Received December 27, 2020, accepted January 7, 2021, date of publication January 25, 2021, date of current version

February 10, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3053998

Image Steganography: A Review

of the Recent Advances
NANDHINI SUBRAMANIAN 1 , (Member, IEEE), OMAR ELHARROUSS1 ,
SOMAYA AL-MAADEED 1 , (Senior Member, IEEE), AND
AHMED BOURIDANE 2 , (Senior Member, IEEE)
1 Department of Computer Science and Engineering, Qatar University, Doha, Qatar
2 Department of Computer and Information Sciences, Northumbria University, Newcastle upon Tyne NE1 8ST, U.K.
Corresponding author: Nandhini Subramanian (ns1808900@qu.edu.qa)
This work was supported by the Qatar National Research Fund (a member of Qatar Foundation) under Grant NPRP11S-0113-180276.
Open Access funding provided by the Qatar National Library.

ABSTRACT Image Steganography is the process of hiding information which can be text, image or video
inside a cover image. The secret information is hidden in a way that it not visible to the human eyes.
Deep learning technology, which has emerged as a powerful tool in various applications including image
steganography, has received increased attention recently. The main goal of this paper is to explore and
discuss various deep learning methods available in image steganography field. Deep learning techniques used
for image steganography can be broadly divided into three categories - traditional methods, Convolutional
Neural Network-based and General Adversarial Network-based methods. Along with the methodology,
an elaborate summary on the datasets used, experimental set-ups considered and the evaluation metrics
commonly used are described in this paper. A table summarizing all the details are also provided for easy
reference. This paper aims to help the fellow researchers by compiling the current trends, challenges and
some future direction in this field.

INDEX TERMS Image steganography, GAN steganography, CNN steganography, information hiding,
image data hiding.

I. INTRODUCTION original data is not visible, however, cipher text is visible in

Technology has blitz scaled over the past years leading to a scrambled form to human eyes leading to suspicion and
a wide usage of multimedia for transferring data, especially further scrutiny. A new research topic, steganography, has
Internet of Things (IoT). Usually, the transfer happens over gained acceptance in this context to hide the data that is not
insecure network channels. In particular, the internet has perceptible to human eyes.
gained accelerated popularity for exchanging digital media Information hiding techniques have been available for a
and individuals, private companies, institutions, governments long time but their importance has been increasing recently.
use these multimedia data transfer methods for exchanging The main reason is the increase in the data traffic through the
data. Though there are numerous advantages attached with internet and social media networks. Though the objectives
it, one prominent disadvantage is the privacy and security of of cryptography and steganography are similar, there is a
the data. The availability of numerous readily available tools subtle difference. Cryptography makes the data unbreakable
capable of exploiting the privacy, data integrity and security and unreadable but the cipher text is visible to human eyes.
of the data being transmitted has made the possibility of Steganography, which is used to hide the information in plain
malicious threats, eavesdropping and other subversive activ- sight, allows the use of wide variety of the secret information
ities. The prominent solution is data encryption where the forms like image, text, audio, video and files. Digital water-
data is converted into a cipher text domain using encryption marking is another method where confidential information is
key. At the receiving end, the cipher text is converted into embedded to claim ownership. Cryptography is the popular
plain text using a decryption key. Using data encryption the method used for information hiding, but, steganography is
gaining popularity in recent times.
The associate editor coordinating the review of this manuscript and Steganography can be defined as the process of hiding a
approving it for publication was Li He . secret small multimedia data inside another but much larger

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 23409
 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

multimedia data such as image, text, file or video [1]. Image

steganography is a technique to hide an image inside another
image. In image steganography, the cover image is manipu-
lated in such a way that the hidden data is not visible thus
making it not suspicious as in the case of cryptography.
Inversely, Steganalysis is used to detect the presence of any
secret message covered in the image and to extract the hidden
data [2]. Steganalysis helps in classifying if the image is either
a stego image or a normal image. Apart from classifying
the image, further investigation is carried out to detect the FIGURE 1. General working principle of steganography and steganalysis.
location and the content of the secret image inside the cover The inputs are the cover image and the secret information and an
image. embedding algorithm is used to generate the container stego image. The
extraction algorithm takes the stego image as input to extract the
With the availability of massive amounts of data, deep ingrained secret information.
learning (DL) has become the trend and is extensively used
for many applications. Deep learning is a useful tool in var-
ious applications like image classification, automatic speech either text or image. DL model can be either a CNN-based
recognition, image recognition, natural language processing, or GAN-based. While the steganography block generates the
recommendation systems, processing of medical images [3]. stego image, the steganalysis model takes the stego image as
Though research on steganography is quite recent, it has input to detect and perhaps extract the secret information.
benefited from DL methods including convolutional Neural In some methods the probability score of the input image
Networks (CNNs) Generative Adversarial Networks (GANs) being normal or stego image is given as the output.
based methods and their deployment in both steganography Text data, color or grayscale images are usually used as
and steganalysis. secret media. Two factors - the nature of the secret media,
The main goal of this paper is to review the available and the technique used are considered in classifying the works
methodologies, present trends and discuss the challenges available to different categories as can be seen in the figure 2.
that are currently available in the studies. Along with these From the analysis on figure 2, it is observed that text is
studies, the datasets that are publicly available and commonly the most commonly used secret information and GAN-based
used, the evaluation metrics considered are also discussed. methods are the preferred mode for secret communication for
Finally, a comparison on the performance among the methods text hiding.
and a possible discussion identifying the gaps in the present
studies, pros and cons of the methods are elaborated. A. TRADITIONAL-BASED STEGANOGRAPHY METHODS
The remaining paper is organized as follows. Section II Conventionally, Least Significant Bits (LSB) substitution
summarizes the working principle of the methods grouped method is employed to perform image steganography. Images
into three categories - Traditional methods, CNN-Based are usually of higher pixel quality, out of which not all the
methods and GAN-Based methods. Datasets used com- pixels are used. LSB methods works under the assumption
monly are elaborated in section III along with evaluation that modifying a few pixel values would not show any visible
in section IV. A table with the comparisons of the results changes. The secret information is converted into a binary
from the different methods are provided with the experimen- form. The cover image is scanned to determine the least sig-
tal set-ups generally used in section IV. Finally, the chal- nificant bits in the noisy area. The binary bits from the secret
lenges faced, a brief discussion, and conclusion are added in image are then substituted in the LSBs of the cover image.
section V, VI and section VII respectively. The substitution method has to be performed cautiously as
overloading the cover image may lead to visible changes
II. SUMMARY OF THE METHODS leaking the presence of the secret information [4] and [5].
After reviewing all the frameworks available, the method- With the LSB method as the baseline, a number of related
ologies are primarily grouped into three categories, namely, methods have been proposed. For example, a slight varia-
traditional image steganography methods, CNN-based image tion in converting the secret message into binary codes is
steganography methods and GAN-based image steganogra- undertaken in [6]. A Huffman encoding method is used to
phy methods. Traditional methods are frameworks which use encode the secret message into the binary bits. The encoded
methods that are not related to machine learning or deep bits are then embedded in the cover image using the LSB
learning algorithms. Many traditional methods are based on method. In [7], another version of the LSB method is used
the LSB technique. CNN-based methods are based on deep for RGB images. The cover image is in 3 channels and they
convolutional neural networks for embedding and extracting are bit sliced. The secret message is embedded in all the three
the secret messages and GAN-based methods use some of the planes in the 2:2:4 ratio for R, G and B planes. Not only
GAN variants. Figure 1 gives an overview of a steganography spatial domain, quantum images are also used [8] and [9]. The
and steganalysis architecture. As shown in figure 1, inputs frequency domain is exploited in quantum image domain and
are cover image and the secret information which can be the pixels which are considered to be affecting the color are

23410 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

FIGURE 2. Classification of the existing methods based on the secret media and the method used. First, the methods are classified and
under each category, further classification is done based on the secret media.

