Image Information Hiding Based On Generative Adversarial Networks

Steganography is a technology that hides secret information in a carrier and achieves the purpose of secret communication through the transmission of the stego image.The principle of watermarking technology is similar to steganography,the difference is that its purpose is to use the watermark information embedded in the carrier to protect intellectual property rights or carry out anti-counterfeiting traceability.Traditional steganography and watermarking algorithms depend on the artificial designed complex feature and requires the designer's domain knowledge and accumulated experience.In recent years,researchers have tried to use deep learning,especially Generative Adversarial Nets(GANs),to automatically design steganographic algorithms and robust watermarking algorithms.However,the designed algorithms have weaknesses in terms such as information extraction accuracy rate,embedding capacity,and steganographic security,or watermark robustness and watermarked image quality.Therefore,in view of the above problems,the main research contents of this dissertation include:Firstly,a new end-to-end steganographic model IIH-GAN(Image Information HidingGAN)based on generative adversarial networks is proposed.An encoder and a decoder based on the SE-Res Net block is designed,which can optimize the interdependence between network channels and enhance the automatic selection of global feature,leading to a more accurate and high-quality of information embedding and extraction;we further use a discriminator to be co-trained with the encoder-decoder,and thus maintain the distribution of the carrier image during adversarial training unchanged and enhance the security performance in resisting steganalysis.To solve the problem of decoding real images in real scenes,we add a round layer between the encoder and decoder.To remedy the shortcomings of GAN-based steganography in resisting the powerful state-of-the-art deep learning-based steganalysis algorithms,we add the adversarial examples to the GAN-based steganographic model.Secondly,a new end-to-end robust blind watermarking model IRBW-GAN(Image Robust Blind Watermark-GAN)based on generative adversarial networks is proposed.An encoder and a decoder based on the SE-Res Net block and the differentiable noise layer for resisting noise attacks are designed.The discriminator is used to maintain the invisibility of the watermark image.The joint training based on encoder,decoder,discriminator and differentiable noise layer greatly improves the robustness of the watermark and can resist various types and intensities of noise attacks.In the noise layer,this dissertation considers a variety of noise attack types and high-intensity noise attacks.For non-differentiable JPEG compression noise,a new type of differentiable network layer is proposed for simulation.Finally,the experimental results show that the designed model IIH-GAN has a significant improvement in performance compared with state-of-the-art methods.When adversarial examples are added to our model,the accuracy of our algorithm being detected by the deep learning-based steganalysis is reduced from 97.43% to 48.69%,which means our method greatly improves the steganographic security.Compared with the state-of-the-art deep learning-based watermarking method,whether it is for single noise or combined noise,and when resisting high-intensity noise attacks,our proposed IRBW-GAN model significantly improves the watermarked image quality and watermark extraction accuracy while increasing the watermark embedding capacity.Our simulated JPEG compression network layer is closer to the real JPEG compression,which can achieve better robustness against image compression.