position and hiding strategy,and the hiding capacity is generally low.In recent years,with the development of artificial intelligence technology,researchers have tried to use deep learning to design automated information hiding methods,but the existing methods still have shortcomings in terms of imperceptibility,hiding capacity,and robustness.In response to this problem,this paper combines Generatative Adversarial Network(GAN),Autoencoder(AE),Attention Mechanism and steganalysis to design a robust and safe image steganography scheme.The main work of this project is as follows:(1)First,the general image steganography process is explained,and the information embedding and extraction process of steganography are modeled respectively,and then the problems existing in the existing digital image steganography methods are analyzed and summarized,such as achieving a balance between the visual quality of the stego image and capacities,and the robustness of reconstructing secret information when the stego image is subjected to different attacks(cropping,scaling,pixel loss,JPEG compression,etc.).(2)Aiming at the problem that it is difficult to balance the embedding capacity and imperceptibility of the existing digital image steganography methods,a digital image steganography model based on Autoencoder and Generatative Adversarial Network is proposed.The actual secret image is converted to the latent space through the pre-trained autoencoder.The embedded information is the sample feature information of the real secret image,which is helpful to realize the high-embedding steganography model.Use an optimized Generative Adversarial Networks to embed the secret image into the cover image to obtain a higher perceptual quality of the secret image.At the same time,an extraction model including visual system,this paper designs a mixed loss function combining L1,L2,MSE and SSIM,and uses the perceptual loss based on VGG-19 to improve the authenticity of the stego image and the extracted image.The experimental results show that this model performs well in terms of imperceptibility and embedding capacity on public datasets such as LFW,PASCAL-VOC 2012 and COCO,achieving an embedding efficiency of 44.17db and an extraction efficiency of 39.11db,ensuring that the stego image and the extracted secret image is authentic.In addition,compared with the current image steganography algorithm based on carrier synthesis,this method does not need to create a carrier image library,and the obtained stego image is closer to the real natural image,and a higher steganography capacity is realized.(3)Aiming at the problem of weak robustness of existing steganography methods that use deep learning to concatenate secret data with cover images,a highly robust digital image steganography method that combines attention mechanism and anti-learning ideas is proposed.Through the attention mechanism to learn the probability distribution of each pixel data of the cover image,select the amount of embedded bits,and embed low-intensity and high-intensity messages into the visually significant and visually insignificant areas of the image,which helps to enhance the robustness of the model.To resist various image distortions.At the same time,the discriminator network and noise layer are added,and noises such as cropping,resizing,pixel loss,and JPEG compression are introduced into the encoder-decoder architecture,so that the model can still reconstruct the secret information robustly under various attacks,and the model is easier to train.The experimental results show that under the premise of ensuring the quality of the stego image,the algorithm is more robust,and the extracted messages can still be recovered with high accuracy under various attacks such as cropping,resizing and pixel loss,especially for resizing and pixel loss attacks,the average bit error rate can reach 0.008.