Ghosal 2018

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

JID: JISA

ARTICLE IN PRESS [m5G;April 21, 2018;14:46]


Journal of Information Security and Applications 0 0 0 (2018) 1–11

Contents lists available at ScienceDirect

Journal of Information Security and Applications


journal homepage: www.elsevier.com/locate/jisa

On the use of the Stirling Transform in image steganography


S.K. Ghosal a, J.K. Mandal b,∗
a
Department of Computer Science & Engineering, Future Institute of Technology, Boral 700154, India
b
Department of Computer Science & Engineering, University of Kalyani, Kalyani, Nadia 741235, India

a r t i c l e i n f o a b s t r a c t

Article history: Stirling Transform (ST) is primarily used in combinatorial mathematics to convert a polynomial sequence
Available online xxx of n non-negative integers into another integer sequence. The source sequence can also be re-generated
from the target sequence based on the principle of inverse Stirling Transform (IST). This property is con-
Keywords:
Steganography sidered as an interesting study and has been exploited for steganography in digital images. In this paper,
Stirling transform the cover image is decomposed into the pair (n = 2) of neighboring pixel components in row major order.
Inverse stirling transform Stirling Transform (ST) converts the pair of pixel components into the pair of transformed components
Payload corresponding to red, green and blue channels. The obtained transformed components are always non-
One time password negative, which fabricates varying sizes of the secret image based on the specified payload in the range
MD5 of 0.5 to 3 bpB (bits per Byte). To address the security issue, the secret image is scrambled through a
simple scrambling process prior to embedding. The pixel adjustment process ensures the non-occurrence
of overflow/underflow during embedding. Inverse Stirling Transform (IST) re-computes the pair of pixel
components from the embedded pair of transformed components. This process is repeated till the secret
information is concealed and the stego-image is produced. At the recipient end, the reverse procedure
is applied to extract the concealed information from the stego-image based on the following parame-
ters: shared PIN, system time and the one time password (OTP). Message digest obtained through MD5
algorithm is used to verify the authenticity. Simulation results ensures that the proposed scheme out-
performs Chang et al.’s [17], Xia et al.’s [18], 3-LSB [2] and hybrid GEMD [7] schemes in terms of image
quality with respect to increasing payloads.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction proposed so far of which the simplest is the Least Significant Bit
(LSB) [2]. The technique is implemented by fabricating the secret
Steganography is a technique of fabricating useful digital in- bits at the least significant bit position of each pixel component. To
formation into the cover media (images, audios and videos etc.) embed k number of bits in each pixel component, the LSB scheme
in an imperceptible manner. The main motivation of steganogra- is renamed as k-LSB. It offers high fidelity but suffers from lack
phy is to hide a large amount of secret data (often referred to of security and payload. In 2003, Wu et al. [3] proposed a novel
as payload) with acceptable quality distortion in the stego-image. approach based on Pixel Value Differencing (PVD) where the dif-
The major applications of steganography include secret communi- ference of each pair of pixels is computed. The difference belongs
cation, copy-right protection, integrity verification and authentica- to a predefined range where larger difference ensures more num-
tion. Since, the volume of payload is usually very high, any sort of bers of bit embedding and smaller difference results less numbers
alteration made to the stego-image is easily detectable. As a conse- of bit embedding. In 2006, Zhang and Wang [4] proposed Exploit-
quence, steganography is considered as an excellent authenticating ing Modification Direction (EMD) based data hiding scheme where
tool to restrain the problem of digital piracy. The authentication secret data is fabricated in the range of 0 to 2n by modifying only
process is operated either by incorporating a cryptographic ap- one of n pixels. That means to fabricate n bits only one pixel is
proach such as Hashing (MD5, SHA-1), Digital Signatures etc. [1] or increased or decreased by 1 among n pixels. The method has good
by setting up a threshold value for embedding secret bits. image fidelity but the capacity is only 1.161 bits per Byte (bpB). In
Steganographic techniques are available both in spatial and 2007, Lee et al. [5] proposed a technique which enhanced the pay-
transform domain. Several spatial domain techniques have been load capacity from 1.161 bpB to 1.51 bpB by considering two neigh-
bouring pixels at a time. In 2009, Chao et al. [6] proposed the di-
amond encoding scheme to fabricate secret information into each

Corresponding author. pair of cover pixels based on the (2n2 + 2n + 1)-ary notational sys-
E-mail address: jkm.cse@gmail.com (J.K. Mandal).

https://doi.org/10.1016/j.jisa.2018.04.003
2214-2126/© 2018 Elsevier Ltd. All rights reserved.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

