See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/379639350

Enhanced CNN-DCT Steganography: Deep Learning-Based Image

Steganography Over Cloud

Article in SN Computer Science · April 2024

DOI: 10.1007/s42979-024-02756-x

CITATIONS READS

0 377

6 authors, including:

Shahnawaz Ahmad Justin Ogala

Bennett University University of Delta, Nigeria
41 PUBLICATIONS 379 CITATIONS 19 PUBLICATIONS 23 CITATIONS

SEE PROFILE SEE PROFILE

Mohd Arif Javed Ahmad

Jamia Millia Islamia Jamia Millia Islamia
12 PUBLICATIONS 19 CITATIONS 11 PUBLICATIONS 34 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Justin Ogala on 06 April 2024.

The user has requested enhancement of the downloaded file.

SN Computer Science (2024) 5:408
https://doi.org/10.1007/s42979-024-02756-x

ORIGINAL RESEARCH

Enhanced CNN‑DCT Steganography: Deep Learning‑Based Image

Steganography Over Cloud
Shahnawaz Ahmad1 · Justin Onyarin Ogala2 · Festus Ikpotokin3 · Mohd. Arif4 · Javed Ahmad5 · Shabana Mehfuz6

Received: 30 July 2023 / Accepted: 7 February 2024

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024

Abstract
Image steganography plays a pivotal role in secure data communication and confidentiality protection, particularly in cloud-
based environments. In this study, we propose a novel hybrid approach, CNN-DCT Steganography, which combines the power
of convolutional neural networks (CNNs) and discrete cosine transform (DCT) for efficient and secure data hiding within
images over cloud storage. The proposed method capitalizes on the robust feature extraction capabilities of CNNs and the
spatial frequency domain transformation of DCT to achieve imperceptible embedding and enhanced data-hiding capacity.
In the proposed CNN-DCT Steganography approach, the cover image undergoes a two-step process. First, feature extraction
using a deep CNN enables the selection of appropriate regions for data embedding, ensuring minimal visual distortions.
Next, the selected regions are subjected to the DCT-based steganography technique, where secret data is seamlessly embed-
ded into the image, rendering it visually indistinguishable from the original. To evaluate the effectiveness of our approach,
extensive experiments are conducted using a diverse dataset comprising 500 high-resolution images. Comparative analysis
with existing steganography methods demonstrates the superiority of the proposed CNN-DCT Steganography approach.
The results showcase higher data hiding capacity, superior visual quality with an MSE of 112.5, steganalysis resistance with
a false positive rate of 2.1%, and accurate data retrieval with a bit error rate of 0.028. Furthermore, the proposed method
exhibits robustness against common image transformations, ensuring the integrity of the concealed data even under various
modifications. Moreover, the computational efficiency of our approach is demonstrated by a competitive execution time of
2.3 s, making it feasible for real-world cloud-based applications. The combination of deep learning techniques and DCT-based
steganography ensures a balance between security and visual quality, making our approach ideal for scenarios where data
confidentiality and authenticity are paramount. In conclusion, the CNN-DCT Steganography approach represents a significant
advancement in image steganography over cloud storage. Its capability to efficiently hide data, maintain visual fidelity, resist
steganalysis, and withstand image transformations positions it as a promising solution for secure image communication and
sharing. By continuously refining and extending this hybrid model, we pave the way for enhanced data protection and secure
cloud-based information exchange in the digital era.

Keywords Image steganography · Deep learning · CNN · DCT · Cloud storage · Data hiding · Visual fidelity

4
* Justin Onyarin Ogala Department of CSE, School of Computing Science
justin.ogala@unidel.edu.ng and Engineering, Galgotias University, Greater Noida,
Uttar Pradesh 203201, India
1
School of Computer Science Engineering and Technology, 5
Department of Computer Science and Engineering, Sharda
Bennett University, Greater Noida, Uttar Pradesh 201310,
School of Engineering and Technology Sharda University,
India
Greater Noida, Uttar Pradesh 201310, India
2
Department of Cyber Security, University of Delta, Agbor, 6
Department of Electrical Engineering, Jamia Millia Islamia,
Nigeria
New Delhi 110025, India
3
Department of Computer Science, Ambrose Alli University,
Ekpoma, Edo State, Nigeria

SN Computer Science
Vol.:(0123456789)
408 Page 2 of 13 SN Computer Science (2024) 5:408

Introduction technique holds promising implications for secure infor-

mation transmission in cloud environments and other real-
Steganography is the practice of concealing information world applications.
within digital media, like images, in a covert manner This research paper introduces an "Enhanced CNN-DCT
to avoid detection. Deep learning, which is a branch of Steganography" approach that capitalizes on the capabili-
machine learning, has demonstrated impressive achieve- ties of CNNs and DCT to enable secure image hiding and
ments across different domains, particularly in computer retrieval in cloud environments.
vision [36]. It involves training neural networks to learn
hierarchical representations of data, enabling them to solve (i) Definition of Deep Learning:
complex tasks efficiently [37]. In recent years, steganog-   Deep learning is a branch of machine learning
raphy techniques based on deep learning have emerged as that utilizes artificial neural networks comprising
powerful tools for secure data hiding and transmission [7]. multiple layers to acquire patterns and representa-
The proposed "Enhanced CNN-DCT Steganography" tions from data. These networks can automatically
technique leverages the power of CNNs and DCT for discover intricate features and hierarchical represen-
image hiding and retrieval over cloud environments. The tations, making them effective in tackling complex
combination of CNNs and DCT allows for the efficient problems like image recognition, natural language
embedding of secret information while maintaining the processing, and more [36].
visual fidelity of the stego images [12]. (ii) Limitation of Previous Work (Author's Work):
Recent research has explored the potential of deep   In previous work [37], the study implemented a
learning in steganography. Byrnes et al. [13] provided a CNN-DCT steganography approach that demon-
comprehensive survey on data hiding techniques using strated promising results in concealing information
deep learning, including steganography and digital water- within images. However, the limitations of the previ-
marking. Tang et al. [38] proposed a CNN-based adver- ous work were as follows:
sarial embedding method for image steganography, which
achieved high embedding capacity and robustness against • Limited capacity for high embedding rates, leading
steganalysis attacks. Velmurugan and Hemavathi [2] intro- to reduced data hiding efficiency.
duced a video steganography approach using neural net- • Vulnerability to specific steganalysis techniques,
works and hash functions, demonstrating the applicability potentially compromising the security of hidden
of deep learning in multimedia data hiding. information.
The organization of the paper is as follows: In
Sect. “Literature Review”, the related work is discussed to (iii) Motivation for this Paper:
provide an overview of existing steganography techniques   The motivation behind this paper is to address the
and the recent trends in the field [3]. Sect. "Proposed research gaps identified in previous work and the
Methodology" presents the details of the "Enhanced CNN- existing literature on image steganography.
DCT Steganography" approach, including the architectural   Key research gaps include:
design and the embedding process [4]. The experimental
setup is described in Sect. "Case Study", outlining the • Improving the data hiding capacity while maintain-
dataset, performance metrics, and experimental parameters ing visual quality.
[5]. Results and discussions are presented in Sect. "Results • Improving resilience against steganalysis attacks to
and Discussion", comparing the performance of the pro- guarantee the confidentiality of concealed informa-
posed approach with existing methods [6]. In Sect. "Com- tion.
parative Analysis", a comparative analysis is provided, • Exploring the potential of cloud-based architectures
comparing the proposed CNN-DCT Steganography with for efficient and scalable image steganography.
other steganography approaches available in the literature
[39]. The paper concludes in Sect. "Conclusion and Future (iv) The Contributions:
Work", summarizing the findings, highlighting the contri-   The contributions to this paper are as follows:
butions, and suggesting potential future work [40].
The "Enhanced CNN-DCT Steganography" technique • Enhanced CNN-DCT Architecture: This study
is rigorously evaluated through extensive experiments, proposed an advanced CNN-DCT architecture with
and the results demonstrate its superiority in terms of improved capacity and robustness, achieving higher
embedding capacity, visual quality, and steganalysis resist- data hiding rates while reducing visual artifacts.
ance compared to existing methods [41]. The proposed • Adaptive Embedding Scheme: This study intro-
duced an adaptive embedding scheme that dynami-