used to hide the secret bits. A combination of cryptography method is used. The secret data is encoded first and then
and steganography is utilized where the LSB of the cover converted to binary bits before embedding it in the cover
image is replaced with the most significant bits of the secret video.
image [10]. The pseudo random number generator is used to Table 6 depicts the detailed review on the traditional image
select the pixels and the key is encrypted using rotation every steganography methods. Another classical method used in
time. A k-LSB method is proposed where the k least bits image steganography field is the Pixel Value Differencing
are replaced with the secret message [11]. For steganalysis, (PVD). PVD works by taking the difference of consecutive
an entropy filter is used to detect and uncover the secret pixels to find the locations for hiding the secret bits in such
image [11]. a way the consistency of the cover image is maintained. For
The LSB methods are used in hiding the secret information every 8 bits, a combination of LSB on the first two bits and
inside videos also. Videos are sequences of images called PVD on the remaining six bits is designed [18]. In addition,
the video frames. Each video is dissected into image frames some other techniques used are the coverless steganography
and the binary bits of the secret information is hidden in where the cover image is not given rather it is generated based
the LSB of the image frames of the video. A basic form on the secret information. The secret information is taken
of LSB substitution method [12] and a combination of the and a relationship management is performed to produce the
huffman encoding and LSB substitution methods is used on cover image using the object detection method. Similar cov-
videos [13]. Another interesting approach is where along erless steganography is proposed where the Local Binary Pat-
with the image frames of the video, audio is also used to terns (LBP) features of the cover image and the secret images
enhance the hiding [14]. Besides the LSB methods, [15], are hashed firstly. Later, the hashes are matched to create the
has proposed a combination of Discrete Cosine Transfor- stego image [19]. Similarly, instead of LBP, the edges of the
mation (DCT) and Discrete Wavelet Transformation (DWT) color cover images are obtained. Then, the binary bits of
for hiding the secret message inside a cover video. To find the secret information is hidden in the edges discovered in
the regions of interest, the multiple object tracking (MOT) the cover images [20].

VOLUME 9, 2021 23411

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

TABLE 1. Summary of the details on the traditional methods.

Other methods of the traditional steganography methods A Separable Convolution with Residual Block (SCR)
are described below. Medical JPEG images are used and is used to concatenate the cover image and the secret
the embedding is implemented by taking the difference in image [25]. The embedded image is given as the input to the
the DCT coefficients of the pixels between two consecutive encoder for constructing the stego image which is fed to the
blocks by considering the pixels at the same positions in the decoder to output the decoded secret image. ELU (Exponen-
two blocks [21]. A novel method called the Pixel Density tial Linear Unit) and Batch normalization are used. A new
Histogram (PHD) is proposed for halftone images [22]. The cost function to reduce the effect of noise in the generated
pixel density of the images are calculated and a pixel den- container image called the variance loss is proposed [26].
sity histogram is formed to hide the secret information. The An encoder-decoder architecture was proposed by
Poisson distribution is utilized to get the burst error and the Rahim et al. in [29]. This method differs from the others in
reconstruction of the images that are compression-resistant the way the inputs are given. The encoder part consists of two
is done using the STC decoding [23]. For more clarifi- parallel architectures each for the cover and the secret image.
cations and explanations on the traditional steganography Features from the cover image and the secret images are
methods, [24] can be referred. extracted through the convolutional layer and concatenated.
The concatenated features are used to construct the stego
image.
B. CNN-BASED STEGANOGRAPHY METHODS A slightly different approach is proposed by Wang et al.
Image steganography using CNN models is heavily inspired in [30] by considering the styling image along with the
from the encoder-decoder architecture. Two inputs – cover secret information and cover image. The created stego image
image and the secret image are fed as the input to the encoder is converted into the style image given as an input. The
to generate the stego image and the stego image is given as reveal network is used to decode the secret information
input to the decoder to output the embedded secret image. from the stego image created. Similar to other methods,
The basic principle is the same except different methods have an auto encoder-decoder architecture with VGG as base is
tried different architectures. The way the input cover image used. An arbitrary image size for the secret information and
and the secret image are concatenated are also different in styling image are taken using the adaptive instance normal-
different approaches while the variations in the convolutional ization (AdaIN) layer and the output is the size of the cover
layer, pooling layer are expected. The number of filters used, image. Pixel distribution of the cover image are got by using
strides, filter size, activation function used and loss function a pixel CNN by Yang et al. in [31]. The secret information
vary from method to method. One important point to note is then embedded in the pixel distribution evenly by reduced
here is the size of the cover image and the secret image has sampling.
to be same, so every pixel of the secret image is distributed in Three networks namely prep-network, hiding network
the cover image. and reveal network are proposed by Baluja et al. in [32]
Wu et al. [25] and [26] have proposed a encoder-decoder based on auto-encoder architectures as a single model. The
architecture. U-Net based encoder-decoder architecture is Prep-network is used to prepare the secret image before
used for hiding and a CNN with 6 layers for extraction is feeding it as input to the hiding network, which takes the
proposed by Duan et al. in [27]. The input shape of the output of the prep-network and the cover image to produce the
U-Net is modified to accept 256 × 256 and 6 channels. The container image. The reveal network decodes the secret image
secret and cover images are concatenated to give the input from the container image by uncovering the cover image.
and hence 6 channels. A U-net based Hiding (H-net) and Two losses are calculated between the cover and the con-
revealing (R-net) network are used by Van et al. in [28]. Batch structed container and between secret and the decoded image.
normalization and ReLU activation are used. The cover and The model is evaluated using the structure similarity index
the secret images are concatenated before being sent to the (SSIM). An extension to [32] is proposed by Zhang et al.
network. Two optimization losses using SSIM and MSE are in [33] where the cover image is converted into an YCrCb
used to reduce the loss and hence improve the performance. image format and the secret image is hidden to only the

23412 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

TABLE 2. Summary of the details on the CNN-based steganography methods.

Y channel as all the semantic and color information are with minimum noise. The discriminator model is trained to
present in the Cr and Cb channels. Also, to reduce the effectively find out the fake images. Many variations on GAN
payload by two thirds, the secret image is converted into have been proposed ever since, making it more powerful and
grayscale image format. The Y channel of the cover image suitable for synthetic image generative tasks.
and the grayscale secret image are given as the input to the GANs are known for their good performance in the image
encoder-decoder network for constructing the stego image. generation field. Image steganography can be considered as
The output stego image is the Y channel which is combined one such image generation task where two inputs – the cover
with the Cr and Cb channels of the cover image to produce image and the secret image are given to generate one output –
the stego image in the YCrCb color space. To extract the stego image. The existing methods used for image steganog-
secret image, the Y channel of the stego image is again raphy using a GAN architecture can be grouped into five
given to the revealing network to output the grayscale secret categories - a three network based GAN model, cycle-GAN
image. Two different variations – basic and residual models based architectures, sender-receiver architecture using GAN,
are also used in the generative models. A mixed loss function coverless model where the cover image is generated ran-
which is more suitable for the steganography using SSIM and domly instead of being given as input and an Alice, Bob and
its variant multi-scale structure similarity index (MS-SSIM) Eve based model. Details on all the categories and how they
are also used. For the concealing network, ISGAN architec- are executed are given below.
ture is used. Table 2 summarizes the review on CNN-based Generally, a GAN model consists of two main components:
steganography methods. the generator and the discriminator. In the context of image
Not only image steganography, but also, video steganog- steganography, a new network named the steganalyzer is
raphy is tried using CNN. Usually, 2D convolutional layers introduced in some of the methods. The main functions of
are used for images whereas 3D convolutional layers are these three components are,
used for videos. Temporally connected cover and secret video • A generator model, G, to generate stego images from the
frames are given as the input to autoencoder network based cover image and the random message.
VStegNET [34] to produce the container video. Each frame of • A discriminator model, D, to classify the generated
the cover image is concatenated with every frame of the secret image from the generator as either real or fake.
video to produce the container video. An identical network • A steganalyzer, S, to check if the input image has a
architecture is used to reveal the hidden secret video. confidential secret data or not.
The three models, G, D and S are made to compete against
C. GAN-BASED STEGANOGRAPHY METHODS each other to produce realistic images close to the cover
General Adversarial Networks are a type of deep CNNs image given as input. The errors of D and S along with
introduced by Goodfellow et al. [35] in 2014. A GAN uses the parameter alpha between [0, 1] are used to produce the
the game theory to train a generative model with adversarial realistic image and their quality for steganalysis. One main
process for image generation tasks. Two networks – generator difference from the GAN here is that G is updated in order to
and discriminator networks compete against each other to maximize not only the error of D, but to maximize the error
generate a perfect image in GAN architecture. The generator of the linear combination of the classifiers D and S.
model is given the data and the output is the close approxi- Volkhonskiy et al. [36], [37] has introduced DCGAN [38]
mation of the given input image. The discriminator networks based Steganographic GAN (SGAN) which is a simple
classifies the images generated as either fake or true. The DCGAN with three modules - G, D and S. Similar to [36],
two networks are trained in such a way that the generator Shi et al. [2] and [39] has proposed a three component GAN
model tries to imitate the input data as close as possible architecture. The only difference is the architecture used by