2 S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11

tem. The diamond characteristics value can alleviate the degrada- tra level of security is incorporated into the proposed technique
tion and elevate the payload drastically. In 2010, Kim et al. [7] pro- by scrambling the secret image.
posed the 2-EMD and the EMD-2 schemes which can alter the LSB • Discrete Cosine Transform (DCT) is usually applied over 8 × 8
values of two pixels at most. The EMD-2 is superior over the orig- blocks of the cover image. Transformed coefficients are usu-
inal EMD and 2-EMD with similar quality degradation under the ally fractional and hence, the computational complexity of
identical conditions. In the same year, Chen et al. [8] proposed the the embedding process is significantly affected. On the other
flexible EMD-based technique which takes the overflow problem hand, Stirling Transform (ST) produces non-negative trans-
into the consideration. To address the overflow issue, the danger formed components as output. These components are positive
pixel components which may result in overflow problem are iden- integer values that lead to the generation of non-fractional co-
tified and are altered to safe area. In 2011, Kieu and Chang [9] ex- efficients on embedding. In addition, the integer calculation im-
ploited eight modification directions instead of four directions for proves the performance of the embedding process in terms of
embedding secret digits into two neighbouring gray-level pixels execution efficiency.
of the cover image. This scheme provides improved payload over • The complex nature of Discrete Fourier Transform (DFT) sup-
Zhang et al.’s EMD scheme [4] with acceptable visual impercep- ports fabrication of secret information either in phase or in
tibility. In 2013, Kuo and Wang [10] proposed a generalized ex- the magnitude of the coefficients. A serious drawback of both
ploiting modification direction method which is superior over all phase and magnitude properties of DFT is the reconstruction
EMD based methods in terms of payload capacity. In this scheme, of the real image subsequent to IDFT since the complex conju-
the embedding secret data based from (2n + 1) to 2n+1 and hidden gate symmetries of the DFT coefficients must be preserved. In
data capacity to (n + 1) binary bits embedded into n adjacent pix- contrast, the Stirling Transform (ST) generates real transformed
els. In 2014, Kuo et al. [11] proposed a hybrid GEMD data hiding coefficients instead of complex coefficients.
method which is improved in terms of data hiding capacity. The • Most of the information embedding techniques is based on
scheme changed the variable code into fixed that eliminated ad- gray-scale images whereas very few of them address color im-
ditional communication and the number of projection vectors has ages. The color images offers high payload within acceptable
been minimized. distortion by considering red, green and blue (RGB) compo-
In contrast to spatial domain schemes discussed so far, trans- nents instead of considering only the luminance component.
form domain techniques works by partitioning the cover images Hence, the secret information is fabricated separately into the
into m × n blocks in a sliding window manner. Each block is con- red, green and blue channels and the payload becomes three
verted into transform domain followed by embedding of secret in- times of that of the gray-scale image by keeping the quality
formation. Inverse transform on the identical block yields the gen- metrics almost intact.
eration of the stego-image. It is apparent from the above litera-
ture that the exploiting modification direction (EMD) based spatial Sections 2 and 3 of the paper deal with the formulation
domain techniques such as EMD, FEMD, GEMD and hybrid GEMD of Stirling Transform and the novel scrambling process. Compu-
[12] etc. are superior over the traditional bit-replacement policies tational complexities and the proposed technique are discussed
(including LSB, PVD etc.) in terms of payload and stego-image qual- in Sections 4 and 5. Results and discussions are elaborated in
ity. In 2013–2014, Ghosal and Mandal [12,15] proposed two novel Section 6. Conclusions and future scope are drawn in Section 7 and
embedding techniques based on Stirling Transform (ST) and Bino- the references are given in the end.
mial Transform (BT) to verify the authenticity of color images. The
limitation of both techniques is the inability of supporting variable 2. Formulation of Stirling Transform
payload since it is applicable only for 2 bpB.
In this study, the advantages of exploiting modification direc- In combinatorial mathematics, Stirling Transform (ST) deals
tion (EMD) and Stirling Transform (ST) are taken into account to with the transformation of a polynomial sequence of n non-
propose a novel steganographic technique. Hence, the concept of negative integers into another sequence. In digital image process-
designing and implementing the proposed technique is considered ing, this transform can be used to convert the pixel components
as the novel idea. The technique summarizes the following bene- {ak } into the transformed components {bk } using the formula given
fits: in Eq. (A.1).
 

n
n
• Embedding of secret bits is done on varying proportion into bn = a (A.1)
k k
the pair of transformed components as the proposed scheme k=1
offers payload in the range of 0.5 to 3 bpB. In contrast, three
n
well known EMD based schemes such as Zhang et al.’s [4], Lee Where, k = 1, 2, 3,…. and { } is the Stirling number of the second
k
et al.’s [5] and Kim et al. [7] methods offered fixed payloads of
kind, also denoted S(n,k) (with a capital S). It can be expressed us-
1.161, 1.51 and 2 bpB, respectively.
ing Eq. (A.2).
• There is a provision of embedding only one bit in each pair of
components unlike the hybrid GEMD [11] method. One bit of  
1 
k
k
insertion into each component ensured only 0.5 bpB of payload; S (n, k ) = (−1 )i i (k − i )n (A.2)
k!
however, as far as the image quality is concerned, the stego- i=0
image is least degraded. By the formulation of pair of pixel components (a1 , a2 ), the
• Overflow and underflow are considered as two serious prob- transformed components (b1 , b2 ) are obtained as follows:
lems in steganography. Since the secret information is con-
cealed into the transformed components in Stirling Transform b1 = S ( 1, 1 ) × a1 [k = 1, and n = 1]
(ST) domain, there is a possibility of overflow/underflow. To ad- b2 = S ( 2, 1 ) × a1 + S ( 2, 2 ) × a2 [k = 1, 2, and n = 2]
dress the issue, a pixel adjustment process has been applied on
On computing S(n, k), the values of S(1, 1), S(2, 1) and S(2, 2)
the pre-embedded pixel components to retain the values in the
becomes one for all three cases. Hence, the formulated expression
range of 0 to 255 followed by embedding.
of b1 and b2 becomes:
• The secret image is fabricated into the transformed compo-
nents for the purpose of information security. However, an ex- b1 = a1 b2 = a1 + a2

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11 3

Similarly, the Inverse Stirling Transform (IST) is used to convert M = 2 and N = 2) can be formulated as follows:
back the transformed components {bk } into pixel components {ak }
1 
1 1
as given in Eq. (A.3). F (u, v ) = f (x, y ) (A.8)
  4

