SSRN Id3171494

3rd International Conference on Internet of Things and Connected Technologies
A Survey on Digital Image Steganography: Current Trends and

Challenges
Debina Laishrama, Dr. Themrichon Tuithungb

a
Research Scholar,Department of Computer Science and Engineering, National Institute of Technology, Dimapur, 797103, Nagaland, India
b
Professor and Director, Department of Computer Science and Engineering, National Institute of Technology, Dimapur, 797103, Nagaland, India
Abstract:
Steganography and Steganalysis is an information hiding technique for data protection and security from unauthorized access and use. Steganography is
defined as the art and science of covert information in plain sight in various media sources such as text, images, audio, video, network channel etc. so, as
to not stimulate any suspicion; while steganalysis is the science of attacking the steganographic system to reveal the secret message. Steganography is
often designed as a trade-off between robustness, imperceptibility and message payload hiding capacity. The paper provides detailed review of the current
trends in digital image steganography, where the secret message is embedded in image cover source and analysis of the different approaches along with
the impact on the choice of embedding algorithm, image cover selection, the image performance metrics with extensive focus given to adaptive image
steganography, which is currently the most secure steganographic system applicable in spatial, transform and side-informed steganography. The research
contributions have been summarized, with directions to future improvement.
Keywords:
Information hiding; Digital image steganography; Evaluation criteria; Image quality metrics; Adaptive steganography; Steganalysis
successfully used in various businesses, multimedia companies apart from

1. Introduction known defense, intelligence, medical, forensic image applications. The
paper begins by discussing the general problem of steganography, the
The security of information across the globe is provided by using evaluation criteria of an image for cover selection, steganographic
information hiding techniques that make it possible to hide some embedding capacity for a given message payload, security against attacks,
information within the digital/non-digital data in such a way that the then the available image quality metrices that can be used to fulfill these
modifications are undetectable to a certain extent. The various schemes requirements. We continue with an elaborate study on the different
available are cryptography, watermarking and steganography all of which domains in which image steganography is applied, their advantages,
provides means for secret communication. Digital watermarking is a disadvantages, and the suitable embedding algorithms for each with more
technique to protect intellectual property from illegal copying to identify emphasis on the latest adaptive steganography. A review of some of the
its rightful owner. Steganography employs information communication current work is summarized. We suggest the direction for future
techniques that hide the very existence of the message itself. The enhancement. Lastly, we conclude our work.
steganographic information hiding process begins by identifying the cover
image’s redundant bits, which can be modified without destroying its 1.1. Problem definition and General model of Steganography
integrity both by Human Visual System (HVS) and statistical methods.
On the other hand, Cryptography aims at marking the hidden information The steganographic framework is formulated in terms of the prisoner’s
unintelligible to unauthorized person who might intercept them by using problem where the two inmates, Alice and Bob who wishes to
encryption and decryption techniques. Digital Image steganography are communicate the escape plan. All the communication and messages must
characterized into two types based on the number of source image where go through the warden, Wendy who at the slightest suspicion of covert
the secret information is hidden; if a single source image is used it is information will put them in solitary confinement [1] [17]. Hence, in the
known as single cover based steganography; if message is hidden in general model of steganography, Alice hides the secret message “M” into
multiple covers, it is known as batch steganography. The work presented a cover-image “C” and sends the stego image “S” [12], which is sent
here, revolves around digital image steganography based on single cover through a public channel as in Fig 2. Thus, we have the following
and does not discuss any other methods. Although, cryptography, definitions:
steganography and watermarking are subcategories of information hiding • Cover-image: the carrier image used as the medium to hide the
as shown in Fig 1, the objectives, conditions and techniques vary. Their secret message using some embedding algorithm.
differences are listed in Table 1 for better understanding. The goal of • Stego image: the cover image after embedding the secret message is
image steganography is to keep the very existence of the message referred to as stego image. The goal of image steganography is to
imperceptible; but often because of its invasive nature the embedding produce a stego image which could successfully embed in the cover
process tends to leave specific detectable statistical artifacts behind in the image without distorting the cover image quality.
cover image. The process of modeling such statistical artifacts in the cover
image as distortions to break the steganography system and reveal the The embedding mapping: C × M × k→S, s=Emb(x,m,k)
information is known as steganalysis. Steganography has been The extraction mapping: C × k→M, Ext (s, k) =mˊ
Electronic copy available at: https://ssrn.com/abstract=3171494

Table 1: The distinction between different information security technique: cryptography, watermarking and Steganography
Scale Cryptography Watermarking Steganography
Source used Text or Image in some cases to generate Images and audio to generate Digital media such as images, audio,
cipher text watermarked file video, texts to generate stego-files
Aim/Objective Protection of information/ data. Preserving copyright to identify the Invisible secret communication to
Robustness is the criteria for rightful owner. Robustness is the criteria achieve high robustness, total
performance. for performance impartibility and embedding capacity
Attacks Cryptanalysis Image processing Steganalysis
Authentication Full retrieval of secret data using Full retrieval of secret data at the
Using cross correlation methods
decryption decoder
Criteria of failure Secret data is deciphered Removed, replaced and modified Secret message is detected
Videos
Cryptography
Others
(Linguistics Digital Images batch Steganography
steganography, network
steganography)
Steganography
Technical Single Cover

Audio
Steganography Steganography
Information Hiding Text
Imperceptible
Robust Finger print
Watermarking
Fragile Visible
Fig. 1 - Different Information Hiding disciplines
Fig. 2 - General model of Steganography

The Warden, Wendy could be either passive or active depending on the 2.2. Performance of the system under Steganalytic attacks
kind of action she performs on suspicion. A passive warden would simply
suppress or ignore while an active warden would alter the message to foil • Confusion Matrix: The steganography is at risk if the warden has
the escape plan. We assume in our study that Wendy is a passive warden false positive probability [6] (Type 1 error) α less than some
and cannot distinguish between cover image and stego image. In this parametric constant α* symbol and false negative (Type || error)
context, steganalysis is the technique that helps Wendy in extricating the probability β less than β*, which becomes criteria for steganalysis.
cover image from the stego image. She analyses the embedding True Positive (TP) → stego image is correctly determined as stego
distortions and other statistical artifacts introduced in the image post image, False Negative (FN) → stego image is incorrectly predicted
embedding to derive at the technique to break the steganographic method. as cover image, True Negative (TN) → a cover mage is correctly
It can be rightly pointed that development of steganography and classified as cover image and False Positive (FP) → cover image
steganalysis go parallel. medium is wrongfully determined as stego image as seen in Fig 3.
𝑇𝑃
𝑇𝑃 𝑟𝑎𝑡𝑒 = (5)
(𝑇𝑃+𝐹𝑁)
2. Performance Evaluation 𝐹𝑃 𝑟𝑎𝑡𝑒 =
𝐹𝑃
(6)
𝑇𝑁+𝐹𝑃
(𝑇𝑃𝑠+𝑇𝑁𝑠)
A steganographic system must ensure “covert” communication; it is 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (7)
𝑇𝑃𝑠+𝑇𝑁𝑠+𝐹𝑃𝑠+𝐹𝑁𝑠
considered secure if the steganalyer in general is unable to differentiate 𝑇𝑃𝑠
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (8)
between the cover image and stego image. The performance evaluation 𝑇𝑃𝑠+𝐹𝑃𝑠
schemes presented so far is limited in their applicability towards a specific

image domain or to a definite embedding scheme. However, the basic
requirement of the evaluation criteria has been identified as security
against steganalysis attacks, capacity and imperceptibility [9][37].
Deliberating their rules, strength and weaknesses may lead to
improvement of the steganographic system design. Security of a
steganographic system can be studied under the below two categories.
2.1. Security of the steganographic method
The security of a steganographic system can be analyzed by considering

that the warden has complete knowledge of both the stego and the cover
image known as stego cover attack. The security is evaluated using
information theory to determine whether the system is “information
theoretically” secure [3-5] [51]. The entropy of H(E) which is “the
uncertainty” about the embedding algorithm must not decrease even when Fig. 3 - Confusion matrix
S and C are known to the steganalyser.
• Receiver Operating Characteristics (ROC) Curve: ROC is used
𝐻(𝐸|(𝑆, 𝐶)) = 𝐻(𝐸), 𝑤ℎ𝑒𝑛 𝐼(𝐸, (𝑆, 𝐶)) = 𝐻(𝐸) − 𝐻(𝐸|(𝑆, 𝐶)) = 0 (1) to test the steganalytic classifier or model in which TP rate is plotted
along the Y-axis and FP rate is plotted along the X axis as in Fig 4.
Suppose that the embedding information is not known. Given three steganalytic methods A, B, C. Method C is said to have
better steganalytic performance than method B and B is better than
𝐻(𝑆|𝐶) = 𝐻(𝐶|𝑆) = 0 (2) method A because the area under the ROC (AUC) curve is larger for
A. Given the ROC curve, the Stegoanalytic method is said to be
H(S|C)=H(C|S)>0, if known (3) r- secure if |TP rate – FP rate| <= r and γϵ[0,1].
If the mutual information after embedding is as below:
𝐼(𝐸, (𝑆, 𝐶)) = 𝐻(𝐸) − 𝐻(𝐸|(𝑆, 𝐶)) > 0 (4)
Then it indicates insecure system. Therefore, the necessary condition for

secure steganography is H(S/C) = H(C/S) = 0 [4]. Security can be boosted
if indeterminism is introduced using preprocessed cover image.
Preprocessing techniques that increase the amplitude of pixel intensities
brings variation to the cover image and makes it more suitable to host the
message. Such methods include successive mean quantization transform
(SMQT), linear stretching, histogram equalization.
Fig. 4 - ROC curve example of three steganalytic methods A, B and C

