You are on page 1of 33

Journal of Ambient Intelligence and Humanized Computing (2023) 14:7761–7793

https://doi.org/10.1007/s12652-023-04591-z

ORIGINAL RESEARCH

Image steganalysis using deep learning: a systematic review and open


research challenges
Numrena Farooq1 · Arvind Selwal1

Received: 12 July 2020 / Accepted: 16 March 2023 / Published online: 31 March 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
Image steganography involves the process of concealing sensitive information in the cover image to achieve secret com-
munication. The counterpart of steganography is image steganalysis, which is used to detect any hidden information that
is being communicated among different entities. The image steganalysis has gained much attention in the recent past from
the information law enforcement as well as with the advancements in communication and information technology, the tech-
niques available for steganography make it more challenging to detect steganographic content. Several steganalysis techniques
are available as reported in the literature and each technique relies on the underlying steganographic method used. In this
study, we expound an in-depth review of existing state-of-the-art steganalysis approaches and also discuss the evolving intel-
ligent approaches i.e. deep learning (DL) that may be used for image steganalysis. We also present a comparative analysis
of various benchmark datasets and evaluation metrics that are commonly used for image steganalysis algorithms. Besides,
we also illustrate some openly available tools that may be used for the task of image steganalysis. Our analysis identifies
several open research issues in this active field of DL-based image steganalysis and we also present different frameworks
that exist in the literature. This study can serve as a reference document to the investigators for future directions in research
on deep learning-based steganalysis.

Keywords Steganalysis · Image steganography · Deep learning (DL) · Convolution neural networks (CNN)

1 Introduction because cryptography has been limited or forbidden by law


in various countries (Karampidis et al. 2018; Ke et al. 2018).
Recent advancements in the field of communication Steganography aims at safeguarding the information by cov-
and information technology have facilitated information ering it in advanced media, for example, digitized content,
exchange including images among remote users. However, designs, pictures, video, etc. to keep away illicit individu-
the users require security for the exchanged information als from finding sensitive information. One of the popular
over the internet so that the information reaches safely to steganographic techniques, known as image steganography,
the intended users. The techniques like steganography and hides secret information by covering it in images (Surse and
cryptography have been designed for protecting sensitive Vinayakray-Jani 2018). The process of image steganography
information in transit from unauthorized access (Kadhim is illustrated in Fig. 1 in which the confidential informa-
et al. 2019). Steganography protects the transmitted informa- tion is encapsulated in a cover image using a data embed-
tion by hiding it in digital media while cryptography uses ding algorithm and the final image is communicated to the
encryption algorithms and keys to manipulate the informa- receiver which uses an extraction algorithm to obtain the
tion being exchanged (Kim et al. 2019; Rasras et al. 2019). actual secret information.
The technique of hiding messages is used more commonly The steganographic process can be carried out in two
domains viz spatial domain and frequency domain. In spa-
tial domain (Ke et al. 2019), some of the pixels of the image
* Numrena Farooq
mirnumrena@gmail.com are changed directly which makes them indistinguishable
to the human eye. The examples of data-hiding algorithms
1
Department of Computer Science and Information used in this domain include HILL (Li et al. 2014), HUGO
Technology, Central University of Jammu, Samba 181143, (Pevný et al. 2010), S-UNIWARD (Holub et al. 2014),
India

13
Vol.:(0123456789)
7762 N. Farooq, A. Selwal

Fig. 1  A schematic representation of the image steganography process

MiPOD (Sedighi et al. 2016) and WOW (Holub and Frid- Transformation (DCT) completely defined in (JinaChanu
rich 2012). Examples of spatial domain steganographic et al. 2012).The commonly used algorithms infrequency
methods include the Random Pixel Embedding (RPE) domain are J-UNIWARD (Holub et al. 2014), UED (Pan
method, Least Significant Bit (LSB) technique, Edge Based et al. 2016), UERD (Guo et al. 2015)and F5 (Westfeld 2001).
data Embedding (EBE) method, Pixel Value Difference The counterpart of steganography is steganalysis (Liu
(PVD) method (Hussain et al. 2018). An another way of et al. 2020) wherein the stego images are tested for any kind
performing steganography is in the frequency domain (Ke of embedded secret information with hardly any understand-
et al. 2019), often referred as JPEG image steganography. ing of the steganographic algorithm used throughout the
In this method, secret information is embedded in the for- process (Kaur and Kaur 2014; Ke et al. 2018). Steganalysis
mat of transform coefficients of the cover image (Djebbar aims at collecting sufficient proof from observed informa-
et al. 2012). This approach is used mainly because it is more tion to verify whether a carrier is a stego or a simple image
resilient to attacks than in the spatial domain (Sajid Ansari depending on the existence or complete absence of hidden
et al. 2019). Examples of frequency-domain transformation data (Chhikara and Singh 2013). The process of steganalysis
systems include Discrete Wavelet Transformation (DWT), is used to find the hidden information from the communi-
Singular Value Decomposition (SVD), and Discrete Cosine cated information which also helps in tracing out the illegal

Fig. 2  An illustration of activities flow in image steganalysis methods

13
Image steganalysis using deep learning: a systematic review and open research challenges 7763

content being exchanged (Nissar and Mir 2010; Jain 2020). c. We provide a detailed review of the existing CNN frame-
The overall process of steganalysis is depicted in Fig. 2 works for image steganalysis with their performance and
wherein a cover image is selected in which a secret mes- detection accuracy.
sage is embedded using one or several steganographic algo- d. An overview of various publicly available benchmark
rithms or using a secret key, so the resultant image becomes datasets and evaluation metrics for image steganalysis
the stego image. This stego image is then transmitted to is also presented.
the intended receiver through the communication channel. e. The prospective open research issues related to image
On the recipient end, the steganalysis process identifies the steganalysis are identified along with their viable future
existence of a concealed message inside the stego image directions.
by using steganalysis algorithm or the equivalent secret key
used during the embedding step. Some acronyms with their meaning used throughout
The process of image steganalysis is usually partitioned in this paper are summarized in Table 7 under Appendix
into feature extraction and binary classification. The for- section. The remainder of this article is organized as fol-
mer depends upon the manual extraction of features, in lows: Sect. 2 presents the overview of existing steganalysis
which the highest outcomes were attained utilizing Rich techniques. Section 3 describes deep learning-based stega-
Models (RM) (Fridrich and Kodovsky 2012).On the other nalysis. Section 4 presents a detailed review of existing
hand, the latter depends on binary classifiers where Ensem- frameworks. An illustration of various benchmark image
ble Classifiers(EC) (Kodovský et al. 2012), Support Vec- steganalysis datasets, evaluation metrics and some open
tor Machines (SVM) (Chang and Lin 2011), or perceptron software tools is presented in Sect. 5. Section 6 presents
(Lubenko and Ker 2012) are commonly used. various identified open research issues and opportunities
Due to recent advances in deep learning (DL) and Graph- in image steganalysis. The concluding remarks and future
ics Processing Units (GPUs) (Tabares-Soto 2020), research- scope is discussed in Sect. 7.
ers are emphasizing to apply DL for steganalysis in order to
acquire better detection rates of steganographic images. DL
has gained a continually growing attention in recent years as
a state-of-the-art classification system. Learning tasks from 2 A classification of steganalysis techniques
events (e.g. face recognition or external appearance recogni-
tion) and using certain specific representation strategies is In this section, we present the classification of the stega-
easier. The major benefit of DL is that the most important nalysis techniques based upon the detection method opted
high-level features of the input data are extracted automati- by steganalyst and the way of detecting the presence of a
cally to enhance the learning of the targeted task (Ke et al. hidden message. The steganalysis techniques are broadly
2019).With the remarkable advancement in deep learning, classified into seven categories and are discussed below.
convolutional neural systems (CNN) executes well in differ-
ent detectors for steganalysis. Without prior selection of fea-
tures (Zhang et al. 2018), a CNN-based image steganalyzer 2.1 Visual attacks based steganalysis
will automatically integrate the feature extraction and classi-
fication steps in one specific architecture. The constraints are The visual attacks based steganalysis is considered as
modified simultaneously, thus reducing the ambiguity and the simplest type of steganalysis in which the doubtful
dimensionality imposed by manual extraction of the function image can be examined with the naked eye in order to
(Djebbar et al. 2012). It also improves detection accuracy. detect hidden contents of the image (Karampidis et al.
A motivation behind this study is to carry out an intensive 2018). Although many steganographic methods embed
study of existing state-of-art DL based Image steganalysis the message bits in sequential manner or randomly but
methods. Our review will be particularly useful for research in many other steganographic methods the message bits
community who want to work in this active field of research. are selected in a non-adaptive manner independent of the
The main contributions of this study are summarized as content of the image (Westfeld and Pfitzmann 2000). If
follows: the image contains homogeneous areas or color-saturated
areas at either 0 or 255, we may use simple visual inspec-
a. We present an in-depth review and analysis of various tion to look for doubtful artifacts. Although we can’t read-
existing state-of-the-art techniques that are used for ily see the artifacts, we can plot one bit-plane (the LSB
image steganalysis. plane, for example) and examine just that bit-plane. The
b. We explored the usage and scope of deep learning mod- whole assault particularly refers to the palette images of
els such CNN in image steganalysis. LSB pallets embedded in indices. Where the information
is simultaneously consecutively embedded, a fair case can

13
7764 N. Farooq, A. Selwal

be made for the existence of steganographic messages in This steganalysis technique is simple but has proven to be
an image (Laskar and Hemachandran 2014). less effective. Though, it is broadly used and independent
of the steganographic algorithms. It also incorporates the
2.2 Statistical steganalysis accompanying two stages like extracting features from the
information and arranging them through two different sets
In this steganalysis method, the secret information inser- viz feature extraction and classification (L and B 2017).
tion process is analyzed and the resultant modified statistics
are observed to extract the hidden information (Sabnis and 2.5 Pair of values (PoVs) method
Awale 2016). Statistical analysis is carried out on the pixels
to determine the occurrence of secret messages. Therefore, Throughout the LSB replacement process, a fixed set of
it is categorized as steganalysis of the spatial domain and Pairs of Values (PoVs) is flipped into one another when
steganalysis of the transform domain (Farsi and Shahi 2014). embedding the message. For instance: 0–1, 2–3,… 254-255-
The pair of pixels are considered in the spatial domain, and 255. 2 Is never going to become 1 or likewise. Pfitzman and
the difference between them is calculated. The pair can be Westfield (Westfeld and Pfitzmann 2000) developed a tech-
adjacent pixels of any two. They can otherwise be picked nique of steganalysis (a strong statistical attack) which can
across the two blocks within a row. Eventually, the histo- be extended towards any steganographic technique where a
gram which shows the hidden message’s presence is plotted. fixed set of Pairs of Values (PoVs) is flipped into one another
The frequency distribution of coefficients is determined in to embed the bits of the message. This approach is defined as
the transform domain, and then the histogram calculation a statistical study of the PoVs exchanged throughout embed-
is conducted with the help of which, the stego and cover ding messages. When the pixel density with which LSB has
images are distinguished. This approach does not, however, been substituted grows, the frequencies within each PoV of
include details about embedding algorithms. To resolve this both values tend to be equivalent. The aim of the statistical
issue, we can choose feature based steganalysis (Cramer attack is to equate the potentially predicted frequency distri-
et al. 2021). bution in the steganographic image for any sample distribu-
tion seen in the likely altered carrier medium (Chen 2013).
2.3 Feature‑based steganalysis
2.6 Structural steganalysis
In present feature-based steganalysis, an image model
is obtained wherein steganalyzers are constructed using Structural steganalysis targets the statistical properties of
machine learning techniques. Generally, this model is chosen image structures (such as pixel pairs, triplets, etc.).Structural
not only by the attributes of the cover image but also based steganalysis relates to a family of methods attempting to
on the embedding effects (Fridrich and Kodovsky 2012). locate hidden messages in spatial domain images by evalu-
However, in traditional feature-based steganalysis tech- ating the statistical properties of contiguous pixel groups.
niques, the inputs, present in JPEG format, are first changed Such methods have been efficient in finding randomized LSB
into the spatial domain (without rounding off the last step). embedding even at low embedding (Powell 2020). Structural
The features of the image are extracted and go through sev- steganalysis methods are intended to capture advantage of
eral filters to pick and preserve necessary information. The the availability of the steganographic algorithm used. When
presence of hidden messages in an image is detected using embedding the secret information each steganographic algo-
these extracted features. The extracted features can also be rithm retains structure characteristics in a cover image. The
used in training of classifiers (Xu 2017). structural steganalysis identifies the presence of hidden mes-
sages by analyzing the structural characteristic changes. RS
2.4 Universal/blind steganalysis analysis is an instance of structural attacks, as defined below.
The images are scanned in structural steganalysis to check
Using newly developed steganographic methods, Universal whether they include some of the defined side-effects for dif-
steganalysis systems can expose the secret message that is ferent steganographic algorithms. The images having these
inserted into the image. In universal steganalysis, firstly the properties are often susceptible to additional examination.
features are extracted from the input image, then the classi- In certain cases, the image can have indications of steg-
fier is used and then the model design training is performed anography when it can be completely innocent. Therefore,
(Broda et al. 2014; Pathak and Selvakumar 2014). However, a detailed investigation becomes necessary after the occur-
in universal based steganalysis technique, the steganalysis rence of a structural attack. So usually a steganalyst would
algorithm isn’t perceived by everybody, thus, anyone can normally have to obtain a larger collection of features which
suggest a predictor to determine the presence of the con- reflect ambiguous images (Laskar and Hemachandran 2014).
cealed message that won’t rely upon steganalysis algorithms. Structural steganalysis works exclusively on the image under