VOLUME 9, 2021 23413

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

FIGURE 3. General working overflow of SGAN and [40]. The generator takes the cover image and the secret information (usually text) as
the input to produce the stego image. The discriminator competes against the generator to classify the generated image as either real or
fake. A steganalyzer is added to extract the secret information from the stego image generated.

these methods. A DCGAN is used in [36], and [37] whereas a double tanh function [42] and a tanh function [41] because
four fractionally convolutional layer followed by a functional the tanh is differentiable and will preserve the gradient loss
layer with Hyperbolic tangent activation with base WGAN during back propagation. The discriminator acts as a stegan-
is used in [2] and [39]. Three models compete against each alyzer with multiple high pass filters for pre-processing to
other in such a way that the generator produces the stego improve the detection accuracy. There are two losses to be
images and the discriminator decodes and recovers the secret considered here – the discriminator loss and the generator
message while the stegananlyzer eavesdrops on the generator loss. The discriminator loss is utilized to distinguish between
produces the probability. Figure 3 represents the working the cover image and the stego image while the generator loss
principle of SGAN. considers the target payload also. Therefore, the generator
An embedding simulator is sometimes used in the place loss is made up of two losses – the entropy loss to guarantee
of the steganalyzer [41], [42] and [43]. Yang et al. [41] the embedding capacity and the adversarial loss to increase
and [42] presented a GAN based image steganography with the anti-detectability rate.
three modules – Generator, Embedding Simulator and Dis- Automatic Steganographic Distortion Learning framework
criminator. The authors in [41] have called the architecture with GAN (ASDL-GAN for short) was introduced by Tang
a UT-GAN since they used the U-Net based architecture in et al. in [43]. In this architecture, the generator is used to learn
the generator network. The generator is used to create the the probabilities for each pixel from the input cover image and
probability map, P, for the input cover image, C and U-Net is the authors have proposed to use a novel activation function
considered because of its effective performance in pixel wise Ternary Embedding Simulator (TES) for generating the stego
segmentation. The U-Net used here also has an expanding images from the generated probabilities. The discriminator
and contracting convolutional layers with increasing filter helps in differentiating between the real and fake images.
number for convolution and decreasing filter numbers for XuNet based architecture is used for discriminator D.
deconvolution layers. Also, the layers I from the down sam- A sender-receiver type of architecture is another way
pling part is concatenated with the layers L - i from the up for image steganography using GAN [33], [43] and [44].
sampling part to help in backpropagation process. The prob- SteganoGAN, a GAN model with encoding network for
ability map, P, from the generator and the random message creating the stego images and decoding network to get the
are given as the input to the embedding simulator to output secret message from the stego image has been proposed by
the modification map, M. The embedding simulator uses a Zhang et al. in [45]. The secret message is embedded in

23414 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

the cover image using the encoder and the secret message is adversarial training is generated. The input image is first
recovered back by the decoder and a critic is used to evaluate converted to a target domain image and then back to the
the quality of the images generated. Three variants of the source image eradicating the necessity for output image. The
encoder – basic, residual and dense are also discussed to deal original cycleGAN method has been modified a little to fit
with different payload capacity. the image steganography methodology perfectly for hiding
Three models – one basic model and two enhanced models secret message inside the cover image [49]–[51], and [52].
are introduced by Chen et al. in [43]. All the three models Firstly, the RGB cover image is converted to grayscale image
are intended to perform steganography using the encoding and then the luma regions in the image are extracted [49].
network and the safe retrieval of the secret image using decod- The secret message is embedded in the LSB bit of the luma
ing network where the enhanced models are more secure field. The generator then creates the new image with the
and robust. The basic model uses the encoding network to secret message embedded. The generated image and the cover
hide grayscale secret image into the color channel B of the image are given to the discriminator to classify it either fake
cover image to produce the stego image. In the same way, or real. A denoiser is introduced between the two generators
the decoding network reveals the secret image from the color of the cycle GAN to reduce and filter the low amplitude,
channel B of the stego image. The reasons for using color high frequency messages and noise that the discriminator
B channel is that the impact of blue color on human eyes is cannot see [52]. A cycle GAN is used for steganography and
less than red and green. A steganalysis network and attack steganalysis in a covert communication to prevent the security
network are added in the basic model to make it more secure breach and privacy preservation in IoT by Meng et al. in [52].
and robust in enhanced models. XuNet with a Spatial Pyramid ACGAN [53] can generate realistic images for a given
Pooling layer is used in the steganalysis network. label and also recognize the label of the generated images.
Generator model is used as sender and the discriminator Three steps are followed to hide information which are gen-
is used as the receiver by Naito et al [44]. The generator erated by the model and extract the hidden information [54]
in the sending end generates the stego images and the dis- and [55]. First, a word segmentation dictionary and image
criminator eliminates the images that are not realistic. At the database is established. A generative hiding model named
receiving end, the discriminator classifies the sender of the Stego-ACGAN is developed. Finally, a hiding and extract-
stego image received to being true sender or third party to ing algorithm for hiding and retrieving the information is
prevent the generator from wasting time on third party stego designed. A database is set-up pairing the word segmentation
images. The generator in the receiving end decodes the stego as the label and the corresponding images. Like in any other
image and uncovers the secret data. Both the sender and the ACGAN, stego-ACGAN also has three neural networks,
receiver shares the same trained generator and discriminator namely. The generator, the discriminator and the auxiliary
for consistent results. classifier. After training the stego-ACGAN, a secret channel
HIDDeN is a GAN based method with four main com- is used to share the model parameters and the constructed
ponents – encoder, decoder, discriminator and a noise layer image and word segment database to the other party. The
proposed by Zhu et al. [46]. The encoder takes the cover secret information is divided into segments and a binary code
image and the secret message as input and creates an is generated. Each code is then given a label and the mode
encoded image which is fed to the noise layer to produce generates a sequence of images for the segments. The image
the noised image. The decoder decodes the secret message sequences and the input noise are given to the generative
from the noised image and the discriminator helps in giving model to generate the stego images. At the receiving end,
the probability of the given image being encoded or not. the noise is removed and the image sequences are extracted.
Ke et al. [47] have proposed a slightly different architecture Then the image sequence is given to the auxiliary classifier
by using a Generative steganography with Kerckhoffs’ prin- to get the secret information.
ciple (GSK). Instead of modifying the cover image and con- Similar to [54] and [55], the authors in [56] Duan et al.
verting it into stego image, a new stego image with the secret have proposed a coverless steganographic method using a
message is generated. The extraction key and the generated generative model. Instead of transmitting the secret image as
stego image are given to the discriminator for decoding. With- such, a new meaning-normal image totally not related to the
out extraction key the discriminator cannot decode the secret cover image is generated using the generative model. At the
image. The generator is publicly available, the sender first receiving end, the transmitted meaning-normal image is fed
gives the cover image to get the extraction key. Then the cover to produce the secret image. A WGAN [57] is used as the
image and the extraction key are sent to the receiver. In the generative model by Li et al. in [58]. A framework where a
case where the cover image or extraction key are sent alone, textural image is generated by a generative model and acts
only noise is given as output. The discriminator can output the as a cover image is proposed. Then, this cover image and the
secret message only when both the image and the extraction secret image is given to the concealing network for hiding the
key are present. secret image inside the cover image. So the final image is a
CycleGAN [48] is well-suited for image steganography texture based image concealed with secret information.
where the input image is given and the output similar to An adversarial learning based method with three compo-
the given input image but with hidden information using the nents – Alice, Bob and Eve is proposed by Hayes et al. [40].