• Maximum Mean Discrepancy (MMD) Security: It is a popular
Steganographic benchmark using two sample non parametric tests a
based on the ensemble of distances representing the samples
probability distributions over an analytic function such as density
function, cumulative distributive function, characteristic function or
mean embedding. X=xi,..,xn and Y=yi,..,yn is the probabilistic
distribution function (pdf) of X and Y. MMD is better than SVM
since only report the minimal total average probabilistic distribution
function of error.
1
𝑀𝑀𝐷(𝐹, 𝑋, 𝑌) = 𝑠𝑢𝑝𝑓∈𝐹 ( ∑𝑁 𝑁
𝑖=1 𝑓( 𝑥𝑖 ) − 1/𝑁 ∑𝑖=1 𝑓( 𝑦𝑖 )) (9)
𝑁
• Kullback – Leibler Divergence: The security of steganographic

system against attacks and performance comparison with other
steganalytic methods can be estimated using a relative entropy b
D(X||Y) using KL divergence (KLD). The objective is to determine
the Steganographic embedding method with least KLD.
(𝑃𝑋 (𝑔)
𝐷(𝑃𝑋 ||𝑃𝑌 ) = ∑𝑔∈𝐺 𝑃𝑋 (𝑔) 𝑙𝑜𝑔 (10)
(𝑃𝑌 (𝑔))
• Linear Classifier with Higher- Order Statistics: The KL distance

(KLD) is used to distinguish between PDFs of the estimated stego
image cover and image denoted as Px and Py because of the potential
problems of bin-width selection for numerical integration. The
selection is to map Px and Py to another domain such as applying real
value Kernel and Taylor series expansion and then use Euclidian
distance as the distance measure. Here, M(x) is the moment
generating function and exp(tx) is the Taylor series expansion.
∞
𝑀(𝑥) = ∫−∞ 𝑝(𝑡) 𝑒𝑥𝑝(𝑡𝑥) 𝑑𝑡 (11) Fig. 5 - (a) RS analysis indicates detectability (b) RS analysis
indicates robustness
• Fisher’s Linear Discriminant: Correlation based security analysis

embedding rate showing small differences and the length of the message
is to determine the effect of change in a pixel/DCT coefficient on its
cannot be detected [8].
neighbors with the help of co-occurrence matrix. However, it fails to
substantially capture the inter-block and intra-block properties. To
2.3. Steganographic Capacity
solve the problem, we can use Fisher’s linear discernment classifiers
in the lower-dimensional space to determine the separability of
A secure steganographic capacity is not so simple to state because the idea
pixels. FLD uses F-score for evaluation. Detectable parts in the
of a “Secure” system is dependent on the application and what may be an
stego image are determined by FLD [53].
acceptable risk level in one scenario maybe insecure in another [5]. The
𝐹𝑠𝑐𝑜𝑟𝑒(𝑑) = (𝜇𝑑 𝑐 − 𝜇𝑑 𝑠 )2 /(𝜎𝑑𝑐 )2 − (𝜎𝑑 𝑠 )2 (12) security of steganography is the factor of the maximum amount of secret
message payload that can be embedded in the host cover medium so that
𝜇𝑑 𝑐 , 𝜇𝑑 𝑠 = 𝑚𝑒𝑎𝑛 𝑜𝑓 𝑑𝑡ℎ 𝑓𝑒𝑎𝑡𝑢𝑟𝑒 𝑖𝑛 𝑐𝑜𝑣𝑒𝑟 𝑎𝑛𝑑 𝑠𝑡𝑒𝑔𝑜 𝑖𝑚𝑎𝑔𝑒 𝑠𝑒𝑡 there is no significant visual or statistical distortion in the stego image or
𝜎𝑑 𝑐 , 𝜎𝑑 𝑠 = 𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑑𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑐𝑜𝑣𝑒𝑟 𝑎𝑛𝑑 𝑠𝑡𝑒𝑔𝑜 𝑖𝑚𝑎𝑔𝑒 in other words [9], the largest payload that can be indistinguishably
transmitted over a channel [10]. There are three significant parameters on
• RS Analysis: RS analysis is used to test the resistance of the which the stenographic capacity depends on namely, Cover Image
steganographic system against statistical attacks. The difference Selection, Embedding Capacity, and Embedding Efficiency. The idea is to
between RM, R_M and SM, S_M should increase with increase minimize the embedding changes for a desired payload and the above-
embedding rate in Stego image. RS Diagram [7] is used to detect the mentioned parameters must be designed considering the embedding
length of the embedding message. In Fig 5, (a) the difference domain or the model (feature) space [14]. The capacity of message that
between RM (SM) and R_M (S_M) increases with increased can be embedded into the cover image is log2|M| bits. If the distance
embedding rate. Hence it can be detected. In (b), the difference d(x,y) → [0, ∞ ) is the distance between cover and stego image
between RM (SM) and R_M (S_M) stays unchanged with increased pixels/transform coefficients at same coordinate (x,y). Then the
embedding efficiency is defined as the embedding bitrate per unit

distortion which is log2|M|/E(d(x,y)) where E is the expected value of d 3. Image Quality Measures
and M is the secret message [11] [16]. However, this simple approach is
not sufficient as different embedding algorithms will have different We have seen in our study of steganographic system requirements that
approach and effects. The popular Markov cover sources accounts for the analysis of imperceptibility, security, cover image selection, embedding
probabilistic impact of mutually independent embedding using KL- capacity and its efficiency rely heavily on the use of Image Quality
Divergence subject to the square root law. It states that payload which metrics (IQM). It is categorized into subjective and objective as seen in
increases at a rate greater than √𝑛 can be detected easily and which Fig 6. Human being’s perception for image quality assessment is termed
increases at a slower rate than √𝑛 is -secure [32-34], where n is the no of as subjective, whereas objective methods are based on theorems and
elements. Another approach is to use the Human Visual System (HVS) procedures to assess the quality automatically using computation rules [8].
properties of grey-scale variation of upper and left neighboring pixel The objective IQA metrices for image steganography is further classified
intensities. into full reference (FR), reduced reference (RF) and no reference (NR)
based on the availability of the cover image. We have emphasized on FR
2.4. Cover Image Selection objective IQA is considered as the most appropriate image fidelity
measure [57] because the distorted stego image quality is assessed in
The steganographic system must have good visual statistical comparison with the cover image which is assumed to be superior in
imperceptibility, i.e., the stego image must not be severely distorted post quality [27][31].
embedding of payload. Imperceptibility or high fidelity of stego image is
dependent on the cover image chosen and image quality metrics used to • Mathematical Metrics: The image is viewed as 2-Dimentional
determine the appropriate cover for a specific application at an accepted signal and the similarity (or dissimilarity) between the cover and the
risk level. The cover image selected must satisfy the image complexity distorted stego image is calculated as distortion or the quality
parameters to determine the embedding region. Ideally the embedding measure [8]. Minskowski measure determines the distance between
region should be adaptively selected part of the cover image adjusted to cover and distorted stego where xi and yi are the ith pixel/DCT
the type of embedding according to the cover image properties determined Coefficient for the cover image, X and the stego image, Y. Here N is
by appropriate image quality measures [12]. Image regions with complex the total number of pixels and Mϵ [1, ∞). If m=2, we get Mean
textures of varying intensities have higher embedding capacity and Square error (MSE).
improve security than smooth regions of simple and constant intensity so
1⁄
that the chance of detection false positive is high i.e. the steganography 𝐷𝑚 = 1/𝑁 ∑𝑁 𝑚
𝑖=1|𝑥𝑖 − 𝑦𝑖 | )
𝑚 (13)
should select a cover image that results in a least detectable stego image.
The selected cover image should guarantee high quality of stego image
which allows high embedding capacity, imperceptibility and enhanced We can obtain PSNR (Peak Signal to Noise Ratio), which is another
security and minimum statistical artifacts. For any cover selected, three every common IQM used in image steganography.
scenarios should be considered:
𝑃𝑆𝑁𝑅 = 10𝑙𝑜𝑔10 (𝛾 2 /𝑀𝑆𝐸) (14)
• S1: Alice has no knowledge of Warden and considers the general

image features which would change during the embedding process Laplacian MSE, absolute error, peak MSE are other metrics used.
such no non-zero AC DCT coefficients or the image pixels in spatial The major drawback of these simple mathematical IQM is the lack
domain. of correlation with perceived quality measurement.
• S2: The steganographer has partial knowledge about the input,
output and statistical characteristics of the steganalysis. • Human Visual System (HVS) IQM: The error signal which is the
• S3: The steganographer has full knowledge about the input, output difference, D(x,y) between the cover and the stego image is
and statistical steganalysis method. normalized according to the visibility perception of human. HVS
[11] follows the four-stage framework for assessing the image
In scenario S3, the steganographer before embedding could use a cover quality. The cover-stego image pair is preprocessed using various
image database to train for a similar steganalysis method and identify methods before decomposing into sub bands for improved
cover images classified as false positive. In SVM based classification performance of quality measurement such as cortex storable
technique the image whose probability distribution function doesn’t pyramid transform, wavelet transform and DCT transform. Details
change significantly after payload embedding is selected as the cover of some of the approaches are in Table 2.
image. Prior available approaches for cover image selection are based on
block texture similarity, in which the cover pixel blocks similar to secret • Structural Information Change (SSIM): In this approach, the
message blocks are replaced and the secret message block’s indices were highly structural nature of the natural scene information is measured
stored in the cover image [16]. Other methods include similarity measures using the highly adaptable nature of HVS. The measure of structural
based on Gabor Filter. The most important factor while choosing the information change between the cover image and the stego image
cover image is the visual quality of the generated stego image i.e, how (SSIM) [8] offers a good estimation of perceived image distortion.
imperceptible it really is. HVS assessment metrices influence adaptive
region selection. Recent approaches include genetic optimization
algorithms to determine the embedding location.