13
Image steganalysis using deep learning: a systematic review and open research challenges 7765

scrutiny and doesn’t need any training stage, which may con- of the main benefits of using steganalysis based on super-
ceivably be an issue as the finder may be “over-trained” to a vised learning are:
particular cause of cover images (Ker 2007).
a. Creating universal detectors for steganalysis through
2.7 RS steganalysis (regular and singular groups) teaching strategies.
b. Many widely accessible internet software applications
RS Steganalysis is a consistent and precise technique may be used explicitly to train a steganalysis detector
proposed by Fridrich et al. (2001) for distinguishing and (Abiodun et al. 2018).
evaluating the least significant bit embedded in images of
color and grayscale. The length of the hidden message is 2.9 Deep learning‑based steganalysis
extracted by examining the lossless capacity i.e. capacity
with regards to lossless information embedded in LSBs. Deep learning is one of the emerging technology that
The embedded data is said to be lossless if the stego image attempts to extract the maximum appropriate high-level fea-
can be returned to an exact copy of the cover image after tures from a given input data in order to improve the learning
extracting the embedded bits. Randomizing the LSBs (e.g. of the desired task (Couchot et al. 2016).Deep learning is
by replacing LSBs) decreases the lossless capacity of the not a new technology but another form of existing neural
LSB plane to embed. The lossless power is thus used by networks or a large number. The rapid increase in computer
(Fridrich et al. 2001) as a sensitive measure of embedded hardware technology enables to compute of deeper levels of
bits in the LSB plane. Even though the LSB plane seems neural networks and remarkable performances are achieved
arbitrary, it is identified with the other bit planes. This (Kim and Lee 2018). Deep learning systems are enormous
relationship is nonlinear and the lossless limit is utilized neural systems that can straightforwardly accept informa-
to quantify this relationship genuinely well (Fridrich and tion as input. In image processing, the system is straight-
Goljan 2002). forwardly taken care of with pixels. A deep learning system
combines two stages in one (feature extraction and classifi-
cation) (Pibre et al. 2016).The most widely used neural net-
2.8 ANN‑based steganalysis works are Deep Neural Network (DNN), Convolution Neural
Network (CNN), Restricted Boltzmann Machine (RBM),
ANN is characterized as a data preparing model that mim- Recurrent neural network (RNN), and Deep Belief Network
ics biological neuron approaches and it incorporates a few (DBN) (Kim and Lee 2018). Convolution Neural Networks
stages to group the information (Kumar and Kumar 2020). (CNN) have shown tremendous success in detecting and rec-
Feedforward and backpropagation neural systems are regu- ognizing objects. In (Zhang et al. 2018; Xu 2017; Xu et al.
larly utilized in a grouping. The classification procedure 2016; Ye et al. 2017; Qian et al. 2018), the use of CNNs for
has two stages viz the training stage and the testing stage. steganalysis has thus been extensively examined. CNN can
During the training stage, the neural system partners the select and learn features that are suitable for the classifica-
results with the given input designs, by altering loads of tion of steganographic methods, automatically. CNN’s can
sources of input. While in a testing stage, the information also be employed in image steganalysis, similar to the way
design is recognized and the related outcome is evaluated they are employed in conventional classification problems
(L and B 2017). The ANN classification model is com- such as object detection and recognition (Kim et al. 2020).
monly used to determine the presence of concealed infor- In the past few years, deep learning, including its capabil-
mation. Using training examples, supervised learning tech- ity to inevitably learn successful feature representation from
niques create a classifier in order to distinguish between information, also made accomplishments in numerous fields,
steganographic images and non-steganographic images. for example, computer vision and natural language process-
Supervised learning methods including neural networks as ing. Simultaneously, deep learning-based image steganaly-
classification methods have achieved considerable promi- sis has to turn out to be an interesting area in the field of
nence in latest steganalysis studies. Some feature vectors steganalysis. It has the capability to naturally extricate noise
are first retrieved in supervised learning system steganaly- residuals from the stego image and can additionally combine
sis, and given to a machine learning model as training the two stages of feature extraction and classification into
data. Such sources include messages from both stego and one system (Tian and Li 2019). As observed the benefits
non-stego. The classification concept of the learning clas- that are likely to be achieved from using DL in image stega-
sifier is updated sequentially, depending on its analysis and nalysis, we will explore it further in the following section.
input image. Ultimately the last stego classifier is achieved
that decides whether a secret message exists or not. Many

13
7766 N. Farooq, A. Selwal

3 Deep learning‑based steganalysis et al. 2018). CNN is like a neural network with a specific
network structure as a constituent to an in-depth learning
Deep Neural Network (DNN) models are more effective at method. It includes local connection, weight sharing, down-
approximating extremely complex functions than deeper sampling, and several other ideas for structural design. It can
ones. This skill suggests that the complex statistical proper- learn more effective expression of features directly from the
ties of natural images can be captured by a very deep neural data. The convolution structure in CNN is suitable from the
network, which could be useful for classification of images. Steganalysis point of view to capture the domain-specific
Recent studies also illustrate that using deep neural networks statistical details of the association between pixel values,
brings substantial improvement than traditional approaches which is quite important for steganalysis. Around the same
in many applications. Although tremendous success in time, weight sharing, pooling, as well as other mechanisms
image recognition has been attained for very deep neural often significantly decrease the training examples, allowing
networks, existing networks for image steganalysis are still the network model capable of handling broad and numerous
deeper ones (Wu et al. 2017; Chaumont 2019). images (Zou et al. 2019). Consequently, CNN is generally
Steganography was fairly straightforward in the initial chosen as the basic model in the context of deep learning as
periods and therefore the dimensions of the features in stega- shown in Fig. 3. A layered architecture of CNN model along
nalysis were small and computationally lightweight. Con- with a set of layers is depicted in Fig. 4.
ventional systems demonstrate good performance in particu- Many of convolution neural networks (CNNs) popu-
lar conditions, however, the results in realistic circumstances lar Architectures are LeNet (LeCun et al. 1995), AlexNet,
are low as there are many types of manipulated images (or VGG, Inception, GoogleNet, ResNet, etc. These architec-
steganography) in the real world, generated with various tures represent as overall design standards which will then
methods of image compression, each one with various be adapted by machine-learning professionals to address
properties. Researchers have suggested methods to resolve difficult computer vision tasks. These architectures act as
this constraint using convnets (a deep learning model) (Park rich feature extractors that can be used for classification,

Input Feature Maps Feature Maps Feature Maps Feature Maps Fully Connected
24×24 20×20×4 10×10×4 8×8×8 8×8×4 Layer

OUTPUT

Convoluon Pooling Convoluon Pooling

Fig. 3  A basic CNN-based layered architecture for image steganalysis

Fig. 4  An architecture of convolutional neural network model Adapted from (Bashkirova 2016)

13
Table 1  A comparison among performance of SOTA CNN architectures (Reporting the top error rates (Khan et al. 2020))
Year CNN model Main contribution Number of layers Top-5 error rate Number of parameters Depth ImageNet References
accuracy (%)
Top-1 Top-5

1998 LeNet Foremost general CNN 8 MNIST: 0.95 60 thousand 5 Lecun et al. (1998)
architecture
2012 AlexNet Wider and broader than the 7 ImageNet: 16.4 60 million 8 63.30 84.60 Krizhevsky et al. (2017)
LeNet
Uses Relu activation func-
tion, dropout and overlap
Pooling
GPUs of NVIDIA GTX
580
2013 ZFNet Intermediate layers visuali- ImageNet: 11.7 60 million 8 Zeiler et al. (2014)
zation.
2014 GoogLeNet Block concept introduced 19 ImageNet: 6.7 4 million 22 74.80 92.20 Szegedy et al. (2015)
Split the concept of trans-
forming and merging
2014 VGG Net Topology is homogeneous 16 ImageNet: 7.3 138 million 19 74.40 91.90 Simonyan and Zisserman
Uses kernels of small (2015)
dimensions
2015 ResNet Includes residual learning 152 ImageNet: 3.6 25.6 million 152 76.0 93.0 He et al. (2016)
Skip links for identity CIFAR-10: 6.43 1.7 million 110
mapping
2015 Inception-V3 Controls the issue of a bot- 48 ImageNet: 3.5 23.6 million 159 78.80 94.40 Szegedy et al. (2016)
Image steganalysis using deep learning: a systematic review and open research challenges

tleneck representation Multi-Crop: 3.58


Substitute filters of large Single-Crop: 5.6
dimensions with small
filters
2016 Inception-V4 Break the idea to transform ImageNet: 4.01 35 million 70 80.0 94.0 Szegedy et al. (2017)
and merge
Makes use of asymmetric
filters
2017 Xception Convolution in depth is 36 ImageNet: 0.055 22.8 million 126 79.0 94.50 Chollet et al. (2017)
followed by convolution
in point
2017 Residual Attention Neural Added a mechanism for CIFAR-10: 3.90 8.6 million 452 80.50 90.20 Wang et al. (2017)
Network focusing attention CIFAR-100: 20.4
ImageNet: 4.8
2017 Squeeze and Excitation Models feature-map inter- ImageNet: 2.3 27.5 million 152 Hu et al. (2020)
Networks dependence
2017 DenseNet Information flow across 6,12,24,32– ImageNet 34 million 190 77.90 93.90 Huang et al. (2017)
layers CIFAR-10 + : 3.46 25.6 million 250
CIFAR-100: 19.64 15.3 million

13
7767
Table 1  (continued)
7768

Year CNN model Main contribution Number of layers Top-5 error rate Number of parameters Depth ImageNet References
accuracy (%)

13
Top-1 Top-5

2017 PolyNet Structure diversity was 8 to 152 ImageNet: Single:4.25 92 million 81.30 95.80 Zhang et al. (2017)
tested Multi:3.45
Poly Inception is a new
module that has been
added
Polynomial compositions
are used to generalize the
residual unit
2018 Convolutional Block Atten- Takes the advantage of both 1001 ImageNet: 5.59 48.96 million 101 Woo et al. (2018)
tion Module (ResNeXt101 spatial and feature-map
(32 × 4d) + CBAM) data
2018 Competitive Squeeze and Feature-map is rescaled CIFAR-10: 3.58 36.92 million 152 Hu et al. (2018)
Excitation Network using both residual and CIFAR-100: 18.47 36.90 million 152
CMPE-SE-WRN-28 identity mappings
2018 MobileNet-v2 Residual structure is 3 ImageNet 3.4 million 53 72.19 90.53 Sandler et al. (2018)
inverted
2018 CapsuleNet Attaches importance to 3 MNIST:0.00855 35.4 million 3 Sabour et al. (2017)
feature-to-feature con-
nections
2019 BiT-L (ResNet) It makes use of large-scale ImageNet 928 million 87.50 98.50 Kolesnikov et al. (2020)
pre-training JFT-300
To balance complexity and
performance, it uses a
simple training and fine-
tuning architecture with
a small number of well
selected components
2019 FixResNeXt-101 32 × 48d Improves the performance ImageNet 829 million 86.4 98.0 Touvron et al. (2019)
of the classifier by using
varied train and test
resolutions
2019 AdvProp (EfficientNet-B8) To avoid overfitting, it ImageNet 88 million 85.5 97.3 Xie et al. (2020a)
interprets adversarial
cases as additional exam-
ples For adversarial cases,
it employs a different aux-
iliary batch norm
N. Farooq, A. Selwal
Table 1  (continued)
Year CNN model Main contribution Number of layers Top-5 error rate Number of parameters Depth ImageNet References
accuracy (%)
Top-1 Top-5

2020 NoisyStudent EfficientNet- Use of equal-or-larger ImageNet-A 480 million 88.4 98.7 Xie et al. (2020b)
L2 student models expands
on the indication of self-
training and concentration
2020 HRNetV2 Representations with a high 1043 ImageNet: 5.4 128.1 million Wang et al. (2020)
level of detail
2021 BEiT-L To pretrain vision Systems, 12–24 ViT ImageNet-22 K 331 million 88.60 98.66 Bao et al. (2021)
a masked image simula- pretrain
tion project was devised
2021 ALIGN (EfficientNet-L2) Using a contrastive loss, ImageNet 480 million 88.64 98.67 Jia et al. (2021)
a basic dual-encoder
architecture learns to
align visual and verbal
representations of image
and text pairs
2021 Meta Pseudo Labels Produce pseudo labels on ImageNet 480 million 90.20 98.8 Pham et al. (2021)
unlabeled data
Feedback from the learner
learning on the labelled
dataset is regularly used
to improve Meta Pseudo
Labels
Image steganalysis using deep learning: a systematic review and open research challenges

13
7769
7770 N. Farooq, A. Selwal

object detection, image segmentation and many more machine learning is to disintegrate them, however, CNN
advanced tasks (Maeda-Gutiérrez et al. 2020; Khan et al. has been utilized in the spatial domain image steganalysis
2020; Alom et al. 2018). Table 1 summarizes the archi- and gained extraordinary ground, therefore the researchers
tectural descriptions of the state-of-the-art CNN models, propose to apply CNN to jpeg domain image steganalysis
their parameters, and benchmark data output. (Tang et al. 2018).
The latest steganalysis theory focuses on deep learn-
ing, that requires a high-level feature dimension and mas- 3.3 Mathematical model for image steganalysis
sive computing (Dengpan et al. 2019). Especially when
timelines for information need to be preserved, real-time Given a training dataset ‘Dt’, composed of images of both
image steganalysis is important. If there are huge quan- cover and stegano classes, the image steganalysis model S ­M
tities of users in the detection scene, deep learning can learns the dataset ­Dt by using the classification algorithm C.
be used to improve the efficiency of image steganalysis, The classes of the model ‘C’ is labled as ‘1’ and ‘0’ respec-
learning from its high computational capability. The real- tively for stegano or cover images (i.e. C ∈ (1, 0) ) in the
time image steganalysis method achieves greater precision dataset. The image steganalysis model is essentially a binary
and efficiency with the usage of deep learning techniques pattern classifier that either uses extracted features of train-
(Ruan et al. 2019). ing images along with labels or direct images are fed in the
Image steganography essentially incorporates two types CNN-based models. Suppose that the features of image χj
of research classes: spatial domain and frequency domain, are denoted by F­ i then the handcrafted features-based model
therefore deep learning-based steganalysis is respectively ­SH is predicted by using Eq. 1.
divided into spatial domain steganalysis and frequency { ( )
(JPEG) domain steganalysis as described below. ifSH( F)i = 1then𝜒 ∈ stego
(1)
( )
Pred_class 𝜒j =
ifSH Fi = 0, then𝜒 ∈ cover

The deep learning-based image steganalysis model ‘SD’


3.1 Deep learning‑based steganalysis in spatial
effectively predicts a presented instance of image χ from
domain
testing dataset using Eq. 2.
Researchers mostly enhance the steganalysis execution by
{
ifSD (𝜒) = 1then𝜒 ∈ stego
structuring another system model, combining extra data or Pred_class (𝜒) =
ifSD (𝜒) = 0, then𝜒 ∈ cover (2)
implementing innovative learning methodology, so deep
learning-based spatial area steganalysis is partitioned in
network structure method, data fusion techniques, and 3.4 DL‑based frameworks for steganalysis other
learning methodology-based strategies. than CNN
a. Steganalysis strategies dependent on deep learning: it As deep learning frameworks use self-learning representa-
includes various network designs which are different tions, they rely on artificial neural networks (ANNs) that
architectures of CNN. mimic how the brain processes information. Frameworks
b. Steganalysis on information fusion methods: to further utilize unknown elements in the input distribution to extract
boost the efficiency of steganalysis, some researchers are features, organize objects, and identify important data pat-
applying fusion techniques to methods of steganalysis terns throughout the training phase. This happens at vari-
focused on deep learning. There have been three major ous levels, employing the techniques to develop the models,
categories: fusion strategies with previous knowledge, much like training machines for self-learning.
fusion strategies with features and fusion strategies with Several frameworks are used in deep learning models.
the framework. While no network is flawless, some frameworks are better
c. Learning procedure-based methods: it includes training suited to specific activities than others. It is beneficial to
of a model by utilizing various methodologies. A large have a thorough understanding of all key frameworks in
number of researchers used a variety of methods, func- order to select the best ones. Some of the most popular DL
tions, strategies to advance the network performance. frameworks for steganalysis other than CNN are listed below
in detail.
3.2 Deep learning‑based approaches
in the frequency domain

JPEG is by far the most widely known image file for-


mat. Traditionally, the initial phase in utilizing JPEGs for

13
Image steganalysis using deep learning: a systematic review and open research challenges 7771

3.4.1 Long short‑term memory networks (LSTMs) algorithms and the growth in computational power have
made them useful to important machine learning tasks.
LSTMs are a form of Recurrent Neural Network (RNN) They’ve recently gained a lot of attention after being offered
capable of learning and remembering long-term dependen- as building blocks for multi-layer learning systems known
cies. The default behavior is to recall past information for as deep belief networks. Dimensionality reduction, classi-
a long time. Over time, LSTMs keep track of information. fication, regression, content—based filtering, pattern rec-
As they recall past inputs, they’re valuable in time-series ognition, and topic modelling are all applications of RBMs
prediction. LSTMs feature a chain-like structure, with four (Fischer and Igel 2012).
interconnected layers that communicate in a unique fash- For the last five years, deep learning has progressed, and
ion. LSTMs are commonly employed for voice recognition, deep learning frameworks have become extensively used
music creation, and drug research, in addition to time-series in a variety of sectors. Table 2 lists some of the DL-based
predictions (Hochreiter and Schmidhuber 1997). frameworks used for image, text, and speech steganalysis.

