International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

IMAGE STEGANOGRAPHY USING CNN

Shourya Chambial1, Dhruv Sood2
1School of Information Technology and Engineering, Vellore Institute of Technology, Vellore, 632014, India
2School of Information Technology and Engineering, Vellore Institute of Technology, Vellore, 632014, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The unfamiliarity and curiosity in the subject
were the key reasons for choosing steganography from the
list of available project ideas. Another argument is that it
ensures data security. Even if a hacker (or intruder) has
access to our image files, they will be unable to access our
information; that is, even if the intruder manages to hack
and gain access to our system, they will be unable to access
our data. This feature adds a lot to the appeal. The project is
Fig -1: Overview
about utilising images to learn about text stenography. For
images, image steganography is used, and the relevant data The first image in the above example is the secret image
is decrypted in order to get the message image. Image that must be kept concealed, while the second image is the
steganography is researched, and one of the methods is used cover image that will be used to conceal the secret image.
to demonstrate it, because it may be done in a variety of The third image, which is essentially a repainted cover
ways. Steganography is the practise of concealing data, such image, is the result of the hiding procedure. Some of the
as text, photos, or audio files, within another image or video most common steganographic approaches involve the
file. In a word, steganography's fundamental goal is to manipulation of the least significant bits (LSB) of the
conceal the desired information within any image, audio, or images to place the secret information, which can be
video that does not appear to be hidden just by looking at it. achieved adaptively or even uniformly through simple
Image-based Steganography is based on a basic concept. replacement or through methods or through methods that
Images are made up of digital data (pixels) that define are more advanced. In our project we will be using
what's inside the image, which is usually the colours of all convolutional neural networks (CNN) for hiding an image
the pixels. Since we know that every image is made up of inside another image. The advantage is a more efficient
pixels, each of which has three values (red, green, blue). We steganography.
anticipate that the image we want to conceal will be totally
absorbed by the cover image, making it invisible to There are disadvantages to this approach also, as on
intruders. enhancing the image by a big factor, one can find out the
original image but it is very hard to do this and to
Key Words: Image Steganography, Steganography, completely reconstruct the original image from the output
Steganography using CNN, Deep learning, GAN, LSB image. We have used the TINY IMAGENET DATABASE.
[Link] 2. Related work
A technique used for hiding secret data within a standard, 1. Summary of various methods
public file or message to avoid detection is known as
Steganography. The secret data is then extracted once it Following a review of all available frameworks, the
arrives at its destination. Steganography, in conjunction methodologies are primarily divided into three categories:
with encryption, can be used to further conceal or traditional image steganography methods, CNN-based
safeguard data. image steganography methods, and GAN-based image
steganography methods. Traditional methods are
Using deep convolutional neural networks, a full-sized frameworks that employ methods unrelated to machine
colour image is concealed inside another image (called learning or deep learning algorithms. The LSB technique is
Cover image) with minimum changes in appearance. The used in many traditional methods. CNN-based methods
hidden image will then be revealed by combining a use deep convolutional neural networks to embed and
"reveal" network with the created image. extract secret messages, whereas GAN-based methods use
some GAN variants.

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 755
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

from the secret image. The substitute approach must be

used with caution, as overloading the cover picture may
result in noticeable alterations that reveal the secret
information's presence.

Table -1: Traditional Methods

Dataset Metrics Advantag Disadvantag

es es
Fig -2: Classification of current methods on basis of the
method and media use Lena PSNR Time to It is insecure.
compute is
2.2 Review on various schemes reduced.
The image
Image steganography is the process of concealing is a coded
information in a cover image, which might be text, image, message in
and
or video. The secret information is concealed in such a way of itself.
that it cannot be seen by human sight. Deep learning
technology, which has recently gained popularity as a
strong tool in a variety of applications, including image
Lena and PSNR and Time to When
steganography, has gotten a lot of attention. The primary compute is compared to
goal of this study is to analyze and give an explanation for Baboon MSE
reduced. deep
the numerous deep studying algorithms to be had inside The image learning
the field of photograph steganography which we came is a coded approaches,
across in our referred papers. Traditional methods, message in security is a
and concern.
Convolutional Neural Network-based methods, and of itself.
General Adversarial Network-based methods are the three Any image
types of deep learning approaches utilized for image format
steganography. This study includes a detailed explanation is
of the datasets used, experimental set-ups explored, and acceptable.
commonly used assessment measures, in addition to the
methodology. For simple reference, a table summarizing
all of the details is also supplied. 1 PSNR and Time It is insecure.
RGB Time to Text is used
compute is to
Image reduced. It communicat
embeds e
data with secret
ease. information.
Steganogra
phy and
steganalysi
s are not
reliant on
one other.