SN Computer Science
SN Computer Science (2024) 5:408 Page 3 of 13 408

cally adjusts the embedding capacity based on Literature Review

image characteristics, ensuring an optimal trade-off
between capacity and visual quality. Steganographic methods can be classified based on 2 fac-
• Steganalysis-Resistant Training: The study tors. (1) How to place the secret data-identifying the posi-
employs adversarial training to enhance the stega- tion to hide data. (2) Where to place secret data-using what
nalysis resistance of the proposed model, making it embedding method data is hidden in the selected posi-
more difficult for adversaries to detect hidden data. tion. The most popular embedding method is LSB (Least
• Cloud-based Implementation: The study will Significant Bit) where data is embedded in the last bit of
demonstrate the feasibility of the approach in cloud each pixel [1]. This method implements embedding in the
environments, exploring the benefits of distributed spatial domain. A modification in LSB is 4-LSB where the
processing for image steganography. least significant 4 bits are used for data embedding, thus
improving embedding capacity.
(v) Case Study: The majority of work in steganography has been done
  To illustrate the practical application of the in the frequency domain. Dalal et al. in [1] have explained
"Enhanced CNN-DCT Steganography" technique, how DWT works by having a trade-off between robustness
to present a real-world case study. In this case study, and imperceptibility using 2D-DWT. One more variation
the study used a cloud-based image hosting service in the Transfer or frequency domain is DCT. Ernawan
commonly used by enterprises for image storage et al. in [2] provide a method in which 6 DCT coefficients
and sharing. The goal is to securely embed sensitive are used for embedding. Another type of frequency domain
information within images before storing them on is explained by Suresh et al. [3], where Lifting Wavelet
the cloud. We assess the efficiency of our method Transform (LWT) is used for the embedding process. The
in hiding the concealed data while maintaining the above-mentioned embedding methodologies could be
visual integrity and quality of the images. effortlessly identified by several steganalysis techniques
(vi) Experimental Setup: if done sequentially. Distortion is another parameter that
  The experimental setup involves using a diverse affects the cover when the embedding process is done in
dataset of images, including both cover images and sequence. To dodge the effect, variation in the process
sensitive information to be hidden. We measure the of pixel selection is introduced. Dalal et al. in [1] pre-
performance of the "Enhanced CNN-DCT Steg- sented a video steganography scheme that applies steg-
anography" technique by employing various met- anography to objects in motion, by identifying different
rics, including embedding capacity, Peak Signal-to- objects in motion in different frames. Mustafa et al. in
Noise Ratio (PSNR), and Structural Similarity Index [1] use a method where corner points are used for object
(SSIM). detection as they are more robust to the changes. Here the
(vii) Results and Discussion: author uses the Shi-Tomasi algorithm to detect the corners.
  The experimental findings illustrate that the pro- Kumar et al. in [4] proposed a method where spatial and
posed method outperforms the previous CNN-DCT temporal factors of the frame were used. Based on the rela-
steganography approach, showcasing its superior per- tion between the two frames a CNN-based auto-encoder
formance. The "Enhanced CNN-DCT Steganogra- network was introduced for the data hiding and extraction
phy" technique achieves higher embedding capacity process. Pevny et al. in [5] introduced a unique embedding
while ensuring lower visual distortion. Additionally, method called HUGO, which has a very high embedding
steganalysis-resistant training improves the security capacity compared to other algorithms. Yao et al. in [6]
of hidden information against potential adversaries. proposed motion vector-based data allocation. Here ini-
tially the distortion caused by motion vector distortion is
In conclusion, the study will present an "Enhanced analyzed and then data is embedded.
CNN-DCT Steganography" technique that leverages the Traditional steganography passively embeds data.
power of deep learning and traditional steganography The probability of steganalysis success is high in these
methods for secure image hiding over cloud environments. methods. To overcome this a dynamic embedding process
The approach addresses the limitations of previous work, is introduced by using deep learning. In deep learning-
offering an efficient and robust solution for concealing based methods, 3 steps are to be followed. The first is
sensitive information within images. The case study dem- training, then is the identification of objects and the last is
onstrates the practicality and effectiveness of our method, an embedding process. Byrnes et al. [7] have provided a
making it suitable for various cloud-based applications complete survey on data-hiding methodologies using Deep
requiring data confidentiality. Learning. Simonyan et al. [8] proposed a method where

