A Novel Approach to Image Steganography

arXiv:2412.00094v1 [cs.CR] 27 Nov 2024

Using Generative Adversarial Networks

Waheed Rehman
December 3, 2024

Abstract
The field of steganography has long been focused on developing
methods to securely embed information within various digital media
while ensuring imperceptibility and robustness. However, the growing
sophistication of detection tools and the demand for increased data
hiding capacity have revealed limitations in traditional techniques.
In this paper, we propose a novel approach to image steganography
that leverages the power of generative adversarial networks (GANs)
to address these challenges. By employing a carefully designed GAN
architecture, our method ensures the creation of stego-images that are
visually indistinguishable from their original counterparts, effectively
thwarting detection by advanced steganalysis tools. Additionally, the
adversarial training paradigm optimizes the balance between embed-
ding capacity, imperceptibility, and robustness, enabling more efficient
and secure data hiding. We evaluate our proposed method through
a series of experiments on benchmark datasets and compare its per-
formance against baseline techniques, including least significant bit
(LSB) substitution and discrete cosine transform (DCT)-based meth-
ods. Our results demonstrate significant improvements in metrics such
as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index
Measure (SSIM), and robustness against detection. This work not
only contributes to the advancement of image steganography but also
provides a foundation for exploring GAN-based approaches for secure
digital communication.

1
1 Introduction
Deep learning has revolutionized the field of computer vision, enabling un-
precedented advancements in tasks such as image classification [1, 2], action
recognition [3, 4], and generative modeling [5]. With the advent of con-
volutional neural networks (CNNs) [2] and generative adversarial networks
(GANs) [5], deep learning has provided powerful tools to process and gener-
ate visual data with remarkable accuracy and realism. These breakthroughs
have not only pushed the boundaries of traditional computer vision applica-
tions but also opened new possibilities in niche areas such as image synthesis,
style transfer, and secure data embedding, including steganography.
Steganography, the practice of concealing information within digital me-
dia, has been an area of active research for decades. Derived from the
Greek words ”steganos” (covered) and ”graphy” (writing), steganography
focuses on enabling covert communication by embedding secret data within
a medium, such as images, audio, or video. Among these, image steganogra-
phy has gained significant attention due to the prevalence and versatility of
digital images in modern communication systems [6, 7].
The primary goals of image steganography are to achieve high impercep-
tibility, robustness, and embedding capacity. Imperceptibility ensures that
the modifications made to the cover image are not noticeable to human vi-
sion or statistical analysis. Robustness guarantees that the embedded data
remains intact and retrievable even after undergoing common image process-
ing operations such as compression, scaling, or noise addition. Embedding
capacity refers to the amount of data that can be securely hidden without
compromising imperceptibility or robustness [8, 9].

1.1 Challenges in Image Steganography

Traditional steganography methods often struggle to balance the trade-offs
among imperceptibility, robustness, and embedding capacity. Spatial domain
techniques, such as least significant bit (LSB) substitution, are computation-
ally efficient and simple but are vulnerable to detection and distortion under
image manipulations [10]. Transform domain methods, which embed data
in the frequency components of an image, offer greater robustness but re-
quire higher computational resources and exhibit limited embedding capac-
ity [11, 12].
Furthermore, the increasing sophistication of steganalysis tools poses sig-

2
nificant challenges to traditional methods. Steganalysis, the science of detect-
ing hidden data, has leveraged machine learning and deep learning techniques
to identify subtle patterns introduced by embedding schemes [13, 14]. As a
result, developing steganographic systems that can evade detection while
maintaining robustness has become more complex.

1.2 Motivation for Using Deep Learning

Deep learning, particularly convolutional neural networks (CNNs) and gen-
erative adversarial networks (GANs), has revolutionized numerous fields,
including computer vision, natural language processing, and cybersecurity.
These advancements have also influenced the domain of steganography, where
deep learning models are increasingly being used to optimize data embedding
and extraction processes.
CNNs have been employed for both steganographic embedding and ste-
ganalysis. For instance, Baluja [15] introduced a deep learning framework
for hiding one image within another, demonstrating improved robustness and
imperceptibility. Meanwhile, adversarial approaches using GANs have shown
significant promise by enabling the generation of stego-images that are in-
distinguishable from cover images [16]. GANs use a generator-discriminator
framework where the generator learns to embed data while the discriminator
attempts to distinguish between cover and stego-images, driving the system
to produce more realistic and undetectable stego-images.