n
n
x=0 y=0
an = b (A.3)
k k Where, u = 0 to 1 and v = 0 to 1.
k=1
Now, if the summation expression 1x=0 1y=0 f (x, y ) is not get-
n ting divisible by 4 for arbitrary values of u and v, then that fre-
where, [ ], also denoted as |s(n,k)| (with a lower-case s) is a signed
k quency component becomes fractional.
Stirling number of the first kind. It can be computed based on the In 2015, Gashkov [16] showed that the computational complex-
following recurrence relation. ity of nth-order binomial transformation in the additive basis B + is
computed as n(n−1)/2 i.e., (n2 ). Similarly, one can also note that
s(n, k ) = s(n − 1, k − 1) + (n − 1 )s(n − 1, k ) for n ≥ 1 (A.4)
the computational complexity of the nth row of the Stirling trian-
where, the initial conditions are s(0, 0) = 1 and s(n, 0) = s(0, n) = 0, gle of the second kind in the basis B+ is O(n2 ). This is also true for
for all, n ≥ 1. the triangle containing the moduli of Stirling numbers of the first
−1
By the formulation of pair of transformed components (b1 , b2 ), kind. By means of the identity snk = snk−1 − (n − 1 )snk −1 in the basis
the pixel components (a1 , a2 ) are obtained as follows: B+ ,∗ , one can also show that, the complexity of the nth-order Stir-
ling transformation of the first kind is O(n2 ). The same is also valid
a1 = s ( 1 , 1 ) × b1 [k = 1, and n = 1] for the monotone Stirling transformation of the first kind.
a2 = s ( 2 , 1 ) × b1 + s ( 2 , 2 ) × b2 [k = 1, 2, and n = 2] By evaluating the Eq. (A.1) with the help of Eq. (A.2) for the
On computing |s(n, k)|, the values of |s(1, 1)|, |s(2, 1)| and |s(2, 1 × 2 block of pixel components (ai ), one can formulate the Stril-
2)| becomes 1, −1 and 1, respectively. Hence, the formulated ex- ing Transform (ST) to obtain the transformed components (bi ) as
pression of a1 and a2 becomes: follows:

a1 = b1 a2 = −b1 + b2 b1 = a1 b2 = a1 + a2

In matrix form, the Stirling Transform (ST) and its inverse (IST) Compared to the computational complexity of two dimensional
are represented as given in Eqs. (A.5) and (A.6). Discrete Fourier Transformations (2D-DFT), the 1 × 2 block of the
Stirling Transformation (ST) yields no fractional transformed com-
  
yields 1 1 yields b1 = a1 ponents (bi ) as there is no pending division operation to be per-
b = aS → b1 b2 = a1 a2 → formed.
0 1 b2 = a1 + a2
(A.5) 4. Scrambling

   The scrambling of secret image (W) prior to embed it into the


−1 yields 1 0 yields a1 = b1 cover image is an interesting study as far as the security is con-
a = bS → a1 a2 = b1 b2 →
−1 1 a2 = b2 − b1 cerned. Arnold’s cat map [13] is widely acceptable scrambling pro-
cess that enhances the security by shuffling the pixel positions of
(A.6)
the secret image. In this study, a simple scrambling process is pro-
where, b and a are 1 × 2 sub-matrices corresponding to the pixel posed to improve the security of the secret image, which is invert-
and transformed components; S and S−1 are the 2 × 2 Stirling ible in nature.
Transform and its inverse sub-matrices. Let (p1 , p2 ) be the non-overlapping pair of pixels of the u × v
secret image and the parameter r is initialized as 0. The modified
3. Computational complexities pair of pixel components (p’1 , p’2 ) is obtained as follows:
p 1 = p1  r
(A.7)
Before analyzing the computational complexity of Stirling p 2 = p 1 + ( p1  p2 )
Transform (ST), we have investigated the complexity of two dimen-
sional Discrete Fourier Transformations (2D-DFT) as most of the To keep the pixel values in the valid range [0, 255], the second
existing steganography schemes were proposed based on it. The modified pixel (p’2 ) may be adjusted further as follows:
two dimensional Discrete Fourier Transformation (2D-DFT) for the 
p2 : i f p2 ≤ 255
image (I) of size M × N can be defined as given in Eq. (A.7). p2 = (A.8)
p2 − 256 : i f p2 > 255
1 
M−1 N−1
F (u, v ) = f (x, y )e− j2π (ux/M+vy/N ) (A.7) For successive cases, the parameter r is assigned as, r = p’2 . The
MN above set of steps is executed until and unless each pair of pixels
x=0 y=0
is scrambled.
Where, u = 0 to M−1 and v = 0 to N−1. Now, Let us consider the scrambled image W’, which is decom-
The variable u and v are the frequency variables, x, y are the posed into successive pairs of pixels (q1 , q2 ) at the recipient end.
spatial variables and f(x,y) is pixel intensity value in spatial do- The pixel value q2 may be adjusted as follows:
main. 
To compute the DFT, using the Eq. (A.7), one need to do an O(1) q2 : i f q2 ≥ q1
q2 = (A.9)
calculation for every input value and for every output value. Hence, q2 + 256 : i f q2 < q1
the complexity for 2D DFT is computed as O((M × N)2 ), because the
number of input pixels is M × N and the number of output pixels To re-compute the original pair of pixels (p1 , p2 ), the following
is also M × N. If the image is a square image i.e., M = N, then the operations are to be performed:
time complexity would be O((N)4 ). p1 = q1  r
(A.10)
Apart from the computation complexity, the two dimensional p2 = p1  ( q2 − q1 )
Discrete Fourier Transformation (2D-DFT) with size 2 × 2 (i.e.,

