Study Of Image Steganography Based On Deep Learning

With the development of the times,the progress of science and technology and the increasing improvement of Internet Technology,the information security has attracted more and more attention.Steganography,mainly researches on how to embed secret information into digital objects without influence its sensory characteristics and value,its principle is embedding the secret information which needed to transmit into the redundant information by the insensitive redundant information of human perception system in common cover,which can be transmitted securely through public channel.The corresponding steganalysis which is to detect the steganography is also attracting more and more attention.Steganalysis is essentially a binary classification problem,which is used to distinguish covers and stegos.Under the development of deep learning in recent years,steganalysis based on the technology has been proposed and made great breakthroughs.Traditional steganography method became more and more difficult.This paper mainly based on image steganography in deep learning,mainly carried out the following research:(1)This paper studies automatic steganography based on generative adversarial networks(GAN).The generator is used to find appropriate hidden locations in the image,the discriminator is used to evaluate whether it is suitable for information hiding,nets adaptive learning the distortion cost.For the existing methods are less secure than traditional steganography,We build a more suitable generator structure and propose a simulator to increase embedding rate based on probability map which simulate high embedding rate of probability map inside the network,and then the high embedding rate stego is obtained to assist the training of the network.The method learns the distortion cost in steganography.Embedding and extracting can be done by traditional adaptive information hiding.The experimental results show that the proposed method is more secure.(2)This paper studies generating covers based on GAN.At the same time,the generator generates the natural image,and the training is guided by the discriminator of steganalysis,which makes the generated image more suitable for information hiding.For the shortcoming of the Least Significant Bit(LSB)embedding in the existing method framework,We propose the adaptive information hiding embedding method which can replace LSB embedding in the framework.The experimental results show that the proposed method is more secure,the covers generated by this method can resist the detection of steganalysis more effectively.