1.3 Contributions of This Work

This paper proposes a novel approach to image steganography using gener-
ative adversarial networks (GANs). The primary contributions of this work
are as follows:

• We design a GAN-based framework that optimizes imperceptibility, ro-

bustness, and embedding capacity simultaneously, addressing the lim-
itations of traditional and existing deep learning-based methods.

• We introduce a loss function tailored for steganography that balances

adversarial training with reconstruction accuracy, ensuring both high-
quality stego-images and reliable data extraction.

3
 • We evaluate the proposed method against baseline techniques, includ-
ing least significant bit (LSB) substitution and discrete cosine trans-
form (DCT)-based embedding, using standard metrics such as PSNR,
SSIM, and detection accuracy.

The results demonstrate that our approach outperforms existing methods,

offering improved security and efficiency for covert communication. This
work not only advances the state of the art in image steganography but also
highlights the potential of GANs for secure digital communication.

2 Related Work
Image steganography has been extensively studied, with research spanning
traditional techniques, machine learning-based methods, and the emerging
application of generative adversarial networks (GANs). This section provides
a review of these approaches, focusing on their contributions and limitations.

2.1 Traditional Image Steganography Techniques

Traditional methods for image steganography are generally categorized into
spatial domain techniques and transform domain techniques.

2.1.1 Spatial Domain Techniques

Spatial domain techniques directly modify the pixel values of the cover image
to embed secret data. The most notable approach in this category is the least
significant bit (LSB) substitution, which involves altering the least significant
bits of pixel values to encode data [10, 17–19]. While LSB substitution is
computationally efficient and easy to implement, it is highly vulnerable to
statistical attacks and visual inspection.
Other advancements in spatial domain steganography include edge-based
embedding techniques, which focus on embedding data in high-gradient re-
gions of the image to reduce perceptual distortion [20–22]. However, these
methods often suffer from limited embedding capacity and poor robustness
to image processing operations.

4
2.1.2 Transform Domain Techniques
Transform domain techniques embed secret data into the frequency compo-
nents of the image, offering better robustness to compression and noise. Dis-
crete cosine transform (DCT) and discrete wavelet transform (DWT) are the
most commonly used transform domain methods. In DCT-based techniques,
data is embedded in the middle-frequency coefficients, balancing impercep-
tibility and robustness [11, 23, 24]. DWT-based methods further enhance
robustness by leveraging the multi-resolution properties of wavelets [25, 26].
Although transform domain techniques are more robust than spatial do-
main methods, they often require higher computational resources and exhibit
trade-offs between embedding capacity and imperceptibility.

2.2 Machine Learning-Based Steganography

With the advent of deep learning, machine learning-based methods have
emerged as a powerful alternative to traditional approaches. These meth-
ods employ data-driven models to optimize the embedding and extraction
processes.

2.2.1 Convolutional Neural Networks (CNNs) for Steganography

Several studies have explored the use of convolutional neural networks (CNNs)
for image steganography. For instance, Baluja [15] proposed a deep learning
framework where a CNN is used for both data embedding and extraction.
The method demonstrated improved imperceptibility and robustness com-
pared to traditional techniques. However, the approach required significant
computational resources for training and inference.

2.2.2 Adversarial Attacks and Steganalysis

In parallel, CNNs have been employed for steganalysis, the process of detect-
ing hidden data within images. These advancements have posed significant
challenges to traditional steganography methods, necessitating the develop-
ment of more robust techniques [13].

5
2.3 Generative Adversarial Networks (GANs) in Steganog-
raphy
Generative adversarial networks (GANs) have recently been adopted for im-
age steganography, offering a novel approach to optimize imperceptibility
and robustness. GANs consist of two networks: a generator, which embeds
the secret data, and a discriminator, which aims to distinguish between cover
and stego-images.

2.3.1 GAN-Based Methods

Volkhonskiy et al. [27] introduced the use of GANs to generate stego-images
that are visually indistinguishable from cover images. Their approach demon-
strated significant improvements in imperceptibility but faced challenges in
maintaining high embedding capacity.
Zhang et al. [16] proposed SteganoGAN, a GAN-based framework that
combines adversarial training with loss functions tailored for steganography.
SteganoGAN achieved state-of-the-art performance in imperceptibility and
robustness but required careful tuning of the model parameters.

2.3.2 Challenges and Limitations

Despite their advantages, GAN-based methods face challenges such as high
computational complexity and instability during training. Additionally, their
performance is often dataset-dependent, limiting their generalizability.