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

4 S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11

For the initial pair of pixels, r = 0; however, for successive cases, formed components. A novel scrambling process has been used
the parameter r preserves the old copy of q2 pixel. to scramble the secret image (W) as discussed in Section 4. The
method is implemented by partitioning each pair of transformed
5. Proposed technique components into two parts VCA (Vector Co-ordinate Area) and
VMA (Vector Modification Area). A reference value (f) is computed
This paper highlights a novel steganographic scheme in Stir- and the secret bits in varying proportion are embedded into the
ling Transform (ST) domain to authenticate color images. Two ma- VMA part of the optimal pair of transformed components (which
jor parameters which keep the security of fabricated secret infor- offers minimum deviation on embedding) for the payload range
mation are the shared PIN (Z) and the system time (X), respec- 0.5–3 bpB. Inverse Stirling Transform (IST) is applied on each
tively. An OTP (One Time Password) is generated from the above pair of embedded components to generate the pixel components
mentioned metrics at the sender end which in turn is transmitted in spatial domain. The process is repeated until and unless the
into the recipient mobile number. However, a copy of system time entire secret information is fabricated and the stego-image (I’) is
(X) is preserved in a non-editable data store (D1) of a database produced.
server in public domain. At the recipient end, the system time is
re-computed (denoted as Y) from the shared PIN (Z) and the re- Algorithm:
ceived OTP, respectively. The sender’s system time (X) is compared Input: Cover image (I), scrambled image (W’) and message digest (MD).
Output: Stego-image (I ).
with the recipient’s system time (Y) to ensure the extraction of se-
Step 1. The cover image (I) is partitioned into non-overlapping
cret information (W). The parameters Z and OTP are kept as pri- (Red\Green\Blue) pair of pixel components (pi , pi+1 ) in a sliding
vate by both parties however, the system time (X) is public in na- window manner.
ture. The block diagram of the authenticating system is shown in Step 2. A pre-embedded adjustment is applied on each pixel component p, if
Fig. A.1. necessary, to avoid overflow and underflow on embedding an average
 payload of B bits per Byte of scrambled data as given in Eq. (A.11).
(a) A message digest (MD) is computed from the secret data ( 28 − 2B ) : p ≥ ( 28 − 2B )
p= (A.11 )
2B : p ≤ 2B
(W) by applying MD5 algorithm.
Step 3. Stirling Transform (ST) is applied on each pair of pixel components
(b) For successive pairs of pixel components of the secret data (pi , pi+1 ) to generate transformed components (ti , ti+1 ) by Eq. (A.5).
(W), it is scrambled as discussed in Section 3. Step 4. N bits of secret bit-stream (scrambled) are fabricated on each pair of
(c) Entire scrambled data and the message digest (MD) are em- transformed components (ti , ti+1 ) where, 1 ≤ N ≤ 6. The embedding
rules are discussed below:
bedded in a block-wise manner into the cover image (I)
a) Secret data (d) is a decimal number obtained by retrieving N number of
which results the stego-image (I’) at the output. bits from the scrambled secret bit-stream (W) as given in Eq. (A.12).
(d) The OTP is generated from the shared PIN (Z) and the sys- d = (b0 b1 b2 b3 . . . bN−1 )2 (A.12 )
tem time (X) as OTP = X  Z. The system time (X) of succes- b) Repeat step 4(b) to 4(h) to select the pair of transformed components
sive authenticating session can be stored at the data store which offers minimal changes followed by embedding from a set of
(2┌N/2┐ + 1)2 pairs of transformed components by taking all the
(D1) of a database server.
possible combinations in the range of (tk −2┌N/2┐−1 ) to (tk +2┌N/2┐−1 ).
(e) The stego-image is transmitted to the intended recipient c) Assign PVMA = N for VMA and the number of pixel components, n = 2.
through the unreliable medium whereas the OTP is delivered d) Allocate the VCA evenly based on n, and generate two vectors as (g1 ,
to the intended recipient in the form of SMS on mobile. g2 ).
e) Allocate N bits of VMA into two halves i.e., N1 and N2 where,
(f) The recipient accepts the OTP in his/her mobile number and
N1 = └ N/2┘ , N2 = N−N1 .
provides the shared PIN to re-compute the system time (Y) f) Calculate the reference value f using Eq. (A.13) with respect to (g1 , g2 )
as, Y = OTP  Z. For ith authentication session, the value of to embed N secret bits where, n = 2.
n
X is fetched from the data store D1 and is compared against f ( g1 , g2 ) = gi (2i − 1 )mod 2n (A.13 )
Y to ensure the extraction process. i=0
g) Compute the summation SV = (d + f) % 2 N.
(g) Only identical values of X and Y allows the extraction of h) Obtain the pair of embedded components (t’i , t’i+1 ) by replacing the
the scrambled data (along with the message digest MD). The VMA with SV as.
scrambled data is unscrambled by applying reverse process. 
t i = (ti & (29 − 2N1 ) ) + SVN 
2 2

(h) Another message digest (MD’) is recomputed at the recip- t  i+1 = (ti+1 & (29 − 2N2 ) ) + (SV %2 2  )
N

ient end which in turn is compared against the recovered Step 5. The Inverse Stirling Transform (IST) is applied on each pair of
embedded components (t’i , t’i+1 ) by Eq. (A.6).
message digest (MD). If both are same then authentication
Step 6. Repeat step 1 to step 5 to complete the process of embedding.
is successful otherwise, the authentication is considered as Successive embedding operation produces the stego-image (I ).
unsuccessful. Step 7. Stop.

Three levels of security are achieved in proposed scheme. The


first level ensured scrambling of secret image through the simple
scrambling process. Second level revealed the OTP sharing process 5.2. Extraction
between the authorized sender and receiver. However, the third
level of security is accomplished for the authenticating process At the recipient end, the stego-image (I ) is received over in-
based on the message digest. ternet, and the OTP is received in his/her private mobile number.
The embedding and extraction phases are elaborately described Re-computed system time (Y), as obtained from the shared PIN (Z)
in Sections 5.1 and 5.2 respectively. and the OTP, is compared against the server’s system time (X). If
X and Y are identical, the stego-image (I ) is processed for carry-
5.1. Embedding ing out the extraction process. The stego-image (I ) is partitioned
into non-overlapping pair of pixel components. Each pair of pixel
The carrier image is partitioned into non-overlapping components are converted into transform domain based on Stir-
(red\green\blue) pair of pixel components in a sliding window ling Transform (ST). The method is implemented by partitioning
manner. A pre-embedded adjustment has been incorporated to each pair of transformed components into two parts VCA (Vec-
avoid overflow and underflow. Stirling Transform (ST) is applied tor Co-ordinate Area) and VMA (Vector Modification Area). As per
on each pair of pixel components (red/green/blue) to obtain trans- the specified payload, varying number of secret bits correspond-

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11 5

SENDER RECIEVER
Secret Data (W) OTP Calculate
System Time

Shared PIN (Z)


Compute Message
Digest Y = OTP ⊕ Z
D1 X
Server
Message Digest (MD)
n X=Y?
Wrong OTP !!!