VOLUME 9, 2021 23415

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

TABLE 3. Summary of the details on the GAN-based steganography methods.

The general working principle is that Eve eavesdrops between same secret message is given twice. This diverts any suspicion
Alice and Bob to check if there are any secret message created and enhances the security of the model. Eve and Dev
embedded in the communication channel between them. The gets both the real image and the generated image. While Eve
authors in [40] have used neural networks to train all the three helps in distinguishing the stego and the cover image, Dev
components. For the steganographic scenario, Alice is trained helps in classifying the image as real or fake. A detailed
to create the steganographic image while Bob recovers the summary of the methods reviewed under GAN-based method
secret message from the stego images. Eve helps Bob by topic is given in table 3.
giving the probability of the given image being a stego image.
A model with four parts – Alice, Bob, Dev and Eve has III. DATASETS USED
been proposed by Wang et al. in [59]. Since the model is There exist one dataset, BOSSBase, that was specifically
an unsupervised generative model, the authors have named created to deal with the problems of steganography. To further
the model Self-supervised Steganographic GAN (SSteGAN). evaluate the performances of the algorithms some existing
Like in any other communication security paradigm, Alice datasets, which are used for other purposes including object
and Bob try to communicate secretly while Eve eavesdrops recognition and face recognition, are re-modeled to fit for the
on the communication channel. As such, Alice acts as the purpose of our experiments. A detailed explanation on the
generator, Bob as the decoder, eve as the steganalyzer to datasets are described in table 4.
classify if the given image is normal or stego image, and
Dev here acts as the discriminator actively competing against A. BOSSBASE
Alice. Along with secret message, input noise is also given Break Our Steganographic System (BOSS) [60] is the first
as input to Alice, to avoid generating identical images if the scientific challenge conducted to take image steganography

23416 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

TABLE 4. Information details of the dataset used in the literature of steganography.

from being a research topic to a practical application. The than 500 to 1000 images. ImageNet does not have any copy-
main aim of the competition was to develop a better ste- rights to the image and contains only the links or thumbnails
ganalysis method that can break the steganographic images to the original image. The dataset consists of images of vary-
created by the HUGO (Highly Undetectable steGO) algo- ing size. Based on the requirement, the number of images,
rithm [61]. The dataset consists of a training set and test- classes they belong to, background and the image size can be
ing set along with the HUGO algorithm that can be used selected from the wide range available.
to create the steganography images. The training dataset
consists of 10,000 grayscale cover images with dimensions D. MNIST HANDWRITTEN DIGITS
512 × 512. The testing set consists of 1000 grayscale images Modified National Institute of Standards and Technology
with dimensions 512 × 512. There is an option to down- database (MNIST) [64] is another dataset that can be used for
load the datasets with steganography images solely for the various computer vision and image processing applications.
purpose of steganalysis. Firstly, the raw images are captured MNIST handwritten dataset consists of a training and testing
using 7 different cameras and they are converted to PGM set with images of handwritten digits 0 to 9. Images in this
images. The links to download the raw images, PGM images, dataset are normalized, black and white with dimensions
the script used to convert the raw images into PGM, EXIF 28 × 28 pixels. The training set consists of 60,000 images
data of the raw images can be found in the official website.1 and testing set consists of 10,000 images.

B. CELEBA E. COCO
Large-scale CelebFaces Attributes dataset, also known as Common Objects in Context (COCO) dataset [65] was
CelebA dataset [62], is a vast dataset with more than 200K mainly developed for object detection, segmentation and
images that can be used for face recognition, face detec- image captioning purposes. This again is a huge dataset with
tion, face localization and other face-related operations. The images from 80 object categories. Each class contains at least
dataset consists of images from various sources, locations, 5 images. This dataset comes along with the class annota-
background and poses and is best suitable for steganography tion and the segmentation annotation. There is no predefined
also. The probability of using a photo/face image as the cover training and testing split. The dataset split can be carried out
for hiding secret images is very high. Along with the images, based on the research topic and the user convenience.
there are 40 different annotations available like with/without
glasses, emotions, hair styles, other accessories like hat. F. OTHERS
Other datasets that can be used for image steganography and
C. IMAGENet steganalysis are Div2K, SZUBase, USC-SIPI, DTD, LFW,
ImageNet [63] is also a very large dataset containing images and Pascal VOC. Div2K [66] is a commonly used dataset
from the WordNet hierarchy with each node containing more for Single Image Super-Resolution introduced in the NTIRE
2017 Challenge on Single Image Super-Resolution. It has
1 http://agents.fel.cvut.cz/boss/index.php?mode=VIEW&tmpl=materials a total of 1000 images split into 800 for training, 100 for

VOLUME 9, 2021 23417

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

FIGURE 4. Sample images from datasets (a) BOSSBase (b) MNIST (c) ImageNet (d) LFW (e) Div2K (f) COCO and (g) CelebA.

validation and 100 for testing. Images with high resolution [32], [56] and [28] or MATLAB [16], [17], [49], and [10].
(1024 × 678) and three grades of low resolution can be found Another popular library used is the tensorflow [40], [52], [55]
in this dataset. SZUBase is a dataset collected in [67] with and [39]. Chainer library from python is also used [44].
40000 grayscale images of size 512 × 512. USC-SIPI [68]
has a variety of different resolution images and sizes for B. METRICS USED
Image processing, image analysis, and machine vision pur- 1) PSNR
poses. Describable Textures Dataset (DTD) [69] is used to Peak Signal to Noise Ratio (PSNR) is used to determine the
analysis the textural components containing 5640 jpg images quality, robustness and invisibility of the proposed steganog-
of two sizes - 300 × 300 and 640 × 640. LFW dataset is used raphy method. PSNR is the ratio between the maximum
for face recognition and verification tasks [70]. PASCAL quality representation of the cover image and the stego image.
Visual Object Classes (VOC) is a challenge conducted for In the case of steganalysis, it is the ratio of the maximum
object detection and classification [71]. Figure 4 below repre- quality measurement of the original secret image and the
sents some of the images collated from the dataset described. extracted secret image. PSNR is used to measure the peak
error of the proposed method. The value of PSNR has to be
IV. EVALUATION high which implies that the quality of the reconstructed stego
Evaluation metrics are used to measure the invisibility, secu- image is good. Mean Squared Error (MSE) is another metric
rity, robustness and capacity of the proposed methods. The to measure the quality of the stego image reconstructed. MSE
most commonly used metrics and the functionality they mea- is the cumulative squared error between the stego image and
sure are described below. the original cover image. For better quality images, PSNR has
to be high whereas the value of MSE has to be low indicating
that the error is low. The formulas for calculating MSE and
A. EXPERIMENTAL SETUP
PSNR are given below.
Image steganography models use two inputs and two output
images mostly. In addition, the datasets used are usually very
2
P
large. A GPU-based computer with a powerful graphic card M ,N [I1 (m, n) − I2 (m, n)]
MSE = (1)
is required to perform the training and testing. The trained M ∗N
model can then be deployed in a CPU or hand held standalone
computers for deployment. All the experiments are conducted where, M and N is the number of rows and columns in the
using either python 3.5 and above versions with pytorch [27], input image respectively.