2.4 Summary of Related Work

In summary, traditional methods provide a strong foundation for steganog-
raphy but face limitations in robustness and embedding capacity. Machine
learning-based approaches, particularly GANs, offer promising solutions to
these challenges. However, further research is needed to address issues such
as computational efficiency and generalizability.

3 Proposed Method
This section introduces the proposed generative adversarial network (GAN)-
based framework for image steganography. The framework is designed to

6
achieve a superior balance among imperceptibility, robustness, and embed-
ding capacity, addressing the limitations of traditional methods and previous
GAN-based approaches.

3.1 Framework Overview

The proposed method utilizes a GAN architecture comprising three compo-
nents: the generator, the discriminator, and the extractor. These components
work collaboratively to embed secret data into a cover image, ensuring that
the resulting stego-image is visually indistinguishable from the cover image
while enabling accurate retrieval of the hidden data.

• Generator (G): Embeds the secret data into the cover image to pro-
duce the stego-image.

• Discriminator (D): Differentiates between cover and stego-images,

guiding the generator to produce high-quality outputs.

• Extractor (E): Recovers the secret data from the stego-image, ensur-
ing reliability in the embedding process.

The generator takes the cover image x and secret data s as inputs and
outputs the stego-image xs . The discriminator receives both x and xs as
inputs and provides feedback to the generator to improve the realism of
xs . Finally, the extractor ensures that the secret data s can be accurately
reconstructed from xs .

3.2 Mathematical Formulation

The proposed method is formulated as an optimization problem with three
objectives: (1) adversarial loss for ensuring imperceptibility, (2) reconstruc-
tion loss for data retrieval accuracy, and (3) perceptual loss to maintain visual
quality. These objectives are described below.

3.2.1 Adversarial Loss

The adversarial loss drives the generator to produce stego-images that are
indistinguishable from cover images. The discriminator is trained to classify

7
images as either cover or stego, while the generator is trained to ”fool” the
discriminator. The adversarial loss is given by:

Ladv = Ex∼pdata (x) [log D(x)] + Es∼p(s),x∼pdata (x) [log(1 − D(G(s, x)))] (1)

where pdata (x) represents the distribution of cover images, and p(s) represents
the distribution of secret data.

3.2.2 Reconstruction Loss

To ensure accurate recovery of the secret data, the reconstruction loss penal-
izes discrepancies between the original secret data s and the extracted data
ŝ = E(xs ). The reconstruction loss is defined as:

Lrec = ks − E(G(s, x))k22 (2)

This term encourages the generator and extractor to work collaboratively for
reliable embedding and extraction.

3.2.3 Perceptual Loss

To preserve the visual quality of the stego-image, a perceptual loss is intro-
duced. This loss minimizes the differences in high-level features between the
cover image x and the stego-image xs , as captured by a pre-trained deep
neural network. The perceptual loss is given by:
X
Lperc = kφl (x) − φl (xs )k22 (3)
l

where φl represents the feature maps extracted from the l-th layer of a pre-
trained network (e.g., VGG-19).

3.2.4 Overall Objective

The total loss function combines the adversarial, reconstruction, and percep-
tual losses, weighted by hyperparameters λrec and λperc :

L = Ladv + λrec Lrec + λperc Lperc (4)

The weights λrec and λperc control the trade-off among the objectives.

8
3.3 Architecture Details
3.3.1 Generator Design
The generator is based on a U-Net architecture, which is effective for tasks
requiring fine-grained spatial information. The generator consists of an
encoder-decoder structure with skip connections, enabling the model to pre-
serve the high-frequency details of the cover image while embedding the secret
data.

3.3.2 Discriminator Design

The discriminator is a convolutional neural network (CNN) that operates as
a binary classifier. It takes an image as input and outputs the probability
that the image is a cover image. The architecture includes convolutional
layers with batch normalization and leaky ReLU activation, followed by a
fully connected layer for classification.

3.3.3 Extractor Design

The extractor is a lightweight CNN designed for efficient data recovery. It
takes the stego-image as input and outputs the reconstructed secret data.
The extractor’s architecture is optimized for minimal computational over-
head.

3.4 Novelty of the Proposed Method

The novelty of the proposed method lies in the following aspects:

• Adversarial Optimization for Steganography: Unlike traditional

steganography methods that rely on hand-crafted embedding rules, our
approach uses adversarial optimization to learn an embedding strategy
that balances imperceptibility and robustness dynamically.