Scrambling y

Stego-image (I') Extraction

Scrambled Data (W’)

Message Digest (MD) Scrambled Data (W')


Embedding
Cover Image (I)

Unscrambling

Stego-image (I’)

Unscrambled Data (W)


Shared PIN (Z)

One Time Password Re-compute


Generation Message Digest

Message Digest (MD')


System Time (X) OTP = X ⊕ Z

Sends OTP to MD = MD' ? n


Recipient’s Mobile No
y
D1 X
Server Stop Authenticated Unauthenticated

Fig. A.1. Block diagram of proposed steganographic technique based authenticating system at sender and recipient ends.

ing to the message digest (MD) and the scrambled image (W’) cussed in Section 4. Hence, the secret image (W) is reconstructed
are extracted. Successive extraction operation ensured the retrieval at the recipient end. Message digest (MD’) is re-computed and the
of message digest (MD) and the secret image in scrambled form same is compared against the retrieved message digest (MD) to
which in turn is unscrambled based on the reverse process as dis- verify the authenticity of the color image.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

6 S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11

Algorithm:
Input: Stego-image (I ) and size of secret data (X).
Output: Scrambled image (W’) and the message digest (MD).
Step 1. The stego-image (I ) is partitioned into non-overlapping (Red\Green\Blue) pair of pixel components (pi , pi+1 ) in a sliding window manner.
Step 2. Stirling Transform (ST) is applied on each pair of pixel components (pi , pi+1 ) to generate transformed components (ti , ti+1 ).
Step 3. N bits of the fabricated secret bit-stream (scrambled) are extracted from each pair of transformed components (ti , ti+1 ) where, each transformed component
is considered as a 9-bit representation and 1 ≤ N ≤ 6. The extraction rules are discussed below:
a) Assign PVMA = N for VMA and the number of pixel components, n = 2.
b) Allocate the VCA evenly according to n, and generate 2 vectors denoted as (g1 , g2 ).
c) N1 and N2 bits are extracted from each pair of transformed components (ti , ti+1 ) i.e., from VMA where, N1 = └ N / 2┘ , N2 = N−N1 . Intermediate data (d) is a
decimal number obtained by concatenating (‘+’) two extracted bit sequences (b’0 b’1 …b’N1−1 )2 and (b”0 b”1 …b”N2−1 )2 from ti and ti+1 as given in Eq. (A.14).
d = b 0 b 1 ...bN1−1 + b 0 b 1 ...b N2−1 (A.14 )
2 2
d) Calculate the reference value f by using Eq. (A.15) with respect to vectors (g1 , g2 ) to extract N secret bits where n = 2.
n
f ( g1 , g2 ) = gi (2i − 1 )mod 2n (A.15 )
i=0
e) Secret data (d’) of the scrambled bit-stream consisting of N bits are recovered from each pair of transformed components (ti , ti+1 ) where, d’ = (2N – f+d) % 2N.
Step 4. The inverse Stirling Transform (IST) is applied on each pair of embedded components (t’i , t’i+1 ) to generate the non-negative pixel components (p’i , p’i+1 ).
Step 5. Repeat step 1 to step 5 as per the size of the secret image (X) to complete the extraction of scrambled image (W’) which in turn is unscrambled.
Step 6. Stop.
Example 4.1.1: Two pixels (174, 105) have been considered for fabricating secret data 14. Apply proposed steganographic technique to re-compute the embedded pair
of pixel components.
Solution:
Embedding:- The following steps are followed to embed the secret data:
Step 1: Pixel pair (p0 , p1 ) = (174, 105), secret data (d) = 14 = (1110)2 , (N) = 4 (no. of bits to be embedded). Since, no overflow and underflow conditions are arises,
no pre-embedded adjustment is needed in this example.
Step 2: Convert the pair of pixel components into transformed domain using Stirling Transform (ST) as given below:
t0 = p0 = 174; t1 = p0 + p1 = 174 + 105 = 279.
Step 3: 25 pairs of transformed components are obtained from (t0 , t1 ), as (172, 277), (172, 278), (172, 279), (172, 280), (172, 281), (173, 277), (173, 278), (173, 279),
(173, 280), (173, 281), (174, 277), (174, 278), (174, 279), (174, 280), (174, 281), (175, 277), (175, 278), (175, 279), (175, 280), (175, 281), (176, 277), (176, 278),
(176, 279), (176, 280) and (176, 281). It can be examined by repeating steps 4(a) to 4(h) of algorithm for embedding that the optimal pair for embedding is
(t0 , t1 ) = (176, 280).
Step 4: PVMA = 4, n = 2, (t0 , t1 ) = (176, 280) = (010110 0 0 0, 10 0 0110 0 0)2 .
Step 5: Generate two vectors (g1 , g2 ) = (010110 0, 10 0 0110)2 = (44, 70).
Step 6: Compute, f (g1 , g2 ) = (44 × 1 + 70 × 3) % 22 = 254% 4 = 2 = (0010)2 .
Step 7: Calculate the summation SV = (14 + 2) % 24 = 0 = (0 0 0 0)2 .
Step 8: Since, VMA = 00 || 00, N1 = └ 4 / 2 ┘ = 2, N2 = 4−2 = 2. Replace it by SV as given below:
t  i = (176&(29 − 22 ) ) + 0/24/2 = (176&(512 − 4 ) + 0/4 = (176&508 ) + 0 = 176
t  i+1 = (280&(29 − 22 ) ) + 0%24/2 = (280&(512 − 4 ) ) + 0% 4 = (280&508 ) + 0 = 280
Step 9: Embedded pair of components, (t’0 , t’1 ) = (176, 280) = (010110 0 0 0, 10 0 0110 0 0)2 .
Step 10: Inverse Stirling Transform (IST) is applied to re-generate the pair of pixel components, (p’0 , p’1 ) = (176, (280−176)) = (176, 104).
Extraction:- The following steps are followed to recover the secret data:
Step 1: Pair of pixel components (p0 , p1 ) = (176, 104), no. of bits to be extracted (N) = 4.
Step 2: Convert the pair of pixel components into transformed domain by applying Stirling Transform (ST) as given below:
t0 = p0 = 176; t1 = p0 + p1 = 176 + 104 = 280.
Step 3: PVMA = 4, n = 2, (t0 , t1 ) = (176, 280) = (010110 0 0 0, 10 0 0110 0 0)2 .
Step 4: Generate 2 vectors (g1 , g2 ) = (010110 0, 10 0 0110)2 = (44, 70).
Step 5: Compute, f (g1 , g2 ) = (44 × 1 + 70 × 3) % 22 = 254% 4 = 2 = (0010)2 .
Step 6: VMA = (0 0)2 || (0 0)2 . Intermediate data (d) = (00)2 + (00)2 = (0000)2 = 0.
Step 7: The recovered secret data, d’ = (24 – 2 + 0) % 24 = (16 – 2 + 0) % 16 = 14 = (1110)2 .
Step 8: Inverse Stirling Transform (IST) is applied on each pair of transformed components to re-compute the pair of pixel components, (p’0 , p’1 ) = (172,
(276–172)) = (172, 104).

