Research On Large-capacity Reversible Image Steganography Based On Deep Neural Network

Reversible image steganography is a research hotspot in the field of information hiding.Unlike robust watermarking,reversible image steganography emphasizes the extraction of secret data and the high quality of host image restoration.The current reversible image steganography technology embeds secret information in a carrier image according to a specific rule by slightly modifying the carrier image,and allows a legitimate user to recover the original carrier image without distortion after extracting the secret information.However,the traditional reversible image steganography method has a low hidden capacity and Image distortion caused by large embedded capacity.In view of the recent breakthrough in combining deep neural networks with steganalysis,there have been relatively few attempts to apply neural networks to reversible image steganography.Therefore,this paper uses the depth model to perform reversible image steganography,implicitly simulates the distribution of natural images using a deep convolutional neural network,and embeds more information(full-size images)into the carrier image.The deep neural network determines the hidden location of the secret information on the carrier image and how to effectively encode it,and the hidden messages are scattered among the bits in the carrier image.Secondly,the codebook database of the image steganography is trained using the Boundary Equilibrium Generative Adversarial Networks(BEGAN)in the generated model,and then the codebook database is used to perform reversible image restoration on the secret image to be transmitted.The main work and achievements are summarized as follows:(1)Aiming at the low image hiding capacity and obvious visual cues in the traditional image steganography method,a reversible image steganography method based on U-Net structure is proposed.First,in the form of paired training,the trained deep convolutional neural network includes a hidden network and an extraction network;then,the sender uses the hidden network to hide the secret image into another full-size image without any modification and send it to the receiving end;finally,the receiving end displays the secret image and the carrier image using the extraction network.The experimental results show that the proposed method compresses and distributes the hidden secret image information on all available bits on the image of the carrier.This image hiding method not only solves the obvious visual cues,but also improves the capacity of information transmission.(2)Under the premise of ensuring large-capacity steganography,in order to get rid of steganalysis and realize high-quality reversible image restoration,a reversible image steganography method based on BEGAN is proposed.First,the communication parties use the BEGAN model to train the model parameters according to the pre-agreed rules,and save the model weight parameters of each training sample as a codebook to form a codebook database;then,the sender uses the BEGAN model to train the secret image to be transmitted,generates a steganographic image,and transmits it to the receiving end;finally,after the receiving end receives the visually normal but does not contain any secret information,it only needs to find the corresponding codebook in the previously saved codebook database,the codebook and the transmitted steganographic image generate a secret image to achieve reversible recovery of high quality secret images.The experimental results show that the steganographic image generated by the proposed method can completely avoid the steganalysis and realize the large-capacity reversible image steganography.