Research On Information Hiding And Extraction Algorithms Based On GAN

In recent years,with the development of network technology,the Internet is more and more deeply into people's daily life.And the issue of network information security has also received more and more attention.Simultaneously,deep learning technologies represented by convolutional neural networks and generative adversarial networks have developed rapidly in recent years,and have achieved great success in many fields such as computer vision and speech recognition.Information hiding is an important method to solve the problem of network information security.It protects secret information while hiding the transmission process.It has some special application fields.Information hiding technology and steganalysis technology are mutually opposing and promoting each other.A good steganography analyzer can help to propose a better information hiding technology,which is highly consistent with the idea of generative adversarial network discriminators to guide generator training.Based on the characteristics of generative adversarial network,this paper studies the information hiding technology based on generative adversarial network.The main research work and innovations of this paper are as follows: 1.An information hiding algorithm “IHGAN” based on generative adversarial network(GAN)was proposed.This paper uses the optimized training method of WGAN-div to overcome the shortcomings of using the original GAN loss function to be difficult to train.At the same time,Dense Net is used to form a network to reduce network parameters and speed up network training.The experimental results show that this method can adaptively hide data in the cover image,compared with the existing embedding capacity of the hidden algorithm of 0.4bpp,IHGAN can increase the hidden capacity to 3bpp while maintaining the distortion of the stego image.When the embedded capacity is 1bpp,it can still guarantee the effective resistance to the steganalysis analyzer.2.An information hiding algorithm “Secure-IHGAN” based on encryption mechanism is proposed.At present,deep learning-based information hiding algorithms including IHGAN are composed of an autoencoder,which is based on the premise of secure transmission of stego images,without considering the security issues in the transmission process.Aiming at this problem,this paper adds an encryption key to IHGAN's autoencoder and adds a discriminator to steal secret messages.By using confrontation training,IHGAN learns to encrypt secret messages.The results show that IHGAN learned how to encrypt information after training.Only the decoder with the key can extract the secret information completely and correctly.At the same time,even if the embedded capacity of Secure-IHGAN is as high as 3bpp,the stego images by Secure-IHGAN generates are not visually distorted.Experiments prove that the Secure-IHGAN model guarantees the secret transmission of messages while maintaining the good information hiding effect of IHGAN.3.A digital watermarking algorithm “Robust-IHGAN” based on GAN is proposed.Digital watermarking technology is an important branch of information hiding technology,which mainly emphasizes the robustness and steganographic capacity of the algorithm.Inspired by the idea of a denoising autoencoder,this paper adds a noise layer to IHGAN to simulate several image distortions that may be encountered in the transmission process of the stego image,so as to improve the robustness of the model and increase the accuracy of the information extracted by the decoder.Experiments show that after training,the RobustIHGAN model with noise layer is more than 90% accurate when faced with pixel loss;the accuracy is up to 98% when decoding stego images with Gaussian noise;even after complex JPEG compression stego image,the Robust-IHGAN model still has 80% decoding accuracy.Experiments show that Robust-IHGAN can effectively resist the designed noise in the noise layer and has strong robustness.