6. Results and discussions Performance of the proposed technique has been evaluated
based on the standard metrics such as Peak Signal to Noise Ratio
The results and discussions for proposed technique have been (PSNR), Mean Squared Error (MSE), Image Fidelity (IF), Structural
elaborated in this section. The variation of quality level of the Similarity index (SSIM) and Universal Image Quality index (UIQ)
stego-image with respect to increasing payload is the fundamental with respect to 0.5, 1.0, 1.5, 2.0, 2.5 and 3 bits per Byte (bpB) of
study of the performance evaluation for proposed technique. The payloads. It can also be observed that the proposed method gives a
quality level is analyzed by means of Mean Squared Error (MSE), uniform PSNR, MSE, IF, SSIM and UIQ and as the payload increases
Peak Signal to Noise Ratio (PSNR), Image Fidelity (IF), Univer- the quality drops subsequently. It is seen from Table A.1 that the
sal Image Quality index (UIQ), Structural Similarity index (SSIM), minimum PSNR obtained is 37.45 dB at 3 bpB for the “Pepper” im-
Standard Deviation (SD) and Standard Deviation Error (SDE), re- age whereas the maximum PSNR is 52.41 dB at 0.5 bpB for the
spectively. Ten benchmark (BMP) color images [14] of dimension “Foster City” image. The MSE is minimum at payload of 0.5 bpB for
512 × 512 labeled as: (i) Lena, (ii) Baboon, (iii) Pepper, (iv) Airplane, “Foster City” and maximum at 3bpB for “Pepper” which is 0.3730
(v) Sailboat, (vi) Earth, (vii) San Diego, (viii) Splash, (ix) Oakland and 11.68 respectively. The other quality metrics such as IF, SSIM
and (x) Foster City are taken to embed varying sizes of secret im- and UIQ analysis has also been made to measures the rate of qual-
age (i.e. the gold-coin) as shown in Fig. A.2. ity degradation, pixel’s independence and structural distortion in
In this proposed scheme, the secret image has been scrambled the stego-image as compared to the original image. Minimum val-
based on a novel scrambling process as discussed in Section 4. The ues of IF, SSIM and UIQ are 0.9990 (Pepper and Splash), 0.9780
different states of modifications of carrier/cover images viz. “Lena”, (Splash) and 0.6906 (Splash) respectively at 3 bpB however, the
“Baboon” and “Airplane” (before and after embedding the scram- maximum values of IF, SSIM and UIQ at 0.5 bpB are 0.9999 (for
bled image) are shown in Fig. A.3. all ten images), 0.9999 (Baboon) and 0.9983 (Baboon) respectively.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11 7

Fig. A.2. Different cover images of dimension 512 × 512 along with the secret image.

Fig. A.3. Different states of modification for Cover, Stego and scrambled images in the proposed technique.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

8 S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11

Table A.1
Simulation results for the carrier/cover images of dimension 512 × 512 with respect to payload vari-
ation of 0.5–3 bpB.