Fig -3: General principle of steganography

2.4 GAN–BASED methods used in Steganography
2.3 Steganography techniques that are based on
traditional methods General Adversarial Networks (GAN) are a type of deep
CNN. For image generation tasks, a GAN employs game
Image steganography is traditionally done using the Least
theory to train a generative model with an adversarial
Significant Bits (LSB) substitution method. The pixel
process. In GAN architecture, two networks – generator
quality of images is normally higher, but not all of the
and discriminator networks – compete to generate a
pixels are used. The LSB approach is based on the notion
perfect image. The data is fed into the generator model,
that changing a few pixel values will not result in
and the output is a close approximation of the given input
noticeable changes. The secret data is transformed into a
image. The discriminator networks classify the generated
binary format. The least significant bits in the noisy area
images as either false or true. The two networks are
are determined by scanning the cover image. The LSBs of
trained in such a way that the generator model attempts to
the cover picture are then replaced with the binary bits

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 756
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

imitate the input data as closely as possible with as little computational

noise as possible. cost
and overhead.

Image steganography existing methods using a GAN

GAN CelebA Unlimited Generator and
architecture can be categorized into five types: a three- number discriminator
network-based GAN model, cycle-GAN-based of cover
are shared.
images are
architectures, sender-receiver GAN architectures, be Images
coverless models in which the cover image is generated generated that
have
randomly rather than being given as input, and an Alice,
been
Bob, and Eve based model. generated are
not natural.
DCGAN CelebA Image creation It
that is more provides
realistic probabilistic
Extremely safe information
rather
than secret
information.
There is no
information on
how to obtain
the secret data.

2.5 Steganography techniques that are based on CNN

The encoder-decoder architecture is greatly influenced by

Fig -4: GAN Overview steganography employing CNN models. The encoder
receives two inputs: the cover picture and the secret
A GAN model is made up of two main components: the image, which are used to construct the stego image, and
generator and the discriminator. In some of the methods the decoder receives the stego image, which outputs the
for image steganography, a new network called the embedded secret image. The core principle remains the
steganalyzer is introduced. These three components' same, however different techniques have experimented
primary functions are as follows: with various structures. Different algorithms change in
how the input cover picture and the secret image are
•A model, G, for generating stego images from the cover concatenated, while variations in the convolutional layer
image and the random message. and pooling layer are to be expected. The number of filters
used, the strides taken, the filter size utilised, the
•A discriminator model, D, to determine whether the
activation function used, and the loss function used differ
generated image from the generator is real or fake. from one approach to the next. One thing to keep in mind
•A steganalyzer, S, to determine whether or not the input is that the cover image and the secret image must be the
same size, so that every pixel of the secret image is
image contains confidential secret data.
dispersed throughout the cover image.
Table -2: GAN Based Methods
Convolution is a type of linear operation that expresses
the degree of overlap between two functions when they
Architecture Dataset Advantages Disadvantage
are shifted over each other. Convolutional networks are
Alice, Bob, Eve BOSSbase and Embedding Messages are simple neural networks with at least one layer that uses
celebA does used instead of convolution instead of ordinary matrix multiplication. The
not necessitate
domain images. following are some of the benefits of using a CNN-based
knowledge. Grid search for architecture for encoding and decoding:
embedding
scheme
selection
● CNN extracts visual features automatically.
consumes
time. ● CNN essentially down samples the image using
DCGAN nearby pixel information, first by convolution, and
celebA Game- Steganalyzer is
and BOSSbase theoretic used then by using a prediction layer at the end.
formulation isto
used provide ● CNN is more accurate and performs effectively.
the
probability by One can get a decent idea of the patterns of natural images
itself,
which by using a deep neural network, in this case a CNN. The
lead
in
© 2022, IRJET | Impact Factor value: 7.529
additional | ISO 9001:2008 Certified Journal | Page 757
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

