Professional Documents
Culture Documents
Steganalysis Techniques A Systematic Review 2023
Steganalysis Techniques A Systematic Review 2023
V. HemaMalini
Research Scholar, Research Department of Computer Science, SDNB Vaishnav College
for Women, ‘University of Madras’, Chennai, India, hemaveera123@gmail.com
Abstract
World at present is full of data. Data plays a major role in all aspects in development in any field,
future predictions, decision making etc. Securing data is very important not only to avoid breaches
but also to shield the confidential messages. There comes the concept of Steganography.
Steganography is the art of covering data up under an ordinary text or image or video. When
steganography is misused by hiding malware in safe files, there comes the art of detecting
steganography called Steganalysis. Some known methods in Steganalysis are Stego-only, Known
cover, Known message, Chosen stego and lastly Chosen message attacks. This work presents a
survey on recent trends in steganography and use of machine learning algorithms in steganalysis.
Keywords: Steganography, Steganalysis, Cryptography, Machine learning.
1 Introduction message itself. Though cryptography and
steganography sound similar, they are two
Immense data is produced electronically every different concepts with the same goal of
second. It is no surprise that protecting the data protecting the message(Dickson, 2020). In
is very important. Lots of methods evolved in cryptography, the information is transformed to
this regard. Steganography is one such method a cipher text using an encryption key. Then the
in which other than the sender and the intended cipher text is transformed back to the initial
recipient, no one could suspect the existence of message using a decryption key by the receiver.
the message. It is a Greek word meaning The middle men would know that some
“Hidden writing”. Steganography was traced encryption form has been applied. But in
long back in history around 440 BC in Greece steganography, the existence of the message is
when people wrote messages on wood and concealed. Data is not usually altered in
covering them with the cover medium like wax, Steganography. But nowadays, data is first
they passed the message. Invisible inks were encrypted and then concealed for increasing
used by Romans, whose writings were security.(Margie Semilof, 2021). The encoding
deciphered using light or heat. Microdots were system of Steganography is its key feature. In
introduced by Germans. Null ciphers were also the data protection field, Steganography is the
used to hide secret messages. In 1499AD from flourishing concept.
a disguised book named Steganographia, we
could find the visibility of the concept 2 Steganography process
(wikipedia, n.d.). Steganography is a
partner to Cryptography. While cryptography In steganography, use of key is optional but
protects a message, steganography hides the when used, gives more security. There will be
a cover file to shield the secret message. The
244
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
245
Steganalysis Techniques: A Systematic Review
U-net structure as the base, authors (Duan et al., divided as normal or steganographic packets
2019)gave a vital image steganography using machine learning algorithms in(Cho et
procedure. Two networks are included in the al., 2018). Random Forest algorithm is found to
trained neural network. They are hiding and be better than Naïve Bayes and Support Vector
extraction network. In the hiding network, the Machines.
encoder network does the work of encoding
secret image under the cover image. The A publicly available cover data is easily
combination is concatenated into 6-channels accessible to anyone. So a unique cover data
using U-net based CNN which is modified to could avoid this problem. (Naito & Zhao,
accept 256*256 cascaded channels. This is fed 2019)used GAN in this context. Two neural
to the hiding network. In the extraction networks are utilized.
network, decoder network is utilized for A Generator, which creates virtual
extracting secret image from stego image. images for cover data,
There are 6 Convolutional layers with 3 *3
sized Convolutional Kernels. Using image-to- A Discriminator, which checks the
image translation, steganography and naturalness of the virtual images.
steganalysis modules are added in CycleGAN, At the gaining end, the generator opens the
a steganographic algorithm is introduced secret data by decoding the stego image. The
in(Meng et al., 2019) for secure Discriminator discards the non-realistic
communications in IOT. The generated image images. It also checks if the sender is a true
quality is guaranteed. The resisting capacity of party or not. The trained generator and the
the stego-image to steganalysis is improved to discriminator are split between the sender and
some extent. This method proved better to the the receiver.
State-of-art method and is named as S-Cycle
GAN. The stego-images were very near to the Some identity of steganography is left behind
cover images by the facility of Cycle in stego files while embedding a message. This
consistency. is due to some allotted bytes used for the
purpose of copyright safety of information.