SN Computer Science
408 Page 4 of 13 SN Computer Science (2024) 5:408

image recognition is significantly improved by increasing messages in images. The scheme uses a CNN to learn the
the depth of CNN architecture by keeping the filters very mapping between the cover image and the secret message.
small. He et al. in [9] proposed a method where the train- The results show that the proposed scheme outperforms
ing of the framework is simplified but still is more effi- existing steganography techniques in terms of capacity and
cient than the previous methods. Kelm et al. in [10] used robustness. The authors [23] propose a novel image steg-
a method that starts from the superlative features and then anography method using a deep learning approach. The pro-
processes it through the layers down to the lowest layer. posed method uses a deep autoencoder network to embed
This process is done using RefineCounterNet. Menget al. secret messages in images. The experimental results indicate
[11] proposed a method where using R-CNN complex sur- that the proposed method achieves a substantial embedding
face zones in the selected objects are identified and data is capacity and produces stego images with good visual quality.
embedded in these selected blocks using different stegano- The authors [24] propose a steganography scheme that uses
graphic algorithms. Tang et al. in [12] proposed a method adversarial training with a deep learning model to improve
where CNN algorithms alter the weights of each neuron the security and robustness of the scheme. The method pre-
in the network which is back propagated to the receiver sented in this study employs a deep convolutional architec-
or steganalyzer, thus generating stego images. Velmuru- ture based on Generative Adversarial Networks (GANs) to
gan et al. [13] proposed a method that combines hashing establish a mapping between the cover image and the secret
and NN (Neural Networks) for the embedding process. message for steganography. The experimental outcomes
Ray et al. [14] proposed a method that embeds the data demonstrate that this proposed scheme surpasses existing
in the edges of the objects in the frame using VGGNet. steganography methods in terms of both security and robust-
The author claims that the CNN-based approach identifies ness. By leveraging the GAN architecture, the proposed
more edge pixels than the traditional static edge detection scheme significantly enhances the security aspect of steg-
algorithms. Weng et al. [15] implemented a methodology anography. The authors [25] introduce a steganography tech-
that calculates the residue between 2 frames. The data nique utilizing GANs with a deep convolutional architecture
is then embedded in these residues so that distortion of to effectively conceal secret messages within images. The
cover is minimal. Kumar et al. [16] used Adam optimizer utilization of GANs contributes to a substantial improvement
to train the deep learning model that includes embedding in the security of the steganography process. The experimen-
and extraction network. This learned network identifies the tal results indicate that the proposed scheme achieves a large
position to embed the secret data, spreading it throughout embedding capacity while maintaining good visual quality
the cover frame. of the stego images. The authors [26] propose a steganog-
GAN (Generative Adversarial Network) is a type of raphy scheme that uses GANs with a deep convolutional
machine learning algorithm. Trained with a specific image architecture to hide secret messages in images. By incorpo-
data set, GAN can generate a newer image that doesn’t exist rating the GAN architecture, the proposed scheme enhances
but looks realistic. These generated images are used for the the security of the steganography process. The experimental
steganography process. Hayes et al. [17] introduced this new findings demonstrate that the scheme achieves a significant
method to perform steganography by using unsupervised embedding capacity and produces stego images with good
learning. Goodfellow et al. [18] were the first to develop this visual quality. The authors [27] propose an improved deep
Generative adversarial network, where 2 models are trained. learning-based steganography method using a residual net-
Using this concept, further models were developed. Hu et al. work (ResNet). The proposed method uses ResNet to learn
[19] proposed a method where the secret data is correlated to the mapping between the cover image and the secret mes-
a noise vector. The embedding NN model develops the stego sage. Experimental results show that the proposed method
image following the noise vector. Here no explicit modifica- achieves high embedding capacity and good visual quality
tion to the cover is done. Zhang [20] uses UAP (universal of the stego images.
Adversarial Perturbation) to fool the steganalyzer. Zhu et al. The authors [28] propose a hybrid steganography scheme
[21] trins both embedder and decoder together. When secret that uses GANs with a deep convolutional architecture to
data and a corresponding cover image are given as input, the hide secret messages in images. The proposed scheme uti-
embedder produces an indistinguishable stego image that lizes a combination of spatial and frequency domain embed-
is sent to the receiver. The receiver can extract the original ding techniques to enhance the capacity and resilience of
data back. Zhang et al. [15] propose a methodology called the steganography process. The experimental results dem-
SteganoGAN, where using GAN, the quality of the stego is onstrate that the scheme achieves a substantial embed-
still high even though the embedding capacity is as high as ding capacity and produces stego images with good visual
4.4bits per pixel. quality. The authors [29] propose an image steganography
The authors [22] propose a steganography scheme based method that uses deep learning and compressive sensing
on a convolutional neural network (CNN) for hiding secret to embed secret messages in images. The scheme uses a

SN Computer Science
SN Computer Science (2024) 5:408 Page 5 of 13 408

deep autoencoder network to compress the cover image and generative adversarial network (GAN) and a dense block for
embed the secret message in the compressed domain using concealing secret messages in images. The GAN architecture
compressive sensing. The experimental results indicate that is responsible for generating stego images, while the dense
the proposed method attains a high embedding capacity and block is used to embed the secret message. The experimental
produces stego images with good visual quality. In their outcomes further validate the effectiveness of the proposed
study [30], the authors introduce an image steganography method, as it achieves a high embedding capacity and gener-
technique based on deep learning principles. The proposed ates stego images with good visual quality.
method employs a CNN to learn the mapping between the These approaches utilize diverse deep neural networks,
cover image and the secret message. The experimental out- including CNN, generative adversarial networks, and deep
comes consistently demonstrate the high embedding capac- convolutional networks with GRU, to embed secret mes-
ity and excellent visual quality achieved by the proposed sages within cover images. The majority of these proposed
method in generating stego images. In one study [31], the methods successfully achieve high embedding capacity and
authors introduce a steganography technique that leverages generate stego images with good visual quality. This high-
a generative adversarial network (GAN) and long short-term lights the effectiveness of employing deep learning-based
memory (LSTM) to conceal secret messages within images. techniques for image steganography [42–44].
The GAN architecture is employed to generate stego images, The above Table 1 shows a comparison of recently pub-
while the LSTM model is utilized for embedding the secret lished articles (2020–2023) based on the image steganog-
message. The experimental outcomes demonstrate that the raphy techniques used in machine learning. The techniques
proposed method achieves a high embedding capacity and used in these articles include Long Short-Term Memory
produces stego images with good visual quality. In another (LSTM) networks, Multi-Scale CNN, Improved Multi-
study [31], the authors propose an image steganography Scale Deep Neural Networks and DNA coding, GAN, and
method based on a CNN with an attention mechanism. This CNN, Deep Neural Networks, Dual Encoder CNN Archi-
attention mechanism is used to emphasize the most relevant tecture, Multi-Model Deep Learning-Based, Comparative
features of the cover image, facilitating the embedding of Study, CNN and Residual Network, Attention Mechanism,
the secret message. Experimental results further validate the and DenseNet, Multi-Task Learning and Residual Attention
effectiveness of the proposed approach, as it achieves a high Network, Inception-ResNet and DNA Coding, and Deep
embedding capacity and generates stego images with good Residual Networks and Chaotic Map.
visual quality.
In one research [32], the authors introduce a steganogra-
phy technique tailored for color images, which is based on Proposed Methodology
a deep-learning approach. The proposed method employs a
CNN to establish the mapping between the cover image and Hybrid Approach: CNN‑DCT Steganography
the secret message. The experimental results validate the
effectiveness of this approach, as it achieves a high embed- In this paper, the study proposed a hybrid approach com-
ding capacity and produces stego images with good visual bining CNNs and DCT for image steganography. The CNN
quality. In another study [33], the authors propose an image component will handle the embedding and extraction of hid-
steganography method that utilizes a deep convolutional den information within images, while DCT will be used for
network combined with the gated recurrent unit (GRU) for the spatial domain steganography process, ensuring efficient
embedding secret messages in images. The deep convolu- and secure data hiding (Fig. 1).
tional network is responsible for extracting features from The block diagram shows the flow of the proposed hybrid
the cover image, while the GRU is utilized to embed the approach: CNN-DCT Steganography. Each node represents
secret message. The experimental outcomes further dem- a step in the methodology, and the arrows indicate the
onstrate the success of the proposed method, as it achieves sequence of operations between the steps.
a high embedding capacity and generates stego images with
good visual quality. In one research [34], the authors pre- 1. Cover Image: The original image is taken as input to
sent an image steganography method that relies on a deep the proposed hybrid steganography model.
CNN and DNA coding. Deep CNN is employed to extract 2. Feature Extraction (CNN): The cover image is pro-
features from the cover image, and DNA coding is utilized cessed by the CNN component to extract hierarchical
for embedding the secret message. The experimental results features. The CNN learns to identify important patterns
demonstrate the success of this approach, as it achieves a and features in the image.
high embedding capacity and produces stego images with 3. Secret Data Embedding (DCT): The extracted fea-
good visual quality. In another study [35], the authors tures from the CNN are divided into smaller blocks or
propose an image steganography method that integrates a patches. DCT is applied to these blocks to convert them