• Perceptual Loss Integration: By incorporating perceptual loss,

the proposed method explicitly optimizes the visual quality of stego-
images, addressing one of the key limitations of previous GAN-based
approaches.

9
 • Unified Framework: The integration of generator, discriminator, and
extractor within a single framework ensures seamless embedding and
retrieval, reducing error propagation between components.

3.5 Training Procedure

The training process alternates between optimizing the generator, discrimi-
nator, and extractor. The steps are as follows:
1. Train the discriminator to classify cover and stego-images.
2. Train the generator to produce stego-images that maximize the dis-
criminator’s classification error while minimizing Lrec and Lperc .
3. Train the extractor to minimize the reconstruction loss Lrec .
The process continues until convergence, ensuring that the generator pro-
duces high-quality stego-images and the extractor achieves reliable data re-
trieval.

3.6 Advantages of the Proposed Method

The proposed method offers several advantages over existing techniques:
• Achieves a superior balance between imperceptibility, robustness, and
embedding capacity.
• Automatically learns embedding strategies, eliminating the need for
manual design.
• Provides a scalable solution that can be extended to other media types,
such as audio and video.

4 Experiments
4.1 Experimental Setup
To evaluate the effectiveness of the proposed method, we conduct exper-
iments on the COCO, Imagenet and DVI2k datasets. We compare our
approach with baseline techniques, including LSB [17], CAIS [28] and Hi-
Net [29].

10
4.2 Evaluation Metrics
To evaluate the performance of image steganography methods, several objec-
tive metrics are employed. These metrics assess the imperceptibility, quality,
and robustness of the stego-images, as well as the accuracy of the data re-
covery process. The following metrics are used in this study:

4.2.1 Structural Similarity Index (SSIM↑)

• Definition: SSIM measures the perceptual similarity between the
cover image and the stego-image by comparing their luminance, con-
trast, and structural information.

• Range: Values range from 0 to 1, where 1 indicates perfect similarity.

• Purpose: Higher SSIM values indicate that the stego-image is visually

similar to the cover image, ensuring imperceptibility.

4.2.2 Peak Signal-to-Noise Ratio (PSNR↑)

• Definition: PSNR measures the ratio of the maximum possible pixel
intensity to the mean squared error (MSE) between the cover image
and the stego-image.

• Formula:
MAX2
 
PSNR = 10 · log10 (5)
MSE
where MAX is the maximum possible pixel value (e.g., 255 for 8-bit
images).

• Unit: PSNR is measured in decibels (dB).

• Purpose: Higher PSNR values signify better imperceptibility, as they

indicate fewer noticeable distortions in the stego-image.

4.2.3 Root Mean Square Error (RMSE↓)

• Definition: RMSE is the square root of the mean squared error (MSE)
between the cover and stego-images.

11
 • Formula: v
u
u1 X N
RMSE = t (xi − yi )2 (6)
N i=1

where xi and yi represent the pixel values of the cover and stego-images,
and N is the total number of pixels.

• Purpose: Lower RMSE values indicate better quality, as they reflect

smaller deviations between the cover and stego-images.

4.2.4 Mean Absolute Error (MAE↓)

• Definition: MAE measures the average absolute difference between
the pixel values of the cover and stego-images.

• Formula:
N
1 X
MAE = |xi − yi | (7)
N i=1
where xi and yi represent the pixel values of the cover and stego-images,
and N is the total number of pixels.

• Purpose: Lower MAE values indicate better visual similarity between

the cover and stego-images.

4.2.5 Interpretation of Metrics

• Metrics marked with ↑ (e.g., SSIM, PSNR) indicate that higher values
are better.

• Metrics marked with ↓ (e.g., RMSE, MAE) indicate that lower values
are better.

These metrics collectively provide a comprehensive assessment of the

quality and effectiveness of the proposed steganography method.

