Image Style Transfer Based On Convolutional Neural Network And Its Application In Coverless Steganography

With the dramatic advancement of graphics processing unit(GPU),deep learning and application have become hot research direction in various fields of information technology.It is possible to search massive image feature matching based on deep learning.Therefore,the research on coverless steganography based on deep learning for massive image feature matching search has been rapidly developed.Coverless steganography does not modify the carrier image.It constructs an index using the characteristics of the image and the secret information to be transmitted.In order to ensure that various secret information can be matched with the image,the design of image database is particularly important.If the image database is not perfect,there are secret information cannot find the corresponding with the secret image.At present,existing algorithms often need a large image database to suit the needs of coverless steganography.As a new method of image synthesis,image style transfer can produce rich artistic images.The naturalness and difference of these image style transfer results are suitable for image database of coverless steganography.In order to improve the diversity of image style transfer results,the VGG-19 architecture was changed.In this paper,multilevel noise coding mechanism and diversity loss are proposed.This makes the result of parametric image style transfer more suitable for coverless steganography.In addition,an adaptive information hiding matrix is proposed.Under the constraint of the adaptive information hiding matrix,the non-parametric image style transfer is carried out.It can avoid the problem that the secret information cannot be matched in the image database.The main work of this paper is as follows:(1)Parametric image style transfer:In this paper,a diversity image style transfer network using multilevel noise encoding(MNE-Style)is proposed.At the same time,a new coverless steganography based on MNE-Style is designed.While improving the VGG-19 architecture,a generator was designed.Therefore,the image style transfer results are more diverse and different.In the generator,a multilevel noise encoding mechanism matching the scale of the subsequent VGG-19 network is adopted,and the diversity loss is increased in the loss network.At the same time,the residual learning is introduced so that the training speed of network is significantly improved.Experiments show that the MNE-Style can generate natural results in a short period of time.The more natural,the better for the safe transmission of secret information.These image style transfer results can be integrated into our coverless steganography scheme.The performance of our steganography scheme is good in steganographic capacity,anti-steganalysis,security,and robustness.(2)Non-parametric image style transfer:In order to avoid secret information being unable to match corresponding secret images,an arbitrary image style transfer network CSST-Net used in coverless steganography is put forward.Coverless steganography and non-parametric image style transfer combined,an adaptive scheme for secret information encoding and adjustment according to the concrete input image is designed,and an adaptive steganography matrix is generated in CSST-Net.Arbitrary image style transfer is performed instructed by the adaptive steganography matrix,and the image style transfer result driven by secret information is directly synthesized.That is,image style transfer results are generated by the process parameters of secret information control to be transmitted.Therefore,there is no problem that secret information cannot match the corresponding secret image.Because image style transfer is not suitable for facial images,an CSST-Net extended network that is suitable for facial images is put forward.The categories of images applied have been expanded.The shortcomings of existing methods have been made up.Experiments show that CSST-Net can not only synthesize any image style transfer results with good visual effect,but also achieve good performance in capacity,anti-steganalysis and security.