Image quality
parameters
Subjective Objective
Reduced
Double stimulus Full reference No Reference
Reference
Structural
Mathematical
Single stimulus HVS based Information Graphical based
metric
change(SSIM)
Visible Difference Hosaka

MSE, RMSE,PSNR, Visual Fidelity
prdictor, perceptual Plot,Histogram
wPSNR information(VIF)
image distortion,JND Plot
Universal Image model,Transform based
Quality index(UQI) metrics,
MSSIM
Fig. 6 - Image Quality Metrics (IQM) classification.
SSIM(x, y) = (2𝜇𝑥 𝜇𝑦 + 𝐶1 )(2𝜎𝑥𝑦 + 𝐶2 )/(𝜇𝑥2 + 𝜇𝑦2 + 𝐶1 )( 𝜎𝑥2 + 𝜎𝑦2 + 𝐶2 ) Transform (DWT), Hadamard Transform, Dual tree DWT, Double
(15) Density Dual Tree DWT (DD DT DWT) Transform, Integer Wavelet
𝜇𝑥 , 𝜇𝑦 = 𝑚𝑒𝑎𝑛 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑖𝑒𝑠 𝑜𝑓 𝑐𝑜𝑣𝑒𝑟 𝑎𝑛𝑑 𝑠𝑡𝑒𝑔𝑜 𝑝𝑖𝑥𝑒𝑙 Transform, Ridgelet Transform, Curvelet Transform etc which are then
𝜎𝑥 , 𝜎𝑦 = 𝑚𝑒𝑎𝑛 𝑖𝑛𝑡𝑒𝑛𝑠𝑖𝑡𝑖𝑒𝑠 𝑜𝑓 𝑐𝑜𝑣𝑒𝑟 𝑎𝑛𝑑 𝑠𝑡𝑒𝑔𝑜 𝑖𝑚𝑎𝑔𝑒 𝑏𝑙𝑜𝑐𝑘 subjected to an optimal embedding technique to modify the coefficients.
Soft computing tools, genetic algorithm and game theory is often
• Graphical Measure: It is a bivariate image quality assessment considered for this purpose [2] [18]. The transform domain methods are
metric which calculates the distance between the singular values of more immune to image processing operations and hence prove less
the cover image blocks and the distorted stego image block. [7][11] susceptible to steganalytic attacks. Therefore, it is more popular and are
usually preferred over spatial domain methods. Spread spectrum
4. Image Steganography Techniques steganography is a blind scheme that embeds in the noise inherent during
the image acquisition process. It incorporates image restoration and error
Digital image steganography can be broadly categorized into spatial control techniques can be used while extracting the data at the decoder
domain, transform domain, side-informed based steganography, spread side as it doesn’t require the original image while extraction and
spectrum and adaptive steganography as depicted in Fig 8. In spatial outperforms other methods in terms of payload capacity and invisibility,
domain, secret message is embedded in the pixel value of the cover image but tends to be labor intensive. Adaptive based steganography methods
directly [60-63] whereas transform domain methods first transforms the establish a statistical model of the cover image. It is also known as model
cover image from spatial to frequency domain using any one of the based steganography and statistics aware embedding. It has recently
transforms such as Discrete Cosine Transform (DCT), Discrete Wavelet gained widespread popularity because it works adaptively in both spatial
and transform domain. It starts by extracting the statistical global features
of the cover image by estimating the pixel cost change function, before
selecting the locations for data hiding and then actual data embedding
process is carried out accordingly to achieve minimal distortion [19].
Therefore, it provides added layer of security to steganography. The
different techniques are discussed in detail in upcoming subsections.
Table 3 exhibits the relative comparison for some of the major
steganography schemes discussed in literature.
4.1. Spatial Domain Steganography
The most computationally inexpensive and simple image steganography

methods are based on spatial domain steganography. The embedding rate
is measure in bpp (bits per pixel). The techniques can be classified into
direct and indirect. The direct methods work by directly manipulating the
cover image pixel using embedding schemes such as substitution, addition
Fig. 7 - Image Steganography classification. and adjustment with the message bits where as in indirect methods, the

cover image bits are converting to another notational system and the • LSB Matching: In this method, we check whether the LSB of
message bits are hidden or the message bits are hidden in the noise during the cover image pixel Xi matches with message bit. If it doesn’t
image acquisition. According to the embedding principles the below are match the pixel Xi is randomly increased or decreased by one,
the major types of steganography. Basically, there are two trending styles known as (±1 embedding) [23]. The even pixel value is increased
of Least Significant Bit (LSB) modification steganography known as LSB and for odd pixel value it is decreased, using the values with the
replacement (LSB-R) and LSB matching (LSB-M) and its variants. grey level range [0 - 255]. Although, the above two techniques are
very simple, it will introduce some anomalies in the first order
• LSB flipping or Replacement: This method simply replaces the statistics Histogram and will be vulnerable to attacks. The variant
cover image pixels by the secret message bits. Consider a cover LSB-MR (LSB-M revisited) uses a pair of pixels (Xi, Xi+1) as an
image source Xϵ[0, 255] for an 8 bit image and the secret message embedding unit to generate stego pixels (Yi, Yi+1) such that it
Mϵmi, i=1,..n. After embedding, the value of the ith message bit is satisfies the condition.
equal to stego image’s ith pixel Yi. The value of (i+1)th message bit is
embedded in Xi+1 to generate Yi+1 and so on until the entire message 𝐿𝑆𝐵(𝑋𝑖 ) = 𝑀𝑖 (16)
bit mi has been embedded [21][22].
Table 2: IQM available for steganographic image quality analysis
FR IQM Type Key functions Advantages Disadvantages/ Improvements

Mathematical These mathematical models do not correlate
MSR, Peak well with perceived Quality measures [25]. New
Simplicity and mathematical tractability.
MSE,RMSR,PNSR, Based on the similarity or dissimilarity mathematical model such as Universal image
It is used to assess distortions such as
wPSNR, Absolute between the cover image and the Quality index (UQI) should be used, which is
Gaussian Blur, White Noise, Mean shift,
error, normalized distorted stego image. computed using a sliding window approach with
transmission error, Fast Fading, Flat
cross-correlation, etc. a window size of 8x8, leading to a quality map
allocation, JPEG, JPEG 2000 etc.
of the image taking loss of correlation, mean
distortion and variance distortion factors [26].
Visible difference Evaluation of high quality imaging
predictor, system.
Perceptual image Accounts for variation in sensitivity due
The mentioned metrics are tailored for specific
distortion to light level, Spatial frequency (CSF) It is a normalization model for baseline
set of conditions w.r.t scaling and saturation
and contrast masking [27] and use of contrast sensitivity, contrast masking
constants and requires a lot of work to adapt.
steerable pyramid transform for
decomposition to several spatial
frequency and orientation bands [28]
Coder specific The coder model is centred on the
Models, Perceptual frequency analysis and controls the It is suitable for image compression
Image-Coder (PIC) quantization parameters to lessen the applications as it adopts the frequency
Not effective for predicting performance across
perceptibility of errors. Also provides decomposition of a coder with many high
different coders.
better correlation with subjective frequency content and is
evaluation than MSE, MTF weighted Computationally efficient.
MSE [27]
Watson DCT Based A quality metric model that computes
metric, Watson’s the visibility thresholds [27] and the
It computes the perceptually optimal
Wavelet Metric such baseline sensitivity thresholds for the
image dependent quantization matrix for Highly complex HVS causes nonlinearity and
as Shapiro’s EZW DCT coefficients and wavelet
DCT and Wavelet based image coders. existence of supra-threshold problem
algorithm, SPIHT decomposition respectively using
algorithm distortion map [28].
Structural Similarity SSIM is designed as a prototype of the SSIM index map often results in undesirable
approach(SSIM) image distortion as a combination of blockiness when calculated within the window.
The SSIM metric can assess Gaussian
three factors that are loss of correlation, This is overcome using Mean SSIM(MSSIM)
blur, additive Gaussian noise, jpeg and
luminance distortion and contrast evaluates the overall quality of the image by
jpeg 2000
distortion [29]. calculating the average of SSIM values over the
SSIM(x,y)=l(x,y)c(x,y)s(x,y) [30] entire windows

Table 2: Contd, IQM available for steganographic image quality analysis
FR IQM Type Key functions Advantages Disadvantages/ Improvements

Graphical This is a plot considering a pair of
Hosaka Plot & features of the degraded image with the
Good indication of how images are
Histogram Plot corresponding features in the original Both Hosaka plot and Histogram plot do not
degraded using vector error measure and
image. Constructed by plotting the provide a good assessment criterion for other
useful in case the impairment is
number of times a specific value occurs types of distortions.
blockiness, blurriness and fuzziness
in the degraded image versus the value
itself [31].
SVD (Singular Value It is a bivariate image quality measure
It is used to assess distortions such as The computational complexity is of order O(n3)
Decomposition) that computes the distance between the
Gaussian Blur, Gaussian Noise, DC which makes it difficult for use in large images.
singular values of the cover image block
shifting, Sharpening, JPEG, JPEG 2000 If Images are segmented into blocks then
and the singular values of the distorted
etc. processing is much faster
image block [32].
𝐿𝑆𝐵(| 𝑋𝑖 /2| + 𝑋𝑖+1 ) = 𝑀𝑖+1 (17)