23418 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

FIGURE 5. Steganalyser accuracy comparison between different methods.

After calculating MSE, its value is used in the calculation TABLE 6. Confusion Matrix for calculating accuracy.
of PSNR.
(R2 )
PSNR = 10 ∗ log10 (2)
MSE
where R is the fluctuation in the input image. For example,
if the input image has a double-precision floating-point data
type, then R is 1. If it has an 8-bit unsigned integer data type, measure the performance of the model truly.
R is 255. Table 5 summarizes the PSNR, MSE and SSIM
(TP + TN )
values obtained by various methods in the study. Accuracy = (3)
(TP + FP + FN + TN )
TABLE 5. PSNR and MSE values of the methods. The values indicate stego For a stronger steganography method, a lower value of the
images and ∗ for cover and ∗∗ for secret images. steganalysis accuracy indicates higher security. The detection
rate is another term used to calculate the testing accuracy
of the steganalysis model used. It is used to determine the
secrecy of the proposed method. Figure 5 shows the accuracy
of the steganography and steganalysis methods used in the
study.
A comparison on the steganalysis accuracy methods for
different methods for stego images produced by popular
methods like WOW [72], HUGO [61], S-UNIWARD are
given in the figure 6.

2) ACCURACY 3) BPP
Accuracy/bit accuracy is the commonly used metric to mea- Hiding Capacity of the steganography algorithm is calculated
sure the security and robustness of the methods. Stegano- using the Bits Per Pixel (BPP). BPP represents the number
graphic accuracy is defined as the accuracy of the proposed of bits that are hidden in every pixel of the cover image
method correctly identifying the image as steganography to produce the stego image. For a higher hiding capacity,
image or not. Accuracy is calculated using four terms: True the value of BPP has to be high. The below equation is used
Positive (TP), True Negative (TN), False Positive (FP) and to calculate the value of BPP.
False Negative (FN). True positive and true negative are the
L
correct predictions made by the model while False positive BPP = (4)
and false negative are the errors made by the model. The HWC
table 6 is used to calculate the measures from the confusion where L is the length of the hidden message, H is the height,
matrix. W is the width and C is the number of channels of the cover
The accuracy is calculated as the total number of correct image.
predictions made by the model against all classifications. Many of these metrics compete against each other. Models
Though accuracy is a simple and commonly used, it is not which have higher capacity typically sacrifice secrecy, since
a good measure when the data is imbalanced. That is the hiding more information in images naturally leads to larger
cost for both false positives and false negatives are same. image distortions; models that are very robust to noise typi-
In addition to accuracy, other evaluation metrics are used to cally sacrifice capacity or secrecy, since the message must be

VOLUME 9, 2021 23419

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

FIGURE 6. Accuracy comparison between different methods (a) Steganalysis and (b) Steganography.

encoded in the image redundantly. In some sense, steganog- V. CHALLENGES

raphy and watermarking are at two ends of a spectrum The following are some challenges for consideration in image
of problems prioritizing these different axes; steganography steganography problems.
stresses secrecy while watermarking emphasizes robustness. Data Availability - Though image steganography is an
In cases of hiding text inside an image, BPP used are either unsupervised learning and the main goal is image reconstruc-
0.1 or 0.4. tion, there is no proper benchmark dataset available except
BOSSBase [60]. The number of images may be large in the
C. OBSERVATIONS BOSSBase but the images are of grayscale stored available in
The observation made from the results reported by the meth- tiff format. Most of the methods deal with hiding RGB images
ods are delineated here. Basically, the hiding capacity, secu- inside RGB cover images. Finding a suitable dataset can be
rity and robustness factors are taken into account while dis- challenging. ImageNet is the most commonly used dataset
cussing the observations. with a major drawback being the image size. The images are
Hiding Capacity: It is generally noted that the hiding very small in size 64 × 64.
capacity of the methods are in the following order. The Convergence of GAN Convergence is a major drawback
methods with least hiding capacity are the traditional methods for GAN where the model does not converge irrespective of
where text is the primary form of secret communication. Fol- the parameters chosen. Mode collapse also happen often as
lowing that is the GAN-based methods, where only text mes- the generator and discriminator are inter-dependent.
sage is used as secret. Unlike the traditional and GAN-based Comparison with other methods Evaluation metrics used
methods, the hiding capacity of the CNN-based methods are by different methods are different and hence comparing the
far better and is almost 1 [27]. The size of the secret image proposed method with the state-of-the-art methods are not
is same as the cover image. Even when a grayscale image easy.
is used for hiding [33], the size of the secret and the cover Real-time steganography Steganography models are
images are equal. In terms of the hiding capacity, CNN-based trained on a huge amounts of datasets, like in, [27],
methods clearly outperform other methods. 45000 training images are used. However, when it comes to
Security and Robustness: Security is associated with the real-time steganography, it gets difficult. The implemen-
embedding and robustness is associated with the extraction tation of the trained model for performing the steganography
of the secret image. From our observations, CNN-based and steganalysis requires transferring the stego image through
methods and GAN-based methods yield higher security. It is an untrusted channel to the receiving end. The capability of
worth noting that the extraction of the secret images is prone the trained model in dealing with real time live images which
to loss in information in deep learning methods. However, may contain noises, skewing, blurring is not proved. The
in traditional methods, the security is less but the robustness implementation of the model for real-time steganography is
is high. PSNR measure is used to correlate the security and still questionable.
the robustness. From 5 and it can be noted that [49] has
the highest PSNR value explaining the higher security of VI. DISCUSSION AND FUTURE WORKS
the GAN based method. The highest value of PSNR being GAN is the most widely used architecture and specifically
64.7, given by the cycle GAN based image steganography cycleGAN when compared to the CNN based methods. The
method. most important factor to consider is that the GAN is a two
From the observations made, GAN-based deep learning part model where one model is used at the sender end for
methods have the best performance in terms of the hid- embedding the secret information while the second model
ing capacity (1), PSNR (64.7) [49]. The discriminators are is used at the receiving end to extract the secret informa-
trained in a way to overcome any steganalysis attack and tion. Two models that are trained end-to-end under same
has better anti-detection property compared to traditional and training circumstances are required for the whole process
CNN-based methods. of image steganography to work perfectly. Without or the