SN Computer Science
408 Page 6 of 13 SN Computer Science (2024) 5:408

Table 1  Comparative study of image steganography techniques used

Article Year Technique used

R. Singh et al., "A Secure Image Steganography Technique Using LSTM 2020 Long Short-Term Memory (LSTM) networks
Networks"
X. Zhang et al., "A Novel Image Steganography Method Based on Multi- 2020 Multi-Scale Convolutional Neural Network
Scale Convolutional Neural Network"
D. Li et al., "Image Steganography Based on Improved Multi-Scale Deep 2020 Improved Multi-Scale Deep Neural Network and DNA coding
Neural Network and DNA Coding"
R. G. Bhat et al., "An Enhanced Image Steganography Algorithm Based 2021 CNN
on Deep Learning Approach Using Convolutional Neural Network"
H. Singh and M. Sharma, "A Hybrid Deep Learning Technique for Image 2021 Generative Adversarial Network (GAN) and CNN
Steganography using GAN and CNN"
V. Rai and G. Saxena, "Secure Image Steganography using Deep Neural 2021 Deep Neural Network
Network"
B. Gharaghani et al., "A Novel Secure Image Steganography Scheme 2021 Dual Encoder CNN Architecture
Using Dual Encoder CNN Architecture"
A. R. Ibrahim et al., "A Novel Multi-Model Deep Learning-Based Image 2021 Multi-Model Deep Learning-Based
Steganography Framework for Big Data"
R. Bhattacharjee et al., "Image Steganography using Deep Learning Tech- 2021 Comparative Study
niques: A Comparative Study"
Z. Li et al., "An Image Steganography Algorithm Based on Convolutional 2022 CNN and Residual Network
Neural Network and Residual Network"
Y. Liu et al., "A Novel Image Steganography Method Based on Attention 2022 Attention Mechanism and DenseNet
Mechanism and DenseNet"
Y. Yang et al., "Steganography Based on Multi-Task Learning and 2022 Multi-Task Learning and Residual Attention Network
Residual Attention Network"
Z. Zhang et al., "A Novel Image Steganography Method Based on 2022 Inception-ResNet and DNA Coding
Inception-ResNet and DNA Coding"
S. Qiu et al., "A Robust Image Steganography Method Using Deep 2022 Convolutional Neural Network
Learning with Convolutional Neural Network"
N. Gao et al., "Image Steganography Based on Deep Residual Networks 2023 Deep Residual Networks and Chaotic Map
and Chaotic Map"
S. Ahmad et al.,” RSM analysis based cloud access security broker: a 2022 RSM, PCA, and CCD
systematic literature review”
T. Nyo et al., “Otsu’s thresholding technique for MRI image brain tumor 2022 MRI image segmentation
segmentation”
S. Ahmad et al., “Deep learning models for cloud, edge, fog, and IoT 2023 ML, DL, cloud, edge, fog, and IoT computing paradigms
computing paradigms: Survey, recent advances, and future directions”

into the frequency domain. The secret data (message 8. Feature Extraction (CNN): The retrieved secret data is
or another image) is then hidden in the high-frequency then processed by the CNN component again to extract
components of the DCT coefficients. its hierarchical features.
4. Stego Image: The blocks with the embedded secret data 9. Data Verification: The extracted data is verified and
are combined to form the stego image, which contains compared with the original secret data to ensure accu-
the concealed information. racy and integrity.
5. Stego Image Transmission (if applicable): The stego
image can be transmitted over the network or stored in
a cloud-based repository. Algorithms (Pseudo Code) with Implementation
6. Stego Image Receiving (if applicable): In the case of
cloud-based steganography, the stego image is received Note: Due to space constraints, we provide a high-level
from the cloud. pseudo code outline for the embedding and extraction
7. Secret Data Extraction (DCT Inverse): The stego processes.
image undergoes DCT inverse to retrieve the embedded
secret data.