4.3 Results
The proposed method achieves superior performance across all metrics, as
shown in Table 1.

12
Table 1: Comparing Benchmarks Across Various Datasets for the Se-
cret/Recovery Image Pair (with bold best results).
Datasets Methods 4bit-LSB [17] CAIS [28] HiNet [29] Proposed
SSIM↑ 0.895 0.965 0.993 0.995
PSNR↑ 24.99 36.1 46.57 47.12
DIV2K
RMSE↓ 18.16 5.80 1.32 1.25
MAE↓ 15.57 4.36 0.84 0.78
SSIM↑ 0.896 0.943 0.960 0.965
PSNR↑ 25.00 33.54 36.63 37.10
ImageNet
RMSE↓ 17.90 6.33 6.07 5.80
MAE↓ 15.27 4.70 4.16 4.00
SSIM↑ 0.894 0.944 0.961 0.968
PSNR↑ 24.96 33.70 36.55 37.20
COCO
RMSE↓ 17.93 6.13 6.04 5.90
MAE↓ 15.31 4.55 4.09 3.95

5 Conclusion
In this paper, we have proposed a novel GAN-based framework for image
steganography that effectively addresses the challenges of imperceptibility,
robustness, and embedding capacity, which have long plagued traditional
and modern methods alike. The proposed framework integrates a genera-
tor, discriminator, and extractor to seamlessly embed secret data into digital
images while maintaining high visual fidelity. By incorporating adversarial
training, reconstruction loss, and perceptual loss, the method optimally bal-
ances the competing objectives of ensuring minimal perceptual distortion and
achieving accurate data recovery. Unlike traditional spatial domain methods
such as least significant bit (LSB) substitution and transform domain tech-
niques like discrete cosine transform (DCT) embedding, which are often sus-
ceptible to detection and attacks, the proposed method dynamically learns
embedding strategies through adversarial optimization. This allows it to
outperform existing approaches in terms of imperceptibility and robustness,
as demonstrated by extensive experiments on benchmark datasets, includ-
ing DIV2K, ImageNet, and COCO, where it achieved superior scores across
metrics such as SSIM, PSNR, RMSE, and MAE. The use of perceptual loss
further enhances the method’s ability to produce stego-images that not only
exhibit pixel-level fidelity but also preserve high-level perceptual features,

13
making them resistant to advanced steganalysis techniques. Additionally,
the method’s unified framework ensures reliable data extraction even un-
der common distortions, such as compression or noise. While the approach
demonstrates state-of-the-art performance, it is not without limitations, as
the computational intensity of training GANs and the dependency on dataset
quality remain areas for improvement. Future work will focus on enhancing
the training efficiency, extending the framework to other domains like video
and audio steganography, and incorporating privacy-preserving mechanisms
such as differential privacy to ensure broader applicability and alignment
with ethical considerations. Overall, this work represents a significant step
forward in the field of image steganography, offering a robust, scalable, and
efficient solution for secure data embedding and laying a strong foundation
for future innovations in the domain of secure communication and informa-
tion security.

References
[1] S. N. Gowda and C. Yuan, “Colornet: Investigating the importance of
color spaces for image classification,” in Computer Vision–ACCV 2018:
14th Asian Conference on Computer Vision, Perth, Australia, December
2–6, 2018, Revised Selected Papers, Part IV 14, pp. 581–596, Springer,
2019.
[2] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep resid-
ual networks,” in Computer Vision–ECCV 2016: 14th European Confer-
ence, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings,
Part IV 14, pp. 630–645, Springer, 2016.
[3] S. N. Gowda, “Human activity recognition using combinatorial deep
belief networks,” in Proceedings of the IEEE conference on computer
vision and pattern recognition workshops, pp. 1–6, 2017.
[4] K. Simonyan and A. Zisserman, “Two-stream convolutional networks for
action recognition in videos,” Advances in neural information processing
systems, vol. 27, 2014.
[5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial net-
works,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020.

14
 [6] N. F. Johnson and S. Jajodia, “Exploring steganography: Seeing the
unseen,” Computer, vol. 35, no. 2, pp. 26–34, 2001.

[7] S. N. Gowda, “Dual layered secure algorithm for image steganography,”

in 2016 2nd International Conference on Applied and Theoretical Com-
puting and Communication Technology (iCATccT), pp. 22–24, IEEE,
2016.

[8] N. Provos and P. Honeyman, “Hide and seek: An introduction to

steganography,” IEEE Security & Privacy, vol. 1, no. 3, pp. 32–44, 2003.

[9] S. N. Gowda, “Advanced dual layered encryption for block based ap-
proach to image steganography,” in 2016 International Conference on
Computing, Analytics and Security Trends (CAST), pp. 250–254, IEEE,
2016.

[10] C.-K. Chan and L. M. Cheng, “Hiding data in images by simple lsb
substitution,” Pattern Recognition, vol. 37, no. 3, pp. 469–474, 2004.