23420 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

loss of one model may affect the embedding/ extraction • Experiments related to optimizing the parameters and
as they are interconnected. CNN-based methods use U- decreasing the storage capacities can be further con-
Net/Xu-Net autoencoder-decoder architecture for embedding ducted using various datasets.
and extracting. Some methods use the encoder for embed- • The era of quantum computing is not far away, more
ding and decoder for extracting, whereas, some use one efforts on developing designs on quantum images can
autoencoder-decoder for embedding and another for extract- be explored.
ing. Though there is some inter dependency, it is not totally • To benefit from a combination of methods, an ensemble
linked like GAN methods. of traditional and deep learning methods can be further
Unbalance in the learning of generator-discriminator can studied.
happen where the generator is performing efficiently but the • Efforts can be directed to form a benchmark dataset
discriminator is struggling. Though the overall efficiency containing images from various source cameras, image
will not be affected, either sender or receiver will be prone formats. A compilation of all possible algorithms can
to faults. This can be avoided by choosing the parameters also be done to create the steganography images.
carefully and avoiding over fitting during training. • Many methods have considered the hiding capacity,
Security capabilities of the GAN methods are higher security and robustness as the performance measure.
when compared with CNN architectures and traditional LSB However, there are possibilities for man-in-the-middle
methods. GANs are basically used in the image reconstruc- attacks when the transfer happens through untrusted
tion field which makes it conveniently suitable for image channels. Tampering of the stego image can also happen
steganography compared to convNets. during the transfer. These attacks and the performance
In deep learning methods, the working principle of the of the designed algorithm against these attacks can be
image steganography is to extract features from the cover and considered for evaluation along with other metrics.
secret image and concatenate them to produce an end result
closer to the cover image. However, where and how the secret VII. CONCLUSION
image pixels are embedded cannot be understood clearly. Image steganography is the method used in transmitting
Without the counterpart extracting model trained, it may be secret information by hiding it in plain sight inside a cover
difficult to crack the steganography image. This increases the image. Deep learning methods are widely used in every field
security but becomes difficult when the extraction model is and has been used in the research of steganography. Review
not working or crashed. of all the related works led to categorizing them into three
Some of the major disadvantages are the time taken for groups vastly. Most of the traditional based steganography
training, the computational time during testing and the stor- methods use the LSB substitution and some of its variants.
age capabilities. The models take two images or one image Other than LSB, PVD, DCT and EMD are commonly used.
and a text message converted into bits as input. The fea- The hiding capacity of the traditional methods are limited
tures are extracted from both the inputs which increases the as over burdening the cover image by exploiting more pix-
computational time in both embedding and extraction. The els for hiding the secret message may led to distortions.
number of parameters also increases by double at least when Also, the autoencoder-decoder structure with VGG as base,
compared to a normal architecture which in turn increase the U-Net and Xu-Net are the most prevailing architectures
storage space required by the model. used for CNN-based image steganography methods. More
RGB secret images are used by a handful methods when recently, GAN architecture has gained significant attention
others used gray scale images. When converting the gray for their ability to deal with image reconstruction tasks. Image
scale image to RGB image for better understanding, there steganography can be considered one such image reconstruc-
can be loss of information. Image enhancement techniques tion task where the cover image and the secret information is
are required in addition to understand the secret information taken as input to reconstruct a steganographic image which is
properly. close to the cover image in resemblance.
Some of the aspects that can be considered for future works There is no benchmark image datasets to perform the
are enumerated below, image steganography while most of them use the Ima-
• Usage of popular networks U-Net, cycleGAN and geNet, CelebA or BOSSBase. Each of the methods have
using DCT and DWT have been considered and more their own evaluation methods and metrics and hence there
exploration on the other customized architectures can is no common platform for comparisons. Peak Signal-to-
be attempted. For example, Recurrent Neural Net- Noise Ratio (PSNR) value comparison shown in table 5
works (RNNs) instead of CNNs can be further explored. gives an idea on the security performance of the different
A customized WGAN can be replaced with other vari- methods. By far, the best PSNR value of 64.7 is obtained by
ants of GAN. Kuppusamy et al. [49] using cycle GAN. GAN based meth-
• The majority of image steganography methods use either ods have proved to have better security performance and hid-
text or gray scale image as the secret information and ing capacity. GAN is the extensively used and most preferred
there is a need for more research in hiding image in architecture over the autoencoder and statistical methods.
image and image in video. Traditional methods are less secure as it is only a matter