network will be able to determine which areas are U-Net ImageNet The However,
redundant, allowing additional pixels to be buried in image is a the image is
certain areas. The amount of hidden data can be raised by coded only 64x64
saving space on superfluous areas. Because the structure message in pixels
and of itself. in size,
and weights can be randomized, the network will conceal The which is
data that is inaccessible to anyone who does not have the simplest quite little.
weights. and The images
most basic in the input
Table -3: CNN Based Methods architecture are simply
is chosen. concatenate
d.
Architectu Dataset Advantage Disadvanta
re s ges

Encoder- ImageNet Highly The loss Encoder- ImageNet The image However,
decoder secure and used isn't decoder is a coded the image is
with SCR dependable message in only 64x64
ideal. and of itself. pixels
In black or in size,
white areas, which is
visible noise quite little.
might be
seen.

Encoder- COCO It is High

decoder and not number 3. Background
with necessary of images
VGG- base [Link] to are 1. General Architecture on Topic chosen
have prior required in
domain computatio The entire system consists of 3 CNNs:
knowledge. n purpose.
Since
the created [Link] Preparation Network: This network prepares the
image secret image to be hidden the purpose is to transform the
is unrelated color-based pixels to more useful features so that in can be
to the encoded such as edges. It contains 50 filters of (3X3, 4X4,
secret 5X5) patches there are total 6 layers of this kind.
information
, it
is extremely [Link] Hiding Network: This network will take the output
secure. of the preparation network and will then create a
Container Image. It is a CNN with 5 convolutional layers
that have 50 filters of (3X3, 4X4, 5X5) patches. This layer
consists of total of 15 convolutional layers in this network.
CNN ImageNet The However,
image is a the image is [Link] Reveal Network: Converts the Container image to the
and Holiday
coded only 64x64 original image this network is used for decoding. Which
message in pixels has 5 convolutional layers that have 50 filters of (3X3, 4X4,
and of itself. in size, 5X5) patches. This layer consists of total of 15
The which is
simplest quite little. convolutional layers in this network.
and The images
most basic in the input
architecture
is chosen. are
To speed up simply
training, a concatenate
new error d.
back
propagation
function
is
implemente
d.
Fig -5: Encoder and Decoder Overview

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 758
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

Some small amount of noise is also added to the images 4. Proposed algorithm
that are the output of the second network in order to make
sure that the system does not simply encode the secret The system that we have implemented in this project uses
image in the LSB. a CNN to hide an image inside another image hence
making the secrete image effectively invisible to the
We want to do image steganography, which involves observer. The neural networks that we have used in our
hiding an image inside one cover image. The concealed work determines where in the cover image it can hide the
images included in the cover image must be retrievable information and hence, is a more efficient process than
with minimal loss. The encoded cover image must LSB manipulation. We have also trained a decoder
resemble the original cover image in appearance. network to reveal the secret image from the container
image. This process is done in such a way that the
We transfer secret images over the prep network, then container image does not change much and the changes to
concatenate the resulting data with the carrier image the container image are not discernible.
before sending it over the Hiding network. The idea of
having decoders, one for each secret picture, to obtain the It can be safely assumed that the intruder does not have
secret image from the container image is then access to the original image. The secret image is effectively
implemented. hidden in all 3 colour channels of the container image.