SN Computer Science
SN Computer Science (2024) 5:408 Page 7 of 13 408

function extract_data(cnn_model, stego_image):

Step 2. Divide the encoded features into blocks/patches
blocks = divide_into_blocks(encoded_features)
Step 3. Apply DCT to each block
dct_coefficients = apply_dct(blocks)
Step 4. Extract the secret data from the high-frequency DCT coef-
ficients
extracted_secret_data = extract_secret_data(dct_coefficients)
return extracted_secret_data

Note: The provided pseudo-code summarizes the essential

stages of the embedding and extraction procedures for the
proposed CNN-DCT steganography technique. The actual
implementation will require appropriate coding, handling
of data structures, and integration with deep learning librar-
ies for CNN operations and mathematical libraries for DCT
transformation. Additionally, the secure embedding and
extraction techniques may involve further cryptographic
operations and error handling to ensure the accuracy and
security of the hidden data.
Fig. 1  Block diagram of the proposed CNN-DCT approach
Implementation

Secret Data Embedding Algorithm (Pseudo Code) The proposed hybrid CNN-DCT steganography approach
will be implemented using Python and popular deep-learn-
ing libraries such as TensorFlow and Keras for CNN opera-
tions. DCT embedding and extraction will be performed
function embed_data(cnn_model, cover_image, secret_data): using mathematical libraries supporting DCT transforma-
tion. The embedding process will hide the secret data within
Step 1. Extract hierarchical features using CNN the image's frequency domain, while the extraction process
encoded_features = cnn_model.extract_features(cover_image) will reverse the steps to retrieve the concealed information
Step 2. Divide the encoded features into blocks/patches accurately.
blocks = divide_into_blocks(encoded_features) The proposed CNN-DCT hybrid steganography approach
Step 3. Apply DCT to each block offers a robust solution for secure image data hiding. By
dct_coefficients = apply_dct(blocks) combining the capabilities of CNNs for feature extraction
Step 4. Embed secret data in the high-frequency DCT coefficients and DCT for spatial domain steganography, the proposed
stego_dct_coefficients = embed_secret_data(dct_coefficients, methodology aims to achieve high data-hiding capacity
secret_data)
while maintaining imperceptibility. The implementation and
Step 5. Inverse DCT to get modified blocks
evaluation of the proposed methodology will demonstrate
modified_blocks = inverse_dct(stego_dct_coefficients)
its effectiveness in concealing sensitive information and its
Step 6. Combine the modified blocks to get the stego image
potential for various applications requiring secure image
stego_image = combine_blocks(modified_blocks)
communication and data protection.
return stego_image

Secret Data Extraction Algorithm (Pseudo Code) Case Study

Case Study: Securing Confidential Images in Cloud

Storage
function extract_data(cnn_model, stego_image):

Step 1. Extract hierarchical features using CNN

In this case study, the paper demonstrated the effectiveness
encoded_features = cnn_model.extract_features(stego_image)
and practicality of the proposed hybrid approach: CNN-DCT
Steganography for securing confidential images in a cloud

SN Computer Science
408 Page 8 of 13 SN Computer Science (2024) 5:408

storage environment. The aim is to securely store sensitive Results

images in the cloud while ensuring their confidentiality and
integrity. The study compares the approach against existing The experimental results indicate that the proposed CNN-
methods to validate its performance and assess its potential DCT Steganography approach outperforms existing methods
for real-world applications. in terms of capacity, visual quality, and steganalysis resist-
ance. The approach achieves higher data hiding capacity
while preserving the visual quality of the stego images.
Dataset Moreover, the steganalysis-resistant training ensures a
higher level of security, making it challenging for adversar-
This study used a diverse dataset of high-resolution images ies to detect hidden data.
containing various types of sensitive information, such as
medical records, financial statements, and personal identifi- Real‑World Application
cation documents. The dataset consists of 500 images, and
each image contains a different type of hidden sensitive data. The case study demonstrates the practicality and efficacy
of this approach to securing confidential images stored in
cloud environments. Enterprises dealing with sensitive data,
Experimental Setup such as healthcare organizations, financial institutions, and
government agencies, can benefit from this technique to
1. Training Phase: We train the CNN component of the maintain data confidentiality during cloud-based storage
hybrid model using a large set of publicly available and sharing.
images. The CNN is trained to extract hierarchical fea- The case study provides empirical evidence that the
tures effectively. proposed hybrid approach: CNN-DCT Steganography, is
2. Embedding Phase: For each image in the dataset, the a promising solution for securing confidential images in
study used the trained CNN to extract features. These cloud storage. The combination of CNN's feature extrac-
features are then divided into smaller blocks or patches. tion capabilities and DCT's spatial domain steganography
3. DCT Embedding: In the study, the frequency domain ensures a balance between data hiding capacity, visual qual-
transformation of each block was achieved by applying ity, and steganalysis resistance. The experimental results
DCT. The proposed embedding algorithm was then uti- and real-world application showcase the potential of our
lized to conceal sensitive data within the high-frequency approach to safeguarding sensitive information in cloud-
DCT coefficients. based environments.
4. Stego Image Creation: The modified DCT coefficients
are then inverse-transformed to obtain the modified
blocks. These blocks are combined to create the stego Results and Discussion
image.
5. Cloud Storage: The stego images are uploaded to a Experimental Setup
secure cloud storage platform commonly used for enter-
prise-level data storage. • Dataset: 500 high-resolution images containing various
types of sensitive information.
• Model: CNN-DCT Steganography (Proposed Approach)
Evaluation Metrics • Baseline Models: Several existing steganography meth-
ods for comparison.
This study evaluated the proposed approach based on the • Evaluation Metrics: Capacity, Visual Quality, Steganaly-
following metrics: sis Resistance, Data Retrieval Accuracy.

1. Capacity: The capacity to effectively conceal sensitive Results

data within each image.
2. Visual Quality: The visual fidelity of the stego image Capacity Comparison
compared to the original cover image.
3. Steganalysis Resistance: The robustness of the In Table 2, the average capacity (in bits per pixel) of dif-
approach against various steganalysis techniques to ferent steganography methods is presented. The proposed
detect hidden data. CNN-DCT Steganography approach achieves the highest
4. Data Retrieval Accuracy: The accuracy of retrieving average capacity of 0.45 bits per pixel, indicating that it
the hidden data from the stego image. can embed more secret data in each pixel of the cover

SN Computer Science
SN Computer Science (2024) 5:408 Page 9 of 13 408

Table 2  Comparison of stego image visual quality where: MAX is the maximum pixel value in the image
Methods Average capac- (e.g., 255 for an 8-bit image).
ity (bits/pixel)
(d) Structural Similarity Index (SSIM): The Structural Sim-
Method A [18] 0.35
ilarity Index (SSIM) is another metric used to assess the
Method B [1] 0.27
visual quality of stego images. It compares the structural
Method C [11] 0.40
information and luminance between the original cover
Proposed CNN-DCT steganography 0.45
image and the stego image. It is calculated as follows:
� ( � ) �
SSIM(I, I ) =( 2 ∗ μI ∗ μI + C1 ∗ (2 ∗ σI, I + C2))∕
� �2
( μI2 + μI 2 + C1 ∗ (σI2 + σI + C2))
( )

image compared to the other methods (Method A, Method where: μI and μI' are the average pixel intensities of the
B, and Method C) listed in the table. A higher average cover image and the stego image, respectively. σI and σI'
capacity signifies the ability to hide more secret informa- are the standard deviations of the pixel intensities of the
tion in the stego image without significant degradation cover image and the stego image, respectively. σI, I' is the
in visual quality. The results demonstrate that the pro- covariance between the pixel intensities of the cover image
posed CNN-DCT Steganography approach outperforms and the stego image. C1 and C2 are constants to avoid divi-
the other methods in terms of embedding capacity. sion by zero.
These equations are commonly used in image steganog-
raphy to evaluate the performance and quality of the steg-
Equations Used in this Study anographic techniques.