indicator channel (I-channel) is used to reveal the presence of the
LSB-MR reduces the asymmetrical artifacts introduced by LSB and secret message in the least significant two bits of the other channel.
LSB-M and also reduces the change rate in terms of contrast. The randomization in embedding provides enhanced security to
Another algorithm based on LSB-M known as EPEs where the complicate the detection of secret data. The secret information and I-
embedding process is described with a table of 2n rows for plane are divided into four sub-blocks and each of the message
embedding i=1..n bits into Yi+1, i=1..n bits by ±1 embedding from blocks is embedded into a specific image block by using "Magic
cover pixels. To extract, a function has to be applied to two adjacent LSB" to make the data extraction even more challenging than the
stego image pixels Yi and Yi+1. The histogram preserving LSB+ and LSB techniques.
LSB++ [29] are one of the most successful LSB technique used
today [33][41]. • Pixel Value Differencing (PVD): This technique is based on the
visual effect of human visual perception capacity. The differences of
• Grey Level Modification (GLM): It was first introduced by Potdar grey values in two-pixels blocks of the cover image are clustered
et al. in 2004 [24]. The method works by mapping the grey levels of according to different contrast and smoothness properties and the
pixels with secret message. A predefined sequence of cover pixels is secret message is embedded into these cluster blocks. [34] This
selected and using the notion of odd and even numbers for mapping. technique yields better result than LSB embedding techniques [63].
The image is a 2-D matrix f(x,y) of MxN size. Grey Level We take two consecutive pixels Pi and Pi+1 in the cover image
Modification modifies the intensity of the grey level scale image at having grey values gi and gi+1. The difference in intensity is
coordinates (x,y) where lϵ[0,255] by n units by an arbitrary function. calculated as 𝑑 = |𝑔𝑖+1 − 𝑔𝑖 |, 𝑑 ∈ [−225,255] . The extremely
The sequence of cover image bits location, p should be smooth regions in the image had d value close to zero (li =0) and d
predetermined. close to ui= -255 to 255 representing edged region. The possible
values of |d| are classified into range Ri, i = 1..n where the bounds
𝑝 = {𝑔(𝑥, 𝑦)|0 ≤ 𝑥 ≤ 𝑀, 0 ≤ 𝑦 ≤ 𝑁} (18) are (li, ui) . The n message bits which can be embedded is formulated
as n=log2(li ,-ui+1)[35] and the next distance is calculated using
The following condition must be met for embedding the value of L 𝑑 ′ = 〈𝑙𝑘 + 𝑏, 𝑑 ≥ 0〉and 𝑑 ′ = −〈𝑙𝑘 + 𝑏), 𝑑 < 0〉 where the value of
for any Pi={0,..(MxN)} is odd, then L is incremented by 1 else it b is 𝑏 ∈ [0, 𝑢𝑘 − 𝑙𝑘 ] which is the value of sub stream S from n bit
remains unchanged. The value of L for every Pi is examined again message. An enhanced variant of PVD called Tri-way PVD works
for mapping the message bit Mi to Pi. If l is even at Pi and Mi is 0, by partitioning the cover image into non- overlapping 2x2 blocks of
then L is not changed else it is decremented by 1. Like LSB 4 pixels [36].
methods, this method is also of low complexity with high payload
embedding capacity [30]. • Pixel Pair Matching (PPM): The technique engages two pixels as
an embedding component to hide a message digit Sn in N-notational
• Pixel Indicator Technique (PIT): This is a modification of LSB system. Here, the LSB of the first pixel is used for embedding one
techniques which aims to enhance security. It works by dividing the message bit and a binary function carries another bit [40]. The
three channels of RGB image into an indicator channel and data improvised PPM techniques are exploiting modification direction
channels. The indicator channel decides which data channel to use (EMD) in which one pixel in the pixel pair is changed by one grey-
for data hiding [33]. The least two significant bits of a specific color scale unit and the message digit is represented in 5-notational
system.

• Predictive coding: The technique embeds the secret message into A variant of QIM is called Dither modulation (DM). It adds a dither
the compression codes during image compression. The technique signal to the input signal before quantization stage and subtracted
works in a two-stages. The first stage is a prediction stage and the after quantization covering all the values of the input signal and di is
second is an entropy coding stage. The secret data is concealed into the dithering signal which is uniformly distributed.
the difference value of the cover image after the prediction stage
using prediction error values [38]. This technique integrates image 𝑦𝑖 = 𝑄𝑚 (𝑥𝑖 + 𝑑𝑖 ) − 𝑑𝑖 Over[−∆/4, ∆/4] (24)
compression and steganography to overcome the bandwidth limited
in the network which results in increased data hiding efficiency [48]. • Histogram Shifting: The process of embedding a MxN grey
The various existing coding techniques are Gradient-adjusted scale image with values on l=[0,255]. The objective is to find the
prediction predictor, the median edge detector [18]. If 𝑥̅ is the minimum zero-point h(z), where zϵ [0,255] and a maximum point
predicted value of x [38], then the predictive rule of the predictor is h(m), where mϵ [0,255]. If the h(z)>0 then coordinates (x,y) is
as follows. recorded and h(z) is set to zero. The part of the histogram H(i) with
iϵ [z, m] is shifted by 1 unit towards right. If the message bit is 1, the
𝑥̅ = {min(𝑥, 𝑦) , 𝑖𝑓 𝑐 ≥ max(𝑥, 𝑦) (19) grey scale pixel value is changed to m+1. IF the message but is 0, m
𝑥̅ = {max(𝑥, 𝑦) , 𝑖𝑓 𝑐 ≤ min(𝑥, 𝑦) (20) remains unchanged. The embedding capacity denoted by C=h(m)-0,
indicates the amount of overhead information [59].
𝑥̅ = {x + y − c, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
4.2. Transform Domain Steganography
Where min(x,y) and max(x,y) function computes the minimum and
maximum values of x and y respectively. Gap Predictor estimates Spatial domain techniques performance in terms of embedding capacity is
the intensities in horizontal and vertical directions represented by gh high. However, it is not robust against statistical attacks. This prompted
and gv. Finally, embedding is carried out in raster scan order from the need for enhanced and secure transform domain steganography. The
top to bottom and from left to right. techniques to implement digital image steganography by converting the
image from spatial to transform domain is termed as transform domain
• Multiple Bit Plane Steganography: The LSB embedding is steganography. It started by adopting the transform domain techniques
extended to hide data in multiple bit planes. It doesn’t consider the used in robust watermarking to image steganography for designing large
local property while embedding which decreases its performance in capacity embedding steganography. There are many candidates transforms
terms of imperceptibility as it follows a non-adaptive embedding such as Discrete Cosine Transform, Discreet Wavelet Transform, Discrete
approach. It’s variant bit plane complexity segmentation (BPCS) Fourier Transform, Haar Transform, Integer Wavelet Transform,
represents the raw image in pure binary coding (PBC) and converts Contourlet Transform etc. [9] These techniques are popular because
it to canonical Grey coding (CGC) system. According to the bit- transform coefficients embedding occurs in more robust areas, spreads
plane, the image is decomposed into a set of binary images and across the entire image, there by producing better resistance against
divided into non-overlapping consecutive blocks of 2x2. The attacks than spatial domain techniques. We shall discuss the most
parameter should be chosen with careful selection [21]. The common transform used in the subsequent section [25-27].
complexity of the image block is computed as below.
• Discrete cosine transforms: JPEG uses DCT to convert from
𝛼 = 𝑘/2 × 2𝐿 × (2𝐿 − 1) (21) spatial to transform domain during compression. The DCT based
steganography system integrates with the image-compression
Where 𝛼 > 𝛼° is the predefined threshold, which help select the algorithm to design various JPEG steganographic schemes [57]. The
image block for data embedding. The message bits are also grouped Non-zero AC DCT Coefficients are used for data embedding
into 2Lx2L. If the message block complexity is less than threshold algorithm. DCT breaks an image into low frequency (FL), mid
𝛼° , then the block is processed using conjugation operator, whose frequency (FM) and high frequency (FH) and embeds in mid
complexity will be (1 − 𝛼) larger than 𝛼° , the message block will be frequency range [20]. Given 2-D image f(x,y) of NxN size, the 2-D
replaced with the image block. The embedding rate is observed to be Discrete cosine transform C(u,v) [6] of image f(x,y) is defined as
as high as 4 bpp without causing severe distortion [62][64]. below.
2𝑥+1 2𝑦+1
• Quantization Based Steganography: The embedding technique 𝐶(𝑢, 𝑣) = 𝛼(𝑢)𝛼(𝑣) ∑𝑁−1 𝑁−1
𝑥=0 ∑𝑦=0 𝑓(𝑥, 𝑦) 𝑐𝑜𝑠 [ ] × 𝑐𝑜𝑠 [ ] (25)
2𝑁 2𝑁
used in digital watermarking known as Quantization Index
Modulation(QIM) is extended to image steganography by quantizing √1
, 𝑓𝑜𝑟 𝑢 = 0
𝑁
an input signal x to y using a set of quantizers Qm(.). The choice of Where 𝐶(𝑢, 𝑣) = { }
√2
, 𝑓𝑜𝑟 𝑢 = 1,2, . . 𝑁 − 1
quantizer is depended on the secret message bit m. QIM [24] is 𝑁
generally employed to transform domain coefficients before

The process begins by dividing the cover image into non-
quantization because it produces signs of discreetness in first order
overlapping blocks and 2D-DCT is applied on each block following
statistics, the histogram in spatial domain when quantization scales
which the DCT coefficients are quantized according to the
greater than two.
quantization table and the secret message bits are then embedded
1 into the quantized DCT Coefficients which are coded using run
𝑦𝑖 = 𝑄𝑚 (𝑥𝑖 ) = ∆[(𝑥𝑖 /∆) + ( )] , if 𝑚𝑖 = 0 (22)
2 length coding and Huffman coding. The High frequency
∆ components of DCT are better suited for embedding because they
𝑦𝑖 = 𝑄𝑚 (𝑥𝑖 ) = ∆[(𝑥𝑖 /∆) + ( )], if 𝑚𝑖 = 1 (23)
2
are visually more resistant to noise than their low frequency