To make our image retrieval model more secure, we 3 different activation functions were used to test the
expand on the concept of putting instead, a secret image network along with various learning rates.
with noise in the original cover image placing the secret
photos on the original cover's LSBs image. The activation functions used were:

Following is a brief overview of the encoder/decoder [Link] Linear Unit (RELU) {de-facto for most CNN
framework used in our technique: networks}

ENCODER: A prep network that corresponds to secret 2. Tanh

image input. The outputs of the prep network are
combined with the cover picture and passed via the Hiding 3. Scaled Exponential Linear Unit (SELU) {has been used to
network. avoid the vanishing gradient problem caused by RELU
activation function}
DECODER: The decoder network is made up of reveal
networks, each of which has been taught to decode its own Advantages: -
message.
1. CNNs learn the most efficient way to hide the secret
The following is the typical framework: image inside the container image which is unpredictable
for us. As a result, it adds to the security of steganography
Prep Networks: Each prep network is made up of two by making the process unpredictable.
layers stacked on top of each other. Each layer is made up
of three independent Conv2D layers. These three Conv2D [Link] container image is note altered a lot which makes
layers have 50, 10, and 5 channels, respectively, with the changes harder to spot.
kernel sizes of 3, 4, and 5 for each layer. Along both axes,
the stride length remains constant at one. To preserve the [Link] changes can only be spotted by the decoder network
output image in the same dimensions, appropriate that is specifically trained to do it.
padding is provided to each Conv2D layer. After each
4. The process is flexible that means that before feeding to
Conv2d layer, a Relu activation is applied.
network, some changes can be made to the image to make
Concealing Network: The concealing network is a three- the process more secure.
layer aggregation. Each of these layers is made up of three
Disadvantages: -
individual Conv2D layers. The Conv2D layers in the hidden
network have a similar basic structure to the Conv2D
1. This is not a full proof method and the changes can be
levels in the Prep Network.
spotted by a specially trained network and the original
image can be somewhat retrieved.
Reveal Network: The reveal network has a similar basic
design to the hidden network, with three levels of Conv2D We have chosen this method because the advantages
layers that are comparable in shape. clearly outweigh the disadvantages.

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 759
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

4.1 Our dataset

We will be using the Tiny ImageNet which contains

100000 images divided into 200 classes (500 images per
class) and downsized to 64 x 64 coloured images. It
consist of 500 training images, 50 validation images, and
50 test images in each class.

Fig -6: Architecture Diagram

Fig -7: Flow Chart

5. Results

ImageNet is a massive dataset that contains images from

the WordNet hierarchy, with each node holding hundreds
of images. The image on ImageNet has no copyrights and
just contains links or thumbnails to the original image.
Images of various sizes make up the dataset. The number
of photographs, the classes they belong to, the
background, and the image size can all be customized
based on the requirements.

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 760
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