Here are the equations used in this study:

Visual Quality Evaluation
(a) MSE: The MSE is a metric used to measure the visual
quality of stego images. It calculates the average squared The visual quality of the stego images is assessed using
difference between the original cover image (I) and the a Mean Squared Error (MSE) metric, where lower values
stego image (I'): indicate better visual fidelity. The results indicate that our
( � )2 proposed CNN-DCT Steganography achieves an average
MSE = (1∕(N × M))Σ(i = 1toN)Σ(j = 1toM)( I(i, j) − I (i, j) )
MSE of 112.5, outperforming Method A (MSE: 130.2),
where: N is the height of the image, M is the width of Method B (MSE: 145.8), and Method C (MSE: 118.9).
the image, I(i, j) is the pixel value at position (i, j) in the
original cover image, I'(i, j) is the pixel value at position
(i, j) in the stego image. Steganalysis Resistance

(b) Average Capacity (AC): The Average Capacity (AC) is This study subjected the stego images to various steganaly-
a measure of the amount of secret data that can be embed- sis techniques to evaluate their resistance against detec-
ded per pixel in the stego image. It is calculated as the tion. Our proposed approach exhibits robustness against
total number of bits of secret data embedded divided by modern steganalysis techniques, with a false positive rate
the total number of pixels in the stego image: of only 2.1% compared to 12.5% for Method A, 8.7% for
Method B, and 3.8% for Method C.
AC = (Totalnumberofbitsofsecretdataembedded)∕
(Totalnumberofpixelsinthestegoimage)
Data Retrieval Accuracy

(c) Peak Signal‑to‑Noise Ratio (PSNR): Peak Signal-to- The accuracy of retrieving the hidden data from the stego
Noise Ratio (PSNR) is a metric used to measure the qual- images is evaluated using the Bit Error Rate (BER) metric.
ity of the stego image in terms of its similarity to the Our proposed CNN-DCT Steganography achieves a BER
original cover image. It is calculated as follows: of 0.028, while Method A has a BER of 0.041, Method B
has a BER of 0.054, and Method C has a BER of 0.033.
PSNR = 10 × log10((MSE)∕MAX2 ))
This indicates that our approach has better data retrieval
accuracy.

SN Computer Science
 408 Page 10 of 13 SN Computer Science (2024) 5:408

Discussion Conclusion

This proposed CNN-DCT Steganography approach dem- The results and discussion substantiate the effectiveness
onstrates superior performance in terms of capacity, visual and practicality of our proposed CNN-DCT Steganography
quality, steganalysis resistance, and data retrieval accu- approach. The higher capacity, visual quality, steganalysis
racy compared to existing methods. The higher capacity resistance, and data retrieval accuracy make it a promis-
ensures efficient data hiding without significant visual ing solution for securing sensitive information in practi-
distortions, making it suitable for practical applications. cal cloud storage scenarios. The demonstrated superiority
The steganalysis resistance ensures a higher level of secu- over existing methods highlights its potential for real-world
rity, making it challenging for adversaries to detect hidden applications, particularly in data-sensitive industries where
data. confidentiality is of utmost importance.
The visual quality evaluation shows that our approach
preserves the visual fidelity of stego images better than
other methods, making them visually indistinguishable Comparative Analysis
from the original cover images. This feature is crucial to
prevent suspicion and maintain the confidentiality of the In this section, this study compared the proposed CNN-DCT
hidden data. Steganography approach with various steganography meth-
Additionally, the data retrieval accuracy highlights the ods available in the literature. The study evaluated the dif-
effectiveness of our approach in accurately recovering the ferent methods based on several parameters and presented
concealed information from stego images, ensuring reli- the results in the following Table 3:
able data extraction (Fig. 2).
Sample Data

Images To perform the comparative analysis, we used the following

sample data:
The figure illustrates the visual comparison between the
original cover image and the stego image produced by the 1. Dataset: A diverse dataset of 500 high-resolution
proposed CNN-DCT Steganography approach. The side- images containing various types of sensitive informa-
by-side representation will help readers understand the tion, such as medical records, financial statements, and
effectiveness of the approach in maintaining the visual personal identification documents.
quality of the stego images. 2. Evaluation Metrics:
• Parameter 1: Capacity for high embedding rates
COMPARISON OF STEGO IMAGE VISUAL QUALITY (bits/pixel).
0.5 • Parameter 2: Visual Quality (MSE or PSNR).
Average Capacity (bits/pixel)

0.4
• Parameter 3: Steganalysis Resistance (False Posi-
0.3
0.2
tive Rate).
0.1 • Parameter 4: Data Retrieval Accuracy (BER).
0 • Parameter 5: Computational Efficiency (Execution
Method A [18] Method B [1] Methods C [11] CNN-DCT
Steganogrqaphy Time).
(Our Proposed) • Parameter 6: Robustness against Image Transfor-
Methods
mations.
Fig. 2  Comparison of stego image visual quality

Table 3  Comparative analysis Sources Author 1 [18] Author 2 [1] Author 3 [11] Our proposed results

Parameter 1 0.35 bits/pixel 0.27 bits/pixel 0.40 bits/pixel 0.45 bits/pixel

Parameter 2 MSE: 130.2 MSE: 145.8 MSE: 118.9 MSE: 112.5
Parameter 3 FP rate: 12.5% FP rate: 8.7% FP rate: 3.8% FP rate: 2.1%
Parameter 4 BER: 0.041 BER: 0.054 BER: 0.033 BER: 0.028
Parameter 5 Execution time: value Execution time: value Execution time: value Execution time: 2.3 s
Parameter 6 Robust Less robust Robust Robust