components. [18] The major steganographic tools based on DCT are of Transforms Domain Steganography are Low hiding capacity and
JPEG/JP Hide which uses LSB embedding to replace LSBs of non- Complex computations.
zero quantized AC coefficients which are randomly selected with the
secret bits. YASS (Yet another Steganographic Scheme) divides the 4.3. Adaptive Steganography
image into fixed size of six blocks and within each block, sub-blocks
called Host blocks (H-Block) are randomly selected. The secret Adaptive steganography is presently the most secure technique for
message bits are embedded into the DCT coefficients of the H-Block empirical cover sources for secret data communication [37]
using QIM [6] DCT based steganography received widespread [39][40][42[46][66]. The embedding method embeds the secret message
acceptance with F5 algorithm which embeds by decreasing the while considering costs of modifying the cover image pixel to embed the
absolute value of the coefficient by 1 using matrix encoding. In payload which should minimize a distortion function designed to capture
Outguess, LSB replacement is used for embedding in DCT statistical detectability. There exist two general frameworks for adaptive
coefficients with values not equal to 0 or 1 [18]. Outguess and Steg- embedding paradigm; firstly, by establishing empirical payload–distortion
Hide are statistics preserving steganography which uses statistical bound for additive distortions which uses near-optimal practical coding
restoration but are highly detectable. Steg-Hide avoids changing the schemes known as non-additive distortion. It accounts for the inter-pixel
histogram as DCT coefficients are swapped. In F5, if absolute value correlations and interactions among the embedding modifications.
becomes 0 i.e. when coefficient absolute value is -1 or 1 then it is Example of adaptive non-additive distortion steganography are HUGO
said to have shrinkage and message bit is embedded in next [47], HILL and UNIWARD [53] variants. However, adaptive non-
coefficient. YASS, resists blind steganalysis and embeds at additive distortion function modelling steganography is more challenging
randomized locations [45] because there is no coding technique capable of minimizing an
arbitrary distortion function and hence, the non-additive distortion
• Discrete Wavelet Transform: The DWT hides data in regions that functions is approximated to an additive form such as CMD which uses
are less sensitive to the HVS at the high-resolution detail bands (HL, side informed precover to embed; Secondly, by designing cover image
LH and HH). These techniques increase the robustness while model noise residuals during image acquisition which takes the difference
maintaining high imperceptibility. The higher sub band represents between a pixel and its estimated value, such as MiPOD additive
the finer details of the image and can be used for embedding while steganography [65]. HUGO is the first method based on distortion
the lowest sub band has the most important and relevant information minimization, it reflects on the smallest impact that modified distortions
[18]. DWT provides more compression ratios and avoids of a cover pixel groups have on statistical distribution in SPAM feature
interference from artifacts [6] High Frequency contains edge space. The distortion minimization factor is further extended in Wavelet
information, low frequency contains signal information. DWT is Obtained Weights (WOW) by using correlations between pixels in the
performed in vertical direction followed by Horizontal direction and predicted pixels. In recent times, the most successful heuristic additive
follows these representations LL → for approximation coefficients, distortion bound adaptive approach is UNIWARD [42], which is the
LH → Vertical details and HH →Diagonal details. Secret images are generalization of WOW to be used in any domain (S-UNIWARD, SI-
disintegrated into LH, HL, HH and LL is processed to set the next UNIWARD and J-UNIWARD). UNIWARD distortion function is the
wavelet coefficient values [44]. Along with DWT, several other relative sum of changes rates of cover and stego wavelet coefficients by
transforms such as curvelet transform, Slantlet transform, Integer evaluating their smoothness in multiple direction using 8-tap wavelet
Transform, Contourlet transform etc., are used to generate robust directional filter bank consisting of LH, HL and HH directional high pass
stego systems; each having certain advantage than the other [43]. filter in spatial domain [54].
• (∑3𝑘=1 ∑𝑢𝑣 |𝑊𝑢𝑣(𝑋)

𝑘 𝑘 )
Contourlet Transforms: Contourlet transform has wavelet features − 𝑊𝑢𝑣(𝑌)
𝐷(𝑋, 𝑌) ≜ (26)
and the sub bands at each scale are decomposed into different  +|𝑊𝑢𝑣(𝑋)
𝑘 |
directions. It solves wavelet band mixing problem. Contourlet

transform proves to be more powerful in characterizing images rich k
where Wuv(X) k
and Wuv(Y) are the wavelet coefficients for cover-stego
in directional details and smooth contours. It provides multi-scale image pair (X,Y) and (u, v) is taken over all sub band and stability constant
and multi directional representation of cover image using Laplacian is > 0. In case of J-UNIWARD, first image should be decomposed to
Pyramid. Hence, they are more effective than DWT to capture spatial domain. It follows the same procedure as S-UNIWARD. In side
smooth contours and geometric structures [55]. On decomposition of informed SI-UNIWARD, the distortion is represented as 𝐷 𝑆𝐼 (𝑋, 𝑌) ≜
the cover image by applying contourlet Transform, a low pan and 𝐷(𝑃, 𝑌) − 𝐷(𝑃, 𝑋), where P is the precover used as side information for
high pan sub bands of the image is obtained. Data is embedded in both binary and ternary embedding with syndrome trellis coding. MiPOD
the high pan sub band. Integer Wavelet Transform (IWT) [46][49- is an optimal content adaptive steganographic method whose performance
50] maps an integer dataset. The floating point coefficients of DWT is like empirical detectors built with real image model classifier. The
wavelet filters are the location of embed data embedding. However, image is modelled by mutually independent embedding using LSB-M.
truncation of floating point values to integer causes loss of the The cover image x={x1,..xn} which generates the stego image y={y1,..yn}
hidden information leading to data hiding failure. To avoid, floating [52] to which we apply the probabilistic rule P(yn=xn+1)=βn ,P(yn=xn-1)=βn
point precision of wavelet filters. The LL sub band of IWT is a and P(yn=xn)=1-2βn where the change rate βn=[0,1/3]. The distortion cost
closer copy than LL of DWT of the original image. Disadvantages of MiPOD is pn =ln(1/ βn-2). Steganalysis is performed with maxSRMd2

feature set [49] [51]. The most commonly used coding scheme for near such as BCHopt anf fast VCH-syndrome coding. These heuristic
optimal payload-distortion bound embedding is Syndrome Trellis Codes approaches exploits the rounding error, the quantization step for designing
(STC) [48] [51]. the distortion function, thus resulting in huge improvement in security
The embedding additive distortion function 𝐷(𝑥, 𝑦) = ∑𝑛𝑖 𝑑(𝑥, 𝑦) where against steganalysis attacks. The most attractive feature of the use of side-
d(x,y) is the distortion cause by a single letter distortion. STC is a variant informed steganography is it can be easily integrated in any
of matrix encoding based on convolutional codes such as binary steganographic scheme designed to minimize distortion. The steganalysis
convolutions code denoted by Hϵ{0,1}mxn [43] and message m ϵ{0,1}n. The involves testing the stego image by resizing, color-depth reduction and
embedding and extraction process is determined as below. image color conversion such as color to grey scale conversion in the
𝑬𝒎𝒃: 𝑦 = 𝑎𝑟𝑔𝑚𝑖𝑛 𝐷(𝑥, 𝑤), wϵ{w| Hw=m, w ϵ{0,1}n} and 𝑬𝒙𝒕: 𝑚 = 𝐻𝑦. spatial domain and in the quantization stage during JPEG compression
[42]. The embedding distortion, βij =exp(-aPij))/(1+exp(-aPij)), where Pij is
4.4. Spread Spectrum Based Steganography the cost of each cover element and a>0 is a parameter chosen so that
denominator does not become zero.
The spread spectrum image steganography (SSIS) [54-55] is key-based
blind steganographic scheme, in which the original cover image is not
exactly needed to extract the secret message. Extraction is based on the 5. Possible Research Direction
encryption key. The message is encrypted, encoded with a low rate error
control coding method and embedded in a randomly generated noise using We have discussed how steganography algorithms have advanced over a
a key with the help of a modulator. SSIS is usually modeled as additive period of time in every domain and have given rise to adaptive
white Gaussian noise [21]. The embedded message is passed over to the steganography. In principle, a good steganography algorithm should have
quantizer to generate the stego image, which is send over a secure channel total imperceptibility, high fidelity, maximum embedding capacity
along with the keys. The decoder produces the cover image estimate by without instigating visible and statistical artifacts and acceptable level of
means of image restoration technique using reverse-encoding process. The security. Although the Adaptive approaches provides better performance
imperceptibility factor performance of SSIS scheme is measures with than the traditional spatial and transform domain techniques, it tends to be
signal to noise ratio (PSNR) [58]. A chaos based SSIS scheme has been computationally expensive, complex and lack a benchmark. Some of the
developed where chaotic encryption based on chaotic shift keying is used future research possibility is as listed below:
to encrypt the secret message signal instead of the pseudorandom noise
generator to generate white noise as the embedding channel. At the • Steganography based on empirical method lacks a formal connection
receiver side, chaotic modulation is applied. between the distortion function and statistical detectability which is
the important problem in adaptive steganography. It requires
4.5. Side-Informed Steganography development and analysis of advance coding techniques that are
statistics aware.
In side-informed steganography, the secret message is embedded in the • In SSIS we suggest the use of advanced error correcting codes for
higher quality cover images known as precover while converting to a further improvement. In order to enhance the performance ensure
lower quality cover image using quantization errors. It is capable of that the system’s, bit error rate (BER) is lowered while increasing
embedding the desired payload using binary or ternary embedding the embedding.
technique. Side Information can be included in any of the steganographic • Research focused on additive distortion minimization model that
scheme to increase robustness and imperceptibility. We could also use studies pixel change directions and correlations with neighboring
image acquisition tools (such as camera) and multiple pictures of the same pixels using higher embedding order such as ternary embedding
scene as side-information. However, they are labor intensive and not as should be investigated. The mutual embedding impact of pixels for
computationally efficient as precover approach. The SI-steganography minimizing additive distortion function improves the low
technique has been majorly used in JPEG compression using Perturbed embedding capacity seen in existing adaptive image steganography
Quantization, F5, and nsF5, MME, MMEx and BCHopt, all of these which is an active area of research with many unknown.
methods are centred on the rounding errors of unquantized DCT • Focus should be given to enhance the security of direct embedding
coefficients. In F5 algorithm, the impact of embedding or the embedding by using efficient and less computationally expensive encryption
efficiency is equal for each coefficient so, the total distortion algorithm. This can be achieved by selecting the best cover image
minimization for a given message payload corresponds to the impact of and determining the region based on the cover image properties. Use
minimizing the individual coefficients which is modified. The distortion of optimization algorithm such genetic algorithm, game theoretic
function is designed by computing the impact of embedding costs from approaches along with Markov cover probabilistic distributions
quantized and unquantized image cover sources. The performance of F5 could be explored.
embedding was increased through matrix encoding. This enhanced F5 • Future work could also be aimed at proposing efficient embedding
when incorporated with wet paper code (WPC) resulted in the improved of secret information in precover images which offers higher
version called nsF5 that took care of the shrinkage issue caused by F5 [9] empirical security. The status of side-informed steganography is
embedding originally. The modified matrix embedding (MME) used in limited in the sense that it lacks theoretical analysis. Current work
JPEG steganography made use of the side information of the original uses the difference of the acquired images (white noise), which is
uncompressed cover image to build the distortion function. Here, only challenging and labor intensive. Simple and efficient ways need to
those coefficients with minimum distortion are selected to be modified. be formulated.
Studies have shown that MME performs better in terms of embedding
efficiency, i.e., the number of message bits embedded per embedding
change. SI-steganography evolved with heuristic optimization algorithms

