Analysis And Research Of Image Steganography For Neural Style Transfer

Image steganography is one of the key technologies in the field of information security.which can hide secret information in public images and realize the protection of key information in the fields of copyright,commercial finance and national security.In recent years,with the development of deep learning,steganographic algorithms have made great breakthroughs and improvements in embedding capacity and information extraction and neural style transfer technology is also very popular.However,the widespread popularity of style transfer technology also accelerates the unauthorized tampering,synthesis and dissemination of image resources,which brings great challenges to the application of steganography technology in image copyright protection.The reason is that steganography is less robust in style transfer scenarios,and it is difficult to effectively extract secret information from stylized images.Aiming at the above problems,this thesis proposes a robust image steganography method for style transfer by analyzing the ability of the style transfer model to preserve the semantics of the original image.Without affecting the quality of style transfer,the robustness and anti-steganalysis ability of steganography algorithm in the process of style transfer are improved,and the security and concealment of steganography technology are enhanced.The main contents of the thesis are as follows:（1）A style transfer method with multi-semantic preservation enhancement is proposed to analyze the preservation mechanism of image semantic information in style transfer.This method maps the content image to the YCb Cr color space,and adds multiple semantic loss functions in the training to enhance the semantic preservation ability of the model.Through the analysis of multi-semantic preservation in image style transfer,it provides the basis for the study of style transfer-oriented image steganography methods.（2）Combined with the characteristics of image steganography,a single neural style transfer-oriented image steganography method is proposed to extract the embedded image copyright information from stylized images.The encoder network and the corresponding decoder network are designed,and the encoder network embeds the copyright information of the image into the original image without affecting the visual effect and quality of the image.After the image is style-transferred,the decoder network can extract secret information from the stylized image.Experiments verify the effectiveness of the encoder and decoder networks,and the mean SSIM and BER of the original embedded secret image and decoded graph reach 0.82 and 0.22,respectively.（3）A robust steganography method for arbitrary style transfer is proposed to improve the versatility and anti-steganalysis ability of steganography algorithm in style transfer.By improving the structure of the original encoder network,a residual network block is introduced to replace part of the convolutional layer network,and then the single style transfer is replaced by an arbitrary style transfer model during training for model iteration optimization.The thesis also conducts steganalysis experiments on the two proposed steganography models,the experiments show that the steganalysis detection accuracy rates of the two methods in this thesis are 43.75% and 45.83%,respectively,which are lower than other steganography methods under the same conditions,and the anti-steganalysis ability of the steganographic models have be improved.To sum up,this thesis proposes two robust steganography methods for style transfer,both of which can better extract the embedded secret information from stylized images.At the same time,Due to the introduction of style transfer,the risk of hidden information being detected by the outside world is greatly reduced,the robustness of the model is enhanced,and the anti-steganalysis ability is also better than other steganography methods.Therefore,the method in this thesis has strong generalization and practicability and wide application prospects.