Goal of hiding stego message and cheating These bytes are extracted as finger prints by
CNN based steganalyzer is presented in (Tang (Cao et al., 2019). Three fixed positions are
et al., 2019) with a steganographic scheme and considered. They are File header of the image;
adversarial embedding (ADV-EMB). Keeping Sequential LSBs of the image file data and
in mind the back transmitted gradients from the Random places behind Image file tail. This
target CNN steganalyzer, the adversarial method is named as Signature method and is
embedding adjusts the costs of modifications of proved to be effective. Half tone images are
image elements. This shows that the inverse produced using dots of various sizes. A
sign of gradient and modification direction has steganographic scheme is proposed for such
high chance of being same. This scheme is images by (Lu et al., 2021). Employing the
capable to overcome highly powerful CNN- concept of pixel density, authors constructed a
based steganalysis and it reduces the PDHistogram. Density blocks are selected to
performance of steganalyzers. Storage based improve visual Quality by density pair
network steganography is difficult to be selection. A pixel flip strategy is proposed
detected by the traditional tools. Firstly, the based on PMMTM to increase security against
data packets are collected in network traffic. Steganalysis.
Wireshark tool is used to capture network
packets. New features are proposed to detect
secret information hidden. Then the records are
246
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
247
Steganalysis Techniques: A Systematic Review
composing 4 hidden layers of generator and which is hidden. Guidance on embedding data
three hidden layers of a critic. Classifiers for and recovering image is provided by a synthetic
example RF and XGboost were used to detect image formed.
the flow. Accuracy was 0.99 and 0.97
respectively with the classifiers and WGAN To encipher the secret message in
which are trained on same data sets and two Steganography, a cover medium or cover
separate datasets. Aiming to improve the image is used. This cover image is removed by
steganographic capacity, FC-DenceNet is steganalysis algorithms and secret messages
utilized by (Duan, Nao, et al., 2020). The first are read. To solve this problem, coverless
convolution block’s input channels and the last image steganography method is used by (Saad
convolution block’s output channels are reset. et al., 2021) which is highly robust. It is gleaned
After they are concatenated through the hidden from Optical Mark Recognition and RBML
network to obtain the stego image, the secret i.e.. Rule Based Machine Learning. The
image is enclosed in the bearer image. The advantages of this method are:
object tag category is not used and No secret information to be shared
LogSoftmax() function is deleted.
No requirement of database
Due to the drawback of the payload capacity, a
novel deep convolutional neural network is Image need not be searched in database
introduced by(Duan, Wang, et al., 2020)with so it is time saving
steganography volume of 1 byte/pixel. A Embedding capacity is very high
pyramid pool module is added to the enclosing robust. So can resist attacks
and excluding network based on up-sampling
and down sampling structure with all important No expensive or special tools used in
previous features. This ensured huge-capacity OMR
embedding and the visual effect of the image.
Highly secure
During the process of training, this also
minimized the loss function of the framework. SPAR-RL is the new cost learning framework
STNet is a Style Transfer method where the introduced by(Tang et al., 2021).There is an
content is enclosed in style images and art iterative interaction amidst the agent and the
images are created by (Z. Wang, Gao, et al., surrounding. The agent uses the policy network
2019). This help in resisting steganalysis tools. to breakdown the process inside into pixel-wise
The arbitrary size of the stego images produced actions. The environment utilizes its network
is 0.06bit/pixel. Being an Unsupervised end-to- for pixel-wise assignment of rewards to the
end training model, STNet is the style transfer agent.
based first steganographic method. As the stego
images are art images, they are hard to be The main inputs in the work are:
detected. Style- transfer is done by Adaptive The agents learn the embedded policy
Instance Normalization. Arbitrary bits of to maximize rewards.
unrevealed messages can be embedded. The
regular singular method in steganography has The embedded policy reduces the
been revisited and re-energized and brought via image level into pixel-wise actions.
neutralization to modern generation by (Chang, Pixel-wise reward function helps agent
2020). Adversarial learning is introduced to get to know about the embedding policy.
back the regularity of natural images after the
discrimination done. GAN are trained for The security and stability performance proved
prediction of bit planes to carry the information good when compared to the other cost methods.