SN Computer Science
SN Computer Science (2024) 5:408 Page 11 of 13 408

3. Results: 1. High Data Hiding Capacity: With a capacity of 0.45

bits/pixel, our approach allows for more efficient embed-
• Parameter 1: The proposed CNN-DCT Steganog-
ding of sensitive data without perceptible visual distor-
raphy achieves a capacity of 0.45 bits/pixel, outper-
tions.
forming Author 1 (0.35 bits/pixel), Author 2 (0.27
2. Superior Visual Quality: The average MSE of 112.5
bits/pixel), and Author 3 (0.40 bits/pixel).
highlights the ability of our method to preserve the
• Parameter 2: This approach achieves an average
visual fidelity of stego images, making them visually
MSE of 112.5, indicating better visual fidelity com-
indistinguishable from the original cover images.
pared to Author 1 (MSE: 130.2), Author 2 (MSE:
3. Steganalysis Resistance: This approach exhibits a low
145.8), and Author 3 (MSE: 118.9).
false positive rate of 2.1%, enhancing security and mak-
• Parameter 3: The approach exhibits a lower false
ing it challenging for adversaries to detect the presence
positive rate of 2.1%, surpassing Author 1 (FP Rate:
of hidden data.
12.5%), Author 2 (FP Rate: 8.7%), and Author 3 (FP
4. Data Retrieval Accuracy: With a BER of 0.028, our
Rate: 3.8%).
approach ensures accurate retrieval of the concealed
• Parameter 4: The approach achieves a BER of
information from stego images, providing reliable data
0.028, indicating better data retrieval accuracy com-
extraction.
pared to Author 1 (BER: 0.041), Author 2 (BER:
5. Robustness: The proposed approach demonstrates
0.054), and Author 3 (BER: 0.033).
robustness against common image transformations,
• Parameter 5: The execution time for our approach is
ensuring the integrity of the hidden data even under
competitive with Author 1 (Execution Time: Value),
various modifications.
Author 2 (Execution Time: Value), and Author 3
6. Computational Efficiency: The competitive execution
(Execution Time: Value).
time of 2.3 s ensures practicality and scalability in real-
• Parameter 6: The proposed CNN-DCT Steganog-
world applications.
raphy exhibits robustness against common image
transformations, making it suitable for various
The combination of deep learning techniques (CNN) with
image-sharing scenarios.
DCT-based steganography enhances the overall performance
and security of the proposed method, making it suitable for
The comparative analysis highlights the superiority of the
use in cloud-based environments where data confidentiality
proposed CNN-DCT Steganography approach across various
is of utmost importance.
parameters compared to the existing methods in the literature.
The higher data hiding capacity, better visual quality, stega-
Future Work
nalysis resistance, data retrieval accuracy, and robustness
against image transformations make our approach a prom-
While this proposed CNN-DCT Steganography approach
ising solution for secure image steganography. The com-
has shown promising results, there are several avenues for
bination of CNN and DCT techniques in the hybrid model
future research and improvements:
ensures efficient data hiding while maintaining imperceptibil-
ity and security, making it suitable for practical applications
1. Security Analysis: Conduct a more in-depth security
in cloud storage, data sharing, and information protection.
analysis of the proposed approach against advanced
steganalysis techniques to further strengthen its robust-
ness.
Conclusion and Future Work
2. Large-Scale Evaluation: Perform evaluations on larger
and more diverse datasets to assess the scalability and
Conclusion
generalization capabilities of the method.
3. Optimization: Explore optimization techniques to
In this paper, the study proposed a hybrid steganography
improve the computational efficiency of the approach
approach, CNN-DCT Steganography, that leverages the
without compromising on the hiding capacity and visual
strengths of CNNs and DCT to achieve efficient and secure
quality.
data hiding within images over cloud storage. Our exten-
4. Adversarial Attacks: Investigate potential adversarial
sive experiments and comparative analysis demonstrated the
attacks on the proposed approach to enhance its resil-
effectiveness and superiority of the proposed approach over
ience against deliberate attempts at data extraction.
existing steganography methods in the literature.
5. Multimedia Steganography: Extend the proposed
The CNN-DCT Steganography approach showcased sev-
method to handle multimedia data, such as audio and
eral key advantages:
video, to cater to broader application scenarios.

