Information hiding is one of the important ways to ensure data security during network communication. It not only ensures the security of the secret information itself, but also ensures the security of data transmission. As the main technology in the field of information hiding, steganography has received extensive attention and in-depth research by domestic and foreign scholars. The current spatial adaptive steganography methods mostly rely on human experience for the selection of pixels to be changed, which requires a lot of time and effort. In recent years, with the rapid development of deep learning, the strong representation learning ability for complex data makes it used in the field of steganalysis, so that the steganalysis models based on deep learning can quickly extract high-dimensional features and simultaneously optimize the classifier. The rapid improvement of the performance of the steganalysis model weakens of the security of steganography, which makes a great challenge to the development of steganography. In 2014, the generative adversarial networks (GAN) was proposed, which provided a valuable opportunity for the combination of deep learning and information hiding. Until 2016, a steganographic model based on deep learning--SGAN was proposed firstly. Since then, a large number of steganographic models using various deep learning networks have emerged, which lead to a great improvement in steganography in terms of steganographic capacity, anti-detection, and stego-image's quality. First of all, this paper explains the importance of information hiding for data security and summarizes the historical development of image-based steganography briefly; The second one, this paper discourses the fourth types of steganographic models based on deep learning: 1) the steganography based on deep learning of generating cover images, which generates cover images that are more secure and suitable for hiding by using deep learning networks; 2) the steganography based on deep learning of embedding information in cover images, which replaces the steganography algorithms designed by human experience with deep learning networks to embed and extract secret messages automatically; 3) the steganography based on deep learning of synthesizing cover images, which performs secondary modification and synthesis on the original cover images, and the sender can more securely hide secret messages in the changed area; 4) the steganography based on deep learning of mapping relations, which builds a mapping relationship between a secret message and a cover image(or a noise vector used to generate a cover image), and uses deep learning networks and mapping relationships to extract secret messages almost completely. Then, this paper analyzes and summarizes various type of steganographic models in detail from the aspect of steganographic capacity, anti-detection, and stego-image's quality, and discuss the similarities and differences of various steganographic models. Next, this paper explores the different problems of various steganographic models. Furthermore, this paper proposes an improved method based on adversarial samples for the security problems of large-capacity steganographic models based on deep learning, and briefly describes the method. Last but not least, this paper summarizes the advantages and disadvantages of the current steganographic models based on deep learning and looks forward to its future development directions. © 2020, Science Press. All right reserved.