248
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
(Cao et al., 2019) 500 images Stego tools fingerprint Stego tools which
(Google (Message extraction done embed message in
website) smuggler, for fixed arbitrary position
Image hide, embedding in 3 cannot be applied
Invisible ways
secrets)
(Yang et al., 2020) BOSSBase, U-Net, UT- Three module GAN based method
SZUBase GAN, XU- frameworks are in Steganography
NET proposed. in JPEG domain
can be applied.
249
Steganalysis Techniques: A Systematic Review
(Tang et al., 2021) SZUBase SPAR- More secure and Other forms of
BOSSBase Reinforcement stable. reward functions
Learning should be
investigated;
3.2 Audio steganography updated by (J. Wu et al., 2020). There are two
steps involved. Firstly, training the CNN based
When the secret data is hidden within a cover series of steganalyzers iteratively and secondly,
audio, it is termed as Audio Steganography. stegos are generated according to the already
Multiple steganography detecting methods, trained steganalytic network. An audio
based on Bayesian network code elements, are steganographic framework is proposed by (L.
proposed in low bit rate to flattened speech by Chen et al., 2021)which learns automatically to
(Yang et al., 2019). Code element perspective create steganographic cover audio for inserting
is considered in developing this method, as messages. The framework has three
there exist the spatio-temporal correlations components. They are the generator, the
between them. Using Dirichlet distribution discriminator and the steganalyzer. Here the
prior for learning network parameters and stego audio is created by implanting
Bayesian inference based implementations confidential message into steganographic cover
Steganalysis is taken forward. For digital audio, audio. Perception quality is high. Traditional
a new steganography method based on LSBM algorithm is applied to embed untold
adversarial attacks, initial costs are iteratively message to steganographic cover audio.
Table 2. Audio Steganography Techniques
Title & publication Data set Techniques/ Advantage
Algorithms
(Yang et al., 2019) 700 speech Bayesian network code Low bit rate
segments element, Dirichlet compressed speech
from nternet. distribution
250
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
3.3 Text Steganography the secret binary digit string produced from a
secret, a binary search tree is initialized
Text Steganography is the process of covering depending on the text from internet and
secret information within a text. corresponding text is searched by exploiting
Sending messages through social media has space mapping concept. Hiding secret
taken increasing popularity with the advent of messages into ebooks with Electronic
smart phones and high speed mobile data. PUBlication (EPUB) format is taken as study
During the transfer of text messages through by (D. C. Wu & Su, 2020). By using the
social media or SMS, end to end safety is very lexographical orders of the selectors and CSS
much necessary. To accomplish this, (Taleby rule declarations, embedding messages into
Ahvanooey et al., 2018) proposed AITSteg, a CSS file if e book is done. To increase the
novel text Steganography method. A trusted capacity of data embedding in CSS file,
scenario is considered while evaluating artificial fake rules are created. For a self-
AITSteg technique. The parity of Chinese generative cover text, NLP based Markov chain
characters’ stroke number is analyzed by (K. model is proposed by (Gurunath et al., 2021).
Wang & Gao, 2019). A binary digit can be Between two RNN models .e RNN- Stega and
expressed by employing the parity and a binary RNN- Generated Lyrics, embedding rate,
digit string is expressed by its combination. capacity and many other attributes are
Space mapping concept is exploited. By using contradicted LSTM Neural Network is
followed by the RNN model.
Table 3. Text Steganography techniques
Title & publication Data set Techniques/ Advantage
Algorithms
(K. Wang & Gao, 2019) Chinese Coverless method based on High enclosing rate
characters from PCCSN feature and good
Internet concealment.
(D. C. Wu & Su, 2020) 40 CSS files Lexographical orders of Feasible and robust
from ebooks in selectors, CSS rules
EPUB format
251
Steganalysis Techniques: A Systematic Review
(Gurunath et al., 2021) Twitter posts, NLP based Markov chain High volume and
news, movie model high rate of
reviews and messages.
lyrics
(J. Wang et al., 2019) Video databases High efficient Good video quality, easy
containing traffic, park, video coding implementation. Cover
party race etc containing based on IPM choosing rule can be
23 YUV sequences combined to other HEVC
IPM Steganography.