3.4.2 Recurrent neural networks (RNNs)


4 State‑of‑the‑art DL‑based steganalysis
RNNs are a type of feedforward neural network that has approaches
the ability to transfer information across time steps. It is a
diverse group of models capable of almost any calculation. This section presents the existing works in the literature that
In practical, the capacity to describe temporal relationships have been carried out to address the different research issues
enables recurrent neural networks particularly well adapted in using DL in steganalysis.
to problems involving non-independent sequences of points To train a Convolution Neural Network (CNN), Tan et al.
(Lipton et al. 2015). Time-series analysis, handwriting iden- (Tan and Li 2014) used an unsupervised learning technique
tification, image captioning, natural-language processing, through a pile of Auto-Encoders. After pre-processing the
and machine translation are all popular examples for RNNs. image, supervised learning was then used with a High Pass
Filter (HPF) to expand the steganographic noise power
incorporated during the embedding phase. Steganographic
3.4.3 Radial basis function networks (RBFNs)
image recognition rates were estimated17% less than the
one for Spatial Rich Models (SRM) (Fridrich and Kodovsky
RBFNs are unique examples of artificial neural networks in
2012), and approximately 11% greater than the Subtractive
the sense that each neuron is made of a Radial-Base Func-
Pixel Adjacency Matrix model (SPAM) (Pevny et al. 2010).
tion (RBF), just one hidden neural layer is present, only one
In 2015, Qian et al. (2015) proposed a modified CNN to
output node is available and a two-stage training technique
learn features consequently utilizing DL. The model auto-
is available. Radial basis functions are employed as activa-
matically records complex situations and acquires feature
tion functions in RBFNs, which are typically utilized for
representation with five convolution layers by utilizing a
classification, time-series prediction, and regression. RBFNs
supervised learning approach with a new activation func-
have a strong fitting capability and converge quickly to the
tion called Gaussian activation. The detection rate acquired
optimal positions (Ekblad et al. 2004).
was around 4% lower than SRM (Fridrich and Kodovsky
2012) and approximately 10% higher than SPAM (Pevny
3.4.4 Multi‑layer perceptrons (MLPs) et al. 2010). Pibre et al. (2016) followed up the investiga-
tion of Qian and proposed two neural networks (CNN and
MLPs are categorized as neural feedforward networks with FNN). The steganography consistently utilizes the embed-
several layers of perceptrons having activating functions. ding key for embedding with a test system (simulator) with
MLPs are made up of two fully connected layers: an input diverse images and due to inborn joint minimization and
layer and an output layer. The input and output layers are the the protection of spatial information, the outcomes acquired
same, although they can use several, hidden layers to provide from CNN and FNN excel the traditional use of rich models
voice recognition, machine translation, and image recogni- with an ensemble classifier. Experiments were carried out to
tion software (Marius-Constantin et al. 2009). find the best shape, in a clairvoyant scenario and in a cover
source mismatch sets. The outcomes revealed a more than
3.4.5 Restricted Boltzmann machines (RBMs) 16% reduction in classification error and naturally robust to
mismatch problem with CNN and FNN.
Restricted Boltzmann machines (RBMs) are probabilistic Xu et al. (2016) describe a CNN architecture consisting
graphical systems that can be used to simulate stochas- of domain information about steganalysis. It consists of five
tic neural networks. The advancement of faster learning convolution layers in which the absolute value of elements

13
7772 N. Farooq, A. Selwal

Table 2  A summary of various DL-based frameworks for image steganalysis


Year Key points and purpose Steganaly- Type of framework Reference
sis based
on

2019 Uses a combination of CNNs and LSTM networks to capture both Text LSTM Bao et al. (2020)
local and long-distance contextual information to enable easier
exploitation of the semantic feature in texts
Recognizes and attends to crucial clues within dubious sentences
using attention mechanism
2020 Detecting dynamic text steganographic techniques with densely Text LSTM Yang et al. (2020)
linked LSTM with feature pyramids that can integrate more low
level characteristics
2021 Introduced an innovative, practical solution for steganalyzing Image Hybrid bidirectional LSTM Amina et al. (2021)
self-assertive-size images that recognizes relationships between
picture sub-locales and steganography activity
2018 To identify QIM steganography Speech RNN Lin et al. (2018)
In VoIP streams, four significant code word correlation patterns
are discovered, that are altered by embedding with concealed
data
2019 After concealing messages, it learns the inherent distribution of DNA RNN Bae et al. (2019)
coded and non-coded patterns and uses distribution deviations to
discover concealed messages
2019 As concealed information is incorporated in automatically created Text RNN Yang et al. (2019)
steganographic texts, the conditional probability distribution of
every word is altered which is detected
2008 Radial basis function is used in conjunction with Fisher’s linear Image RBFN Baragada et al. (2008)
discriminant (FLD) function to detect the existence of secret
data in a carrier image
2012 Karhunen-Loève (KL) transform coefficients are used that com- Image RBFN Hasoon and Khalifa (2012)
bines an ANN as a classifier to discover information hidden in
colored and grayscale images
2012 GLCM characteristics that are different between the cover picture Image MLP Ghanbari et al. (2012)
and the stego image are extracted
2017 Instead of replacing the EC with a deep nonlinear CNN, shallow Image MLP Zheng et al. (2018)
nonlinear neural networks, i.e. MLP, were used to detect secret
content
2016 RBMs are employed in the initial phase of a steganalysis method Speech RBM Paulin and Selouani (2016)
for speech/audio files, and they are trained using Evolutionary
Algorithms (EAs)

present in the feature maps is used to strengthen the statisti- to features obtained from complicated texture areas, while
cal model for improved outcomes. In deeper layers, a 1 × 1 features obtained from smooth areas are assigned as small
convolution kernel is used to avoid overfitting by restraining weights. The maximum value for weight is taken as 9 × 9
the collection of data values having TanH saturation area. In neighborhood of embedding probability matrix in proposed
order to achieve a much better efficiency, Qian et al. (2016) CNN. For JPEG steganalysis, Zeng et al. (2018) presented a
designed a methodology focusing on transferring learning to hybrid deep learning architecture that consists of the domain
improve CNN steganalysis training. The authors describe a information within rich steganalysis models. It entails two
pre-train CNN with feature representations to distinguish a phases in which the first step correlates with convolution and
high payload steganographic algorithm and are effectively the quantization (Q&T) and truncation step of rich models
transformed to outdo the learning features to recognize the while the second phase consists of deep neural networks
same low payload steganographic algorithm. The experi- with several deep subnetworks and parameter values learned
mental results depict a significant improvement as compared during the training process. The authors carried out ample
to existing CNN models that do not use transfer learning experiments using the dataset from ImageNet (Krizhevsky
technique. Yang et al. (2017) set forward a CNN archi- et al. 2017). Results of various Q&T combinations show
tecture called as maxCNN, having the information of the successive accuracy changes and better performance than
selection channel joined into CNN. Large weights are given prior state of art.

13
Image steganalysis using deep learning: a systematic review and open research challenges 7773

Ye et al. (2017)put forward a fallback way for steganalysis Enormous studies showed the prevailing execution of such
of digital images focused on CNN, that is manifested to be a system with considerable improvement, particularly in the
capable of reproducing and optimize the basic steps for a JPEG region.
confederate architecture and learn hierarchical representa- Dengpan et al. (2019) investigated that depth of model
tion directly from raw images. The developed scheme (CNN) and backpropagation makes each module hard to learn and
varies entirely against the one used in standard computer devour assets on GPU, therefore to beat this calculation
vision (CV) tasks, and weights are assigned with a simple and quicken the training procedure the authors character-
high pass filter set for the early layers. They used a new ized an architecture that joins batch normalization through
truncated linear unit (TLU) activation function to accu- shallow layers. Also, the depth in-network is diminished
rately represent the structure of low signal-to-noise (SNR) and width is increased to minimize the loss of small data
embedding signals. The performance of their proposed CNN in steganalysis. The network consists of five layers CNN
based steganalyzer has been improved by subsuming selec- in which there are two convolution layers and three fully
tion channel information in it. Wu et al. (2017) presented connected layers with two transfer learning plans under
a different CNN model focused on profound deep residual various payloads to improve overall efficiency. Ni et al.
learning-based network (DRN). The proposed model con- (2019) developed a selective ensemble method based on
tains countless network layers which end up being effec- deep Q-learning in image steganalysis that consolidates
tive in capturing complex digital image statistics. However, reinforcement learning with CNN and rarely perceived
residual learning in DRN allows stego signals originating in ensemble pruning. The main task of this method is to
through secret messages that are very useful for segregating enhance the generalization efficiency of the model and
cover and stego images. diminish the size of the ensemble as well in order to
Yedroudj et al. (2018) designed a spatial domain CNN select an alternate classifier or remove an alternate clas-
model in which they utilized a number of input filters for the sifier. Ke et al. (2019) proposed an easy and productive
pre-preparing stage of a pre-defined linear SRM. It includes multi-column CNN (MCNN) on steganalysis architecture.
five convolutional layers, BN, and TLU activation function The authors allowed input image of any size or resolution
which replaces the TanH used in the first two initial con- by using filters of various sizes and the features acquire
volution layers. Also, the size of the preparation database variations in the payload by each column. The model is
is increased to show signs of improvement in results from also capable of detecting internet steganographic software
those detailed in the literature. Zhang et al. (2018) proposed using the same secret embedding key. Hu et al. (2019)
an improved CNN named as “Zhu-net” based on fixing the performed steganalysis dependent on visual attention
issue of reducing sign to noise proportion of input image and (VA) and deep reinforcement learning to figure out how to
performing arbitrary-sized image steganalysis. The model identify JPEG created adaptive steganography. The steps
comprises of 3 × 3 kernels rather than conventional 5 × 5 ker- involved include selecting an image region by VA pro-
nels and upgrades convolution kernels in the pre-processing cess, then making continuous choices to generate summary
layer. It utilizes various convolutions in order to acquire resi- region, and ultimately enhancing the value of training set
due channel associations and spatial associations. The pro- and recognition capability by exchanging misclassified
posed CNN executes better in both detection precision and training images with their analogous summary region.
similarity when contrasted with different CNNs. Tian et al. Table 3 presents the summarized characteristics, archi-
(2019) trained several CNN’s using the transfer learning tecture that each author has introduced, the database used
approach for steganalysis. A Gaussian-high pass filter was in the training, validation, and testing, the domain of experi-
developed for pre-processing the images so as to improve ment (spatial or frequency), the steganographic algorithms
weak steganographic noise in steganographic images and used in the steganalysis, and reasonable outcomes. It is
ameliorated inception V3 model is used. The framework is noted that the research was carried by the implementation
divided into three phases i.e., the image processing, incep- of an Auto-Encoders stack utilizing unsupervised learning.
tion V3 layer, and re-training layer. The effectiveness of Research on Supervised Learning has continued since fol-
the proposed model is evaluated using spatial domain con- lowing three basic principles for steganalysis: reinforcement
tent-adaptive steganographic algorithms and the outcome learning of steganographic noise, utilizing a fixed high-pass
obtained depicts that developed CNN attains improved per- filter, extraction of characteristics and classification; each
formance at low embedding rates. Boroumand et al. (2019) combined together under single architecture that simultane-
presented a deep residual architecture intended to limit the ously improves its parameters. The first progress has been
utilization of heuristic and outer implemented components made mostly on the issue in the spatial domain and, later the
for both spatial and JPEG domains. The utilization of fixed researchers went into the frequency domain (JPEG).
or compelled pre-processing kernels instated to SRM filters The review of the above mentioned literature survey uses
or DCT bases will increase overall system performance. mostly the BOSSBaseV1.01 (Bas et al. 2011) database that

13
Table 3  An analysis with characteristics, databases, steganographic algorithms, and results of the various DL-based image steganalysis techniques
7774

Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

13
Qian et al. (Qian CNN Gaussian Spatial BOSSBase ImageNet HUGO Error detection Comparable perfor-
et al. 2015) 2015 1 pre-processing WOW Using 0.4bpp on BOSSBase mance than SRM
layer S-UNIWARD CNN = HUGO(28.29),WOW(29.3),S-
5 convolution layers UNIWARD(30.39)
Average pooling SRM = HUGO(25.2),WOW(25.7),S-
3 completely con- UNIWARD(26.3)
nected layer Using 0.4 bpp on ImageNet
1 Softmax CNN = HUGO(33.6),WOW(34.1),S-
UNIWARD(34.7)
SRM = HUGO(32.5),WOW(34.7),S-
UNIWARD(34.4)
Tan et al. (Tan and Li CNN Cost function Spatial BOSSBase HUGO Error detection Better execution
2014) 2016 9 convolution layers Using 0.4 bpp on BOSSBase compared to conven-
subdivided into 3 CNN = 31 tional CNN for the
stages SPAM = 42 detection of HUGO
Max pooling SRM = 14 generated stego
1 completely con- images in BOSSBase
nected layer image database
1 Softmax
Pibre et al. (Pibre CNN ReLU Spatial BOSSBaseV1.0(10 k S-UNIWARD Error Detection Using cropped BOSS-
et al. 2016) 2016 1 pre-processing gray level images) Using 0.4bpp on BOSSBase Base with embed-
layer LIRMMBase(1008 Gy Clairvoyant Scenario ding S-UNIWARD
2 convolution layers level images) CNN = 7.4 at 0.4 bpp the
No pooling FNN = 8.66 classification error
2 completely con- SRM = 24.67 by CNN and FNN
nected layers Using 0.4bpp on BOSSBase(Train) has reduced 3 folds
1 Softmax And LIRMMBase(Test) Cover-Source to EC with SRM
FNN Mismatch Scenario features. Also testing
1 pre-processing CNN = 5.16 on LIRMMBase
layer FNN = 5.89 CNN exhibits natural
2 completely con- SRM = 48.39 invariance to Cover
nected layers Source Mismatch
1 Softmax with a classification
error of 5.16%
N. Farooq, A. Selwal
Table 3  (continued)
Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