Images Payloads (bpB) PSNR (dB) MSE IF SSIM UIQ

Lena 0.5 52.3854 0.3754 0.9999 0.9998 0.9911


1.0 48.6962 0.8779 0.9999 0.9996 0.9800
1.5 46.4729 1.4648 0.9999 0.9990 0.9676
2.0 43.9828 2.5989 0.9998 0.9987 0.9475
2.5 41.0293 5.1304 0.9997 0.9956 0.9096
3.0 38.5650 9.0485 0.9994 0.9949 0.8613
Baboon 0.5 52.3904 0.3750 0.9999 0.9999 0.9983
1.0 48.7029 0.8766 0.9999 0.9999 0.9962
1.5 46.4944 1.4576 0.9999 0.9996 0.9937
2.0 43.9521 2.6174 0.9999 0.9995 0.9889
2.5 41.0459 5.1108 0.9997 0.9984 0.9799
3.0 38.5620 9.0547 0.9995 0.9982 0.9675
Pepper 0.5 51.9951 0.4107 0.9999 0.9985 0.9851
1.0 48.5137 0.9156 0.9999 0.9981 0.9771
1.5 45.1303 1.9955 0.9998 0.9910 0.9620
2.0 42.9095 3.3276 0.9997 0.9895 0.9458
2.5 39.2033 7.8117 0.9993 0.9790 0.9115
3.0 37.4555 11.6824 0.9990 0.9784 0.8715
Airplane 0.5 52.3427 0.3792 0.9999 0.9998 0.9722
1.0 48.6965 0.8779 0.9999 0.9995 0.9444
1.5 46.4708 1.4655 0.9999 0.9988 0.9181
2.0 43.9514 2.6178 0.9999 0.9985 0.8762
2.5 41.1240 5.0197 0.9999 0.9950 0.8207
3.0 38.6191 8.9365 0.9997 0.9944 0.7487
Sailboat 0.5 52.3292 0.3803 0.9999 0.9998 0.9923
1.0 48.7079 0.8756 0.9999 0.9997 0.9833
1.5 46.3994 1.4898 0.9999 0.9992 0.9748
2.0 43.8894 2.6555 0.9999 0.9990 0.9578
2.5 40.9377 5.2398 0.9997 0.9965 0.9330
3.0 38.4999 9.1853 0.9995 0.9960 0.8976
Earth 0.5 52.3946 0.3746 0.9999 0.9998 0.9959
1.0 48.6964 0.8779 0.9999 0.9997 0.9907
1.5 46.5259 1.4471 0.9999 0.9992 0.9851
2.0 43.9668 2.6085 0.9998 0.9990 0.9738
2.5 41.1148 5.0304 0.9997 0.9967 0.9527
3.0 38.5808 9.0156 0.9995 0.9962 0.9207
San Diego 0.5 52.4064 0.3736 0.9999 0.9999 0.9989
1.0 48.6994 0.8773 0.9999 0.9999 0.9974
1.5 46.4905 1.4589 0.9999 0.9997 0.9957
2.0 43.9709 2.6061 0.9999 0.9996 0.9928
2.5 41.0860 5.0638 0.9998 0.9988 0.9868
3.0 38.5859 9.0052 0.9997 0.9986 0.9782
Splash 0.5 52.1497 0.3964 0.9999 0.9979 0.9539
1.0 48.5684 0.9041 0.9999 0.9974 0.9222
1.5 45.5149 1.8264 0.9998 0.9896 0.8907
2.0 43.2052 3.1085 0.9997 0.9880 0.8423
2.5 39.6977 6.9712 0.9993 0.9790 0.7708
3.0 37.7319 10.9620 0.9990 0.9780 0.6906
Oakland 0.5 52.3019 0.3827 0.9999 0.9998 0.9977
1.0 48.6541 0.8865 0.9999 0.9997 0.9952
1.5 46.0359 1.6199 0.9999 0.9991 0.9918
2.0 43.5880 2.8463 0.9998 0.9989 0.9861
2.5 40.4133 5.9122 0.9997 0.9965 0.9745
3.0 38.2077 9.8245 0.9995 0.9962 0.9575
Foster City 0.5 52.4138 0.3730 0.9999 0.9997 0.9884
1.0 48.7416 0.8688 0.9999 0.9994 0.9744
1.5 46.3083 1.5214 0.9999 0.9985 0.9544
2.0 44.0106 2.5824 0.9999 0.9982 0.9314
2.5 41.1739 4.9624 0.9998 0.9942 0.8843
3.0 38.6306 8.9129 0.9997 0.9935 0.8192

Usual values of IF, SSIM and UIQ lie between 0 and 1. The closer et al.’s scheme [17], the PSNR values are computed at 1 bpB of
the IF, SSIM and UIQ to one is considered as the more similar is payload for both schemes and at the same time, proposed method
the test to the original image. Simulation results of Table A.1 en- offers higher PSNR for the following test images: ‘Lena’, ‘Baboon’,
sured that the deviation of stego-image quality is very less and an ‘Pepper’, ‘F16’ and ‘Boat’. Fig. A.4 illustrates that the PSNR enhance-
observer cannot easily detect any difference between the original ment is around 1.5 dB for all of these test images except the ‘Pep-
and stego images. per’. Similarly, the proposed method has been compared against
In the comparative analysis part, proposed method has been the Xia et al.’s scheme [18] and STCIA scheme [12] in terms of
compared against the existing steganography schemes such as PSNR at 2 bpB of payload. Fig. A.5 shows that the PSNR enhance-
Chang et al.’s [17], Xia et al.’s [18], STCIA [12], 3-LSB [2] and hybrid ments for all five test images are around 2.5 dB and 0.5 dB which
GEMD [7] with respect to different payloads. Compared to Chang are quite good from the perspective of image quality. Moreover, the

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11 9

46.5
46.3
46.1
45.9

PSNR (dB)
45.7
45.5
45.3
45.1
44.9
44.7
44.5

Chang et al.'s scheme [17] Proposed method


Fig. A.4. Comparisons of PSNR (dB) improvement between Chang et al.’s scheme [17] and proposed scheme at 1.5 bpB of payload.

44.5
44
43.5
43
PSNR (dB)

42.5
42
41.5
41
40.5
40
39.5

Xia et al.'s scheme [18] STCIA [12] Proposed method


Fig. A.5. Comparisons of PSNR (dB) improvement between Chang et al.’s scheme [18] and proposed scheme at 2 bpB of payload.

38.6

38.4

38.2
PSNR (dB)

38

37.8

37.6

37.4

3-LSB method [2] Hybrid GEMD [7] Proposed method


Fig. A.6. Comparisons of PSNR (dB) improvement between 3-LSB method [2], hybrid GEMD method [7] and proposed scheme at 3 bpB of payload.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

10 S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11

56

54

52

Standard Deviation 50
Lena
48 Baboon
46 Airplane

44

42
0 0.5 1 1.5 2 2.5 3
Payload (bpB)
Fig. A.7. Comparative analysis of standard deviation between original and stego images with respect to varying payloads.