(Mstafa et al., 2020) 15 video sequences from LSB, Shi-Tomasi Secure from salt and pepper,
252
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
253
Steganalysis Techniques: A Systematic Review
the algorithm. Three different data hiding trained networks with same architecture, a
methods are considered. A new CNN network popular CNN model including ensemble
structure is designed, where deduction classification strategy is presented by (Su &
accuracy of special domain steganography is Zhao, 2019). The method can be employed in
improved by (Zhang et al., 2020). In the pre- boosting detection accuracy in spatial and
processing layer, convolution kernels are JPEG steganography by some steganalysis
optimized by using 3X3 Kernels. Separable based on CNN with the given architecture. An
convolutions are used. This helps to: optimized novel voting fusion structure is also
proposed. This method upgrades the execution
Apply Channel correlation of residuals. of CNN based steganalysis. Space to train well
Image substance compression. designed base learners is also constructed.
Based on deep Q network, a selective collective
Signal -To-Noise ratio increased process in image steganalysis is given by (Ni et
Spatial pyramid pooling is used to enhance the al., 2019). This method combines
ability of representation. Data augmentation is reinforcement learning with CNN and they are
used to raise performance of network. not seen in collective pruning. Deep Q Network
is an ensemble optimization method. Feature
A novel steganalysis method is introduced to extraction methods such as GFR, SRM and
detect all steganography types with Q factor maxSRMd2 are employed. When images are
between 99 and 100. Rounding errors when intentionally resized by nearest-neighbour
decompressing JPEG image, exhibits Gaussian interpolation, performance of CNN is
distribution with 1/12 as variance which disturbed. To solve this, a pre-processing is
increased on steganographic embedding on proposed for intentional image down sampling
DCT co efficient. This helps in accurate by (Kato et al., 2020). In both resized original
detection. The approach from(Butora& and steganographic images, additional
Fridrich, 2020)is named as “reverse JPEG steganographic signals are embedded. These
Compatibility attack”. The attack is examined signals get embedded in same pixels of resized
based on statistical signal detection using images. So the special frequency differences
reasoning under simple assumptions. Machine are apparent. This helps CNN to learn features.
learning tools provide best detection practice. As the vast existing deep learning based
The method proves good with a variety of steganalysis resulted in huge cost and storage,
JPEG compressions on diverse datasets of color the authors (Tan et al., 2021)proposed a
and grey scale images and with five Channel pruning drawn deep residual network
steganographic schemes. Importance of data architecture, where the network structure is
security is explained in terms of steganography shrinked. This is done by combining two
and steganalysis by (Jung, 2019). Categories of network pruning schemes with a hybrid
steganography based on the cover media used criterion in a data-driven manner. The network
and divisions of steganalysis like visual, resulted seemed to be in a slim and bottleneck
statistical, structural and learning are discussed. like structure.
The authors also reviewed on the frameworks
like Tensor flow, Keras, Scikit learn which are GBRAS-net is a novel CNN architecture from
used to develop machine learning algorithms. (Reinel et al., 2021). The Softmax activation
Finally the process of machine learning for function is applied towards the end here. The
steganalysis is briefly unfolded. pre-processing step uses filter banks to increase
steganographic noise. The feature extraction
As the existing CNN based steganalysis step uses depth-wise separable convolutional
algorithms combine the output of separately layers. To improve the traditional text
254
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
(Wen et al., 2019) COCO, Simple CNN Decision strategy is Advanced frame
Gutenberg architecture is proposed to increase like RNN and
designed using the long text performance Attention
principle of Mechanism can be
Occam’s Razor used to design new
steganalysis
architecture.