Xu et al. (Xu et al. CNN TanH Spatial BOSSBaseV1.0(10 k S-UNIWARD Error detection Detection performance
2016) 2016 1 pre-processing ReLU gray level images) HILL Using 0.1 bpp on BOSSBase is competitive w.r.t
layer CNN = S-UNIWARD(42.67),HILL( SRM with EC
5 convolution layers 41.56)
Average pooling SRM = S-UNIWARD(40.75),HILL(
Batch Normalization 43.56)
BN Using 0.4 bpp on BOSSBase
ABS absolute value CNN = S-UNIWARD(19.76),HILL(
layer 20.76)
2 completely con- SRM = S-UNIWARD(20.47),HILL(
nected layers 24.53)
1 Softmax
Qian et al. (Qian CNN Gaussian Spatial BOSSBaseV1.0(10 k WOW Error detection Shows better perfor-
et al. 2016) 2016 1 pre-processing gray level images) S-UNIWARD Using 0.4bpp on BOSSBase mance in terms of
layer CNN = WOW(21.95), S-UNI- error detection with
5 convolution layers WARD(22.05) low payload than
Average pooling SRM = WOW(20.67), S-UNI- traditional SRM
2 completely con- WARD(20.55)
nected layers
1 Softmax
Yang et al. (Yang maxCNN TanH Spatial BOSSBaseV1.0(10 k WOW Detection accuracy on detecting WOW Improvement over
et al. 2017) 2016 1 pre-processing ReLU gray level images) Using 0.1 bpp on BOSSBase previous CNN is
filter maxCNN = 66.90 2.28% for 0.1 bpp
6 convolution layers maxSRMd2 = 69.75 0.83% for 0.4 bpp.
Image steganalysis using deep learning: a systematic review and open research challenges

ABS absolute value Using 0.4 bpp on BOSSBase Also comparable


layer(only after the maxCNN = 85.08 performance with
first convolution maxSRMd2 = 84.73 maxSRMd2 with
layer) Detection accuracy of the ensemble EC when payload is
Batch normalization method on detecting WOW 0.4bpp
BN Using 0.1 bpp on BOSSBase
Average pooling maxCNN = 71.52
1 fully connected maxSRMd2 = 69.75
layer Using 0.4 bpp on BOSSBase
1 Softmax maxCNN = 87.60
maxSRMd2 = 84.73

13
7775
Table 3  (continued)
7776

Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

13
Zeng et al. (Zeng CNN ReLU JPEG ImageNet(500 k cover J-UNIWARD Results are in graphical form Provides a boost of
et al. 2018) 2017 3 convolution layers images) UERD performance with
Average pooling UED quantitative metrics
Batch normalization
BN
ABS absolute value
layer
3 completely con-
nected layers
1 Softmax
Ye et al. (Ye et al. CNN TLU Spatial BOSSBaseV1.0(10 k WOW Error Detection Shows improved
2017) 2017 Contains 10 layers ReLU gray level images) HILL Using 0.1 bpp on BOSSBase + BOWS detection accuracy
First layers of 30 fil- BOWS2(cropped gray S-UNIWARD (Train and Test) performance by a
ters whose weights scale images) CNN = WOW (24.42), S-UNIWARD large margin com-
are not initialized (32.20), HILL (38.80) pared to previous
randomly, but with SRM = WOW (31.63), S-UNIWARD CNNs, conventional
the values of high (38.06), HILL (38.94) SRMs and maxS-
pass filter used in Using 0.4 bpp on BOSSBase + BOWS RMd2
SRM (Train and Test)
8 convolution layers CNN = WOW (9.59), S-UNIWARD
Average pooling (12.81), HILL (17.08)
1 completely con- SRM = WOW (15.36), S-UNIWARD
nected layers (21.36), HILL (24.10)
1 Softmax
Wu et al. (Wu et al. DRN ReLU Spatial BOSSBase v1.01 (10 k S-UNIWARD Error detection Can learn more effec-
2017) 2017 Subnetwork uses 64 gray level natural WOW Using 0.4 bpp on BOSSBase tive features as com-
convolution filters images) HILL DRN = WOW (4.3), S-UNIWARD pared to SRM. Error
Maximum pooling MiPOD (6.3), HILL (10.4), MiPOD (4.9) detection rate outper-
Batch normalization SRM = WOW (20.1), S-UNIWARD forms than SRM and
For non-bottle neck (20.3), HILL(24.2), MiPOD (22.1) maxSRMd2
block 2 convolution MaxSRMd2 = WOW (15.2), S-UNI-
layers and for bot- WARD (18.8), HILL (21.6), MiPOD
tleneck 3 convolu- (20.4)
tion layers
Average pooling
1 fully connected
layer
N. Farooq, A. Selwal
Table 3  (continued)
Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

Yedroudj et al. CNN TLU Spatial BOSSBase WOW Error Detection Create a clear strategy
(Yedroudj et al. 30 pre-processing ReLU S-UNIWARD Using 0.2 bpp on BOSSBase to beat the state-of-
2018) 2018 filters based on CNN = WOW(27.80), S-UNI- the-art strategies
SRM WARD(36.70) without awareness of
5 convolution layers SRM = WOW(36.50), S-UNI- the selection channel
1 ABS only after first WARD(36.60) in a classic clairvoy-
layer convolution Using 0.4bpp on BOSSBase ant scenario
Average pooling CNN = WOW(14.10), S-UNI-
3 fully connected WARD(22.80)
layers SRM = WOW(25.50), S-UNI-
1 Softmax WARD(24.70)
Zhang et al. (Zhang CNN ReLU Spatial BOSSBase v1.01 (10 k WOW Error Detection Excellent detec-
et al. 2018) 2018 1 layer of 30 filters gray level images) S-UNIWARD Using 0.2bpp on BOSSBase tion accuracy than
whose weights BOWS2 (10 k gray CNN = WOW(23.33), S-UNI- SRM,Ye.net,Xu.net
are not initialized level images) WARD(28.50) and Yedroudj.net
randomly,but with SRM = WOW(36.50), S-UNI-
the values of high WARD(36.60)
pass filter used in Using 0.4bpp on BOSSBase
SRM CNN = WOW(11.80), S-UNI-
2 separate convolu- WARD(15.30)
tion layers to obtain SRM = WOW(25.50), S-UNI-
channel correlation WARD(24.70)
and spatial correla- Using 0.2 bpp on BOSS-
tion residue Base + BOWS2(train)
Image steganalysis using deep learning: a systematic review and open research challenges

4 convolution layers BOSSBase(test)


BN from 2 to 7 layers CNN = WOW(13.1), S-UNI-
Average pooling of WARD(17.1)
4–6 layers Using 0.4 bpp on BOSS-
1 Spatial pyramid Base + BOWS2(train)
pooling (SPP) BOSSBase(test)
2 fully connected CNN = WOW(6.5), S-UNIWARD(8.1)
layers
1 Softmax

13
7777
Table 3  (continued)
7778

Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

13
Tian et al. (Tian and CNN TLU Spatial BOSSBase v1.01 (10 k WOW Detection accuracy Executes well under
Li 2019) 2018 1 Image processing gray level images) S-UNIWARD Using 0.1bpp on BOSSBase low embedding rates
layer CNN = WOW(67.9), S-UNI- than SRM with EC
5 convolution layers WARD(65.1) and low dimensional
Average pooling SRM = WOW(60.3), S-UNI- SPAM with Gauss-
3 different structures WARD(59.25) ian SVM
corresponding to SPAM = WOW(53.03), S-UNI-
mixed1 mixed2 WARD(55.7)
and mixed3 in Using 0.3bpp on BOSSBase
Inception-V3 CNN = WOW(69.0), S-UNI-
1 fully connected WARD(67.2)
layer SRM = WOW(68.8), S-UNI-
1 Softmax WARD(65.8)
SPAM = WOW(57.8), S-UNI-
WARD(60.)
Using 0.4 bpp on BOSSBase
CNN = WOW(71.4), S-UNI-
WARD(69.8)
SRM = WOW(74.3), S-UNI-
WARD(73.7)
SPAM = WOW(61.8), S-UNI-
WARD(64.9)
Bourmand et al. SRNet ReLU Spatial BOSSBase v1.01 (10 k J-UNIWARD Error Detection for SRnet trained Satisfactory results in
(Bourmand et al. 12 layers (1 to7 layer JPEG gray level images) S-UNIWARD on one algo and tested on other at detecting low false
2019) 2018 uses unpooled BOWS2 (10 k gray HILL 0.4bpp alarm levels
feature maps) level images) WOW WOW = WOW(8.93), HILL(32.28),S-
Average pooling with MiPOD UNIWARD(15.52),MiPOD(28.7)
stride 2 applied to UED HILL = WOW(17.42), HILL(14.14),S-
layer 8 to 11 UNIWARD(10.23),MiPOD(21.80)
Batch normalization S-UNIWARD = WOW(11.02),
BN HILL(24.83),S-
1 fully connected UNIWARD(10.23),MiPOD(21.16)
layer MiPOD = WOW(14.73),
1 Softmax HILL(18.88),S-
UNIWARD(15.59),MiPOD(14.97)
N. Farooq, A. Selwal
Table 3  (continued)
Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

Dengpan et al. CNN ReLU Spatial BOSSBase v1.01 WOW Error Detection Improved performance
(Dengpan et al. 2 convolution layers S-UNIWARD Using 0.2bpp on BOSSBase and higher utilization
2019) 2019 Batch normalization CNN = WOW(29.04), S-UNI- over SRM Ye.net in
BN WARD(36.45) some unique Payload
Maximum pooling SRM = WOW(38.53), S-UNI-
2 fully connected WARD(38.23)
layers Ye.net = WOW(28.08),S-UNI-
2-way Softmax WARD(33.18)
Using 0.4bpp on BOSSBase
CNN = WOW(16.54), S-UNI-
WARD(18.49)
SRM = WOW(28.87), S-UNI-
WARD(28.05)
Ye.net = WOW(20.44),S-UNI-
WARD(23.74)
Ni et al. (Ni et al. DQN Action value function Spatial BOSSBase v1.01 (10 k J-UNIWARD Testing accuracy Accuracy is improved
2019) 2019 Combination of (Reinforcement JPEG gray level images) S-UNIWARD Using 0.1bpp on BOSSBase by 0.34% on average
Q-learning and learning) DQN = J-UNIWARD(52.20)with the to EC when the num-
CNN number of classifier 101, S-UNI- ber of classifiers was
Uses Tensor flow WARD(59.51) with the number of 36 and 0.37% when
to build neural classifier 115 the number of classi-
network part PEP = J-UNIWARD(51.79) with the fiers was 42
number of classifier 101, S-UNI-
WARD(59.21) with the number of
Image steganalysis using deep learning: a systematic review and open research challenges

classifier 115

13
7779
Table 3  (continued)
7780

Author and year Network architecture Activation function/ Domain Database Steganographic Performance error/detection accuracy Analysis
function algorithm

13
Ke et al. (Ke et al. MCNN ReLU BOSSBase v1.01 (10 k HUGO Detection Accuracy Higher performance of
2019) 2019 3 parallel CNNs gray level images) S-UNIWARD Using 0.4 bpp on BOSSBase under detection accuracy
All columns having RAISE (8156 high- clairvoyant scenario and cover source
the same network resolution images) PibreCNN = HUGO(93.03), S-UNI- mismatch than SRM
structure WARD(91.47) with EC. It reduces
No pooling operation MCNN = HUGO(94.83), S-UNI- classification error
Less number of filters WARD(95.13) by 10% or more
1 convolution layer Using 0.1 bpp on BOSSBase under
before fully con- clairvoyant scenario
nected layer Pibre CNN = HUGO(72.17), S-UNI-
1 fully connected WARD(70.35)
layer MCNN = HUGO(77.07), S-UNI-
1 Softmax WARD(74.25)
Using 0.4 bpp on BOSSBase(Training)
and RAISE(testing) undercover
source mismatch
Pibre CNN = HUGO(89.72), S-UNI-
WARD(87.02)
MCNN = HUGO(92.83), S-UNI-
WARD(90.45)
Using 0.1 bpp on BOSSBase(Training)
and RAISE(testing) under cover
source mismatch
Pibre CNN = HUGO(65.46), S-UNI-
WARD(61.45)
MCNN = HUGO(68.87), S-UNI-
WARD(65.05)
Hu et al. (Hu et al. Discriminant model Reward function JPEG BOSSBase v1.01 (10 k J-UNIWARD Error Detection Increase the consist-
2019) 2019 for the brain(CNN) (Reinforcement gray level images) Quality factor = 75 ency of the training
Visual attention learning) Using 0.1bpp on BOSSBase set and eliminate
model for eyes CNN = J-UNIWARD(34.18), unfavorable training
Region generation Xu = J-UNIWARD(34.54) process features and
model for evalua- Using 0.2bpp on BOSSBase enhance accuracy in
tion of minds, CNN = J-UNIWARD(20.93), detection
Selecting and replac- Xu = J-UNIWARD(21.47)
ing the model Quality factor = 95
Using 0.3bpp on BOSSBase
CNN = J-UNIWARD(32.29),
Xu = J-UNIWARD(32.58)
Using 0.4bpp on BOSSBase
CNN = J-UNIWARD(25.31),
Xu = J-UNIWARD(25.86)
N. Farooq, A. Selwal
Image steganalysis using deep learning: a systematic review and open research challenges 7781

comprises of ten thousand 8-bit Portable Gray Map (PGM) ¡. False Alarm (FA): an inaccurate classification of
images having size 512 × 512. Also the BOWS2 (Bas et al. the cover picture C (negative) as the stego image
2011) database that also comprises of ten thousand 8-bit S (positive).
PGM images having size 512 × 512 is used. Another data- ¡¡. Missed Detection (MD): a stego image S was
base that was used is the vast ImageNet (Krizhevsky et al. misclassified as a cover image C, which is known
2017), which consists of over 14 million images of vari- as a Missed Detection.
ous sizes. Most popular ImageNet (Krizhevsky et al. 2017)
• Receiver operating characteristics (ROC) Curve: the
database was usually used for Frequency Domain (JPEG)
effectiveness of a binary classifier is measured using
research. Some of the CNNs Implementation frameworks
the ROC curve. Plotting the True Positive Rate (TPR)
are in TensorFlow (Abadi et al. 2016) which allows flexible
against the False Positive Rate (FPR) at various threshold
and responsive development of CNNs while some frame-
settings results in this curve. The area under curve refers
works implementation uses LIRMM (Chaumont 2019).
to the region beneath the ROC curve (AUC).
From the table it can be observed that the results of the
• Weighted area under curve (WAUC): it is generally
first CNNs were lower than those reported from normal
acknowledged that minimal false-alarm rates should fol-
calculations, but as research advanced on the construction
low through with reliable steganalysis. As a result, after
of new networks or unique computational devices, the out-
normalizing in 0 and 1, more weight is given to the AUC
comes of these CNNs surpassed the results revealed in the
below the true-positive rate threshold than to the region
literature.
above it. The WAUC is the name given to the modified
AUC in steganalysis research.