6. Conclusion without compromising in security, capacity and imperceptibility. This is
possible because the message is distributed in the entire region of cover
In this paper, we have reviewed the fundamental concepts of digital image image based on cost of change to cause minimal distortion. We have
steganography and have acknowledged that the key requirement in any reviewed some state of the art techniques and image model in spatial,
good steganography system is high imperceptibility, security of the stego transform, spread spectrum, side informed steganography with more
system, resistance against attacks and high payload embedding capacity emphasis on adaptive image steganography. We have seen the use of
which has been discussed elaborately. The high fidelity of stego images is genetic, game theory, quantum algorithms in steganography, however
obtained by analyzing appropriate statistical and visual IQM to select the these algorithms are computationally expensive and complicated. This
best cover image to apply the steganographic algorithm. We have also paper has been written to provide full-fledged knowledge on the trends of
established that the security of the stego system is based on the entropy digital image steganography and the possible research direction that it will
information. Resistance against attacks is investigated using ROC curve, follow in coming times. We would like to emphasize that the choice of
KL- divergence, FLD, RS-diagram etc. The region of embedding in embedding algorithm, the technique is depended on its application, choice
earlier studies was restricted to edges and textured region in earlier studies of IQM and the steganalytic tool used to break it because what works well
but recent development in adaptive steganography have enabled us to for one system, might not be so for another.
embed high payload using both additive and non-additive approaches
Table 3: Assessment of some of the recent steganography schemes discussed in literature
Review Embedding algorithm Performance criteria
Xin Liao, Kaide Li & Jiaojiao Embeds the encrypted data in the higher frequency of both real and It creates Complex matrix which often requires data
Yin, 2017 et al. [58] imaginary matrix is obtained by DFT coefficients division. The low recovery phase to obtain the image. The stego image has
frequency coefficients of matrix MR(MI) constitute a matrix MRL(MIL) PSNR = 37.98 dB. The recovered image after extracting of
with the size of ρn1 × n2, and then the high frequency coefficients additional data with PSNR = 45.75 dB
constitute matrix MRH(MIH) with the size of (1 − ρ)n1 × n2 where 0 < ρ
≤ 0.5. Encryption of message is done using the stream cipher algorithm.
Xiaolong Li, Bin Li, Bin The host image is divided into non-overlapping blocks such that each The PNSR <= 43.24 dB for an embedding rate=0.5B bpp,
Yang & Tieyong Zeng. 2013 block contains n pixels [60-61]. The n-dimensional histogram is the embedding capacity >= 59.37 DB for 10000 message
et al. [59] generated by counting the frequency of the pixel-value-array of each bits. The embedding method introduces capacity distortion
divided block to embed the secret data. Then the pixel-value-array is error which could be eliminated using optimized-shifting.
divided into two disjoint sets where one set is used to carry hidden data
by expansion embedding while the other set is simply shifted to create
vacant spaces to ensure the reversibility.
Chuan Qin, Chin-Chen EMD causes less embedding distortions and achieves greater embedding The stego image1 has PSNR =52.11 dB and image2 has
Chang & Tai-Jung Hsu, 2014 capacity than the LSB method. Data embedding is based on the prediction PSNR=41.62dB. Each pixel pair in one steganographic
et al. [62] errors to improve the performance. A cover image is chosen which image carries one secret digit in the quinary notation
produces two visually similar steganographic images post embedding. system. Hence, the total hiding capacity of the proposed
[63-65]. The first steganographic image shouldn’t be modified no more scheme is 608680 bits and the embedding rate is 1.16 bpp.
than one gray level (3/5th of pixels are not modified) based on the
modification direction of each stego pixel pair of the image, the
modification direction of each pixel pair in the second steganographic
image (13 twenty-fifths of the pixels are modified 1 and two-fifths of the
pixels in are modified by one, two, three, four, or five gray levels with the
equal probabilities) is adaptively modified which guarantee no
confusions in the image recovery procedure.
Jessica Fridrich & Jan The Multivariate Gaussian (MG) algorithm is content-adaptive which The embedding change probabilities aim to minimize KL
Kodovský, 2013 et al. [65] concentrates the embedding changes in edges and textures. The focus is to divergence between the cover and stego images. The
minimize the steganographic Fisher information to maximize the size of detection performance is evaluated using the minimal total
the secure payload that can be embedded at a fixed level of statistical error under equal prior probabilities and Lagrange
detectability using additive distortion. Syndrome Trellis Coding with multiplier is used to derive the optimal embedding change
LSB-M is used as the embedding algorithm. probabilities for a given payload and image. MG algorithm
offers better security than HUGO for payloads larger than
0.3 bpp.

Table 3: Contd, Assessment of some of the recent steganography schemes discussed in literature
Q Shen, G Liu, W Liu & Y Dai, Image block complexity is determined using linear texture and ROC curve is used as IQM metric. LibSVM is used for training
2015 et al. [66] blocks with larger texture complexity is selected depending on a and testing. It shows better resistance performance capability to
threshold. It employs pixel selection and syndrome-trellis codes typical SPAM steganalysis tool than EALSBMR and HUGO,
(STCs) to determine modification directions for embedding the without correction when tested by embedding with relative bpp
secret message using the modified distortion profile definition of payload of 0.05, 0.1 and 0.2. It was found that the method
HUGO algorithm. performs better in terms of detection rate and false alarm rate
Aref Miri & Karim Faez. 2017 et al. Data hiding in done in adaptive wavelet transform domain of a The cost function and performance metric is evaluated using MSE
[67] two dimensional linear frequency modulation (LFM) The mother and PSNR and PSPNR for computing surrounding variance and
wavelet matrix is used using genetic algorithm to match the secret ensures 80 dB for the stego image for different volume of payload
data with the cover image The inverse transform must be integer capacity.
numbers to guarantee image brightness in spatial domain.
Message is embedded by comparing the threshold value with the
average of the coefficients.
K. Muhammad, 2017 et al. [68] Security has been given utmost importance by encrypting the The stego image quality is evaluated using PSNR, NCC, RMSE,
stego key and secret data using a two-level encryption algorithm and SSIM. For payload of 2, 4, 6, and 8 KB, PSNR ranges from
(TLEA) and multi-level encryption algorithm (MLEA). The 61.7022 db to 55.9027 db
encrypted information is embedded in the host image using an
adaptive LSB substitution method that hides the encrypted
message in blue or green channel on the basis of red channel’s
LSB adaptively using the encrypted bits of secret key following
the scanning order of stego key
Soumendu Chakraborty, 2017 et al. An edge area in the cover image is predicted using Modified The computation complexity of the proposed method is O((m × n)
[69] Median Edge Detector (MMED) predictor to embed the payload, + S) for mxn grey scale image with S secret message, M. The
based on three levels of threshold which determines the number stego image has high MSE and PSNR= 53.84db for embedding
of bits to be embedded M=61,656 bits. It is robust against blind steganalysis evaluated
using Regular Singular (RS) and FLD (Fisher Linear
Discriminant) classifier.
Baby Della, 2015 et al. [71] DWT is used to decompose the cover image into 4 non- The average PSNR value is 55.53. The average MSE (Mean
overlapping sub-bands. At N stage, 3N+1 LL sub bands Square Error) value is around 5-10 for the stego image.
consisting of (LLX), (LHX), (HLX) and (HHX) are obtained by
employing Haar filters.
P.D Shah, 2018 et al. [73] and Steganography is modeled as a search and optimization problem ROC curve is used as IQM metric. LibSVM is used for training
Hamidreza Rashidy Kanan, 2014 et to find the best direction to embed. The starting point in host and testing. It offers better resistance performance capability to
al. [74] image for hiding secret is calculated using Genetic algorithm. The typical SPAM steganalysis tool than EALSBMR and HUGO
direction of pixel scanning has 16 possible states, genes of 4 bits without correction when tested by embedded with relative
length where the starting point is X-offset and Y-offset with 8 bits payload bpp of 0.05, 0.1 and 0.2 and and found that the method
length for each of them. Bit-Planes methods are utilized for performs better in terms of detection rate and false alarm rate
determining LSB planes in host pixels for embedding secret data
in host pixels. The genes are separated in two distinctive groups,
first group contains the genes that denote the embedding location,
and the second includes the genes manage adaptability more with
the host image by changing some values.
Tomas Denemark, 2017 et al. [75] It is based on MiPOD, the multivariate Gaussian model is The cost function and performance metric is evaluated using MSE
imposed on noise that estimates its parameters from the available and PSNR and PSPNR for computing surrounding variance and
precover or side information. The embedding rule is constrained ensures 80 dB for the stego image for different volume of payload
to a binary function. The Model based SI-MiPOD is designed to capacity
work for both RGB and JPEG domain. The average statistical
spread in terms of the standard deviation across ten image
databases is 0.0029 with 91% of all spreads falling in the range
0.0010–0.0040

