Research On Large Capacity Steganography For Color Image Based On Generative Adversarial Network

With the popularization of the mobile Internet,a large amount of multimedia information is transmitted on the network,how to ensure the information security during transmission has attracted widespread attention from all walks of life.As an important branch in the field of information security,steganographic technology completes information transmission by concealing communication behaviors,relying on its "non-inductive transmission"characteristics,and has become a research hotspot for many scholars.In recent years,with the rapid development of computer computing power and neural network algorithms,steganographic algorithms based on deep learning have emerged,and have achieved good results in terms of hiding speed and algorithm simplicity.However,these algorithms still have problems such as poor visual quality,low hiding capacity,and poor resistance to steganalysis.In order to solve these problems,this dissertation studies the image steganography algorithm based on Generative Adversarial Networks(GAN),and combines the relevant theories such as attention mechanism,depthwise separable convolution and steganalysis technology to effectively improve image steganography.Write quality and capacity.The main work content of this dissertation is as follows:(1)Firstly,the importance of information hiding in the Internet era is expounded,the development of image steganography algorithms is reviewed,and the research progress of steganography with cover image and coverless steganography is deeply analyzed,and the challenges and shortcomings of existing algorithms are pointed out.(2)Aiming at the problem that the current steganography algorithm with cover image is difficult to balance the image quality and hidden capacity,a high-quality full-size color image steganography algorithm based on generative adversarial network is proposed.The encoder using the U-Net model combined with the attention mechanism effectively guarantees the visual quality of the stego image.In view of the widespread use of mobile devices,the decoder uses depthwise separable convolutions to reduce the number of network parameters.In terms of security,using DFSE-Net as a discriminator improves the anti-steganalysis ability of stego images.In addition,considering that JPEG compression often occurs in network transmission,a simulated JPEG compression layer is added during training to improve the robustness of the model.The experimental results show that the algorithm has achieved good visual quality and hidden capacity on the three data sets of COCO,ImageNet and CelebA.When the color image is hidden to the same size color image,the average PSNR reaches 36.63dB(the cover image and stego image)and 38.33dB(secret image and reconstructed secret image),and can effectively resist the detection of JPEG compression and steganalysis models.(3)Aiming at the defects of small hiding capacity and poor image quality in current coverless steganography algorithm,a coverless steganography algorithm based on cover synthesis is proposed.The algorithm directly converts the secret image into the stego image,which essentially avoids the attack of the current steganalysis algorithm and achieves good visual quality.The encoder designed a convolution module combining Xception and Res2Net for feature extraction,and used the VGG-19 network to extract image content information and fuse it with backbone network features to retain as many secret image features as possible.The discriminator uses the Markov discriminator PatchGAN,which divides the image into multiple regions and then distinguishes them sequentially,effectively improving the resolution of the image.The experimental results show that the coverless steganography algorithm can completely hide the 256*256 color image.On the two data sets of COCO and ImageNet,the average FID can reach 80.04,and the PSNR can reach 28.03dB(secret image and reconstructed secret image).And the algorithm can effectively resist the detection of the classic steganalysis tool StegExpose.