2022 International Conference on Education, Network and Information Technology (ICENIT)

Generative Image Steganography Scheme Based on

Deep Learning
Jingyi Qiu*
2022 International Conference on Education, Network and Information Technology (ICENIT) | 978-1-6654-6307-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICENIT57306.2022.00049

Tongda College of Nanjing University of Post and Telecommunications, Yangzhou 225002, Jiangsu, China
[email protected]
*corresponding author

Abstract—The popularity of the Internet and the negative impact on society. In recent years, terrorists have
development of multimedia technology have made it easier to posed a major hidden danger to national security. Illegal use
store, modify, and share digital multimedia information (text, of image steganography for communication, disrupting
audio, video, etc). Because multimedia data contains some social order, and steganography analysis technology can not
personal information, if this information is intercepted and only combat illegal terrorists, but also expose terrorist
stolen by an illegal third party, resulting in the public information and prevent the leakage of important secrets. At
transmission of the information on the Internet, then the the same time, steganalysis detects image defects and
personal contact between the communicating parties will no performs related evaluations on images, which facilitates the
longer be private. Therefore, how to ensure communication
continuous updating of technology, and is also the main
privacy and information security has always been an
driving force for the continuous development of image
important issue that is difficult to get rid of in the development
of the Internet. This paper studies the generative image
steganography [9]. Some researchers suggest that multiple
steganography scheme based on deep learning, understands considerations must be taken when selecting a carrier object
the related theories of image steganography and deep learning for secure and confidential communication. Regarding
based on the literature, and then designs the generative image carrier formats, there are three main types of carriers: video,
steganography scheme based on deep learning, and the image, and audio. Digitized vector objects will generate a lot
designed scheme is tested, and the test results show that the of unnecessary space and autocorrelation, so these redundant
steganography scheme designed in this paper has better spaces can be used to transmit confidential information. For
performance in accuracy. Compared with the traditional example, the pixels of an image can be represented by 8-bit
scheme, the difference between the two is up to 4%. bins. Different bits have different weights. The content of
the image is mainly determined by 5 binary bits. The least
Keywords—Deep learning, image steganography, CNN significant bit is equivalent to random noise, so the least
algorithm, scheme optimization significant bit can be used to transmit confidential
information. This method is imperceptible to human vision,
I. INTRODUCTIONS so as to achieve the purpose of steganography. For special
With the rapid development of computer technology, the vectors that easily evoke human vision, it is necessary to
Internet has made cultural exchanges and cooperation design suitable special algorithms to achieve safe
between various countries more and more frequent, which steganography [10]. In summary, there are many researches
also illustrates that the Internet has important uses in on image steganography, but there are still problems. For
government and national defense construction [1-2]. At the example, most image sealing algorithms that use deep
same time, network security has become a common concern learning are built into the global integration, so that images
of the military and the government. Topics such as military can be steganographic better, but they are less robust, and
information transmission, copyright protection, electronic denser images are more susceptible to RTS attacks, noise,
cash fingerprints, and system security have emerged [3-4]. filtering, etc. Impact, and important information may be lost
However, the prevention of data intrusion and malicious data during the export process.
leakage for image data files is not taken seriously, because the
security of media data is easier to be overlooked than network This paper studies the generative image steganography
security [5-6]. Multimedia files have the characteristics of scheme based on deep learning, analyzes the image
strong redundancy, large data volume, rich content, powerful steganography framework and the application of deep
functions, and convenient transmission. At present, some learning based on the literature, and then designs the
staff members have proposed to protect multimedia data files generative image steganography scheme based on deep
and have conducted related research on its protection learning. Finally the designed scheme is verified, and
technology [7-8]. relevant conclusions are drawn through the test results.