Table 3: Contd, Assessment of some of the recent steganography schemes discussed in literature
Jiang, 2017et al. [77] The cover image is processed using quantum image processing (QIP) to give The stego image quality is evaluated using PSNR, NCC,
the representation of quantum images, followed by image scrambling to RMSE, and SSIM. For payload of 2, 4, 6, and 8 KB,
transform the cover image into a disordered image by permutations of the PSNR ranges from 61.7022 db to 55.9027 db
pixels into new positions. The 2n CNOT gates are used to test the embedding
position information of I and M representation of quantum representation
(NEQR). Two blind steganography schemes is developed by embedding using
simple LSB in which the message bits substitute for the pixels LSB directly.
Another scheme uses block LSB which embeds a message bit into a number of
pixels that belong to one image block. The message is regained using
extracting circuits only according to the stego cover.
Vahid Sedighi, 2015 et al. The cover pixels probabilities are changed by a certain amount derived from The computation complexity of the proposed method is
[9] and Yuan Bian, 2017 et the cover model to minimize the power of optimal statistical test. This model O((m × n) + S) for mxn grey scale image with S secret
al. [37] driven approach improves the empirical security by allowing the embedding message.
changes in highly textured areas to have larger amplitude to embed larger IQM for stego is PSNR, MSE. High PSNR of 53.84db
payload. The sender first estimates the cover model parameters, the pixel for modification of 61,656 bits. It is robust against blind
variances and models the pixels as a sequence of independent non-identically steganalysis evaluated using Regular Singular (RS) and
distributed generalized Gaussian random variables. Embedding is done using FLD(Fisher Linear Discriminant) classifier
syndrome-trellis codes and pentary embedding, computed by solving a pair of
non-linear algebraic equations.
Bin Li, 2015 et al. [44] The cover image is decomposed into several sub-images, in which message The average PSNR value is 55.53. The average MSE
segments are embedding in local textured regions with well-known schemes (Mean Square Error) value is around 5-10 for the stego
using additive distortion functions by clustering the locations of embedding image.
modifications and directions of embedding modifications. The bias of a pixel
cost is updated based on change of neighboring pixels direction
(positive/negative). Ternary Embedding (±1) is used.
Guo, 2012 et al. [78] Designs an additive distortion metric for JPEG steganography. Embedding is For a secret image of 256x256 embedded in a stego
done uniformly using Syndrome Trellis Coding by modifying nonzero image of 512x512, the average PNSR is 45.13. The
quantized DCT coefficients. Uniformity is made possible by using uniform genetic algorithm parameters used are Population size of
embedding distortion metric (UED) which is based on the magnitude of the 300, Crossover rate =0.7 Mutation rate= 0.04
DCT coefficients and the intra- and inter-block neighbouring coefficients. Replacement rate=0.8. The average embedding capacity
for 0.5 bpp to 3.95 bpp is in the range of 54.25 to 35.42.
References
[1] Kharrazi, Mehdi, Husrev T. Sencar & Nasir Memon. (2007). Image Heidelberg.
Steganography and Steganalysis: Concepts and Practice. Lecture [5] D. Ker. (2009). The Square Root Law requires a linear key.
Notes Series Institute for Mathematical Sciences, National Proceedings of 11th ACM Workshop on Multimedia and
University of Singapore. Security, New Jersey, USA.
[2] C. Vanmathi & S. Prabu. (2018). Image Steganography using [6] Bui Cong Nguyen, Sang Moon Yoon & Heung-Kyu Lee. (2006).
Fuzzy Logic and Chaotic for Large Payload and High Multi Bit Plane Image Steganography. Springer Conference on
Imperceptibility. Springer International Journal of Fuzzy Digital Watermarking (IWDW), Lecture Notes in Computer
Systems, Berlin, Heidelberg. Science, Berlin, Heidelberg.
[3] A. Cheddad, J. Condell, K. Curran & P.M. Kevitt. (2010). Digital [7] Kim-Han Thung & Paramesran Raveendran. (2009). A survey of
image steganography: survey and analysis of current methods, image quality measures. IEEE International Conference for
Elsevier Journal on Signal Processing, 727–752, Northern Ireland, Technical Postgraduates (TECHPOS), Kuala Lumpur, Malaysia.
UK. [8] Linjie Guo, Jiangqun Ni & Yun Qing Shi. (2014). Uniform
[4] C. Cachin. (1998). An information-theoretic model for Embedding for Efficient JPEG Steganography. IEEE Transactions
steganography. Aucsmith D. (eds) Information Hiding, Springer, on Information Forensics and Security, 814-825.
Lecture Notes in Computer Science, 1525, 306–318, Verlag, [9] Vahid Sedighi, Jessica Fridrich, Remi Cogranne. (2015). Content-

adaptive pentary steganography using the multivariate generalized Image Quality and System Performance, California, USA.
Gaussian cover model (MVGG). Electronic Imaging Proceedings [27] Kalpana Seshadrinathan1, Thrasyvoulos N. Pappas, Robert J.
of Media Watermarking Security, and Forensics. Safranek, Junqing Chen, Zhou Wang, Hamid R. Sheikh & Alan C.
[10] Jessica Fridrich, Tomas Penvy & Jan Kodovsky. (2007). Bovik. (2009). Image Quality Assessment. The Essential Guide to
Statistically undetectable jpeg steganography. ACM Proceedings of Image Processing.
the 9th workshop on Multimedia & security, Texas, USA. [28] Thrasyvoulos N Pappas & Robert J Safranek. (2005). Perceptual
[11] Feng, Pan. Jun Li, Xiuguang Li & Yao Guo. (2011). Criteria for Image Quality valuation. Handbook of Image and
Steganography based on Minimizing Embedding Impact function Video Processing.
and HVS. IEEE International Conference on Electronics [29] Karem Qazanfari & Reza Safabakhsh. (2014). A new
Communications and Control (ICECC), Ningbo. China. steganography method which preserves histogram,:Generalization
[12] N. Provos & P. Honeyman. (2003). Hide and seek: An introduction of LSB ++. Elsevier Ariticle in Information Sciences.
to steganography, IEEE journal on Security and Privacy, 32-44. [30] MA Khan, V Potdar & E Chang. (2004). An architecture platform
[13] R. Cogranne, V. Sedighi, T. Denemark, & J. Fridrich. (2014). for grey level modification steganography system IEEE Industrial
Using statistical image model for steganographic distortion Electronics Society, South Korea.
function based on detectability. IEEE Workshop on Information [31] A.M. Eskicioglu & P.S.Fisher. (1995). Image quality measures and
Forensics and Security, Atlanta, USA. their performance. IEEE Transactions on Communications.
[14] A. Ker, T. Pevny, J. Kodovsky & J. Fridrich. (2008). The square [32] Nguyen, Truc & Jong-Myon Kim. (2013). Copyright Protection for
root law of steganographic capacity. Proceedings of 10th ACM Fire Video Images using an Effective Watermarking Method. KIPS
Workshop on Multimedia and Security, 107–116. Transactions on Software and Data Engineering.
[15] Chin-Chen Chang, Yung Chen Chou & Chia Chen Lin. (2009). A [33] Muhammad, Khan, Muhammad Sajjad, Irfan Mehmood, Seungmin
steganographic scheme based on wet paper codes suitable for Rho & Sung Wook Baik. (2015). A novel magic LSB substitution
uniformly distributed wet pixels. IEEE International Symposium method (M-LSB-SM) using multi-level encryption and achromatic
on Circuits and Systems, Taiwan. component of an image. Multimedia Tools and Applications.
[16] T. Filler, A. Ker & J. Fridrich. (2009). The square root law of [34] Wu, D.C & Wen-Hsiang Tsai. (2003). A steganographic method
steganographic capacity for Markov covers. Proceedings of Media for images by pixel-value differencing. Pattern Recognition
Forensics and Security, California, United States. Letters.
[17] Mansi S. Subhedara & Vijay H. Mankarb. (2014). Current status [35] Shrikant S. Agrawal & Samant.R.M. (2011). Data Hiding in Gray-
and key issues in image steganography: A survey. Elsevier Scale Images Using Pixel Value Differencing. Communications in
Computer Science review, 95-113. Computer and Information Science, Springer, Berlin, Heidelberg.
[18] H. Dadgostar & F. Afsari. (2016). Image steganography based on [36] Ki-Hyun Jung & Kee-Young Yoo. (2014). Directional Data Hiding
interval-valued intuitionistic fuzzy edge detection and modified Method for Digital Images, Cryptologia.
LSB", Elsevier Journal of Information Security and Applications. [37] Yuan Bian, G. Tang, Z. Gao & S. Wang. (2017). A distortion cost
[19] Kang Myeongsu, Linh Tran Ho; Yongmin Kim; Jong-Myon Kim; modification strategy for adaptive pentary steganography. Springer
Cheol Hong Kim. (2011). Image watermarking using a Journal on Multimedia Tools and Applications.
dynamically weighted fuzzy c-means algorithm. SPIE Journal on [38] Yuan-Hui Yu,Chin-Chen Chang & Yu-Chen Hu. (2005). Hiding
Optical Engineering. secret data in images via predictive coding. Pattern Recognition
[20] M. Omoomi, S. Samavi & S. Dumitrescu. (2011). An efficient high [39] Weiming Zhang, Zhuo Zhang; Lili Zhang; Hanyi Li & Nenghai
payload ±1 data embedding scheme. Springer Multimedia Tools Yu. (2017). Decomposing Joint Distortion for adaptive
and Applications, USA. steganography. IEEE Transactions on Circuits and Systems for
[21] Sarreshtedari Saeed & Mohammad Ali Akhaee. (2013). One-third video technology.
probability embedding: a new ±1 histogram compensating image [40] Wenbo Zhou, Weiming Zhang & Nenghai Yu. (2017). A New
least significant bit steganography scheme, Journal on IET Image rule for cost reassignment in adaptive steganography. IEEE
Processing. transactions on information forensics and security.
[22] Chang, Chin-Chen & Hsien-Wen Tseng. (2009). Data hiding in [41] Antonio Tasheva, Zhaneta Tasheva & Plamen Nakov. (2017).
images by hybrid LSB substitution. Third International Conference Image based steganography using modified LSB insertion mwthod
on. IEEE Multimedia and Ubiquitous Engineering, China. with contrast stretching. Proccedings of the ACM 18th International
[23] Zhang, Jun, Feng Xiong & Dan Zhang. (2012). Steganalysis for Conference on Computer Systems and Technologies, Bulgaria.
LSB Matching Based on the Dependences Between Neighboring [42] Qingqing Shen, Guangjie Liu, Weiwei Liu & Yuewei Dai. (2015).
Pixels. Journal of Multimedia. Adaptive image steganography based on pixel selection. IEEE
[24] V.M Potdar & E. Chang. (2004). Grey level modification International Conference on Progress in Informatics and
steganography for secret communication. IEEE Conference on Computing (PIC), Nanjing, China.
Industrial Informatics, Berlin. [43] Baby Della, Jitha Thomas, Gisny Augustine, Elsa George & Neenu
[25] Zianou Ahmed Seghir & Fella Hachouf. (2011). Edge-region Rosia Michael. (2015). A Novel DWT Based Image Securing
information measure based on deformed and displaced pixel for Method Using Steganography. Procedia Computer Science.
image quality assessment. Elsevier Journal on Signal Processing: [44] Bin Li, Ming Wang, Xiaolong Li, Shunquan Tan & Jiwu Huang.
Image Communication. (2015). A Strategy of Clustering Modification Directions in Spatial
[26] Shnayderman Aleksandr, Alexander Gusev, Ahmet M. Eskicioglu, Image Steganography. IEEE Transactions on Information
& D. Rene Rasmussen. (2003. An SVD-based grayscale image Forensics and Security.
quality measure for local and global assessment. Proceedings of [45] Li, Jun, Xiaoyuan Yang, Xin Liao, Feng Pan & Minqing Zhang.