5 Evaluation methodologies
5.1.2 Visual metrics
In this section, we discuss various protocols that are widely
used for evaluating the efficacy of image steganalysis It mainly includes checking the quality of the image during
approaches. In addition, some open source tools are also image steganalysis process.
discussed for the task of image steganalysis.
• Image Quality: any modification that is done to an image
5.1 Evaluation metrics has the potential to reduce its quality. There are two
methods for evaluating image quality: subjective and
To analyze the performance of image steganalysis objective. Subjective methods operate without clear ref-
approaches, both quantitative as well as visual measures can erence points and are based on human perception. It is
be used to assess steganography and steganalysis techniques possible to use a range of references, including ground
(Singh 2021). truth or previous knowledge while employing objective
procedures because comparisons using explicit numeri-
cal criteria constitute the foundation of these methods.
5.1.1 Quantitative metrics
In this study, the average Peak Signal to Noise Ratio
(PSNR), the average Structural Similarity Index (SSIM),
The performance of a steganalyzer is demonstrated quanti-
the average Multi-scale Structural Similarity Index (MS-
tatively using a number of measures, including the minimal
SSIM) and the average Visual Information Fidelity are a
detection error probability, Receiver Operating Character-
few examples of the objective methods utilized to assess
istics (ROC), area under curve (AUC), and weighted area
the quality of stego pictures (VIF). Between the original
under curve (WAUC).
cover photos and the matching produced stego images,
these metrics are computed.
• Minimum detection error probability: three basic parts
make up a steganalytic detection system: input, stegana-
The confusion matrix forms the basis for many evaluation
lytic detector, and output. (i) The input image can be a
metrics. The most important measurements used in Image
stego image or a cover (actual image) (encrypted image).
Steganalysis are listed in Table 4.
(ii) A model called the steganalytic detector is trained to
determine if a given input image is a cover or a stego.
The determination of whether the provided image is a 5.2 Benchmark datasets
cover or stego is the steganalyzer’s output (iii) One of the
two errors listed below may occur when a model is tried The benchmark datasets are frequently utilized in a vari-
to categorize a given image: ety of fields, including image processing, AI, and cyber

13
7782 N. Farooq, A. Selwal

Table 4  A summary of various performance evaluation metrics for image steganalysis algorithms
Abbreviation Metric Formula Definition

FPR False positive rate FPR = FP The portion of an attack presentation made
TN+FP
using the same type of PAI is incorrectly
classified as a legitimate presentation in a
certain setting
FNR False negative rate BPCER = FN The portion of a legitimate presentation that
TP+FN
is mistakenly labeled as a presentation
attack in a certain setting
ACER Average classification error ACER = APCER+BPCER APCER and BPCER averaged with a
rate 2
defined threshold or limit
ACA​ Average classification accu- ACA = TP+TN To evaluate and contrast detection efficiency
racy 2

DET curve Decision error threshold The graphical display of a biometric sys-
curve/detection error trade- tem’s performance
off curve

ROC curve Receiver operating character- Any classification model’s performance at


istic curve categorization thresholds is represented
graphically

Accuracy Accuracy Accuracy = TP+TN It is described as the proportion of accurate


TP+TN+FP+FN
forecasts to all of the system’s predictions
Precision Precision Precision = TP It is described as the proportion of correctly
TP+FP
positive findings to those that the classifier
correctly predicted
ACR​ Average classification rate ACR = 100%−APCER+BPCER It is described as the typical percentage of
2
test feature vectors with the correct clas-
sification (Stego or cover)
PSNR Peak signal-to-noise ratio Peak 2
PSNR(x, y) = ­10log10( MSE(x,y) ) Is the most popular metric for assessing
stego image quality. However, the most
accurate way to gauge image quality is by
subjective assessment
SSIM Structural similarity index SSIM(x, y) = l(x, y).c(x, y).s(x, y) to evaluate the steganographic image’s level
of imperceptibility
MS-SSIM Multi-scale structural similar- MS-SSIM (x, y) = [lM (x, y)]𝛼M The image is down sampled by a factor of
ity index . i=1 [Ci (x, y)]𝛽i [Si (x, y)]𝛾j
∏M two and low-pass filtering is applied suc-
cessively after taking the original image
and the distorted image as input. The
maximum scale M and scale 1 are both
indexed for the original image

13
Image steganalysis using deep learning: a systematic review and open research challenges 7783

Table 4  (continued)
Abbreviation Metric Formula Definition
VIF Visual information fidelity I(CN,i ;F N,i �SN,i ) It uses a statistical model to determine the

∑i𝜖subbands
VIF = i𝜖subbands I(CN,i ;EN,i �SN,i ) information fidelity for the entire image.
The first is the statistical comparison of
the visual channel’s initial and end stages
without distortion. The second is the
shared information between the distortion
block’s input and the visual block’s output

security. Researchers can test, evaluate, and recreate previ- detect its presence, and assess its reliability using a variety
ously published or proposed techniques using publicly avail- of various adaptable ways. There are numerous other mod-
able datasets, enhancing academic research transparency and ules in VSL, including a number of distortion techniques,
integrity. There are no datasets for image steganography that which can be used to evaluate the steganographic technique’s
we are aware of. As a result, datasets utilized in previous resistance. The program includes built-in modules that aid
research have often been created by the researchers from with file processing, picture analysis, research, reports, and
public domain images. However, since the absence of steg- other tasks.
anographic material is unknown a priori, employing public
domain images poses a risk of image integrity being com- 5.3.2 Stego spy
promised (Woolley et al. 2017).
The publicly available image dataset datasets cited in The StegSpy program can tell which file contains the data
the literature were NRCS (Pevny et al. 2010), Greenspun and what kind of data is stored there. Only stego files made
(Avcibaş et al. 2005), and ImageNet (Krizhevsky et al. with certain steganographic tools, such JPHideandSeek,
2017). Kharrazi et al. (Sencar 2006) utilized search tools to and Invisible Secrets, can be detected by it; it cannot detect
collect openly accessible images from the Internet to down- all stego files. Michael T. Raggo created the free software
load images. Others relied on well-known steganographic program known as StegSpy V2.1. According to them, it is
databases such as the BOWS2 (Bas et al. 2011) and BOSS- able to recognize steganography sequences from Hider-
Base (Bas et al. 2011) databases. Both the origin and the man, JPHideandSeek, Masker, JPegX, and Invisible Secrets.
clarity of the pictures can affect the success of steganaly- StegSpy was demonstrated by the project owner at InfoSec
sis. As a result, several researchers created their new image 2004, BlackHat 2004, and DefCon 2004. StegSpy’s most
datasets, giving them more flexibility over the images they recent version was created in VisualBasic v2.1. The user can
used for their unique objectives, such as using numerous choose a file to review manually using a graphical interface.
image acquisition sources to better reflect real-world sce-
narios (Woolley et al. 2017). The most often used datasets 5.3.3 Steg detect
for image steganalysis by researchers are briefly mentioned
in Table 5. In conjunction with programs like Outguess 0.1 and Invis-
ible Secrets, the free steganalysis tool Stegdetect can find
5.3 Open source software tools the information encoded in a JPEG image. The OS program
Stegdetect was developed by Niels Provos. With the help of
There are several software tools available for the detection jsteg, jphide (Unix and Windows), Invisible Secrets, Out-
of steganographic content, used by researchers and other guess 0.1, 3b, F5 (page header analysis), Appendix, and
interested groups. Some of the tools are open source and camouflage, it may identify certain portions of the message.
few of these are proprietary software. A brief illustration of
these tools is summarized in Table 6 and these are discussed 5.3.4 Steg break
in the following subsection (Hassan et al. 2000).
Stegbreak is software for communication abstraction and
5.3.1 Virtual steganographic laboratory (VSL) not for detecting the existence of steganography. Against
Jsteg-Shell, JPHide, and Outguess, a dictionary attack is
VSL is a graphical block modeling tool that enables the made. The password and dictionary without a doubt is essen-
complicated use, testing, and modifying of algorithms for tial to its success. A key component of ensuring success
both image steganography and steganalysis. In addition to is adhering to the guidelines for rearranging the words in
modular, plug-in designs, VSL offers an easy-to-use GUI. a dictionary. From a legal perspective, the variations may
The application’s purpose is to conceal data in digital files, be the password being used; research might offer hints as

13
Table 5  A comparison among various benchmark datasets for image steganalysis
7784

Dataset Year Size Size of image usage Performed tasks Dataset variant Best model Total Number of studies
bench-

13
marks

MNIST 1998 Lecun et al. 60 K training set Grey level 28 × 28 Classifying images Image Classifica- MNIST Branching/Merging 37 4305
(1998) and 10 K test that are accessi- tion Sequential MNIST CNN + Homo-
set ble from various Sequential Image geneous Vector
places Classification Capsules
CKCNN
BOWS2 2007 Bas et al. 10 K Grey scale Extracting features Image classifica- BOWS2 SRNET, Bourmant – 5 (up to 2018)
(2011) 512 × 512 from text tion Net
MIRFlickr 2008–2010 Huiskes 25 K and 1 M Color vary For the annotation Image classifica- MIR- Teacher-Student – –
et al. (2008) challenge tion FLICKER-25 k
CIFAR-10 2009 Krizhevsky 60 K (10 classes) Color32 × 32 Train ML and CV Image classifica- CIFAR-10 ViT-H/14 47 6555
et al. (2009) algorithms tion CIFAR-10, 4000 LaplaceNet
Semi-Supervised Labels NAT-M4
image classifica- CIFAR-10
tion
Neural Architecture
Search
CIFAR-100 2009 Krizhevsky 60 K (100classes) Color32 × 32 Train massive neu- Image Classifica- CIFAR-100 EffNet-L2 30 3004
et al. (2009) ral networks tion CIFAR-100, 1000 EnAET
Semi-Supervised Labels
Image Classifica-
tion
ImageNet 2009 Fie-Fie et al. 14,197,122 Color vary Image classifica- Image classifica- ImageNet CoAtNet-7 68 6460
(2010) tion and object tion ImageNet Real Meta Pseudo
detection Image classifica- ImageNet-R Labels
tion RVT-B
Domain Generali-
zation
BOSSbase 2010 Bas et al. 10 K Grey scale Consists of cover Image classifica- BOSSbase v1.01 Tan-Net, Qian Net, – 22 (up to 2018)
(2011) 512 × 512 images only that tion Xu-Net
can be used for
testing purpose
Flickr30k 2014 Young et al. 31 K Color vary Automatic image Node Classification Flickr GCN-GAugM 9 355
(2014) depiction and Image Retrieval Flickr30k and 1k SGRAF
word recognition test
on the basis of
context
COCO 2015 Lin et al. 328 k Color vary Image recognition, Object Detection COCO minival Dual-Swin-L 60 4433
(2014) classification, Keypoint Detection COCO test-dev HRNet
key-point detec- Real-Time Object COCO minival Mask R-CNN
tion, and labeling Detection X-152–32 × 8d
dataset
N. Farooq, A. Selwal
Image steganalysis using deep learning: a systematic review and open research challenges 7785

Table 6  A summary of steganalysis software applications or tools


Number of studies
Name Open source/propri- Developed/maintained
etary By

Virtual stegano- Proprietary Michal Wegrzyn


graphic laboratory

(VSL)
StegSpy Open source Michael T. Raggo
bench-
marks
Total

StegDetect Open source Niels Provos


StegBreak Open source Niels Provos


Stego suite Proprietary Wetstone
Stego hunter Proprietary Wetstone
Best model

Stego watch Proprietary Wetstone


Stego ananlyst Proprietary Wetstone
Stego break Proprietary Wetstone

StegSecret Open source Alfonso Munoz


Ben-4D Open source Alexzaharis
Dataset variant

LIRMM

to the passwords. Stegbreak’s vocabulary can be expanded


to include these suggestions, such as name, birthday, pet’s
name, etc. For Stegbreak to work, a technician must demon-
strate that the deleted bit string has an embedded message
Image classifica-
Performed tasks

rather than just a sound.

5.3.5 Stego suite
tion

The previous 10 years have seen significant advancements


in the technologies used to digitally handle image, video,
and audio data, which has sped up the retention of infor-
Demosaicing

mation within binary files. Many websites provide “stego”


programs that are available for download. The potential
usage

for business espionage, trade secret theft, cyber-weapons


trading, and criminal collaboration is limitless. In order to
detect the usage of digital steganography in all of its forms,
Size of image

numerous governments and commercial institutions are


Color/Grey
512 × 512
256 × 256

looking for technologies. A tool called Stego Suite can be


used against a target file to find steganography even when
there is no previous knowledge of its presence. The Suite of
Stego software is the only one with this capability, known as
“blind” detection of steganography. Active online browsing
is becoming more and more important to prevent unwanted
users from using a website to send innocent video messages
1008
Size

or steal information about the property. Using a web domain,


the Stego Suite program continuously searches a domain
for the existence of confidential data in digital photographs.
LIRMMBase 2016 Chaumont

