Research Of Image Steganography Based On Generative Adversarial Networks

People are paying more and more attention to information security of the communication,as the Internet is used widespreadly and is developed rapidly.The steganography can embed the secret information into a cover,then finish the secret communication through transmitting the cover on public channels.It can not only hide the secret content,but also the secret communication.Conventional steganography algorithms are mostly heuristic methods,plenty of expert knowledge are needed when designing these algorithms.Furthermore,these conventional algorithms cannot be adaptively adjusted due to the existence of different attack algorithms.So there are lots of challenges when designing new steganography algorithms.In this paper,digital images are taken as the research object,and the image steganography based on convolutional neural networks and generative adversarial networks is studied.The main results of the paper are as follows:(1)The image steganography model based on convolutional neural networks is proposed.We follow others'work and design new frameworks.The encoder can embed a gray secret image into a same-size color cover image,then generates an image called stego.The decoder can reveal out the secret image from the stego.In addition,a mixed loss function for the image steganography is put forward by analyzing different metrics.The experiment and results show that the loss function can not only improve the training speed but also strengthen the quality of generated images.(2)An new image steganography model based on Generative adversarial networks(GAN)is proposed.Firstly,we analyze the difference between the stego images and the natural cover images from the respective of probability.Then,we introduce the generative adversarial network to the task of steganography to decrease the difference.After that,the steganography model will modify its parameters to deceive steganalysis algorithms throughout the whole training process.The experiment shows that the adversarial training of the GAN does increase the security.(3)The Generalization and the robustness are improved by fine tuning the model and introducing a noise layer into the model:We train the model with different datasets which contain samples with different sizes,and train the model with different datasets from different sources.A noise layer is introduced into the model to simulate the distortion produced by different attacks.The experiments show that the generalization is improved and the model can resist the attack of cropping.