| Information hiding is an effective technology to ensure the security and availability of data.It mainly hiding secret information in multimedia files in a specific way for covert communication.The traditional information embedding-based methods usually makes slight changes to some attributes of the cover image,and this process will inevitably leave modification traces.With the development of steganalysis technology,the security of embedding-based methods is also facing more and more challenges.To address this issue,the researchers proposed a generative information hiding method.Instead of modifying the existing cover,this method directly generates a new multimedia cover as the stego-image through a secret information-driven model.So,it can effectively achieve anti-detectability to steganalysis.However,the existing generative information hiding methods cannot guarantee the quality of the stego-image,the extraction rate of secret information and the hiding capacity at the same time.To solve this problem,this thesis proposes two generative information hiding methods based on two different Markov chain models.The details are as follows:(1)A contour generative information hiding method based on Generative Adversarial Networks(GAN)is proposed.This method includes a secret information encoding method(CPS-Encoding)base on Contour Point Selection(CPS),and a Contour Generative Adversarial Network(Contour GAN,Ctr GAN).In the stage of secret information hiding,the method first uses CPS-Encoding and Ctr GAN to encode the secret information into disentangled features of the image(image contour),and then converts the contour image into stego-image for covert communication through a reversible generative model.The model used throughout the process is reversible,so the secret information extraction stage can extract the secret information by taking the steps opposite to the hidden stage.This method effectively utilizes the advantages of disentangled features,so it solves the problem that existing methods cannot guarantee high hidden capacity and high extraction accuracy at the same time.Extensive experiments prove that the proposed method still has high extraction accuracy,anti-detectability and image quality under high hidden payload.(2)A generative information hiding method based on Diffusion Model(DM)is proposed.This method includes a secret information coding method based on diffusion model and Huffman coding(Diffusion Model and Huffman Coding,DMHC),and an image-to-text network based on Transformer(Image-to-Text Based on Transformer,ITBT).In the secret information hiding stage,firstly,the secret information is encoded into text by DMHC,and then the text is converted into the corresponding stego-image by diffusion model.In the stage of extracting secret information,firstly,the secret image is converted into text by ITBT,and then the secret information is extracted by the reverse process of hiding stage.This method effectively solves the problem of Markov error accumulation in the GAN-based contour generation information hiding method.Experimental results show that the proposed method ensures both high generated image quality and high information extraction rate. |