(Boroumand et al., BOSSBase SRNet, spatial- Reducing the use of Inner workings of
2019) BOWS2 domain heuristic design steganographic
embedding elements, Using techniques should
methods such as selection channel for be monitored
WOW, HILL, JPEG domain
MiPOD, S-
UNIWARD
(Z. Wang, Li, et al., BOSSbase k-means, Even a very smaller set Should recognize
2019) ver.1.01, SUNIWARD of suspected images, steganography
UCID and WOW(for the method works well users and then
steganography) conduct a
SRM and PSRM Performance
(feature comparison.
extraction)
(Li & Bors, 2020) AIM@SHAP RRFS algorithm Use of three varying Limited
E, FOCUS using PCC and data hiding methods transformation sets
K3D, Shape MIC are used.
retrieval
contest 2007
(Zhang et al., 2020) BOSSBase S-UNIWARD, Number of parameters To measure how
v1.01 WOW and HILL is reduced by better the development in
BOWS2 CNN network- convolution kernels; the network shape
Zhu-Net, Spatial Residual is used in pre- impacts on the
Pyramid Pooling processing layer; working of the
mapping to fixed steganographic
dimensions are done Analyzer.
by SPP-module.
(Butora & Fridrich, BOSSbase Detectors : e- Training the detector For spatial domain
2020) 1.01 SRNet, e-GFR, e- in one embedding embedding,
BOWS2, Hist algorithm makes it decompressed
Embedding applicable in other JPEGs should not
255
Steganalysis Techniques: A Systematic Review
(Y. Niu et al., 2019) Twitter, R-BILSTM-C Both the long term text -
IMDB, combining Bi- semantic information
Gutenberg LSTM and CNN and local features are
captured to improve
efficiency.
(Su & Zhao, 2019) BOSSBase CAECS One model is trained to New ways to create
S-UNIWARD produce a group. subspaces to be
J-UNIWARD Detection accuracy is exploited.
boosted in both spatial
and JPEG
steganography.
(Ni et al., 2019) BOSSbase Ensemble Automatically the -
1.01 optimization machine can find two
method ground on goals to meet the goal
Deep Q Network. state at the same time.
S-UNIWARD
J-UNIWARD
(Kato et al., 2020) BOSSbase CNN based Less training time for More practical way
1.01 SRNet, Xu-Net the pre-processing should be devised,
layer which yields
Comparable
effectiveness.
(Tan et al., 2021) BOSSbase Deep Residual achieve good detection To develop self-
1.01 Network CALPA- execution with a adaptive deep
BOWS2, NET less % of the model learning
ImageNet, size and a minimum steganalyzers
ALASKA proportion of working without introducing
cost. parameters.
(Reinel et al., 2021) BOSSBase CNN Huge capacity to find New techniques to
BOWS2 architechture- adaptive be incorporated and
GBRAS-Net steganography in techniques to
WOW, MiPOD, spatial domain. improve pre-
S-UNIWARD, processing stage
HILL should be worked
HUGO on.
Accuracy percentage of various CNN increasing from the previous one to the other
architectures in Steganalysis is found to be (Refer Table 6).
256
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
Table 6. Percentage of Accuracy in various CNN architectures using BOSSbase 1.01 data
and BOWS 2 for training (Reinel et al., 2021)
Algorithm Payload WOW S-UNIWARD
257
Steganalysis Techniques: A Systematic Review
258
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
259
Steganalysis Techniques: A Systematic Review
260
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
Lu, W., Xue, Y., Yeung, Y., Liu, H., Huang, J., Niu, K., Li, J., Yang, X., Zhang, S., & Wang,
& Shi, Y. Q. (2021). Secure Halftone B. (2019). Hybrid Adaptive Video
Image Steganography Based on Pixel Steganography Scheme under Game
Density Transition. IEEE Transactions on Model. IEEE Access, 7, 61523–61533.
Dependable and Secure Computing, 18(3), https://doi.org/10.1109/ACCESS.2019.29
1137–1149. 02464
https://doi.org/10.1109/TDSC.2019.29336
21 Niu, Y., Wen, J., Zhong, P., & Xue, Y. (2019).
A Hybrid R-BILSTM-C Neural Network
Margie Semilof, C. C. (2021). What is Based Text Steganalysis. IEEE Signal
steganography? Processing Letters, 26(12), 1907–1911.
https://www.techtarget.com/searchsecurit https://doi.org/10.1109/LSP.2019.295395
y/definition/steganography 3
Meng, R., Cui, Q., Zhou, Z., Fu, Z., & Sun, X. Paladion. (2005). Steganalysis.