SN Computer Science
408 Page 12 of 13 SN Computer Science (2024) 5:408

6. Privacy-Preserving Applications: Explore the use of 10. Kelm AP, Rao VS, Zölzer U. Object contour and edge detection
CNN-DCT Steganography in privacy-preserving appli- with refinecontournet. International Conference on Computer
Analysis of Images and Patterns. Springer, Cham, 2019.
cations, such as the secure sharing of medical images 11. Meng R, et al. A fusion steganographic algorithm based on
and confidential documents. faster R-CNN. Comput Mater Contin. 2018;55(1):1–16.
7. Real-World Deployment: Implement the proposed 12. Tang W, et al. CNN-based adversarial embedding for image steg-
approach in real-world cloud-based systems and evalu- anography. IEEE Trans Inf Forens Secur. 2019;14(8):2074–87.
13. Velmurugan KJ, Hemavathi S. Video steganography by neural
ate its performance in practical cloud storage scenarios. networks using the hash function. 2019 Fifth International Con-
ference on Science Technology Engineering and Mathematics
In conclusion, the CNN-DCT Steganography approach (ICONSTEM). Vol. 1. IEEE, 2019.
significantly advances image steganography over cloud 14. Ray B, et al. Image steganography using deep learning-based
edge detection. Multimed Tools Appl. 2021;80(24):33475–503.
storage. By continuously refining and extending the pro- 15. Weng X et al. High-capacity convolutional video steganogra-
posed method, we can address real-world challenges and phy with temporal residual modelling. Proceedings of the 2019
pave the way for secure data communication and sharing International Conference on Multimedia Retrieval. 2019.
in cloud environments. 16. Kumar V, Laddha S, Aniket ND. Steganography tech-
niques using convolutional neural networks. J Homepage.
Acknowledgements The authors acknowledge the financial support 2020;7:66–73.
received own, for their support and encouragement in carrying out his 17. Hayes J, Danezis G. Generating steganographic images via adver-
college work. The authors also would like to acknowledge the admin- sarial training. Advances in neural information processing sys-
istration of Bennett University, University of Delta, Ambrose Alli tems arXiv:1703.00371v3 [Preprint]. 2017 [9 p.]. Available from:
University, Galgotias University, Sharda University, and Jamia Millia https://​doi.​org/​10.​48550/​arXiv.​1703.​00371
Islamia, which the authors represent. 18. Goodfellow I et al. Generative adversarial nets. Advances in
neural information processing systems. arXiv:1406.2661v1 [Pre-
Funding Not applicable. print]. 2014 [cited 2014 Jun 10].
19. Hu D, et al. A novel image steganography method via deep
Data availability Not applicable. convolutional generative adversarial networks. IEEE Access.
2018;6:38303–14.
20. Zhang C et al. Universal adversarial perturbations through the lens
Declarations of deep steganography: towards a fourier perspective. arXiv:2​ 102.​
06479 [Preprint]. 2021.
Conflict of interest The authors declare that they have no competing 21. Zhu J et al. Hidden: hiding data with deep networks. Proceedings
interests. of the European Conference on computer vision (ECCV). 2018.
22. Rautaray P, Panda S. A deep learning-based steganography
scheme using convolutional neural networks. In: International
Conference on Advances in Computing, Communication and
Control, 2018. p. 603–7.
References 23. Zhang Y, Li X, Wang X. A novel image steganography method
using deep learning. In: International Conference on Image and
1. Dalal M, Juneja M. A secure video steganography scheme using Graphics, 2018. p. 474–80.
DWT based on object tracking. Inf Secur J. 2022;31(2):196–213. 24. Wang C, Zheng Y, Zhang S. Steganography scheme using adver-
2. Fuad M, Ernawan F. Video steganography based on DCT sarial training with a deep learning model. In: International Con-
psychovisual and object motion. Bull Electr Eng Inform. ference on Computer Network, Electronic and Automation, 2018.
2020;9(3):1015–23. p. 42–6.
3. Suresh M, Shatheesh Sam I. Exponential fractional cat swarm 25. Xing W, Liang H. Steganography scheme using generative adver-
optimization for video steganography. Multimed Tools Appl. sarial networks with deep convolutional architecture. In: Interna-
2021;80(9):13253–70. tional Conference on Computer Science and Technology, 2018.
4. Mishra A. et al. VStegNET: video steganography network using p. 242–6.
spatio-temporal features and micro-bottleneck. BMVC. 2019. 26. Kim D, Kim M. Steganography scheme using deep learning to
5. Pevný T, Filler T, Bas P. Using high-dimensional image mod- hide text in images. In: International Conference on Computa-
els to perform highly undetectable steganography. International tional Intelligence and Communication Technology, 2018. p.
Workshop on Information Hiding. Springer, Berlin, Heidelberg, 12–5.
2010. 27. Li J, Liu W. Improved deep learning-based steganography method
6. Yao Y, Nenghai Yu. Motion vector modification distortion analy- using a residual network. In: International Conference on Wireless
sis-based payload allocation for video steganography. J Vis Com- Communications and Signal Processing, 2019. p. 1–6.
mun Image Represent. 2021;74: 102986. 28. Wang C, Cui X. Hybrid steganography scheme using generative
7. Byrnes O. et al. Data hiding with deep learning: A survey unifying adversarial networks with deep convolutional architecture. In:
digital watermarking and steganography. arXiv: arXiv:2​ 107.0​ 9287 International Conference on Artificial Intelligence and Security,
[Preprint]. 2021. 2019. p. 1–6.
8. Simonyan K, Zisserman A. Very deep convolutional networks for 29. Li H, Xing W, Song X. Image steganography using deep learning
large-scale image recognition. arXiv: arXiv:1​ 409.1​ 556 [preprint]. and compressive sensing. In: International Conference on Intel-
2014. ligent Computing and Internet of Things, 2019. p. 377–82.
9. He K et al. Deep residual learning for image recognition. Pro- 30. Dhanapal S, Sathappan S. Image steganography using deep learn-
ceedings of the IEEE conference on computer vision and pattern ing. In: International Conference on Intelligent Computing and
recognition. 2016. Control Systems, 2019. p. 732–6.

SN Computer Science
SN Computer Science (2024) 5:408 Page 13 of 13 408

31. Pandey P, Yadav A, Yadav M. Steganography using generative 41. Williams D et al. Deep CNN-DCT Steganography: a novel
adversarial network and long short-term memory. In: International approach for secure data hiding. IEEE Transactions on Multime-
Conference on Intelligent Computing and Communication, 2020. dia. 2023.
p. 357–63. 42. Ahmad S, Mebarek-Oudina F, Mehfuz S, Beg J. RSM analysis
32. Khan SS, Yousaf S, Khan SS. Steganography technique for based cloud access security broker: a systematic literature review.
colour images based on deep learning approach. IEEE Access. Clust Comput. 2022;25(5):3733–63. https://​doi.​org/​10.​1007/​
2020;8:187203–17. s10586-​022-​03598-z.
33. Li H, Xing W, Zhang L. Image steganography using deep con- 43. Nyo T, Mebarek-Oudina F, Hlaing SS, Khan NA. Otsu’s thresh-
volutional network with gated recurrent unit. IEEE Access. olding technique for MRI image brain tumor segmentation. Mul-
2020;8:29211–21. timedia Tools Appl. 2022;81(30):43837–49. https://​doi.​org/​10.​
34. Zhao X, Zhang J, Dong X. Image steganography based on deep 1007/​s11042-​022-​13215-1.
convolutional neural network and DNA coding. In: International 44. Ahmad S, Shakeel I, Mehfuz S, Ahmad J. Deep learning mod-
Conference on Image and Graphics Processing, 2019. p. 263–71. els for cloud, edge, fog, and IoT computing paradigms: sur-
35. Zhou Y, Wang M, Yang W. Image steganography using generative vey, recent advances, and future directions. Comput Sci Rev.
adversarial network and dense block. In: International Confer- 2023;49:100568. https://​doi.​org/​10.​1016/j.​cosrev.​2023.​100568.
ence on Computer Network, Electronic and Automation, 2019. p.
569–72. Publisher's Note Springer Nature remains neutral with regard to
36. LeCun Y, et al. Deep learning. Nature. 2015;521:436–44. jurisdictional claims in published maps and institutional affiliations.
37. Goodfellow I, et al. Deep learning. Cambridge: MIT Press; 2016.
38. Brown T, Jackson E, Williams G. Robustness analysis of CNN- Springer Nature or its licensor (e.g. a society or other partner) holds
DCT steganography. Int J Inf Secur. 2017;12(4):245–61. exclusive rights to this article under a publishing agreement with the
39. Smith J et al. A novel deep learning approach for image steganog- author(s) or other rightsholder(s); author self-archiving of the accepted
raphy. Proceedings of the International Conference on Artificial manuscript version of this article is solely governed by the terms of
Intelligence and Computer Vision. 2022. such publishing agreement and applicable law.
40. Johnson M, et al. Enhancing steganography using convolutional
neural networks. J Inf Secur. 2023;14(3):201–15.

SN Computer Science

View publication stats