Research On Image Steganography Based On Generative Adversarial Networks

Information hiding is one of the important means to ensure data security during network communication.The sender can use keys and specific algorithms to embed secret information into the carrier,and then the receiver can reuse them to extract the secret information.Images have become one of the most widely used carriers due to their easy accessibility and diversity.Information hiding can not only ensure the security of secret information itself,but also ensure the reliable transmission of secret information.Therefore,it has attracted the attention and research of domestic and foreign scholars.Traditional image-based adaptive steganography algorithms mostly rely on human experience design for the selection of pixel positions to be changed,which requires a lot of time and energy.And researchers considered the visual quality of the steganographic images,so that the steganographic capacity was limited.With the development of deep learning rapidly,the strong representation ability of deep learning network for complex data makes it applied to the field of steganalysis,which makes the discriminative ability of steganalysis models greatly improved.This paper constructs imag-steganography model based on deep learning by using the various deep learning networks and their different advantages,which can further improve its security and visual quality of dense images under the premise of expanding its steganography capacity.The research contents of this paper are as follows:(1)An image steganography model based on residual network is proposed.We firstly take the secret RGB images and the carrier RGB images as the input of the encoder composed of residual blocks and convolutional neural network layers,and then the encoder outputs the stetganographic images with low distortion rate.Then the decoder composed of the convolutional neural network layers reconstructs the secret image from the steganographic images.Experiments show that compared with the previous works,the structure of encoder in steganographic model is improved.Continuous residual blocks are added between the downsampling convolution blocks and the upsampling convolution blocks to extract more complex and high-dimensional image features to reduce the distortion rate of steganographic images containing a lot of secret information.(2)An image steganography model based on generate adversarial network called HIGAN is proposed.Inspired by the game theory of generating antagonism network,the steganographic analysis model is used as a discriminator,and the training parameters and methods of the model are constantly adjusted,so that the security of the image steganographic model with large steganographic capacity can be significantly improved.In addition,a multiple loss function is designed to measure the difference between the steganographic images and the carrier images,which can not only guarantee the visual quality of the steganographic images,but also improve the security of the image steganographic model.Finally,the results of experiments prove that the security of the proposed steganography model under large steganography capacity is significantly improved,and the effectiveness of the proposed model is verified by comparison with prior works.