Research On Digital Image Information Hiding Algorithms Based On Adversarial Network

With the rapid development of communication technology,the Internet has become the most important channel for information transmission.As an information transmission carrier,digital images are widely used in all walks of life.To ensure the security of information transmission on the Internet and prevent the leakage of transmitted information,the research of information hiding algorithm based on digital image becomes more and more important.Aiming at the existing problems in the digital image information hiding methods,this thesis has carried out the following three aspects of research:(1)Existing deep learning-based simulate JPEG compression network has some problems such as “not simulating JPEG compression with different quality factors” and “not considering the constraint of loss function”.Therefore,an improved simulate JPEG compression network is proposed in this thesis,which uses loss function to train the network and can simulate JPEG compression with different quality factors.The excellent performance of the proposed analog JPEG compression network is demonstrated by experiments.(2)The following problems exist in the high-capacity steganography algorithm of digital image based on deep learning:(a)With the increase of embedded secret information,the colors of the generated steganography images are distorted;(b)the steganography model does not consider both security against steganalysis and robustness against some attacks during the training phase or only considers security.For problem(a),the Inception module is introduced to optimize the network structure and information is embedded only in the blue channel of color images.For problem(b),steganalysis network and attack network layers(such as noise,filtering and JPEG compression)are added to enhance the security and robustness of the algorithm through adversarial learning.Therefore,a high-capacity robust steganography algorithm based on adversarial networks is proposed.The experimental results on the Image Net dataset demonstrate that the algorithm is superior to the existing algorithm in terms of invisibility,security and robustness.(3)The existing digital image watermarking algorithms based on deep learning have the following problems:(a)the designed network is not an end-to-end model;(b)thewatermarking model does not consider simulating JPEG compression network or the accuracy of the simulated JPEG compression network is not high.For problem(a),the watermarking is embedded by convolution and deconvolution with different strides,and the watermarking is extracted by using a convolutional network.At the same time,the attack network layers are inserted before the watermarking is extracted,eventually forming an end-to-end learning system.For problem(b),the JPEG compression network simulated in(1)and attack network layers such as Gaussian noise and Gaussian smoothing are combined to enhance the robustness of the algorithm through adversarial learning.A digital image watermarking algorithm based on adversarial network is proposed.Experiments demonstrate that the proposed algorithm is superior to the existing watermarking algorithm in invisibility and robustness.