Regarding the study of image steganography, some II. RESEARCH ON GENERATIVE IMAGE STEGANOGRAPHY
researchers have proposed that image steganography is to
A. The Basic Framework of Generative Image
determine whether it contains sensitive information through
Steganography
observation or related professional knowledge when it is not
clear whether the image is embedded. In this case, the Figure 1 shows the basic framework for generating image
process of analyzing the size and location of the confidential steganography. Suppose A and B pass confidential
information and finally extracting the confidential information through a seal. The sending process is as follows:
information. Image steganography technology is used to
detect illegally embedded sensitive information, and to
prevent illegal personnel from using stamped seals to have a

978-1-6654-6307-2/22/$31.00 ©2022 IEEE 191

DOI 10.1109/ICENIT57306.2022.00049
Authorized licensed use limited to: ULAKBIM UASL - Bahcesehir University. Downloaded on July 01,2023 at 19:10:20 UTC from IEEE Xplore. Restrictions apply.
 (1) A prepares an image c without embedded information, (3) Use the K key to reset the scrambled image and
the confidential information I to be sent, and the key K create a SI.
previously agreed with B.
(4) Play images on public channels.
(2) A uses the K key to shuffle the original image c, and
merges the secret information I into C. (5) B uses the matched key K to shuffle the received
encrypted image s to extract the secret information I.

A
B
A A

Information I Information I

Image source c translater Steganographic image s 用户

Steganographic image s decoder

Common channel

Fig. 1. Steganography model

B. Application of Deep Learning extraction. Some CNN models also use large distorted cores
for seal detection. A large rotating core can add a weak
The deep learning theories that have emerged in recent secret information to the image and allow the secret image
years have provided many convenient tools for image (SI) to be detected more accurately. However, most sealing
steganography, and at the same time added more ideas for algorithms embed sensitive information adapted to the
image steganography [11]. The purpose of steganography is content, so a very large distorted core size is inappropriate.
to prevent hidden information from being discovered, but the
integration process requires some changes to some features of (2) Activation mode: The activation mode is very
the carrier image (such as statistical features) to achieve the important for CNN-based image sealing. Currently
purpose of hiding information. The current development is commonly used activation functions include Gaussian
the combination of information retrieval and extraction. The activation function, rectified linear unit (ReLU) activation
more mature technology is the combination of deep learning function and TanH function. ReLU is usually used as an
technology and steganography. It is a popular trend to try to activation function to improve the statistical characteristics
use machine learning technology to offset the statistical of the noise part of the input image.
changes brought about by the integration process.
(3) Batch processing: Xu first uses the BN and TanH
C. Deep Learning Algorithm activation functions in the image sealing network. The role
Sealing with an image as a carrier can be regarded as an of BN is to classify data into unsaturated regions so that the
image classification task [12]. However, the image network can learn more effectively.
classification here is different from the computer vision (4) Merge: For the CNN model, the main function of
image classification, mainly because the two themes have Merge is to reduce the size of the feature map while
different ideas. Tasks such as image classification and object obtaining accurate data features. Most CNN models use
detection focus on key image information, while concentration to extract dense features, and Merge is a
steganography is the opposite. Images contain unnecessary process of losing information. Therefore, the selection of the
information that is not perceivable by humans, and image concentration layer should be considered in combination
steganography algorithms use these unnecessary information with the actual situation.
to hide sensitive information. Image steganography mainly
receives various weak information such as noise residue in III. GENERATIVE IMAGE STEGANOGRAPHY BASED ON
the image, and directly uses these areas, so the existing DEEP LEARNING
classification model is not feasible. Therefore, this article
needs to analyze the CNN model for improvement. A. Network Model Structure
As shown in Figure 2, the proposed image steganography
(1) Combination layer: Most CNN-based steganography framework is different from many common steganography
detection methods use a small distortion core to perform methods, which encode secret information into less important
distortion operations on the input image. The advantage of parts of the carrier image and hide information without the
this method is that it can capture various local correlations carrier. The method in this paper has all the available bits in
between image pixels and promote efficient feature vector image compression and the distribution of SIs.

