Research On Robust Image Steganography Method Based On Generative Adversarial Network

With the rapid development of the Internet and artificial intelligence,image steganography based on deep learning has become one of the important directions in the future development of information hiding.Deep learning network can not only improve the security and steganography capacity of the steganography model but also improve the detection accuracy of the steganalysis model.However,with the rapid improvement of the detection performance of the steganalysis model,the existing steganography is weak in resisting the detection of the steganalysis model,which brings great challenges to the research of steganography.In this thesis,aiming at the problems existing in the steganography model of encoding-decoding network based on Generative Adversarial Network(GAN),such as the residual images of the cover image and the steganography image that will expose the secret image,the poor visual effect of the extracted secret image,the weak detection ability of the steganalysis model,the weak robustness on the condition that the image is attacked and the unguaranteed security of the secret image,the key technologies such as GAN,encoding-decoding network,attention mechanism and chaotic encryption system are used to study the image steganography method of the encoding-decoding network in which gray images are hidden in color images.The main research work is as follows:1.In order to solve the degradation of stego image quality and secret information outline in the stego image in GAN encoding-decoding networks,and improve image security and visual effects,a steganography method based on GAN and the two-dimensional chaotic system was proposed.The secret image of this method is encrypted by the chaotic encryption algorithm and hidden in the Y channel of the cover image.By using the GAN encodingdecoding network,the stego image with better visual effect is obtained by minimizing the probability distribution of the stego image and the cover image.The experimental results show this method can generate high-quality steganography images,which can not only effectively prevent the leak of secret image information,but also avoid the color distortion of steganography images.At the same time,the security of the secret image can be ensured after the steganography algorithm is disclosed.2.In order to solve the problems of poor quality of the steganography image,weak detection ability of anti-steganalysis,incomplete extraction of secret images and privacy leak in the steganography model based on GAN,an image steganography method for encodingdecoding network based on attention GAN and the three-dimensional chaotic system was proposed.In this method,the secret images are encrypted by the three-dimensional Chen chaotic system.The ECA-Net module is introduced into the GAN encoding-decoding network to increase the attention of the channel.In the encoding network,Wasserstein distance is used to reduce the sensitivity of the network architecture and hyperparameter configuration in the training process.Experimental results show that this method can not only generate high-quality stego images that can resist the steganalysis model,but also extract relatively complete secret images,and has good robustness to common image attacks.3.Aiming at the problems of poor quality and insufficient anti-noise ability of stego images in GAN encoding-decoding networks,a steganography method based on attention GAN with strong robustness was proposed.The SE-Res Net attention module is introduced,and the encoding-decoding network is built on this basis,focusing on extracting important information for the image to ensure the better visual quality of the image.In addition,a simulated noise layer is added between the encoding and decoding networks,and the steganography image obtained by using the model trained by the encoding network-noise layer-decoding network can resist common noise attacks.Experimental results show that this method has less loss of steganography image faced with different noise attacks,which further improves the accuracy and robustness of secret information extraction.