VOLUME 9, 2021 23421

 N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

of detection of presence of the secret message. The secret [18] G. Swain, ‘‘Very high capacity image steganography technique using
message can be easily extracted as the embedding used a quotient value differencing and LSB substitution,’’ Arabian J. Sci. Eng.,
vol. 44, no. 4, pp. 2995–3004, Apr. 2019.
statistical method. [19] A. Qiu, X. Chen, X. Sun, S. Wang, and W. Guo, ‘‘Coverless image
In summary, this paper has elaborated on the techniques steganography method based on feature selection,’’ J. Inf. Hiding Privacy
used in the recent times for image steganography, the current Protection, vol. 1, no. 2, p. 49, 2019.
[20] R. D. Rashid and T. F. Majeed, ‘‘Edge based image steganography: Prob-
trends. Along with it, details on the datasets and evaluation lems and solution,’’ in Proc. Int. Conf. Commun., Signal Process., Appl.
metrics are detailed. Challenges faced, some discussions on (ICCSPA), Mar. 2019, pp. 1–5.
the gaps and the scopes for future direction is also evaluated [21] X. Liao, J. Yin, S. Guo, X. Li, and A. K. Sangaiah, ‘‘Medical JPEG image
steganography based on preserving inter-block dependencies,’’ Comput.
in this paper. It can be concluded that deep learning has Electr. Eng., vol. 67, pp. 320–329, Apr. 2018.
tremendous potential in the image steganography field taking [22] W. Lu, Y. Xue, Y. Yeung, H. Liu, J. Huang, and Y. Shi, ‘‘Secure
into consideration that all the challenges and gaps are filled. halftone image steganography based on pixel density transition,’’ IEEE
Trans. Dependable Secure Comput., early access, Aug. 6, 2019, doi:
ACKNOWLEDGMENT 10.1109/TDSC.2019.2933621.
[23] Y. Zhang, C. Qin, W. Zhang, F. Liu, and X. Luo, ‘‘On the fault-tolerant
The findings achieved herein are solely the responsibility of performance for a class of robust image steganography,’’ Signal Process.,
the author. Open Access funding was provided by the Qatar vol. 146, pp. 99–111, May 2018.
National Library. [24] H. M. Sidqi and M. S. Al-Ani, ‘‘Image steganography: Review study,’’ in
Proc. Int. Conf. Image Process., Comput. Vis., Pattern Recognit. (IPCV),
REFERENCES 2019, pp. 134–140.
[25] P. Wu, Y. Yang, and X. Li, ‘‘Image-into-image steganography using deep
[1] Wikipedia. (2020). Steganography. [Online]. Available: https://en.
convolutional network,’’ in Proc. Pacific Rim Conf. Multimedia. Cham,
wikipedia.org/wiki/Steganography
Switzerland: Springer, 2018, pp. 792–802.
[2] H. Shi, X.-Y. Zhang, S. Wang, G. Fu, and J. Tang, ‘‘Synchronized detec-
tion and recovery of steganographic messages with adversarial learning,’’ [26] P. Wu, Y. Yang, and X. Li, ‘‘StegNet: Mega image steganography capacity
in Proc. Int. Conf. Comput. Sci. Cham, Switzerland: Springer, 2019, with deep convolutional network,’’ Future Internet, vol. 10, no. 6, p. 54,
pp. 31–43. Jun. 2018.
[3] N. F. Hordri, S. S. Yuhaniz, and S. M. Shamsuddin, ‘‘Deep learning and its [27] X. Duan, K. Jia, B. Li, D. Guo, E. Zhang, and C. Qin, ‘‘Reversible image
applications: A review,’’ in Proc. Conf. Postgraduate Annu. Res. Informat. steganography scheme based on a U-Net structure,’’ IEEE Access, vol. 7,
Seminar, 2016, pp. 1–6. pp. 9314–9323, 2019.
[4] N. F. Johnson and S. Jajodia, ‘‘Exploring steganography: Seeing the [28] T. P. Van, T. H. Dinh, and T. M. Thanh, ‘‘Simultaneous convolutional neural
unseen,’’ Computer, vol. 31, no. 2, pp. 26–34, Feb. 1998. network for highly efficient image steganography,’’ in Proc. 19th Int. Symp.
[5] S. Gupta, G. Gujral, and N. Aggarwal, ‘‘Enhanced least significant bit Commun. Inf. Technol. (ISCIT), Sep. 2019, pp. 410–415.
algorithm for image steganography,’’ Int. J. Comput. Eng. Manage., vol. 15, [29] R. Rahim and S. Nadeem, ‘‘End-to-end trained CNN encoder-decoder
no. 4, pp. 40–42, 2012. networks for image steganography,’’ in Proc. Eur. Conf. Comput. Vis.
[6] R. Das and T. Tuithung, ‘‘A novel steganography method for image based (ECCV), 2018, pp. 1–6.
on Huffman encoding,’’ in Proc. 3rd Nat. Conf. Emerg. Trends Appl. [30] Z. Wang, N. Gao, X. Wang, J. Xiang, and G. Liu, ‘‘STNet: A style
Comput. Sci., Mar. 2012, pp. 14–18. transformation network for deep image steganography,’’ in Proc. Int. Conf.
[7] A. Singh and H. Singh, ‘‘An improved LSB based image steganography Neural Inf. Process. Cham, Switzerland: Springer, 2019, pp. 3–14.
technique for RGB images,’’ in Proc. IEEE Int. Conf. Electr., Comput. [31] K. Yang, K. Chen, W. Zhang, and N. Yu, ‘‘Provably secure generative
Commun. Technol. (ICECCT), Mar. 2015, pp. 1–4. steganography based on autoregressive model,’’ in Proc. Int. Workshop
[8] Z. Qu, Z. Cheng, W. Liu, and X. Wang, ‘‘A novel quantum image steganog- Digit. Watermarking. Cham, Switzerland: Springer, 2018, pp. 55–68.
raphy algorithm based on exploiting modification direction,’’ Multimedia [32] S. Baluja, ‘‘Hiding images in plain sight: Deep steganography,’’ in Proc.
Tools Appl., vol. 78, no. 7, pp. 7981–8001, Apr. 2019. Adv. Neural Inf. Process. Syst., 2017, pp. 2069–2079.
[9] S. Wang, J. Sang, X. Song, and X. Niu, ‘‘Least significant qubit (LSQb) [33] R. Zhang, S. Dong, and J. Liu, ‘‘Invisible steganography via gener-
information hiding algorithm for quantum image,’’ Measurement, vol. 73, ative adversarial networks,’’ Multimedia Tools Appl., vol. 78, no. 7,
pp. 352–359, Sep. 2015. pp. 8559–8575, Apr. 2019.
[10] N. Patel and S. Meena, ‘‘LSB based image steganography using dynamic [34] S. Islam, A. Nigam, A. Mishra, and S. Kumar, ‘‘VStegNET: Video
key cryptography,’’ in Proc. Int. Conf. Emerg. Trends Commun. Technol. steganography network using spatio-temporal features and micro-
(ETCT), Nov. 2016, pp. 1–5. bottleneck,’’ in Proc. BMVC, Sep. 2019, p. 274.
[11] O. Elharrouss, N. Almaadeed, and S. Al-Maadeed, ‘‘An image steganogra- [35] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
phy approach based on k-least significant bits (k-LSB),’’ in Proc. IEEE Int. S. Ozair, A. Courville, and Y. Bengio, ‘‘Generative adversarial nets,’’ in
Conf. Informat., IoT, Enabling Technol. (ICIoT), Feb. 2020, pp. 131–135. Proc. Adv. Neural Inf. Process. Syst., 2014, pp. 2672–2680.
[12] M. V. S. Tarun, K. V. Rao, M. N. Mahesh, N. Srikanth, and M. Reddy, [36] D. Volkhonskiy, B. Borisenko, and E. Burnaev, ‘‘Generative adversarial
‘‘Digital video steganography using LSB technique,’’ Red, vol. 100111, networks for image steganography,’’ in Proc. ICRL Conf., 2016.
Apr. 2020, Art. no. 11001001. [37] D. Volkhonskiy, I. Nazarov, and E. Burnaev, ‘‘Steganographic genera-
[13] S. S. M. Than, ‘‘Secure data transmission in video format based on LSB tive adversarial networks,’’ in Proc. 12th Int. Conf. Mach. Vis. (ICMV),
and Huffman coding,’’ Int. J. Image, Graph. Signal Process., vol. 12, no. 1, vol. 11433, 2020, Art. no. 114333M.
p. 10, 2020. [38] D. J. Im, C. D. Kim, H. Jiang, and R. Memisevic, ‘‘Generating images
[14] M. B. Tuieb, M. Z. Abdullah, and N. S. Abdul-Razaq, ‘‘An efficiency, with recurrent adversarial networks,’’ 2016, arXiv:1602.05110. [Online].
secured and reversible video steganography approach based on lest sig- Available: http://arxiv.org/abs/1602.05110
nificant,’’ J. Cellular Automata, vol. 16, no. 17, Apr. 2020. [39] H. Shi, J. Dong, W. Wang, Y. Qian, and X. Zhang, ‘‘SSGAN: Secure
[15] R. J. Mstafa, K. M. Elleithy, and E. Abdelfattah, ‘‘A robust and secure video steganography based on generative adversarial networks,’’ in Proc. Pacific
steganography method in DWT-DCT domains based on multiple object Rim Conf. Multimedia. Cham, Switzerland: Springer, 2017, pp. 534–544.
tracking and ECC,’’ IEEE Access, vol. 5, pp. 5354–5365, 2017. [40] J. Hayes and G. Danezis, ‘‘Generating steganographic images via adversar-
[16] K. A. Al-Afandy, O. S. Faragallah, A. Elmhalawy, E.-S.-M. El-Rabaie, and ial training,’’ in Proc. Adv. Neural Inf. Process. Syst., 2017, pp. 1954–1963.
G. M. El-Banby, ‘‘High security data hiding using image cropping and LSB [41] J. Yang, K. Liu, X. Kang, E. K. Wong, and Y.-Q. Shi, ‘‘Spatial
least significant bit steganography,’’ in Proc. 4th IEEE Int. Colloq. Inf. Sci. image steganography based on generative adversarial network,’’ 2018,
Technol. (CiSt), Oct. 2016, pp. 400–404. arXiv:1804.07939. [Online]. Available: http://arxiv.org/abs/1804.07939
[17] A. Arya and S. Soni, ‘‘Performance evaluation of secrete image steganog- [42] J. Yang, D. Ruan, J. Huang, X. Kang, and Y.-Q. Shi, ‘‘An embedding
raphy techniques using least significant bit (LSB) method,’’ Int. J. Comput. cost learning framework using GAN,’’ IEEE Trans. Inf. Forensics Security,
Sci. Trends Technol., vol. 6, no. 2, pp. 160–165, 2018. vol. 15, pp. 839–851, 2020.