(2019). A Steganography Algorithm Based
on CycleGAN for Covert Communication Rabie, T., & Baziyad, M. (2019). The
in the Internet of Things. IEEE Access, 7, Pixogram: Addressing High Payload
90574–90584. Demands for Video Steganography. IEEE
https://doi.org/10.1109/ACCESS.2019.29 Access, 7, 21948–21962.
20956 https://doi.org/10.1109/ACCESS.2019.28
98838
Mstafa, R. J., Younis, Y. M., Hussein, H. I., &
Atto, M. (2020). A New Video Reinel, T. S., Brayan, A. A. H., Alejandro, B.
Steganography Scheme Based on Shi- O. M., Alejandro, M. R., Daniel, A. G.,
Tomasi Corner Detector. IEEE Access, 8, Alejandro, A. G. J., Buenaventura, B. J. A.,
161825–161837. Simon, O. A., Gustavo, I., & Raul, R. P.
https://doi.org/10.1109/ACCESS.2020.30 (2021). GBRAS-Net: A Convolutional
21356 Neural Network Architecture for Spatial
Image Steganalysis. IEEE Access, 9,
Naito, H., & Zhao, Q. (2019). A New 14340–14350.
Steganography Method Based on https://doi.org/10.1109/ACCESS.2021.30
Generative Adversarial Networks. 2019 52494
IEEE 10th International Conference on
Awareness Science and Technology, Saad, A. H. S., Mohamed, M. S., & Hafez, E.
ICAST 2019 - Proceedings, 1–6. H. (2021). Coverless Image
https://doi.org/10.1109/ICAwST.2019.89 Steganography Based on Optical Mark
23579 Recognition and Machine Learning. IEEE
Access, 9, 16522–16531.
Ni, D., Feng, G., Shen, L., & Zhang, X. (2019). https://doi.org/10.1109/ACCESS.2021.30
Selective Ensemble Classification of 50737
Image Steganalysis Via Deep Q Network.
261
Steganalysis Techniques: A Systematic Review
Su, A., & Zhao, X. (2019). Boosting Image Mode. IEEE Access, 7, 119393–119402.
Steganalysis under Universal Deep https://doi.org/10.1109/ACCESS.2019.29
Learning Architecture Incorporating 36614
Ensemble Classification Strategy. IEEE
Signal Processing Letters, 26(12), 1852– Wang, K., & Gao, Q. (2019). A Coverless Plain
1856. Text Steganography Based on Character
https://doi.org/10.1109/LSP.2019.295008 Features. IEEE Access, 7, 95665–95676.
1 https://doi.org/10.1109/ACCESS.2019.29
29123
Taleby Ahvanooey, M., Li, Q., Hou, J., Dana
Mazraeh, H., & Zhang, J. (2018). AITSteg: Wang, Z., Gao, N., Wang, X., Xiang, J., & Liu,
An innovative text steganography G. (2019). STNet: A Style Transformation
technique for hidden transmission of text Network for Deep Image Steganography.
message via social media. IEEE Access, 6, In Lecture Notes in Computer Science
65981–65995. (including subseries Lecture Notes in
https://doi.org/10.1109/ACCESS.2018.28 Artificial Intelligence and Lecture Notes in
66063 Bioinformatics): Vol. 11954 LNCS.
Springer International Publishing.
Tan, S., Wu, W., Shao, Z., Li, Q., Li, B., & https://doi.org/10.1007/978-3-030-36711-
Huang, J. (2021). CALPA-NET: Channel- 4_1
Pruning-Assisted Deep Residual Network
for Steganalysis of Digital Images. IEEE Wang, Z., Li, S., & Zhang, X. (2019). Towards
Transactions on Information Forensics and Improved Steganalysis: When Cover
Security, 16(c), 131–146. Selection is Used in Steganography. IEEE
https://doi.org/10.1109/TIFS.2020.300530 Access, 7, 168914–168921.
4 https://doi.org/10.1109/ACCESS.2019.29
55113
Tang, W., Li, B., Barni, M., Li, J., & Huang, J.
(2021). An automatic cost learning Wen, J., Zhou, X., Zhong, P., & Xue, Y. (2019).
framework for image steganography using Convolutional neural network based text
deep reinforcement learning. IEEE steganalysis. IEEE Signal Processing
Transactions on Information Forensics and Letters, 26(3), 460–464.
