Research On The Theory And Method For Image Information Hiding Based On Deep Learning And Cover Construction

Currently,image information hiding has been widely used for secure transmission in public channel.In the early days,secret information is hidden by modifying the original cover image via the artificially designed encoding function.However,with the rapdid development of computer networks and steganalysis technology,this kind of image information hiding is facing more problems.Firstly,its capacity is limited.Usually,the capacity of this kind of image information hiding is less than 1 bpp,and the security will be greatly reduced if the capacity further increases.Secondly,the capability of resisting steganalysis is insufficient.Since it embeds secret information by modifying the cover image,the steganalysis algorithm may capture the hidden traces.In order to solve the above problems,we focus on three aspects in this paper: end-to-end image information hiding based on deep learning,image information hiding based on mapping rules and generative image information hiding.The main contributions are summaried as follows:1.A new end-to-end image information hiding model based on skip connected dense block and mixed loss is proposed.A new type of hiding/extraction network is first designed by utilizing dense block and skip connection,and then a mixed loss function combining edge differential ratio,mean square error and peak signal-to-noise ratio is defined to measure the difference between the predicted distribution and the real distribution from multiple perspectives.In a covert communication,the sender needs to input a cover image and a secret image to the trained hiding network,and a stego image which has no visual difference from the cover image will be produced.For the receiver,the secret image can be recovered by the extraction network based on the stego image.The whole scheme is a complete end-to-end model.Experimental results and analysis demonstrate the effectiveness of the proposed model,and a secret image with the same size of the cover image can be hidden.Compared with the existing end-to-end image information hiding models based on deep learning,the proposed model can converge very fast.Moreover,the cover image and secret image generated by the proposed model can achieve better visual quality,and the effective capacity is the highest.2.A image information hiding algorithm based on mapping rules with discrete cosine transform and latent dirichlet allocation topic classification is presented.Firstly,latent dirichlet allocation topic model is utilized for classifying the image database.Secondly,the images belong to one topic are selected,and 8×8 block discrete cosine transform transform is performed to these images.Then robust feature sequence is generated through the relation between direct current coefficients in the adjacent blocks.Finally,an inverted index which contains the feature sequence,dc,location coordinates and image path is created.For the purpose of achieving image steganography,the secret information is converted into a binary sequence and partitioned into segments,and the image whose feature sequence equals to the secret information segments is chosen as the cover image according to the index.After that,all cover images are sent to the receiver.In the whole process,no modification is done to the original images.Experimental results and analysis show that the proposed algorithm can completely resist the detection of existing objective steganalysis algorithms.Compared with the existing image information hiding algorithms based on mapping rulesand,the proposed algorithm has better robustness against common image processing and better ability to resist subjective detection.Meanwhile,it is resistant to geometric attacks to some extent.3.A generative image information hiding algorithm based on fractal theory is proposed.Firstly,four fractal image generation methods are analyzed,and the relationship between the generative information hiding and these methods is discussed.Secondly,based on the fractal image generation algorithm,secret information is hidden by controlling pixel rendering during the generation process.The robustness,imperceptibility,and capability of resisting steganalysis are balanced by adjusting the rendering distance.As directly generating stego images,it can resist the detection of most existing objective steganalysis methods.Meanwhile,different capacities can be achieved by adjusting the size of the generated image.Experimental results and analysis show that the proposed scheme can effectively resist steganalysis and has good robustness against various image attacks.Furthermore,it can achieve large capacity.4.A generative image information hiding framework based on Image-to-Image translation is presented.In this framework,the secret image is regarded as an attribute image,and the image whose style is different from secret image,is regarded as the reference image.At the sender,the secret image and reference image are first disentangled into content features and attribute features by image-to-image translation networks.Then,the stego content feature is otained by hiding content feature of the secret image into that of the reference image with the hidden network.Afterwards,the stego image is generated by combining the stego content feature with the attribute feature of the secret image.At the receiver,after getting the stego image,the stego content feature and the attribute of the secret image can be recovered by disentangling the stego image with the image-to-image translation networks,and then the content feature of the secret image can be obtained by inputting the stego content feature into the extraction network.Finally,the secret image can be recovered by inputting the content and attribute feature of the secret image into the image-to-image translation networks.The experimental results show that the proposed framework can generate the stego image with same size as the secret image,and it can recover the secret image from the stego image.The generated image has high visual quality.Meanwhile,since the secret image is hidden by the proposed framework with no modification on any cover image,it can resist the detection of steganalysis methods.