23422 VOLUME 9, 2021

N. Subramanian et al.: Image Steganography: A Review of the Recent Advances

[43] W. Tang, S. Tan, B. Li, and J. Huang, ‘‘Automatic steganographic distortion [66] E. Agustsson and R. Timofte, ‘‘NTIRE 2017 challenge on single image
learning using a generative adversarial network,’’ IEEE Signal Process. super-resolution: Dataset and study,’’ in Proc. IEEE Conf. Comput. Vis.
Lett., vol. 24, no. 10, pp. 1547–1551, Oct. 2017. Pattern Recognit. Workshops (CVPRW), Jul. 2017, pp. 126–135.
[44] H. Naito and Q. Zhao, ‘‘A new steganography method based on genera- [67] W. Tang, S. Tan, B. Li, and J. Huang, ‘‘Automatic steganographic distortion
tive adversarial networks,’’ in Proc. IEEE 10th Int. Conf. Awareness Sci. learning using a generative adversarial network,’’ IEEE Signal Process.
Technol. (iCAST), Oct. 2019, pp. 1–6. Lett., vol. 24, no. 10, pp. 1547–1551, Oct. 2017.
[45] K. Zhang, A. Cuesta-Infante, and K. Veeramachaneni, ‘‘SteganoGAN: [68] M. Zhang and A. A. Sawchuk, ‘‘USC-HAD: A daily activity dataset for
Pushing the limits of image steganography,’’ Jan. 2019, arXiv:1901.03892. ubiquitous activity recognition using wearable sensors,’’ in Proc. ACM Int.
[Online]. Available: https://arxiv.org/abs/1901.03892 Conf. Ubiquitous Comput. (Ubicomp), Workshop Situation, Activity Goal
[46] J. Zhu, R. Kaplan, J. Johnson, and L. Fei-Fei, ‘‘Hidden: Hiding data Awareness (SAGAware), Pittsburgh, PA, USA, Sep. 2012, pp. 1036–1043.
with deep networks,’’ in Proc. Eur. Conf. Comput. Vis. (ECCV), 2018, [69] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, ‘‘Describ-
pp. 657–672. ing textures in the wild,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recog-
nit. (CVPR), Jun. 2014, pp. 3606–3613.
[47] Y. Ke, M. Zhang, J. Liu, T. Su, and X. Yang, ‘‘Generative steganog-
[70] B. G. Huang, M. Ramesh, T. Berg, and E. Learned-Miller, ‘‘Labeled faces
raphy with Kerckhoffs’ principle based on generative adversarial net-
in the wild: A database for studying face recognition in unconstrained
works,’’ 2017, arXiv:1711.04916. [Online]. Available: http://arxiv.org/
environments,’’ Univ. Massachusetts, Amherst, Amherst, MA, USA, Tech.
abs/1711.04916
Rep. 07-49, Oct. 2007.
[48] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, ‘‘Unpaired image-to-image
[71] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman,
translation using cycle-consistent adversarial networks,’’ in Proc. IEEE Int.
‘‘The Pascal visual object classes (VOC) challenge,’’ Int. J. Comput. Vis.,
Conf. Comput. Vis. (ICCV), Oct. 2017, pp. 2223–2232.
vol. 88, no. 2, pp. 303–338, Jun. 2010.
[49] P. G. Kuppusamy, K. C. Ramya, S. Sheebha Rani, M. Sivaram, and [72] V. Holub and J. Fridrich, ‘‘Designing steganographic distortion using direc-
V. Dhasarathan, ‘‘A novel approach based on modified cycle generative tional filters,’’ in Proc. IEEE Int. Workshop Inf. Forensics Secur. (WIFS),
adversarial networks for image steganography,’’ Scalable Comput., Pract. Dec. 2012, pp. 234–239.
Exper., vol. 21, no. 1, pp. 63–72, Mar. 2020.
[50] C. Chu, A. Zhmoginov, and M. Sandler, ‘‘CycleGAN, a master of steganog-
raphy,’’ 2017, arXiv:1712.02950. [Online]. Available: http://arxiv.org/
abs/1712.02950 NANDHINI SUBRAMANIAN (Member, IEEE) received the bachelor’s
degree in electrical and electronics engineering from the PSG College
[51] H. Porav, V. Musat, and P. Newman, ‘‘Reducing steganography in cycle-
consistency GANs,’’ in Proc. CVPR Workshops, 2019, pp. 78–82. of Technology, India, and the master’s degree in computing from Qatar
[52] R. Meng, Q. Cui, Z. Zhou, Z. Fu, and X. Sun, ‘‘A steganography algorithm University, Doha. She is currently working as a Research Assistant with
based on CycleGAN for covert communication in the Internet of Things,’’ Dr. S. Al-Maadeed with Qatar University. Her interests include computer
IEEE Access, vol. 7, pp. 90574–90584, 2019. vision, artificial intelligence, machine learning, and cloud computing.
[53] A. Odena, C. Olah, and J. Shlens, ‘‘Conditional image synthesis with
auxiliary classifier GANs,’’ in Proc. Int. Conf. Mach. Learn., 2017, OMAR ELHARROUSS received the master’s
pp. 2642–2651.
degree from the Faculty of Sciences, Dhar El
[54] Z. Zhang, G. Fu, J. Liu, and W. Fu, ‘‘Generative information hiding method Mehraz, Fez, Morocco, in 2013, and the Ph.D.
based on adversarial networks,’’ in Proc. Int. Conf. Comput. Eng. Netw.
degree from the LIIAN Laboratory, USMBA-Fez
Cham, Switzerland: Springer, 2018, pp. 261–270.
University, in 2017. His research interests include
[55] M.-M. Liu, M.-Q. Zhang, J. Liu, Y.-N. Zhang, and Y. Ke, ‘‘Coverless
pattern recognition, image processing, and com-
information hiding based on generative adversarial networks,’’ 2017,
arXiv:1712.06951. [Online]. Available: http://arxiv.org/abs/1712.06951
puter vision.
[56] X. Duan, H. Song, C. Qin, and M. K. Khan, ‘‘Coverless steganography for
digital images based on a generative model,’’ Comput., Mater. Continua,
vol. 55, no. 3, pp. 483–493, Jul. 2018.
[57] M. Arjovsky, S. Chintala, and L. Bottou, ‘‘Wasserstein GAN,’’ 2017,
SOMAYA AL-MAADEED (Senior Member, IEEE) received the Ph.D. degree
arXiv:1701.07875. [Online]. Available: http://arxiv.org/abs/1701.07875
in computer science from the University of Nottingham, U.K., in 2004. She is
[58] C. Li, Y. Jiang, and M. Cheslyar, ‘‘Embedding image through generated
currently the Head of the Computer Science Department, Qatar University,
intermediate medium using deep convolutional generative adversarial net-
work,’’ Comput., Mater. Continua, vol. 56, no. 2, pp. 313–324, 2018.
where she is also a Coordinator of the Computer Vision and AI Research
Group. She enjoys excellent collaboration with national and international
[59] Z. Wang, N. Gao, X. Wang, X. Qu, and L. Li, ‘‘SSteGAN: Self-
learning steganography based on generative adversarial networks,’’ in institutions and industry. She is a Principal Investigator of several funded
Proc. Int. Conf. Neural Inf. Process. Cham, Switzerland: Springer, 2018, research projects generating approximately five million. She has published
pp. 253–264. extensively pattern recognition and delivered workshops on teaching pro-
[60] P. Bas, T. Filler, and T. Pevný, ‘‘‘Break our steganographic system’: The ins gramming for undergraduate students. She attended workshops related to
and outs of organizing BOSS,’’ in Information Hiding, T. Filler, T. Pevný, higher education strategy, assessment methods, and interactive teaching.
S. Craver, and A. Ker, Eds. Berlin, Germany: Springer, 2011, pp. 59–70. In 2015, she was elected as the IEEE Chair for the Qatar Section.
[61] T. Pevný, T. Filler, and P. Bas, ‘‘Using high-dimensional image mod-
els to perform highly undetectable steganography,’’ in Information Hid-
AHMED BOURIDANE (Senior Member, IEEE)
ing, R. Böhme, P. W. L. Fong, and R. Safavi-Naini, Eds. Berlin,
Germany: Springer, 2010, pp. 161–177.
received the Ingénieur d’Etat degree in electronics
from the Ecole Nationale Polytechnique d’Alger,
[62] Z. Liu, P. Luo, X. Wang, and X. Tang, ‘‘Deep learning face attributes
in the wild,’’ in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), Dec. 2015, Algeria, in 1982, the M.Phil. degree in VLSI
pp. 3730–3738. for signal processing from Newcastle University,
[63] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, ‘‘ImageNet: U.K., in 1988, and the Ph.D. degree in computer
A large-scale hierarchical image database,’’ in Proc. IEEE Conf. Comput. vision from the University of Nottingham, U.K.,
Vis. Pattern Recognit., Jun. 2009, pp. 248–255. in 1992. He joined Queen’s University Belfast,
[64] L. Deng, ‘‘The MNIST database of handwritten digit images for machine in 1994, initially as a Lecturer and then as a Reader
learning research [best of the Web],’’ IEEE Signal Process. Mag., vol. 29, in Computer Science. He is currently a Professor
no. 6, pp. 141–142, Nov. 2012. of Computer Science and leads the Intelligent Systems and Security Group,
[65] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, Northumbria University, Newcastle upon Tyne, U.K. He has authored or
P. Perona, D. Ramanan, C. L. Zitnick, and P. Dollár, ‘‘Microsoft COCO: coauthored over 350 publications. His research interests include imaging for
Common objects in context,’’ in Proc. Eur. Conf. Comput. Vis., 2014, security and medical engineering.
pp. 740–755.

VOLUME 9, 2021 23423