192

Authorized licensed use limited to: ULAKBIM UASL - Bahcesehir University. Downloaded on July 01,2023 at 19:10:20 UTC from IEEE Xplore. Restrictions apply.
 Cover
imare (c)

Netmork 1 :Hiding Containe Netmorh 2:Reva 1 Revea1ed

Netmork (In∞der) imge (C) Netmrk (Decoder) Image (S)

Secret
imere (s)

Fig. 2. Image steganography framework

B. Decoder and Encoder Loss these features are difficult to rediscover and cannot be
displayed. Because steganography needs to eliminate these
In order to minimize the loss of the generated unstable features. In this article, only the feature point with
steganographic image C and the original carrier image c, as the largest width will be selected as the base point, so that
well as the extracted SI s and the original SI, this paper forces there are more feature points with the highest gradient
the hidden network and the extraction network to energy, so more integration regions can be selected.
continuously optimize learning, and finally minimize the
reconstruction error. The loss cost of the model is obtained by (3) Select the area where you want to embed the hidden
minimizing the error between the reconstructed image
steganographic image C and the original carrier image c, and
The SI to be hidden in this article is a high-resolution
the extracted SI s and the original SI s. The backpropagation
grayscale image, so the embedding network used should be a
method is used to perform model parameter Θ = {w i , bi } deep neural network, and the size of the embedding area
constantly adjust, for a set of real image x and image should match the size of the hidden image. You also need to
F j (Y , Θ) select a rectangular area that is the same size as the SI on the
obtained by network reconstruction, this paper image. The SIFT algorithm is used to select the feature point
uses the mean square error MSE as the cost function, see with the largest width. This feature point is the center of the
formula (1): high-resolution positive field as the embedding area for
embedding the hidden image.
n 2
1
L ( Θ) =
n ∑ F j (Y ; Θ) − X j
(4) Integration of SIs
i =1 (1) Once you have completed all the steps above, you can
use the encoder's deep neural network to steganography
Where n is the number of training samples. The training images, and use the coding network trained in this article to
of the network is completed by minimizing the loss of steganography SIs. The specific network structure and
equation (2): related parameters are described in the description of this
article. Specific steps are as follows:
L ( c, C , s , S ) = c − C + α s − S 1) Find the SIFT operating point in the picture.
(2)
2) If it is at the edge, select the feature point with the
Where c and s are the carrier and the SI respectively, α is widest width.
c−C 3) Use a deep neural network to embed a SI with the
how to weigh the reconstruction error, and and
s−S same size as the central area to obtain a steganographic
are the loss costs of the hidden network and the image.
extracted network.
IV. VERIFICATION OF GENERATIVE IMAGE
C. Image Information Shadow Writing
STEGANOGRAPHY SCHEME
(1) Image preprocessing
A. Data Set
The image must be preprocessed to be steganographic.
All experiments in this article are run on the standard
This is an important step in digital image processing. In this
Bosbase v1.01 dataset, which contains 15,000 vector images.
article, high-resolution images will be used, and the image
The encrypted image data set is generated by the sealing
coding network will use grayscale integration of images.
algorithm wow21 and SUNIWARDP21, and the embedding
Convert a 3-channel high-resolution color image to a
rate is between 0.2bpp and 0.5bpp. Each data set is divided
1-channel grayscale image while reducing the complexity of
into training set, validation set and test set at a ratio of 65%,
algorithm application.
45%, and 25%. On this basis, a series of experiments will be
(2) Find and display feature points conducted to verify the effectiveness of the improved method
proposed in this article.
The SFT algorithm is used to find image features in
high-resolution grayscale. The SIFT algorithm shows that
some feature points are unstable. Once the image is attacked,

193

Authorized licensed use limited to: ULAKBIM UASL - Bahcesehir University. Downloaded on July 01,2023 at 19:10:20 UTC from IEEE Xplore. Restrictions apply.
 B. Experimental Design CNN algorithms are introduced for comparison. The
Experiments are carried out on the basis of the experimental results are shown in Table I:
steganography method designed in this paper, and traditional
TABLE I. EXPERIMENTAL TEST RESULTS