A digital crime scene’s following data processing might be


a time-consuming operation. When we consider numerous
(2019)

photographs, audio files, or movies that might provide veri-


Table 5  (continued)
Year

fied evidence, the complexity is also raised. Tools from the


Stego Suite assist the assessment and evaluation of such
time-consuming computing operations used for private data
Dataset

or communications. A monitoring/scanning service that is


made available to others as well as standalone software made

13
7786 N. Farooq, A. Selwal

up of three different types of applications is called Stego pairs, and more. We have included filter options that change
Suite. Customers who want to leave the management of the photos into one of three distinct presentations: intensity,
the programme in the hands of Wetstone personnel experts Saturation, or Hue, allowing investigators to search for fur-
should use the selection authority. Customers who can con- ther signs that steganography is being used. Other filtering
duct the steganography detection function on their own options only show particular LSBs of chosen hues. Since
systems utilize the software. Both options include technical many steganographic methods rely on LSBs to conceal data,
assistance as well as team knowledge from the Stego team. examining the LSBs of an image may occasionally show
One of the most sophisticated commercial steganalysis tools signs of steganography.
is Stego Suite, which can detect the presence of data univer-
sally stored in sound and picture without needing to know
the storage mechanism. Steganalysis can be performed on 5.3.9 Stego break
BMP, GIF, PNG, JPG, and WAV files. Stego Suite, which
is available from Whetstone Technologies, comprises the A built-in tool called Stego Break can be used to discover
PW cracker Stego Break as well as the detection tools Stego the pass phrase used on a file that has been identified as
Watch and Analyst. At the 2005 USA conference, they also containing steganography. Popular password dictionaries
suggest providing training on how to use these technologies are provided with the tool purchase in order to carry out a
collaboratively. dictionary attack. Investigators can also use alternative dic-
tionaries or, if the password has been revealed during ques-
tioning of the subject, they can check the password against
5.3.6 Stego hunter
the identified image or audio recordings.
As a first step in the investigative process, Stego Hunter is
made to detect steganography applications fast, correctly, 5.3.10 StegSecret
and simply. How do you know whether steganography exists
is a common question we receive. Any installed or even pre- A steganalysis open source project called Stegsecret (GNU/
viously installed application’s results can be quickly relayed GPL) makes it possible to find concealed data in various
back to the investigator using Stego Hunter. We go a step digital media. With the use of the most well-known steg-
further than that and identify the suspected carrier types so anographic techniques, concealed information can be found
you know what to search for to continue your investigation. using the multiplatform java-based steganalysis programme
In addition to EnCase, FTK, dd, raw, ISO, and safe back called StegSecret. It can recognise DCTs, LSB, EOF, and
images, we are able to scan forensic photos created by other other techniques. The goal of the StegSecret project is to
well-known forensic programs. gather, apply, and simplify the use of steganalysis tech-
niques, particularly in digital media like as photos, audio,
5.3.7 Stego watch and video. This initiative seeks to raise awareness of the
flaws in a number of steganographic tools and algorithms
The cutting-edge steganography detection program we use that are readily available online.
is called Stego Watch. When suspect carrier files are discov-
ered, Stego Watch may instantly search the entire file sys-
5.3.11 Ben‑4D
tem and return results with suspected files highlighted. Files
that have been marked are found using a blind detection
It is a simple and precise way to identify stego-carrier files
technique that scans the media files for artifacts. A threat
among a group of files. To determine whether a file is a
association and notification of any artifacts discovered are
stego-carrier, the fundamentals of Benford’s Law distribu-
then shown. No prior programming experience with steg-
tion are generalized and applied to the suspected file. It is
anography is required.
utilized by LSB, Invisible Secrets v4.0, and Camouflage
v1.2.1 to identify JPHSWin. It employs JPEGSnoop to scan
5.3.8 Stego analyst certain files or folders and produce reports.

The full-featured imaging and analysis tool Stego Analyst


enables researchers to look for visual indicators that steg-
anography has really been used across both audio and image
files. A file viewing window is included, which shows the
particular file’s image or audio wave as well as the features
of the file, such as image specifics, DCT coefficients, color

13
Image steganalysis using deep learning: a systematic review and open research challenges 7787

6 Open research issues and opportunities larger cover, stronger detection techniques and extended
image features are required for the detection of low payload
Employing deep learning models for image steganalysis has embedding in either case (Al-Manaseer and Al-Jarrah 2019).
opened an interesting way to carry out steganalysis. A large Few CNN focused strategies even can’t have the option to
number of researcher’s attention is attracted to it and gained meet some prospects. The presentation is restricted simply
incredible growth. Though, the present investigation on deep depending on the network framework to detect the poor
learning-based image steganalysis is still in its early stages. signals and depict its complicated measurable (statistical)
There exist different research issues in using DL techniques attributes, also some different estimates, for example, train-
for image steganalysis for detection and retrieval of the hid- ing sample collection, learning methodologies, and different
den information, some of them are discussed in this section. advancements must have to increase the detection accuracy.
In this manner, it is necessary to attempt to join the network
6.1 Open research issues of system model plan, training sample collection, and learn-
ing techniques to support the execution of steganographic
Our analysis of various existing DL-based steganalysis SOA image detection having low payload.
methods has raised some potential problems that are listed
as the key open research issues as follows:
6.1.3 Generalization of steganalysis
6.1.1 Detection and depiction of statistical characteristics
Deep learning-based image steganalysis strategies are pri-
of stego signals
marily in the research center stage that is not the same as
the real application stage comparatively. These distinctions
The hidden content is exceptionally weak with respect to the
introduce a few issues, for instance, steganographic tech-
image content in a steganographic image. Also, the hidden
niques mismatch, image source mismatch, payload mismatch
content is extremely complex and contrasts from various
which reduces the precision of steganalysis by a substan-
images. Although several innovative steganographic tech-
tial factor and thereby affects its applicability. Therefore,
niques are developed which attempt to protect the statistical
it is important to improve the flexibility of steganalysis to
characteristics of the image, and these methods correspond-
embedding rate, image source, and steganalysis algorithm
ingly grow with upcoming or existing steganalysis meth-
that makes it imperative for researchers to pay attention
ods. Thus, making it challenging task for steganalyzers to
towards the generalization of deep learning-based stega-
identify the presence of secret content in the cover image.
nalysis. Although the generalization of steganalysis can be
These developing steganographic methods, algorithms, and
viewed as the problem of domain adaption, which is a part of
techniques are respectively upgrading or changing with the
transfer learning, it represents ways to deal with consolidate
latest steganalysis techniques making it tough to detect and
transfer learning with deep learning-based steganalysis to
depict the characteristics of steganographic signal. Albeit
advance its generalization.
a few research works have examined this issue and gained
some ground, which concludes that it still needs to improve
further. At the same time, as deep learning technology keeps 6.1.4 Quantitative and detecting image steganalysis
on growing, it is very important to know how to apply these dependent on deep learning
most recent innovations (for example, Generative Adver-
sarial Networks (GAN) (Gonog and Zhou 2019), transfer Presently, the deep learning-based steganalysis techniques
learning) to steganalysis, detection and depict of the char- are at a point for identifying the presence of image stega-
acteristics of weak stego signal is still an issue which needs nalysis, yet less research work is done on quantitative and
to be studied in detail. detection of image steganalysis, which are critical issues in
the field. Thus, the use of deep learning technology is worth
6.1.2 Steganographic image detection with low payload exploring these problems.

Payload is defined as the secret message to be inserted into


an image (Chakraborty et al. 2013). One of the interesting 6.2 Research opportunities and future directions
problems in steganalysis is low payload image detection.
This is because low payload steganography has a slight vari- The observations drawn from this review of DL-based Image
ation in the cover, resulting more hard to catch the statis- steganalysis and research challenges for further research are
tical characteristics of these steganographic signals. Even listed below.
when the payload proportion is decreased by utilizing a

13
7788 N. Farooq, A. Selwal

a. To develop new CNNs bringing together the benefits of such size may be appropriate for traditional handcrafted
existing networks or create altogether new frameworks, approaches, but it is inadequate for training deep-learning
(shallow and additionally deeper structures), to advance model. An additional problem involves that, it does not
the detection rates, both in the spatial and JPEG domain. include stego images; therefore, an equivalent stego image
b. Use diverse computerized image databases, taking into corresponding for each cover image is obtained manu-
account, for instance, the utilization of various cameras, ally. The training overhead may be effectively improved
to test more simulations and research rigorously about developing lightweight CNNs or using pre-trained mod-
Cover-Source Mismatch effect. els through the notion of transfer learning. To tackle the
c. To train existing CNNs with enormous scale databases problem of scarcity of dataset, data augmentation meth-
and bigger image sizes. To do this, it is vital to prepare ods are employed. Most recently, Bashir et al. (Bashir and
a CPU and GPU bunch design so as to satisfy the needs Selwal 2021) proposed model named as DeepSteg that is
of preparing and memory. comparatively lightweight with lower number of trainable
d. To train the CNNs using a specified steganographic parameters and it uses data augmentation via operations
algorithm and test on other algorithms in order to exam- such as flipping, scaling, rotation, and etc. The type of
ine the transfer rate among different algorithms. steganography algorithms also plays significant role while
e. To develop new CNNs and computational components evaluating a steganalysis model. Various underlying steg-
that permit to acquire the noise created by the steganog- anography algorithms (i.e. HUGO, WOW, HILL, S-UNI-
raphy procedure in a more efficient manner, to increase WARD, J-UNIWARD) are available for converting a cover
the illustration of characteristics, to categorize images image to its corresponding stego image. Previously, (Ayub
in JPEG or spatial domain. and Selwal 2020) presented a more efficient steganography
f. To measure the performance of filters, such as high pass algorithm that hides data in DCT domain.
filter that are used in the pre-processing phase in cor-
relation with the activation functions employed in deep
learning based steganalysis.
7 Conclusions
Therefore, it is observed that, there are diverse possibili-
ties for future work that spurs researchers to keep working in In this study, we explored and analyzed various DL-based
this direction and motivates new researchers who are inter- steganalysis techniques that exist in the literature. We
ested in DL-based steganalysis. observed that, it is difficult to ascertain that the specific
steganalysis technique is preferably used over others, as the
6.3 Our contributions in image steganalysis steganographic methods employed in concealing the secret
information are diverse. We classified the different stega-
The image steganalysis is a significant research area that nalysis techniques and also discussed the significance of
has drawn the attention of many researchers. According deep learning for steganalysis purpose. We have identified
to a recent work (Taneja et al. 2021), the paradigm has critical research problems in the area of deep learning-based
shifted from traditional handcrafted features to modern steganalysis and also highlighted the possible viable solu-
DL-based in the steganalysis approaches. The success of tions to overcome these issues in further research. DL-based
these mechanisms is mainly dependent on more accurate steganalysis techniques are used to identify the presence of
representation of discriminative image features and clas- concealed messages inserted in the spatial domain and fre-
sifiers. The emergence of CNN models have facilitated quency domain. Even though advancement is made in the
the task of automatic and deep-level features extraction field of image steganalysis by utilizing deep learning, there
for more accurate steganalysis methods with an additional are as yet numerous issues that need to be addressed before
training overhead. In our recently published brief study DL can be efficiently used for the task of steganalysis. Some
(Bashir and Selwal 2021), we have highlighted various of the prevailing issues include detection and depiction of
critical issues and further scope in the deep learning-ena- complex statistical features of weak stego signals, detection
bled image steganalysis approaches. It is clear that major- of low payload steganographic images, steganalysis gener-
ity of the deep learning-based rely on large-scale datasets alization, quantitative image steganalysis, should be settled.
and scarcity of datasets for building such models is a key It is predicted that substantial improvement will be accom-
challenge. Furthermore, the training overhead to learn plished if numerous constructs, for example, network model,
these models requires higher training overhead. This work training sets determination along with learning scheme are
has clearly inferred that majority of the image steganalysis considered when deep learning advancements are applied
methods utilized BOSSbase v1.01 dataset that contains to image steganalysis. More research must focus towards
only 10,000 images of size 512 × 512. However, dataset of designing more efficient and lightweight steganalysis models

13
Image steganalysis using deep learning: a systematic review and open research challenges 7789

by exploring recent conception of domain transfer, active Data availability The datasets generated during and/or analyzed dur-
learning and incremental learning. We need to design and ing the current study are available from the corresponding author on
reasonable request.
develop robust steganalysis models that operate efficiently in
unknown environments (i.e. cross-dataset and cross-sensor Declarations
scenario).
Conflict of interest The authors declare that they do not have any con-
flict of interest.

Appendix
References
See (Table 7).
Abadi M, Barham P, Chen J, et al (2016) TensorFlow: A system for
large-scale machine learning. Proc 12th USENIX Symp Oper
Table 7  Some acronyms and their descriptions Syst Des Implementation, OSDI 2016 265–283
Abiodun OI, Jantan A, Omolara AE et al (2018) State-of-the-art
Abbreviation Full description in artificial neural network applications: a survey. Heliyon
4:e00938. https://​doi.​org/​10.​1016/j.​heliy​on.​2018.​e00938
AI Artificial intelligence Al-Manaseer RM, Al-Jarrah MM (2019) Steganalysis of Color
ANN Artificial neural network Images for Low Payload Detection. 35–38. https://​doi.​org/​10.​
1145/​33839​13.​33839​15
AUC​ Area under curve
Alom MZ, Taha TM, Yakopcic C, Westberg S, Sidike P, Nasrin MS,
ACR​ Average classification rate et al (2018) The history began from AlexNet: a comprehensive
ACER Average classification error rate survey on deep learning approaches. arXiv:​1803.​01164
CNN Convolution neural network Amina S, Mubeena AK, Shireen MT (2021) AAA​AAA​A CNN and
BILSTM based Image Steganalysis. 9:63–67
CV Computer vision
Avcibaş I, Kharrazi M, Memon N, Sankur B (2005) Image steganaly-
DL Deep learning sis with binary similarity measures. EURASIP J Appl Signal
DBN Deep belief network Process 2005:2749–2757. https://​doi.​org/​10.​1155/​ASP.​2005.​
DCT Discrete cosine transformation 2749
Ayub N, Selwal A (2020) An improved image steganography tech-
DNN Deep neural network
nique using edge based data hiding in DCT domain. J Inter-
EC Ensemble classifier discip Math 23:357–366. https://​doi.​org/​10.​1080/​09720​502.​
FNN Fully connected neural network 2020.​17319​49
FPR False positive rate Bae H, Lee B, Kwon S, Yoon S (2019) DNA steganalysis using deep
recurrent neural networks. Pac Symp Biocomput 24:88–99.
FNR False negative rate
https://​doi.​org/​10.​1142/​97898​13279​827_​0009
GPU Graphics processing units Bao Y, Yang H, Yang Z, et al (2020) Text steganalysis with atten-
HPF High pass filter tional L STM-CNN. 2020 5th Int Conf Comput Commun Syst
LSB Least significant bit ICCCS 2020. 138–142. https://​doi.​org/​10.​1109/​ICCCS​49078.​
2020.​91184​10
LSTM Long short term memory networks
Bao H, Dong L, Wei F (2021) BEiT: BERT Pre-Training of Image
MLP Multi-layer perceptron Transformers. 1–16
PGM Portable gray map Baragada SR, Ramakrishna S, Rao MS, Purushothaman S (2008)
PSNR Peak signal-to-noise ratio Implementation of radial basis function neural network for
image steganalysis. Int J Comput Sci Secur 2:12–22
RBF Radial basis function
Bas PL, Filler T, Pevný T (2011) “Break our steganographic sys-
RBFN Radial basis function network tem”: the ins and outs of organizing BOSS. Lect Notes Com-
RBM Restricted Boltzmann machine put Sci (Including Subser Lect Notes Artif Intell Lect Notes
ReLU Rectified linear unit Bioinformatics) 6958 LNCS:59–70. https://​doi.​org/​10.​1007/​
978-3-​642-​24178-9_5
RNN Recurrent neural network
Bashir B, Selwal A (2021) Towards deep learning-based image
RM Rich models steganalysis: practices and open research issues. SSRN Elec-
ROC Receiver operating curve tron J. https://​doi.​org/​10.​2139/​ssrn.​38833​30
SSIM Structural similarity index Bashkirova D (2016) Convolutional neural networks for image stega-
nalysis. Bionanoscience 6:246–248. https://​doi.​org/​10.​1007/​
SNR Signal to noise ratio
s12668-​016-​0215-z
SOA State-of-the-art Boroumand M, Chen M, Fridrich J (2019) Deep residual network for
SPAM Subtractive pixel adjacency matrix model steganalysis of digital images. IEEE Trans Inf Forensics Secur
SRM Spatial rich models 14:1181–1193. https://​doi.​org/​10.​1109/​TIFS.​2018.​28717​49
SVM Support vector machines Broda M, Levicky D, Banoci V, Bugar G (2014) Universal image
steganalytic method based on binary similarity measures. 2014
TLU Truncated linear unit 24th Int Conf Radioelektronika, RADIOELEKTRONIKA 2014
VIF Visual information fidelity - Proc 3–6. https://​doi.​org/​10.​1109/​Radio​elek.​2014.​68284​67