0.1
0.09
Standard Deviaon Error

0.08
0.07
0.06
0.05 Baboon
0.04 Lena
0.03 Airplane
0.02
0.01
0
0.5 1 1.5 2 2.5 3
Payloads (bpB)
Fig. A.8. Comparative analysis of standard deviation error between original and stego images with respect to varying payloads.

performance analysis has also been made among 3-LSB scheme [2], (SDE) is around 0.09 which is negligible as far as the comparison
hybrid GEMD [7] and the proposed method at 3 bpB of payloads. between original and stego images is concerned.
It is clearly observed from Fig. A.6 that the PSNR improvement for
proposed method over the LSB [2] and hybrid GEMD [7] schemes 7. Conclusions and future scopes
is around 1 dB for all three test images.
In Fig. A.7 it is seen that the Standard Deviation (SD) analysis Proposed technique is an authenticating process operated in
has been made for “Lena”, “Baboon” and “Airplane”. The Standard transform domain that offers three levels of security by providing
Deviation is computed by taking the average of Standard Deviation two important parameters: system time (X) and shared PIN (Z). It
values corresponding to red, green and blue channels. Standard De- scrambles the secret image prior to embedding. The quality level is
viation (SD) of the stego-images at varying payload is differing very analyzed for the payload range 0.5 to 3 bpB. The PSNR of all stego-
slightly and one cannot easily detect the differences between U × V images is above 30 dB. Hence, the technique ensured improved se-
source and stego-images and they are seemingly identical. curity with respect to the variable payload and minimum degrada-
The dispersion in Standard Deviation (SD) due to embedding tion in quality. Simulation results also confirms that the proposed
can be analyzed by Standard Deviation Error (SDE) where the er- technique outperforms Chang et al.’s [17], Xia et al.’s [18], STCIA
ror has been computed as the absolute difference of Standard De- [12], 3-LSB [2] and hybrid GEMD [7] methods in terms of image
viation (SD) between the original and the stego images. It can be quality. The inclusion of Stirling Transform (ST) ensures integer cal-
noticed from Fig. A.8 that the stego-images are deviating from the culation; and the color images as the cover provides three times of
original image with respect to increasing payload. It can also be payload as compared to gray-scale image.
seen that the error is around 0.02 up to 2 bpB and then it is In future work, the payload and quality may be improvised by
slightly increasing. The maximum value of standard deviation error embedding more information into the edge areas based on the
classical edge detectors. One dimensional chaotic sequence such as

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003
JID: JISA
ARTICLE IN PRESS [m5G;April 21, 2018;14:46]

S.K. Ghosal, J.K. Mandal / Journal of Information Security and Applications 000 (2018) 1–11 11

Tent Map and Logistic Map may also be investigated to increase [10] Kuo WC, Wang CC. Data hiding based on generalized exploiting modification
the security. direction method. Imag Sci 2013;61(6):484–90.
[11] Kuo WC, Chang SY. Hybrid GEMD data hiding. J Inf Hiding Multimedia Signal
Process 2014;5(3):420–30.
References [12] Ghosal SK, Mandal JK. Stirling transform based color image authentication
(STCIA). In: Proceedings of international conference on computational intelli-
[1] Paar C, et al. Understanding cryptography. 1st ed. Springer; 2010. gence: modeling, techniques and applications (CIMTA-2013), procedia technol-
[2] Bender W, Gruhl D, Morimoto N, Lu A. Techniques for data hiding. IBM Syst J ogy, Elsevier, ISSN: 2212-0173, Kalyani, India; 2013.
1996;35(3–4):313–36. [13] Rawat S, Raman B. A chaotic system based fragile watermarking scheme for
[3] Wu D-C, Tsai W-H. A steganographic method for images by pixel-value differ- image tamper detection. In: International journal of electronics and communi-
encing. Pattern Recognit Lett 2003;24(9-10):1613–26. cations (AEÜ), Elsevier; 2011. p. 840–7.
[4] Zhang X, Wang S. Efficient steganographic embedding by exploiting modifica- [14] Weber A.G., The USC-SIPI image database: version 5, original release: Octo-
tion direction. IEEE Commun Lett 2006;10(11):1–3. ber 1997, Signal and Image Processing Institute, University of Southern Cali-
[5] Lee CF, Wang YR, Chang CC. A steganographic method with high embedding fornia, Department of Electrical Engineering. http://sipi.usc.edu/database/ (ac-
capacity by improving exploiting modification direction. In: Proceedings of the cessed on 23rd September 2014).
3rd international conference on intelligent information hiding and multimedia [15] Ghosal SK, Mandal JK. Binomial transform based fragile watermarking for im-
signal processing, Kaohsiung, vol. 1; 2007. p. 497–500. age authentication. J Inf Secur Appl, Elsevier 2014;19(4–5):272–81.
[6] Chao RM, Wu HC, Lee CC, Chu YP. A novel image data hiding scheme with [16] Gashkov SB, Gashkov IB. On the complexity of computation of differentials and
diamond encoding. EURASIP J Inf Secur 2009;1:658047. gradients. Diskret Mat 2005;17(3):45–67 [DiscreteMath. Appl. 15 (4), 327–350
[7] Kim HJ, Kim C, Choi Y, Wang S, Zhang X. Improved modification direction (2005)].
methods. Comput Math Appl 2010;60(2):319–25. [17] Chang CC, Chou YC, Kieu D. An information hiding scheme using Sudoku. In:
[8] Chen KN, Chang CC, Lin HC. A large payload EMD embedding scheme with Proceedings of the 3rd international conference on innovative computing in-
high stego-image quality. In: Proceedings of IEEE international conference in formation and control, IEEE; 2008.
computational aspects of social networks (CASoN); 2010. p. 126–30. [18] Xia B-B, Wang A-H, Chang C-C, Liu L. An image steganography scheme using
[9] Kieu TD, Chang CC. A steganographic scheme by fully exploiting modification 3D-Sudoku. J Inf Hiding Multim Signal Process 2016;7(July (4)).
directions. Expert Syst Appl 2011;38(8):10648–57.

Please cite this article as: S.K. Ghosal, J.K. Mandal, On the use of the Stirling Transform in image steganography, Journal of Information
Security and Applications (2018), https://doi.org/10.1016/j.jisa.2018.04.003

You might also like