The algorithm of this article Traditional algorithm

0.2bpp 68.21% 64.12%
0.3bpp 73.13% 69.41%
0.4bpp 77.65% 75.12%
0.5bpp 80.12% 78.21%
algorithm, and as the embedding rate is lower, the advantage
It can be seen from Figure 3 that the accuracy of the of the accuracy of the improved algorithm is more obvious.
improved algorithm is higher than that of the traditional

The algorithm of this article Traditional algorithm

90.00%
80.12%78.21%
77.65%
80.00% 75.12%
73.13%
68.21% 69.41%
70.00% 64.12%

60.00%
percentage

50.00%

40.00%

30.00%

20.00%

10.00%

0.00%
0.2bpp 0.3bpp 0.4bpp 0.5bpp
Embedding rate

Fig. 3. Experimental test results

V. CONCLUSIONS [5] Bau D, Strobelt H, Peebles W, et al. Semantic photo manipulation

with a generative image prior. ACM Transactions on Graphics, 2019,
This paper studies the generative image steganography 38(4):1-11.
scheme based on deep learning. After analyzing the relevant [6] Jeon G, Jeong H, Choi J. An Efficient Explorative Sampling
theories, design the generative image steganography scheme Considering the Generative Boundaries of Deep Generative Neural
based on deep learning, and verify the designed scheme, and Networks. Proceedings of the AAAI Conference on Artificial
Intelligence, 2020, 34(4):4288-4295.
get experimental results, the accuracy of the image
steganography scheme in this article is higher than that of the [7] Sae-Tang W, Fujiyoshi M, Kiya H. A New Copyright- and
Privacy-Protected Image Trading System Using A Novel
traditional scheme, and the difference between the two can Steganography-Based Visual Encryption Scheme. Transactions on
reach up to 4%. Electrical Engineering, 2019, 17(1):95-107.
[8] Behrooz, Khosravi, Behnam, et al. A new method for pdf
REFERENCES steganography in justified texts - ScienceDirect. Journal of
[1] Mohsin A H, Zaidan A A, Zaidan B B , et al. PSO–Blockchain-based Information Security and Applications, 2019, 45(C):61-70.
image steganography: towards a new method to secure updating and [9] Rashid R, Majeed T, Sardar L. Multi-Edge Concept used for Image
sharing COVID-19 data in decentralised hospitals intelligence Steganography. Innovaciencia, 2019, 7(1):1-12.
architecture. Multimedia Tools and Applications, 2021(3):1-25. [10] Mahato S, Yadav D K, Khan D A. A novel information hiding
[2] Kuppusamy P G, Ramya K C , Rani S S , et al. A Novel Approach scheme based on social networking site viewers' public comments.
Based on Modified Cycle Generative Adversarial Networks for Information Security Technical Report, 2019, 47(Aug.):275-283.
Image Steganography. Scalable Computing, 2020, 21(1):63-72. [11] Taburet T, Bas P, Sawaya W, et al. Natural Steganography in JPEG
[3] Ren T, Kotera M, Ogawa T, et al. Text-Guided Style Transfer-based Domain with a Linear Development Pipeline. IEEE Transactions on
Image Manipulation Using Multimodal Generative Models. IEEE Information Forensics and Security, 2020, PP(99):1-1.
Access, 2021, PP(99):1-1. [12] Yousif A J. Image Steganography Based on Wavelet Transform and
[4] Krebs J, Delingette H, Ayache N, et al. Learning a Generative Color Space Approach. Diyala Journal of Engineering Sciences,
Motion Model from Image Sequences based on a Latent Motion 2020, 13(3):23-34.
Matrix. IEEE Transactions on Medical Imaging, 2021, PP(99):1-1.

194

Authorized licensed use limited to: ULAKBIM UASL - Bahcesehir University. Downloaded on July 01,2023 at 19:10:20 UTC from IEEE Xplore. Restrictions apply.