13
7790 N. Farooq, A. Selwal

Chakraborty S, Jalal AS, Bhatnagar C (2013) Secret image sharing Guo L, Ni J, Su W et al (2015) Using statistical image model for
using grayscale payload decomposition and irreversible image JPEG steganography: uniform embedding revisited. IEEE
steganography. J Inf Secur Appl 18:180–192. https://​doi.​org/​ Trans Inf Forensics Secur 10:2669–2680. https://​doi.​org/​10.​
10.​1016/j.​istr.​2013.​02.​006 1109/​TIFS.​2015.​24738​15
Chang CC, Lin CJ (2011) LIBSVM: a library for support vector Hasoon S, Khalifa F (2012) Steganalysis using KL transform and
machines. ACM Trans Intell Syst Technol 2:1–39. https://​doi.​ radial basis neural network. AL-Rafidain J Comput Sci Math
org/​10.​1145/​19611​89.​19611​99 9:47–58. https://​doi.​org/​10.​33899/​csmj.​2012.​163670
Chaumont M (2019) Deep Learning in steganography and steganaly- Hassan M, AMIN M, Mahdi S, “Steganalysis Techniques and Com-
sis from 2015 to 2018 To cite this version : HAL Id : lirmm- parison of Available Softwares,” no. 2000, 2020. https://​doi.​
02087729 Deep Learning in steganog-raphy and steganalysis org/​10.​4108/​eai.​28-6-​2020.​22979​70
from He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for
Chen J (2013) Anti PoV-steganalysis data hiding algorithm. ICICS image recognition. Proc IEEE Comput Soc Conf Comput Vis
2013—Conf Guid 9th Int Conf Information, Commun Signal Pattern Recognit 2016-Decem:770–778. https://​d oi.​o rg/​1 0.​
Process 0–4. https://​doi.​org/​10.​1109/​ICICS.​2013.​67829​15 1109/​CVPR.​2016.​90
Chhikara R, Singh L (2013) A review on digital image steganaly- Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural
sis techniques categorised by features extracted. Image (IN) Comput 9:1735–1780. https://​doi.​org/​10.​1162/​neco.​1997.9.​8.​
3:203–213 1735
Chollet F (2017) Xception: Deep learning with depthwise separa- Holub V, Fridrich J (2012) Designing steganographic distortion using
ble convolutions. Proc—30th IEEE Conf Comput Vis Pattern directional filters. WIFS 2012—Proc 2012 IEEE Int Work Inf
Recognition, CVPR 2017 2017-Janua:1800–1807. https://​doi.​ Forensics Secur 234–239. https://​doi.​org/​10.​1109/​WIFS.​2012.​
org/​10.​1109/​CVPR.​2017.​195 64126​55
Couchot J-F, Couturier R, Guyeux C, Salomon M (2016) Steganaly- Holub V, Fridrich J, Denemark T (2014) Universal distortion function
sis via a Convolutional Neural Network using Large Convolu- for steganography in an arbitrary domain. Eurasip J Inf Secur
tion Filters for Embedding Process with Same Stego Key. 1–24 2014:1–13. https://​doi.​org/​10.​1186/​1687-​417X-​2014-1
Cramer R, Damgård IB, Nielsen JB (2001) Multiparty computa- Hu Y, Wen G, Luo M, Dai, D, Ma J, Yu Z (2018) Competitive Inner-
tion from threshold homomorphic encryption. In: Pfitzmann imaging squeeze and excitation for residual network. arXiv:1​ 807.​
B (ed) EUROCRYPT 2001. LNCS, vol. 2045, Springer, Hei- 08920
delberg, pp 280–299 Hu D, Zhou S, Shen Q et al (2019) Digital image steganalysis based on
Dengpan Y, Shunzhi J, Shiyu L, ChangRui L (2019) Faster and trans- visual attention and deep reinforcement learning. IEEE Access
ferable deep learning steganalysis on GPU. J Real-Time Image 7:25924–25935. https://d​ oi.o​ rg/1​ 0.1​ 109/A
​ CCESS.2​ 019.2​ 90007​ 6
Process. https://​doi.​org/​10.​1007/​s11554-​019-​00870-1 Hu J, Shen L, Albanie S et al (2020) Squeeze-and-excitation networks.
Djebbar F, Ayad B, Meraim KA, Hamam H (2012) Comparative IEEE Trans Pattern Anal Mach Intell 42:2011–2023. https://​doi.​
study of digital audio steganography techniques. Eurasip J org/​10.​1109/​TPAMI.​2019.​29133​72
Audio Speech Music Process 2012:1–16. https://​doi.​org/​10.​ Huang G, Liu Z, Van Der Maaten L, Weinberger KQ (2017) Densely
1186/​1687-​4722-​2012-​25 connected convolutional networks. Proc—30th IEEE Conf Com-
Ekblad U, Kinser JM, Atmer J, Zetterlund N (2004) The intersecting put Vis Pattern Recognition, CVPR 2017 2017-Janua:2261–
cortical model in image processing. Nucl Instrum Methods 2269. https://​doi.​org/​10.​1109/​CVPR.​2017.​243
Phys Res Sect A 525(1–2):392–396 Huiskes MJ, Lew MS (2008) The MIR Flickr retrieval evaluation. Proc
Farsi H, Shahi A (2014) Steganalysis of images based on spatial 1st Int ACM Conf Multimed Inf Retrieval, MIR2008, Co-located
domain and two-dimensional JPEG array. J Chin Inst Eng with 2008 ACM Int Conf Multimedia, MM’08 39–43. https://d​ oi.​
Trans Chin Inst Eng A 37:1055–1063. https://​doi.​org/​10.​1080/​ org/​10.​1145/​14600​96.​14601​04
02533​839.​2014.​929711 Hussain M, Wahab AWA, Bin IYI et al (2018) Image steganography in
Fei-Fei L, Deng J, Li K (2010) ImageNet: constructing a large-scale spatial domain: a survey. Signal Process Image Commun 65:46–
image database. J Vis 9:1037–1037. https://​doi.​org/​10.​1167/9.​ 66. https://​doi.​org/​10.​1016/j.​image.​2018.​03.​012
8.​1037 Jain T (2020) Spatial domain steganography techniques and neural
Fischer A, Igel C (2012) An introduction to restricted Boltzmann network based steganalysis with differential storage. Int Conf
machines. Lect Notes Comput Sci (Including Subser Lect Emerg Trends Inf Technol Eng Ic-ETITE 2020:1–4. https://​doi.​
Notes Artif Intell Lect Notes Bioinformatics) 7441 LNCS:14– org/​10.​1109/​ic-​ETITE​47903.​2020.​416
36. https://​doi.​org/​10.​1007/​978-3-​642-​33275-3_2 Jia C, Yang Y, Xia Y, Chen YT, Parekh Z, Pham H et al (2021) Scal-
Fridrich J, Goljan M (2002) <title>Practical steganalysis of digital ing up visual and vision-language representation learning with
images: state of the art</title>. Secur Watermarking Multimed noisy text supervision. In: International Conference on Machine
Contents IV 4675:1–13. https://​doi.​org/​10.​1117/​12.​465263 Learning, pp 4904–4916. PMLR
Fridrich J, Kodovsky J (2012) Rich models for steganalysis of digital JinaChanu Y, Manglem Singh K, Tuithung T (2012) Image steganog-
images. IEEE Trans Inf Forensics Secur 7:868–882. https://​doi.​ raphy and steganalysis: a survey. Int J Comput Appl 52:1–11.
org/​10.​1109/​TIFS.​2012.​21904​02 https://​doi.​org/​10.​5120/​8171-​1484
Fridrich J, Goljan M, Du R (2001) Reliable detection of LSB steg- Kadhim IJ, Premaratne P, Vial PJ, Halloran B (2019) Comprehensive
anography in color and grayscale images. Proc ACM Int Mul- survey of image steganography: techniques, evaluations, and
timed Conf Exhib 27–30. https://​d oi.​o rg/​1 0.​1 145/​1 2324​5 4.​ trends in future research. Neurocomputing 335:299–326. https://​
12324​66 doi.​org/​10.​1016/j.​neucom.​2018.​06.​075
Ghanbari S, Keshtegary M, ghanbari N (2012) New steganalysis Karampidis K, Kavallieratou E, Papadourakis G (2018) A review of
method using GLCM and neural network. Int J Comput Appl image steganalysis techniques for digital forensics. J Inf Secur
42:46–50. https://​doi.​org/​10.​5120/​5708-​6266 Appl 40:217–235. https://​doi.​org/​10.​1016/j.​jisa.​2018.​04.​005
Gonog L, Zhou Y (2019) A review: Generative adversarial networks. Kaur M, Kaur G (2014) Review of various steganalysis. Techniques
Proc 14th IEEE Conf Ind Electron Appl ICIEA 2019 505–510. 5:1744–1747
https://​doi.​org/​10.​1109/​ICIEA.​2019.​88336​86

13
Image steganalysis using deep learning: a systematic review and open research challenges 7791

Ke Y, Liu J, Zhang MQ et al (2018) Steganography security: principle Liu F, Yan X, Lu Y (2020) Feature selection for image steganalysis
and practice. IEEE Access 6:73009–73022. https://​doi.​org/​10.​ using binary bat algorithm. IEEE Access 8:4244–4249. https://​
1109/​ACCESS.​2018.​28816​80 doi.​org/​10.​1109/​ACCESS.​2019.​29630​84
Ke Q, Ming LD, Daxing Z (2019) Image steganalysis via multi-column Lubenko I, Ker AD (2012) Steganalysis with mismatched covers. Do
convolutional neural network. Int Conf Signal Process Proceed- simple classifiers help?. In: Proceedings of the on Multimedia
ings, ICSP 2018-Augus:550–553. https://​doi.​org/​10.​1109/​ICSP.​ and security. pp 11–18
2018.​86523​24 Maeda-Gutiérrez V, Galván-Tejada CE, Zanella-Calzada LA et al
Ker AD (2007) A fusion of maximum likelihood and structural stega- (2020) Comparison of convolutional neural network architectures
nalysis. Lect Notes Comput Sci (Including Subser Lect Notes for classification of tomato plant diseases. Appl Sci. https://​doi.​
Artif Intell Lect Notes Bioinformatics) 4567 LNCS:204–219. org/​10.​3390/​app10​041245
https://​doi.​org/​10.​1007/​978-3-​540-​77370-2_​14 Marius-Constantin P, Balas VE, Perescu-Popescu L, Mastorakis N
Khan A, Sohail A, Zahoora U, Qureshi AS (2020) A survey of the (2009) Multilayer perceptron and neural networks. WSEAS
recent architectures of deep convolutional neural networks. Artif Trans Circuits Syst 8:579–588
Intell Rev. https://​doi.​org/​10.​1007/​s10462-​020-​09825-6 Ni D, Feng G, Shen L, Zhang X (2019) Selective ensemble classifi-
Kim DH, Lee HY (2018) Deep learning-based steganalysis against cation of image steganalysis via deep Q network. IEEE Signal
spatial domain steganography. Proc—2017 Eur Conf Electr Eng Process Lett 26:1065–1069. https://​doi.​org/​10.​1109/​LSP.​2019.​
Comput Sci EECS 2017 1–4. https://​doi.​org/​10.​1109/​EECS.​ 29130​18
2017.9 Nissar A, Mir AH (2010) Classification of steganalysis techniques: a
Kim J, Park H, Il PJ (2020) CNN-based image steganalysis using addi- study. Digit Signal Process A Rev J 20:1758–1770. https://​doi.​
tional data embedding. Multimed Tools Appl 79:1355–1372. org/​10.​1016/j.​dsp.​2010.​02.​003
https://​doi.​org/​10.​1007/​s11042-​019-​08251-3 Pan Y, Ni J, Su W (2016) Improved uniform embedding for effi-
Kim JT, Kim S, Kim K (2019) A Study on Improved JPEG Steganog- cient JPEG steganography. Lect Notes Comput Sci (Includ-
raphy Algorithm to Prevent Steganalysis. ICTC 2019—10th Int ing Subser Lect Notes Artif Intell Lect Notes Bioinformatics)
Conf ICT Converg ICT Converg Lead Auton Futur 960–963. 10039:125–133. https://​doi.​org/​10.​1007/​978-3-​319-​48671-0_​
https://​doi.​org/​10.​1109/​ICTC4​6691.​2019.​89397​63 12
Kodovský J, Fridrich J, Holub V (2012) Ensemble classifiers for Park JS, Kim HG, Kim DG et al (2018) Paired mini-batch training: a
steganalysis of digital media. IEEE Trans Inf Forensics Secur new deep network training for image forensics and steganaly-
7:432–444. https://​doi.​org/​10.​1109/​TIFS.​2011.​21759​19 sis. Signal Process Image Commun 67:132–139. https://​doi.​
Kolesnikov A, Beyer L, Zhai X et al (2020) Big transfer (BiT): general org/​10.​1016/j.​image.​2018.​04.​015
visual representation learning. Lect Notes Comput Sci (Including Pathak P, Selvakumar S (2014) Blind Image steganalysis of JPEG
Subser Lect Notes Artif Intell Lect Notes Bioinformatics) 12350 images using feature extraction through the process of dila-
LCNS:491–507. https://​doi.​org/​10.​1007/​978-3-​030-​58558-7_​29 tion. Digit Investig 11:67–77. https://​doi.​org/​10.​1016/j.​diin.​
Krizhevsky A (2009) Learning multiple layers of features from tiny 2013.​12.​002
images. Master’s thesis, University of Tront Paulin C, Selouani SA (2016) Hervet E (2016) Speech steganaly-
Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classifica- sis using evolutionary restricted Boltzmann machines. IEEE
tion with deep convolutional neural networks. Commun ACM Congr Evol Comput CEC 2016:4831–4838. https://​doi.​org/​10.​
60:84–90. https://​doi.​org/​10.​1145/​30653​86 1109/​CEC.​2016.​77444​09
Kumar N, Kumar D (2020) Classification using artificial neural net- Pevny T, Bas P, Fridrich J, et al (2010) Steganalysis by Subtractive
work optimized with bat algorithm. Int J Innov Technol Explor Pixel Adjacency Matrix To cite this version : HAL Id : hal-
Eng 9:696–700. https://​doi.​org/​10.​35940/​ijitee.​c8378.​019320 00541410 Steganalysis by Subtractive Pixel Adjacency Matrix
L R, B L (2017) Approaches and methods for steganalysis—a survey. Pevný T, Filler T, Bas P (2010) Using high-dimensional image mod-
Ijarcce 6:433–438. https://​doi.​org/​10.​17148/​ijarc​ce.​2017.​6678 els to perform highly undetectable steganography. Lect Notes
Laskar SA, Hemachandran K (2014) A review on image steganalysis Comput Sci (Including Subser Lect Notes Artif Intell Lect
techniques for attacking steganography. Int J Eng Res Technol Notes Bioinformatics) 6387 LNCS:161–177. https://​doi.​org/​
3:3400–3410 10.​1007/​978-3-​642-​16435-4_​13
LeCun Y, Jackel LD, Bottou L et al (1995) Learning algorithms for Pham H, Dai Z, Xie Q, Luong M-T, Le QV (2021) Meta pseudo
classification: a comparison on handwritten digit recognition. labels. Mach Learn. https://​d oi.​o rg/​1 0.​4 8550/​a rXiv.​2 003.​
Neural Netw Stat Mech Perspect 261:276 10580
Lecun Y, Bottou L, Bengio Y, Ha P (1998) Gradient-based learning Pibre L, Pasquet J, Ienco D, Chaumont M (2016) Deep learning is
applied to document recognition. Proc IEEE 86:1–46 a good steganalysis tool when embedding key is reused for
Li B, Wang M, Huang J, Li X (2014) A new cost function for spatial different images, even if there is a cover sourcemismatch. IS
image steganography. 2014 IEEE Int Conf Image Process ICIP T Int Symp Electron Imaging Sci Technol. https://​doi.​org/​10.​
2014 4206–4210. https://​doi.​org/​10.​1109/​ICIP.​2014.​70258​54 2352/​ISSN.​2470-​1173.​2016.8.​MWSF-​078
Lin TY, Maire M, Belongie S et al (2014) Microsoft COCO: common Powell BA (2020) Securing LSB embedding against structural stega-
objects in context. Lect Notes Comput Sci (Including Subser Lect nalysis. J Computer Secur 30:1–11
Notes Artif Intell Lect Notes Bioinformatics) 8693:740–755. Qian Y, Dong J, Wang W, Tan T (2015) Deep learning for stega-
https://​doi.​org/​10.​1007/​978-3-​319-​10602-1_​48 nalysis via convolutional neural networks. Media Watermark-
Lin Z, Huang Y, Wang J (2018) RNN-SM: fast steganalysis of VoIP ing Secur Forensics 9409:94090J. https://​doi.​org/​10.​1117/​12.​
streams using recurrent neural network. IEEE Trans Inf Forensics 20834​79
Secur 13:1854–1868. https://​doi.​org/​10.​1109/​TIFS.​2018.​28067​ Qian Y, Dong J, Wang W, Tan T (2016) Learning and transferring
41 representations for image steganalysis using convolutional
Lipton ZC, Berkowitz J, Elkan C (2015) A critical review of recur- neural network. Proc—Int Conf Image Process ICIP 2016-
rent neural networks for sequence learning. arXiv preprint arXiv:​ Augus:2752–2756. https://​doi.​org/​10.​1109/​ICIP.​2016.​75328​60
1506.​00019 Qian Y, Dong J, Wang W, Tan T (2018) Feature learning for steganal-
ysis using convolutional neural networks. Multimed Tools Appl
77:19633–19657. https://​doi.​org/​10.​1007/​s11042-​017-​5326-1