ImageNet's Tiny ImageNet dataset is a subset of the larger PSNR 56.95 40.66 62.53 76.214
ImageNet dataset. The dataset consists of 100,000 photos
divided into 200 classes (500 images per class) that have SSIM - 0.964 - 0.965
been reduced to 6464 coloured images. There are 500
training images, 50 validation images, and 50 test
photographs in each class.
6. Conclusion and future work
The quality of image steganography is assessed using a
variety of approaches. Each of these methods evaluates a ● The network was constructed and is performing
different component of the steganographic outcome. Mean well. Even though it has shown good performance
Square Error (MSE), Peak Signal to Noise Ratio (PSNR), encoding and decoding, it is not a full proof
and Structured Similarity Index Measure (SSIM) are some system as every technology has its shortcomings.
of the well-known approaches. ● As stated in the disadvantages, there is a lot of
Mean Square Error (MSE):- scope of improvement in this project as there is a
chance that networks can be trained specifically
The averaged value of the square of the pixel-by-pixel for detecting that an image has been hidden inside
difference between the original image and the stego-image a given image.
is the Mean Square Error. It provides a measure of the
cover image error caused by the data embedding ● This project can be improved upon by creating a
procedure. A lower MSE value denotes a high-quality better architecture or by altering the architecture
embedding. of this network and improving its performance
further, making the secret image tougher to
Peak Signal to Noise Ratio (PSNR):- decode by any program other than the decoder
and to make it more difficult to detect the
PSNR is another useful metric for determining the degree presence of any secret image inside the encoded
of embedding-induced distortion in the cover image. It's image.
the ratio of a signal's greatest possible value to the power
of distortion noise (MSE). It is expressed in decibels (dB). REFERENCES
A higher PSNR value denotes a higher-quality embedding.
[1] H. Kato, K. Osuge, S. Haruta and I. Sasase, "A
Structured Similarity Index Measure (SSIM):- Preprocessing by Using Multiple Steganography for
Intentional Image Downsampling on CNN-Based
SSIM is a comparison statistic used to see how similar the Steganalysis," in IEEE Access, vol. 8, pp. 195578-
cover picture and the stego-image are. The perceived 195593, 2020, doi: 10.1109/ACCESS.2020.3033814.
difference between the two images is measured. We were
able to achieve MSE value of 0.006259255611669444, [2] Jin, Z., Yang, Y., Chen, Y. and Chen, Y., 2020. IAS-CNN:
PSNR value of 76.21412244818211 dB, and SSIM value of Image adaptive steganalysis via convolutional neural
0.96512678174912 network combined with selection channel.
International Journal of Distributed Sensor Networks,
In this section the results of the proposed methodology 16(3), p.1550147720911002.
are compared with the existing.
3 Xiang, Z., Sang, J., Zhang, Q., Cai, B., Xia, X. and Wu, W.,
Table -3: Comparative Study 2020. A new convolutional neural network-based
steganalysis method for content-adaptive image
Parameter [31] [32] [33] Proposed steganography in the spatial domain. IEEE Access, 8,
Algorith pp.47013-47020.
m
4 Duan, X., Liu, N., Gou, M., Wang, W. and Qin, C., 2020.
Dataset Lena and ImageNet 1 Tiny SteganoCNN: Image Steganography with
Baboon RGB ImageNet Generalization Ability Based on Convolutional Neural
image Network. Entropy, 22(10), p.1140.
Method GAN GAN LSB CNN
used
5 Duan, X., Guo, D., Liu, N., Li, B., Gou, M. and Qin, C.,
2020. A new high capacity image steganography
Architectur Tradition U NET Tradition Encoding method combined with image elliptic curve
e al Method al Method and cryptography and deep neural network. IEEE Access,
decoding 8, pp.25777-25788.

MSE 0.13 - - 0.006

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 761
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

6 Li, Q., Wang, X., Wang, X., Ma, B., Wang, C., Xian, Y. and Communication in Untrusted Channels (Master's
Shi, Y., 2020. A novel grayscale image steganography thesis).
scheme based on chaos encryption and generative
adversarial networks. IEEE Access, 8, pp.168166- [18] Lu, S. P., Wang, R., Zhong, T., & Rosin, P. L. (2021).
Large-
168176. Capacity Image Steganography Based on Invertible
Neural Networks. In Proceedings of the
7 Tang, W., Li, B., Tan, S., Barni, M. and Huang, J., 2019. IEEE/Computer Vision and Pattern Recognition (pp.
CNN-based adversarial embedding for image 10816-10825).
steganography. IEEE Transactions on Information
Forensics and Security, 14(8), pp.2074-2087. [19] Ruiz, H., Chaumont, M., Yedroudj, M., Amara, A. O.,
Comby, F., & Subsol, G. (2021, January). Analysis of the
[8] Yu, X., Tan, H., Liang, H., Li, C.T. and Liao, G., 2018, scalability of a deep-learning network for
December. A multi-task learning CNN for image steganography “Into the Wild”.Springer, Cham.
steganalysis. In 2018 IEEE International Workshop on
information forensics and security (WIFS) (pp. 1-7). 20 Byrnes, O., La, W., Wang, H., Ma, C., Xue, M., & Wu, Q.
IEEE. (2021). Data Hiding with Deep Learning: A Survey
Unifying Digital Watermarking and Steganography.
[9] Yuan, Y., Lu, W., Feng, B. and Weng, J., 2017, June. arXiv preprint arXiv:2107.09287.
Steganalysis with CNN using multi-channels filtered
residuals. In International Conference on Cloud 21 Zhangjie, F., Enlu, L., Xu, C., Yongfeng, H., & Yuting, H.
Computing and Security (pp. 110-120). Springer, (2021). Recent Advances in Image Steganography
Cham. Based on Deep Learning. Journal of Computer
Research and Development, 58(3), 548.
10 Duan, X., Guo, D., Liu, N., Li, B., Gou, M. and Qin, C.,
2020. A new high-capacity image steganography 22 Chang, C. C., Wang, X., Chen, S., Echizen, I., Sanchez, V.,
method combined with image elliptic curve & Li, C. T. (2021). Deep Learning for Reversible
cryptography and deep neural network. IEEE Access, Steganography: Principles and Insights. arXiv preprint
8, pp.25777-25788. arXiv:2106.06924.