Security, 16, 952–967. https://doi.org/10.1109/LSP.2019.289528
https://doi.org/10.1109/TIFS.2020.302543 6
8
wikipedia. (n.d.). Steganography-wikipedia.
Tang, W., Li, B., Tan, S., Barni, M., & Huang,
J. (2019). CNN-Based Adversarial Wu, D. C., & Su, H. Y. (2020). Steganography
Embedding for Image Steganography. via E-Books with the EPUB format by
IEEE Transactions on Information rearrangements of the contents of the CSS
Forensics and Security, 14(8), 2074–2087. Files. IEEE Access, 8, 20459–20472.
https://doi.org/10.1109/TIFS.2019.289123 https://doi.org/10.1109/ACCESS.2020.29
7 66889
Wang, J., Jia, X., Kang, X., & Shi, Y. Q. Wu, J., Chen, B., Luo, W., & Fang, Y. (2020).
(2019). A Cover Selection HEVC Video Audio Steganography Based on Iterative
Steganography Based on Intra Prediction Adversarial Attacks against Convolutional
262
Journal of Survey in Fisheries Sciences 10(2S) 244-263 2023
Neural Networks. IEEE Transactions on You, W., Zhang, H., & Zhao, X. (2021). A
Information Forensics and Security, Siamese CNN for Image Steganalysis.
15(XX), 2282–2294. IEEE Transactions on Information
https://doi.org/10.1109/TIFS.2019.296376 Forensics and Security, 16, 291–306.
4 https://doi.org/10.1109/TIFS.2020.301320
4
Xu, G., Wu, H. Z., & Shi, Y. Q. (2016).
Structural design of convolutional neural Zhang, R., Dong, S., & Liu, J. (2019). Invisible
networks for steganalysis. IEEE Signal steganography via generative adversarial
Processing Letters, 23(5), 708–712. networks. Multimedia Tools and
https://doi.org/10.1109/LSP.2016.254842 Applications, 78(7), 8559–8575.
1 https://doi.org/10.1007/s11042-018-6951-
z
Yang, J., Liu, P., & Li, S. (2019). A common
method for detecting multiple Zhang, R., Zhu, F., Liu, J., & Liu, G. (2020).
steganographies in low-bit-rate Depth-Wise Separable Convolutions and
compressed speech based on bayesian Multi-Level Pooling for an Efficient
inference. IEEE Access, 7, 128313– Spatial CNN-Based Steganalysis. IEEE
128324. Transactions on Information Forensics and
https://doi.org/10.1109/ACCESS.2019.29 Security, 15, 1138–1150.
39629 https://doi.org/10.1109/TIFS.2019.293691
3
Yang, J., Ruan, D., Huang, J., Kang, X., & Shi,
Y. Q. (2020). An Embedding Cost Zhao, H., Liu, Y., Wang, Y., Liu, S., & Feng,
Learning Framework Using GAN. IEEE C. (2021). A Video Steganography
Transactions on Information Forensics and Method Based on Transform Block
Security, 15, 839–851. Decision for H.265/HEVC. IEEE Access,
https://doi.org/10.1109/TIFS.2019.292222 9, 55506–55521.
9 https://doi.org/10.1109/ACCESS.2021.30
59654
Ye, J., Ni, J., & Yi, Y. (2017). Deep Learning
Hierarchical Representations for Image Zhou, L., Feng, G., Shen, L., & Zhang, X.
Steganalysis. IEEE Transactions on (2020). On Security Enhancement of
Information Forensics and Security, Steganography via Generative Adversarial
12(11), 2545–2557. Image. IEEE Signal Processing Letters,
https://doi.org/10.1109/TIFS.2017.271094 27(c), 166–170.
6 https://doi.org/10.1109/LSP.2019.296318
0
Yedroudj, M., Comby, F., & Chaumont, M.
(2018). Yedroudj-Net: An Efficient CNN
for Spatial Steganalysis. ICASSP, IEEE
International Conference on Acoustics,
Speech and Signal Processing -
Proceedings, 2018-April(1), 2092–2096.
https://doi.org/10.1109/ICASSP.2018.846
1438
263