13
7792 N. Farooq, A. Selwal

Rasras R, AlQadi Z, Sara M (2019) A methodology based on steg- Tian J, Li Y (2019) Convolutional neural networks for steganalysis via
anography and cryptography to protect highly secure messages. transfer learning. Int J Pattern Recognit Artif Intell. https://​doi.​org/​
Eng Technol Appl Sci Res 9:3681–3684. https://​doi.​org/​10.​ 10.​1142/​S0218​00141​95900​67
5281/​zenodo.​25762​30 Touvron H, Vedaldi A, Douze M, Jégou H (2019) Fixing the train-test
Ruan F, Zhang X, Zhu D et al (2019) Deep learning for real-time resolution discrepancy. Adv Neural Inf Process Syst 32:1–14
image steganalysis: a survey. J Real-Time Image Process. Wang F, Jiang M, Qian C, et al (2017) Residual attention network for
https://​doi.​org/​10.​1007/​s11554-​019-​00915-5 image classification. Proc-30th IEEE Conf Comput Vis Pattern
Sabnis SK, Awale RN (2016) Statistical steganalysis of high capac- Recognition, CVPR 2017 2017-Janua:6450–6458. https://​doi.​org/​
ity image steganography with cryptography. Procedia Comput 10.​1109/​CVPR.​2017.​683
Sci 79:321–327. https://​doi.​org/​10.​1016/j.​procs.​2016.​03.​042 Wang J, Sun K, Cheng T et al (2020) Deep high-resolution representation
Sabour S, Frosst N, Hinton GE (2017) Dynamic routing between learning for visual recognition. IEEE Trans Pattern Anal Mach
capsules. Adv Neural Inf Process Syst 2017:3857–3867 Intell 43:3349–3364. https://​doi.​org/​10.​1109/​tpami.​2020.​29836​86
Sajid Ansari A, Sajid Mohammadi M, Tanvir Parvez M (2019) A Westfeld A (2001) F5-a steganographic algorithm high capacity despite
comparative study of recent steganography techniques for mul- better steganalysis. Lect Notes Comput Sci (Including Subser Lect
tiple image formats. Int J Comput Netw Inf Secur 11:11–25. Notes Artif Intell Lect Notes Bioinformatics) 2137:289–302
https://​doi.​org/​10.​5815/​ijcnis.​2019.​01.​02 Westfeld A, Pfitzmann A (2000) Attacks on Steganographic Systems.
Sandler M, Howard A, Zhu M, et al (2018) MobileNetV2: Inverted Springer, Berlin, pp 61–76. https://​doi.​org/​10.​1007/​10719​724_5
Residuals and Linear Bottlenecks. Proc IEEE Comput Soc Woo S, Park J, Lee JY, Kweon IS (2018) CBAM: Convolutional block
Conf Comput Vis Pattern Recognit 4510–4520. https://​d oi.​ attention module. Lect Notes Comput Sci (Including Subser Lect
org/​10.​1109/​CVPR.​2018.​00474 Notes Artif Intell Lect Notes Bioinformatics) 11211 LNCS:3–19.
Sedighi V, Cogranne R, Fridrich J (2016) Content-adaptive steg- https://​doi.​org/​10.​1007/​978-3-​030-​01234-2_1
anography by minimizing statistical detectability. IEEE Trans Woolley C, Ibrahim A, Hannay P (2017) Building a dataset for image
Inf Forensics Secur 11:221–234. https://​doi.​org/​10.​1109/​TIFS.​ steganography. Proc 15th Aust Digit Forensics Conf ADF 2017
2015.​24867​44 36–45. https://​doi.​org/​10.​4225/​75/​5a83a​0541d​284
Sencar HT (2006) Performance study of common image steganography Wu S, Zhong S, Liu Y (2017) Deep residual learning for image
and steganalysis techniques. J Electron Imaging 15:041104. https://​ steganalysis. Multimed Tools Appl. https://​doi.​org/​10.​1007/​
doi.​org/​10.​1117/1.​24006​72 s11042-​017-​4440-4
Simonyan K, Zisserman A (2015) Very deep convolutional networks for Xie C, Tan M, Gong B, et al (2020a) Adversarial examples improve
large-scale image recognition. 3rd Int Conf Learn Represent ICLR image recognition. Proc IEEE Comput Soc Conf Comput Vis Pat-
2015—Conf Track Proc 1–14 tern Recognit 816–825. https://​doi.​org/​10.​1109/​CVPR4​2600.​2020.​
Singh B (2021) Spatial-domain image steganalysis using deep learning 00090
techniques Xie Q, Luong MT, Hovy E, Le Q V. (2020b) Self-training with noisy
Surse NM, Vinayakray-Jani P (2018) A comparative study on recent student improves imagenet classification. Proc IEEE Comput Soc
image steganography techniques based on DWT. Proc 2017 Int Conf Comput Vis Pattern Recognit 10684–10695. https://​doi.​org/​
Conf Wirel Commun Signal Process Networking, WiSPNET 2017 10.​1109/​CVPR4​2600.​2020.​01070
2018-Janua:1308–1314. https://​doi.​org/​10.​1109/​WiSPN​ET.​2017.​ Xu G (2017) Deep convolutional neural network to detect J-UNIWARD.
82999​75 IH MMSec 2017—Proc 2017 ACM Work Inf Hiding Multimed
Szegedy C, Liu W, Jia Y, et al (2015) Going deeper with convolutions. Secur 67–73. https://​doi.​org/​10.​1145/​30820​31.​30832​36
Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit 07-12- Xu G, Wu HZ, Shi YQ (2016) Structural design of convolutional neural
June:1–9. https://​doi.​org/​10.​1109/​CVPR.​2015.​72985​94 networks for steganalysis. IEEE Signal Process Lett 23:708–712.
Szegedy C, Vanhoucke V, Ioffe S, et al (2016) Rethinking the Inception https://​doi.​org/​10.​1109/​LSP.​2016.​25484​21
Architecture for Computer Vision. Proc IEEE Comput Soc Conf Yang J, Liu K, Kang X, Wong E, Shi Y (2017) Steganalysis based on
Comput Vis Pattern Recognit 2016-Decem:2818–2826. https://​doi.​ awareness of selection-channel and deep learning. In: Digital
org/​10.​1109/​CVPR.​2016.​308 forensics and watermarking: 16th international workshop, IWDW
Szegedy C, Ioffe S, Vanhoucke V, Alemi AA (2017) Inception-v4, incep- 2017, Magdeburg, Germany, August 23–25, 2017, Proceedings 16,
tion-ResNet and the impact of residual connections on learning. Springer, pp 263–272
31st AAAI Conf Artif Intell AAAI 2017 4278–4284 Yang Z, Wang K, Li J et al (2019) TS-RNN: text steganalysis based
Tabares-Soto R, Ramos-Pollán R, Isaza G, Orozco-Arias S, Ortíz MAB, on recurrent neural networks. IEEE Signal Process Lett 26:1743–
Arteaga HBA (2020) Digital media steganalysis. In: Digital Media 1747. https://​doi.​org/​10.​1109/​LSP.​2019.​29204​52
Steganography, Academic Press, pp 259–293, Academic Press Yang H, Bao Y, Yang Z, et al (2020) Linguistic Steganalysis via Densely
Tan S, Li B (2014) Stacked convolutional auto-encoders for steganalysis Connected LSTM with Feature Pyramid. IH MMSec 2020—Proc
of digital images. 2014 Asia-Pacific Signal Inf Process Assoc Annu 2020 ACM Work Inf Hiding Multimed Secur 5–10. https://​doi.​org/​
Summit Conf APSIPA 2014. https://​doi.​org/​10.​1109/​APSIPA.​ 10.​1145/​33694​12.​33950​67
2014.​70415​65 Ye J, Ni J, Yi Y (2017) Deep learning hierarchical representations for
Taneja K, Taneja H, Kumar K et al (2021) Data science and inno- image steganalysis. IEEE Trans Inf Forensics Secur 12:2545–2557.
vations for intelligent systems. Data Sci Innov Intell Syst. https://​doi.​org/​10.​1109/​TIFS.​2017.​27109​46
https://​ d oi.​ o rg/​ 1 0.​ 1 201/​ 9 7810​ 0 3132​ 0 80/​ DATA-​ S CIEN​ Yedroudj M, Comby F, Chaumont M (2018) Yedroudj-Net: An Efficient
C E - ​ I N N O V​ AT I O N S - ​ I N T E L ​ L I G E N T- ​ S YS T E ​ M S -​ CNN for Spatial Steganalysis. ICASSP, IEEE Int Conf Acoust
KAVITA-​TANEJA-​HARMU​NISH-​TANEJA-​KULDE​ Speech Signal Process-Proc 2018-April:2092–2096. https://​doi.​
EP-​KUMAR-​ARVIND-​SELWAL-​ENG-​LIEH-​OUH org/​10.​1109/​ICASSP.​2018.​84614​38
Tang YH, Jiang LH, He HQ, Dong WY (2018) A review on deep learning Young P, Lai A, Hodosh M, Hockenmaier J (2014) From image descrip-
based image steganalysis. Proc 2018 IEEE 3rd Adv Inf Technol tions to visual denotations. Trans Assoc Comput Linguist 2:67–78
Electron Autom Control Conf IAEAC 2018 1764–1770. https://​ Zeiler MD, Fergus R (2014) Visualizing and understanding convolutional
doi.​org/​10.​1109/​IAEAC.​2018.​85776​55 networks. Lect Notes Comput Sci (Including Subser Lect Notes
Artif Intell Lect Notes Bioinformatics) 8689 LNCS:818–833.
https://​doi.​org/​10.​1007/​978-3-​319-​10590-1_​53

13
Image steganalysis using deep learning: a systematic review and open research challenges 7793

Zeng J, Tan S, Li B, Huang J (2018) Large-scale JPEG image steganalysis Zou Y, Zhang G, Liu L (2019) Research on image steganography analysis
using hybrid deep-learning framework. IEEE Trans Inf Forensics based on deep learning. J vis Commun Image Represent 60:266–
Secur 13:1200–1214. https://​doi.​org/​10.​1109/​TIFS.​2017.​27794​46 275. https://​doi.​org/​10.​1016/j.​jvcir.​2019.​02.​034
Zhang X, Li Z, Loy CC, Lin D (2017) PolyNet: A pursuit of structural
diversity in very deep networks. Proc-30th IEEE Conf Comput Vis Publisher's Note Springer Nature remains neutral with regard to
Pattern Recognition, CVPR 2017 2017-Janua:3900–3908. https://​ jurisdictional claims in published maps and institutional affiliations.
doi.​org/​10.​1109/​CVPR.​2017.​415
Zhang R, Zhu F, Liu J, Liu G (2018) Efficient feature learning and multi- Springer Nature or its licensor (e.g. a society or other partner) holds
size image steganalysis based on CNN. 1–10 exclusive rights to this article under a publishing agreement with the
Zheng L, Zhang Y, Thing VLL (2018) Understanding multi-layer percep- author(s) or other rightsholder(s); author self-archiving of the accepted
trons on spatial image steganalysis features. Proc-9th Asia-Pacific manuscript version of this article is solely governed by the terms of
Signal Inf Process Assoc Annu Summit Conf APSIPA ASC 2017 such publishing agreement and applicable law.
2018-Febru:1035–1039. https://​doi.​org/​10.​1109/​APSIPA.​2017.​
82821​81

13

You might also like