(2017). A game-theoretic method for designing distortion function watermarking techniques: An overview and a classification.
in spatial steganography. Multimedia Tools and Applications. EURASIP Journal on Information Security, Pages 1–19.
[46] Kavitha, C. T & C. Chellamuthu. (2010). Multimodal medical [62] Chuan Qin, Chin-Chen Chang & Tai-Jung Hsu. (2014). Reversible
image fusion based on Integer Wavelet Transform and Neuro- data hiding scheme based on exploiting modification direction with
Fuzzy. International Conference on Signal and Image, Chennai, two steganographic images. Springer Science and Business Media,
India. New York.
[47] Luo, Xiangyang, Xiaofeng Song, Xiaolong Li, Weiming Zhang, [63] Lee CF, Chen HL & Tso HK. (2010). Embedding capacity raising
Jicang Lu, Chunfang Yang & Fenlin Liu. (2016). Steganalysis of in reversible data hiding based on prediction of difference
HUGO steganography based on parameter recognition of expansion. Journal of Systems and Software.
syndrome-trellis-codes. Multimedia Tools. [64] Chin-Feng Lee, Chin-Chen Chang & Kuo-Hua Wang. (2008). An
[48] Winkler, Antj. (2012). Advances in Syndrome Coding based on improvement of EMD embedding method for large payloads by
Stochastic and Deterministic Matrices for Steganography. pixel segmentation strategy. Image and Vision Computing.
Saechsische Landesbibliothek- Staats- und Universitaetsbibliothek, [65] Jessica Fridrich & Jan Kodovský. (2013). Multivariate Gaussian
Dresden. Model for Designing Additive Distortion For Steganography.
[49] R. O. El Safy, H. H. Zayed & A. El Dessouki. (2009). An adaptive ICASSP.
steganographic technique based on integer wavelet transform. [66] Q Shen, G Liu, W Liu & Y Dai. (2015). Adaptive image
International Conference on Networking and Media Convergence, steganography based on pixel selection. IEEE International
Cario, Egypt. Conference on Progress in Informatics and Computing (PIC),
[50] Feng Gu & Lu J. (2006). A new composite implicit iterative Nanjing, China.
process for a finite family of nonexpansive mappings in Banach [67] Aref Miri & Karim Faez. (2017). Adaptive Image Steganography
spaces. Fixed Point Theory and Applications. based on Transform Domain via Genetic Algorithm. Optik -
[51] Thomas Mittelholzer. (2000). An Information-Theoretic Approach International Journal for Light and electron.
to Steganography and Watermarking. Lecture Notes in Computer [68] K. Muhammad, J. Ahmad & N. U. Rehman. (2017). CISSKA-LSB:
Science, Springer, Berlin, Heidelberg. color image steganography using stego key-directed adaptive LSB
[52] Y. Qi, Y. Wang & J. Yuan. (2009). Audio Steganalysis Based on substitution method. Multimedia Tools Applications.
Co-occurrence Matrix and PCA. International Conference on [69] Soumendu Chakraborty, Anand Singh & JalalCharul Bhatnagar.
Measuring Technology and Mechatronics Automation, 433-436, (2017). LSB based non blind predictive edge adaptive image
Zhangjiajie, Hunan. steganography. Multimedia Tools and Applications, New York.
[53] Vojtěch Holub & Jessica Fridrich. (2013). Digital image [70] Qi Li, Xin Liao, Guoyong Chen & Liping Ding. (2017). A Novel
steganography using universal distortion. ACM workshop on Game-Theoretic Model for Content-Adaptive Image
Information hiding and multimedia security, Montpellier, France. Steganography. IEEE 37th International Conference on
[54] Mohan, Malini, & P.R Anurenjan. (2011). A new algorithm for Distributed Computing Systems Workshops (ICDCSW), Atlanta,
data hiding in images using contourlet transform. Recent Advances GA, USA.
in Intelligent Computational Systems(RAICS) IEEE, Trivandrum, [71] Saidi M, Hermassi H & Rhouma R. (2017). A new adaptive image
India. steganography scheme based on DCT and chaotic map. Multimedia
[55] Hayat Al-Dmour & Ahmed Al-Ani. (2016). A steganography Tools Applications.
embedding method based on edge identification and XOR coding. [72] Mingwei Tang, Shenke Zeng,Xiao liang Chen,Jie Hu & Yajun Du.
Expert Systems with Applications. (2016). An adaptive image steganography using AMBTC
[56] Qingzhong Liu & Zhongxue Chen. (2015). Improved Approaches compression and interpolation technique. Optik -International
with Calibrated Neighboring Joint Density to Steganalysis and Journal for Light and Electron Optics, Pages 471-477.
Seam-Carved Forgery Detection in JPEG Images. ACM [73] P.D Shah & R.S Bichkar. (2018). A Secure Spatial Domain Image
Transactions on Intelligent Systems and Technology, New York, Steganography Using Genetic Algorithm and Linear Congruential
NY, USA. Generator. International Conference on Intelligent Computing and
[57] Chong-Yaw Wee, Raveendran Paramesran & R. Applications, Singapore.
Mukundan,Xudong Jiang. (2010). Image quality assessment by [74] Hamidreza Rashidy Kanan & Bahram Nazeri. (2014). A novel
discrete orthogonal moments. Pattern Recognition. image steganography scheme with high embedding capacity and
[58] Xin Liao, Kaide Li & Jiaojiao Yin. (2017). Separable data hiding tunable visual image quality based on a genetic algorithm. Expert
in encrypted image based on compressive sensing and discrete Systems with Applications.
fourier transform. Springer Science & Business Media, New York. [75] Tomas Denemark & Jessica Fridrich. (2017). Model Based
[59] Xiaolong Li, Bin Li, Bin Yang & Tieyong Zeng. (2013). General Steganography with Precover. Media Watermarking, Security, and
Framework to Histogram-Shifting-Based Reversible Data Hiding. Forensics, Page 56-66.
IEEE Transactions on Image Processing. [76] Ou, Bo, Xiaolong Li & Jinwei Wang. (2016). High fidelity
[60] Zhicheng Ni, Yun-Qing Shi, N. Ansari & Wei Su. (2006). reversible data hiding based on pixelvalue-ordering and pairwise
Reversible data hiding. IEEE Transactions on Circuits and Systems prediction-error expansion. Journal of Visual Communication and
for Video Technology. Image Representation.
[61] R. Caldelli, F. Filippini, & R. Becarelli.(2010). Reversible [77] Jiang, Nan, Na Zhao, & Luo Wang. (2017). LSB Based Quantum

Image Steganography Algorithm. International Journal of
Theoretical Physics.
[78] Guo, Linjie, Jiangqun Ni, & Yun Qing Shi. (2012). An efficient
JPEG steganographic scheme using uniform embedding. IEEE
International Workshop on Information Forensics and Security
(WIFS), Tenerife, Spain.

SSRN Id3171494

Uploaded by

Copyright:

Available Formats

SSRN Id3171494

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3171494

Uploaded by

Copyright:

Available Formats

3rd International Conference on Internet of Things and Connected Technologies

A Survey on Digital Image Steganography: Current Trends and

Debina Laishrama, Dr. Themrichon Tuithungb

successfully used in various businesses, multimedia companies apart from

Electronic copy available at: https://ssrn.com/abstract=3171494

Technical Single Cover

Information Hiding Text

Robust Finger print

Fig. 1 - Different Information Hiding disciplines

Fig. 2 - General model of Steganography

Electronic copy available at: https://ssrn.com/abstract=3171494

schemes presented so far is limited in their applicability towards a specific

2.1. Security of the steganographic method

The security of a steganographic system can be analyzed by considering

If the mutual information after embedding is as below:

𝐼(𝐸, (𝑆, 𝐶)) = 𝐻(𝐸) − 𝐻(𝐸|(𝑆, 𝐶)) > 0 (4)

Then it indicates insecure system. Therefore, the necessary condition for

Fig. 4 - ROC curve example of three steganalytic methods A, B and C

Electronic copy available at: https://ssrn.com/abstract=3171494

• Kullback – Leibler Divergence: The security of steganographic

• Linear Classifier with Higher- Order Statistics: The KL distance

• Fisher’s Linear Discriminant: Correlation based security analysis

Electronic copy available at: https://ssrn.com/abstract=3171494

• S1: Alice has no knowledge of Warden and considers the general

Electronic copy available at: https://ssrn.com/abstract=3171494

Visible Difference Hosaka

Fig. 6 - Image Quality Metrics (IQM) classification.

4.1. Spatial Domain Steganography

The most computationally inexpensive and simple image steganography

Electronic copy available at: https://ssrn.com/abstract=3171494

Table 2: IQM available for steganographic image quality analysis

FR IQM Type Key functions Advantages Disadvantages/ Improvements

Electronic copy available at: https://ssrn.com/abstract=3171494

FR IQM Type Key functions Advantages Disadvantages/ Improvements

𝐿𝑆𝐵(| 𝑋𝑖 /2| + 𝑋𝑖+1 ) = 𝑀𝑖+1 (17)

Electronic copy available at: https://ssrn.com/abstract=3171494

generally employed to transform domain coefficients before

Electronic copy available at: https://ssrn.com/abstract=3171494

• (∑3𝑘=1 ∑𝑢𝑣 |𝑊𝑢𝑣(𝑋)

directions. It solves wavelet band mixing problem. Contourlet

Electronic copy available at: https://ssrn.com/abstract=3171494

Electronic copy available at: https://ssrn.com/abstract=3171494

Table 3: Assessment of some of the recent steganography schemes discussed in literature

Review Embedding algorithm Performance criteria

Electronic copy available at: https://ssrn.com/abstract=3171494

Review Embedding algorithm Performance criteria

Electronic copy available at: https://ssrn.com/abstract=3171494

Review Embedding algorithm Performance criteria

Electronic copy available at: https://ssrn.com/abstract=3171494

Electronic copy available at: https://ssrn.com/abstract=3171494

Electronic copy available at: https://ssrn.com/abstract=3171494

Electronic copy available at: https://ssrn.com/abstract=3171494

You might also like