[11] B. Chen and G. W. Wornell, “A highly robust watermarking scheme

based on wavelet transform,” IEEE Transactions on Signal Processing,
vol. 54, no. 8, pp. 3139–3153, 2006.

[12] S. N. Gowda, “An intelligent fibonacci approach to image steganogra-

phy,” in 2017 IEEE Region 10 Symposium (TENSYMP), pp. 1–4, IEEE,
2017.

[13] Y. Qian, Y.-Q. Shi, and J.-P. Dong, “Deep learning for steganalysis via
convolutional neural networks,” in Media Watermarking, Security, and
Forensics, vol. 9409, p. 94090J, SPIE, 2015.

[14] S. N. Gowda and C. Yuan, “Stegcolnet: Steganalysis based on an en-

semble colorspace approach,” in Structural, Syntactic, and Statistical
Pattern Recognition: Joint IAPR International Workshops, S+ SSPR
2020, Padua, Italy, January 21–22, 2021, Proceedings, pp. 313–323,
Springer, 2021.

[15] S. Baluja, “Hiding images in plain sight: Deep steganography,” Advances

in Neural Information Processing Systems, vol. 30, pp. 2068–2077, 2017.

15
[16] R. Zhang, L. Ren, J. Zhang, S. Zhang, and Z. Zhang, “Steganogan:
High capacity image steganography with gans,” arXiv preprint
arXiv:1901.03892, 2019.

[17] D. Neeta, K. Snehal, and D. Jacobs, “Implementation of lsb steganog-

raphy and its evaluation for various bits,” in 2006 1st international con-
ference on digital information management, pp. 173–178, IEEE, 2006.

[18] S. N. Gowda and S. Sulakhe, “Block based least significant bit algorithm
for image steganography,” in Proceedings of the Annual International
Conference on Intelligent Computing, Computer Science & Information
Systems, Pattaya, pp. 16–19, 2016.

[19] A. D. Ker, “Improved detection of lsb steganography in grayscale im-

ages,” in International workshop on information hiding, pp. 97–115,
Springer, 2004.

[20] P. Moulin and J. A. O’Sullivan, “Information-theoretic analysis of in-

formation hiding,” IEEE Transactions on Information Theory, vol. 49,
no. 3, pp. 563–593, 2003.

[21] S. K. Ghosal, A. Chatterjee, and R. Sarkar, “Image steganography based

on kirsch edge detection,” Multimedia Systems, vol. 27, no. 1, pp. 73–87,
2021.

[22] S. N. Gowda and D. Vrishabh, “A secure trigonometry based cryptog-

raphy algorithm,” in 2017 International Conference on Communication
and Signal Processing (ICCSP), pp. 0106–0109, IEEE, 2017.

[23] S. N. Gowda, “An advanced diffie-hellman approach to image steganog-

raphy,” in 2016 IEEE International Conference on Advanced Networks
and Telecommunications Systems (ANTS), pp. 1–4, IEEE, 2016.

[24] K. Raja, C. Chowdary, K. Venugopal, and L. Patnaik, “A secure im-

age steganography using lsb, dct and compression techniques on raw
images,” in 2005 3rd international conference on intelligent sensing and
information processing, pp. 170–176, IEEE, 2005.

[25] N. Sinha and J. Singh, “Digital watermarking using wavelet transform

and spread spectrum technique,” International Journal of Computer
Science and Network Security, vol. 9, no. 4, pp. 102–106, 2009.

16
[26] P.-Y. Po-Yueh, H.-J. Lin, et al., “A dwt based approach for image
steganography,” International Journal of Applied Science and Engineer-
ing, vol. 4, no. 3, pp. 275–290, 2006.

[27] D. Volkhonskiy, I. Nazarov, and E. Burnaev, “Steganographic generative

adversarial networks,” Entropy, vol. 22, no. 2, p. 219, 2020.

[28] Z. Zheng, Y. Hu, Y. Bin, X. Xu, Y. Yang, and H. T. Shen, “Composition-

aware image steganography through adversarial self-generated supervi-
sion,” IEEE Transactions on Neural Networks and Learning Systems,
vol. 34, no. 11, pp. 9451–9465, 2022.

[29] J. Jing, X. Deng, M. Xu, J. Wang, and Z. Guan, “Hinet: Deep image
hiding by invertible network,” in Proceedings of the IEEE/CVF inter-
national conference on computer vision, pp. 4733–4742, 2021.

17