11 Pibre, L., Pasquet, J., Ienco, D. and Chaumont, M., 2020. 23 Selvaraj, A., Ezhilarasan, A., Wellington, S. L. J., & Sam,
Deep learning is a good steganalysis tool when A. R. (2021). Digital image steganalysis: A survey on
embedding key is reused for different images, even if paradigm shift from machine learning to deep
there is a cover source mismatch. Electronic Imaging, learning based techniques. IET Image Processing.
2016(8), pp.1-11.
24 Chang, C. C. (2021). Neural Reversible Steganography
12 Kim, J., Park, H. and Park, J.I., 2020. CNN-based image with Long Short-Term Memory. Security and
steganalysis using additional data embedding. Communication Networks, 2021.
Multimedia Tools and Applications, 79(1), pp.1355-
1372. 25 Kweon, H., Park, J., Woo, S., & Cho, D. (2021). Deep
Multi-Image Steganography with Private Keys.
13 Zou, Y., Zhang, G. and Liu, L., 2019. Research on image Electronics, 10(16), 2021.
steganography analysis based on deep learning.
Journal of Visual Communication and Image [26] Bashir, B., & Selwal, A. (2021). Towards Deep Learning-
Representation, 60, pp.266-275. Based Image Steganalysis: Practices and Open
Research Issues. Available at SSRN 3883330.
14 Kweon, H., Park, J., Woo, S. and Cho, D., 2021. Deep
Multi-Image Steganography with Private Keys. 27 Salunkhe, S., & Bhosale, S. Feature Extraction Based
Electronics, 10(16), p.1906. Image Steganalysis Using Deep Learning.

[15] You, W., Zhang, H. and Zhao, X., 2020. A Siamese CNN 28 Das, A., Wahi, J. S., Anand, M., & Rana, Y. (2021). Multi-
for image steganalysis. IEEE Transactions on Image Steganography Using Deep Neural Networks.
Information Forensics and Security, 16, pp.291-306. arXiv preprint arXiv:2101.00350.

[16] Zhang, C., Lin, C., Benz, P., Chen, K., Zhang, W., & 29 Cui, J., Zhang, P., Li, S., Zheng, L., Bao, C., Xia, J., & Li, X.
Kweon, I. S. (2021). A brief survey on deep learning (2021). Multitask Identity-Aware Image
based data hiding, steganography and watermarking. Steganography via Minimax Optimization. arXiv
arXiv preprint arXiv:2103.01607. preprint arXiv:2107.05819.

[17] SUBRAMANIAN, N. (2021). Image Steganography [30] Cherian, R. E. Cryptography and Deep Neural
Using Deep Learning Methods to Detect Covert Network: An Art of Hiding Data in Images.

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 762
 International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 [Link] p-ISSN: 2395-0072

31 A. Arya and S. Soni, ‘‘Performance evaluation of

secrete image steganography techniques using least
significant bit (LSB) method,’’ Int. J. Comput. Sci.
Trends Technol., vol. 6, no. 2, pp. 160–165, 2018.

32 X. Duan, K. Jia, B. Li, D. Guo, E. Zhang, and C. Qin,

‘‘Reversible image steganography scheme based on a
U-Net structure,’’ IEEE Access, vol. 7, pp. 9314–9323,
2019.

33 K. A. Al-Afandy, O. S. Faragallah, A. Elmhalawy, E.-S.-M.

El-Rabaie, and G. M. ElBanby, ‘‘High security data
hiding using image cropping and LSB least significant
bit steganography,’’ in Proc. 4th IEEE Int. Colloq. Inf.
Sci. Technol. (CiSt), Oct. 2016, pp. 400–404.

© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 763