A Survey on Digital Image Steganography

: Internet brings us not only the convenience of communication but also some security risks, such as intercepting information and stealing information. Therefore, some important information needs to be hidden during communication. Steganography is the most common information hiding technology. This paper provides a literature review on digital image steganography. The existing steganography algorithms are classified into traditional algorithms and deep learning-based algorithms. Moreover, their advantages and weaknesses are pointed out. Finally, further research directions are discussed.


Introduction
With the rapid development of information processing technology, computer technology and network communication technology, human has entered the digital information age, and the Internet has become a new channel for transmitting information. However, Some security risks have also appeared, such as intercepting information and stealing information. Therefore, ensuring the security of information on the Internet has become one of the problems to be solved urgently [Gupta, Gupta and Singhal (2014)]. In order to ensure the security of information transmitted on the Internet and prevent information leakage, information hiding technology has emerged. Information hiding technology can be divided into two important branches: watermarking and steganography. Digital watermarking hides secret information in digital media (such as pdf documents, word documents, videos, etc.) to prevent piracy and protect copyright of the original. Steganography is mainly used for covert communication, so that the secret information hidden in the carrier data does not attract the attention of external observers. Generally speaking, steganography implements a "peer-to-peer" covert communication. Only the communicating parties know the existence of the secret message, and the receiver can accurately extract the secret information. Steganography is a very important tool in the protection of national information security, especially in the military. Therefore, steganography is a very important research content in the field of information security. Most of image steganography algorithms can be divided into two stages (As shown in Fig.  1): 1) Embedding. In this stage, secret message is embedded into carrier image; 2) Extracting. It realizes the function of recovering secret message from steganographic images. In this paper, some traditional image steganography algorithms and recent algorithms based on deep learning are surveyed. The rest of the paper is organized as follows: Section 2 reviews some traditional algorithms from spatial-domain and transform-domain. Section 3 describes the deep learning-based algorithms. Future research directions are discussed in Section 4. Section 5 summarizes the paper.

Traditional algorithms
In recent years, many researchers have studied and explored steganography techniques and provided many excellent steganography algorithms. From the perspective of embedding domain, traditional steganography algorithms can be divided into spatialdomain algorithms and transform-domain algorithms [Carvajal-Gamez, Gallegos-Funes and Rosales-Silva (2013)].

Spatial-domain algorithms
The spatial-domain algorithms mainly hide secret information by changing the brightness value or chrominance value of carrier image. Least significant bit (LSB) algorithm is a simple and representative spatial-domain algorithm. It first converts secret information into binary data, then replaces the LSB of some pixels of the carrier image [Bender, Gruhl, Morimoto et al. (1996); Lie and Chang (1999)]. The LSB algorithm has low complexity and high embedding capacity, and is easy to implement. However, its robustness against steganalysis is not very strong. Therefore, many improved LSB algorithms have been proposed to solve this problem.  [Tavares and Junior (2016)] reduced the modification of pixel value when hiding information, thus indirectly increased visual quality of the steganographic image. Moreover, some works contribute to improving the security of LSB algorithm by combining some different encryption algorithms [Amirtharajan and Rayappan (2012); Muhammad, Ahmad, Rehman et al. (2017)].  In summary, the spatial-domain steganography algorithm has the advantages of large capacity, good visual quality and easy implementation. Due to these advantages of the spatial-domain steganography algorithm, many steganography algorithms borrow or use spatial-domain steganography or its derived algorithms. Therefore, this kind of algorithm has become the most representative method in digital image steganography. However, spatial-domain algorithms have many advantages, which also have some drawbacks. For example, spatial-domain algorithms are easily detected by corresponding methods. Transform-domain algorithms tend to have better anti-attack capabilities than spatialdomain algorithms, but in general, they have less payload.

Deep learning-based algorithms
In recent years, the development of deep learning has injected new vitality into the field of steganography. Some researchers have been working on improving deep learning network structure with the consideration of the characteristics of steganography. The steganography algorithms based on deep learning are basically realized by adversarial network. Volkhonskiy et al. [Volkhonskiy, Nazarov, Borisenko et al. (2017) Although the above-mentioned adversarial network-based steganography algorithms achieve fine performance in steganography and also in resisting steganalysis, their hiding capacities are very limited. So, some researchers worked on large-capacity steganography algorithms to embed secret images into carrier images with the same size. Rahim et al.
[ Rahim and Nadeem (2017)] proposed an end-to-end framework to embed a secret gray image into a color carrier image with the same size. Their work realizes high-capacity embedding but distorts the steganographic image in color. Then, in the loss function Baluja et al. [Baluja (2017)] considered the correlation between secret image and error image obtaining from steganographic image and carrier image to improve the invisibility of the steganographic image. Zhang et al. [Zhang, Dong and Liu (2018)] proposed ISGAN (Invisible Steganography via Generative Adversarial Networks) by introducing the steganalysis network proposed by Xu et al. [Xu, Wu and Shi (2016)] into their basic model to improve its ability to resist steganalysis. In addition, the SteganoGAN model proposed by Zhang et al. [Zhang, Cuesta-Infante, Xu et al. (2019)] used residual structure to further improve the quality of steganographic image.

Future research directions
The existing algorithms presented in Section 3 have achieved fine performance in steganography, but there still remain some unresolved issues.

How to achieve a balance between security and capacity
For both of traditional algorithms and deep learning-based algorithms, there is such a problem: with the increase of steganography capacity, steganography algorithm is more easily detected by steganalysis. Then, some algorithms reduce the capacity to resist the steganalysis. So, how to achieve a balance between capacity and security? It needs to be explored in future work.

How to improve the quality of steganographic image from the deep learning-based large-capacity steganography algorithm
For the deep learning-based large-capacity steganography algorithm, the quality of both of steganographic images and extracted images need to be improved. For example, in the work of Rahim et al. [Rahim and Nadeem (2017)], there is a color distortion of the steganographic image. Such color distortion is not feasible in the steganography algorithm. Therefore, the quality of steganographic image should be paid more attention.

How to consider robustness
Most of steganography algorithms mainly consider the ability of hiding information and also resisting steganalysis, while they ignore the consideration of robustness. So, the algorithms are not robust. In future work, for the machine learning-based algorithm, the robustness should be considered in both of training phase and testing phase.

Conclusion
This paper starts with introducing the background of image steganography. Then, the state-of-the-art algorithms are summarized from spatial-domain algorithms, transformdomain algorithms to deep learning-based algorithms. What's more, the current existing issues and future research directions are also illustrated in this paper.