You are on page 1of 19

Multimedia Tools and Applications (2021) 80:21751–21769

https://doi.org/10.1007/s11042-021-10784-5

Advanced image steganography based on exploiting


modification direction and neutrosophic set

Randa Atta 1 & Mohammad Ghanbari


2,3
& Ibrahim Elnahry
1

Received: 4 September 2020 / Revised: 1 December 2020 / Accepted: 4 March 2021 /


Published online: 19 March 2021
# The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2021

Abstract
A new variant of image steganography based on exploiting modification directions (EMD)
named advanced EMD (AEMD) is introduced. In this method the secret digits in mn -ary
notional systems are embedded into a group of n pixels of the cover image. To increase data
hiding capacity, edge masking characteristics of human visual system is exploited to embed
more bits at image edge pixels than non-edge pixels. To do this, a neutrosophic set edge
detector, named as MNSED is also introduced to classify cover image into 2 × 2 non-
overlapping edge and non-edge blocks without any overhead information. Hence the secret
digits in two different bases are embedded into the edge and non-edge blocks. Experimental
results show that the proposed method provides higher embedding capacity and better quality
compared to the existing schemes, while its resistance to steganographic attack is still higher.

Keywords Image steganography . Edge detection . Neutrosophic set

1 Introduction

Information security has become a main challenge in communication environment due to the
development of digital media, computer networks and internet. Information hiding including

* Randa Atta
r.atta@eng.psu.edu.eg

Mohammad Ghanbari
ghan@ut.ac.ir; ghan@essex.ac.uk

Ibrahim Elnahry
ibrahim.farouk@eng.psu.edu.eg

1
Electrical Engineering Department, Port Said University, Port Said 42523, Egypt
2
School of Electrical and Computer Engineering, College of Engineering, University of Tehran,
Tehran, Iran
3
School of Computer Science and Electronic Engineering, University of Essex, Colchester CO4 3SQ,
UK
21752 Multimedia Tools and Applications (2021) 80:21751–21769

watermarking and steganography [6, 25] is at the heart of the discipline of information
security. Steganography is used to hide secret messages into a carrier such as a digital image,
making them invisible to ordinary users of public networks.
Three important issues to be considered in designing an image steganography system are:
image fidelity, high embedding capacity, and steganographic security (un-detectability against
steganalysis tests [7, 26, 35]). However, achieving these issues simultaneously is a challenging
task, and various image steganography methods have been reported in the literature [3, 5, 8,
11, 13, 16, 18–20, 23, 24, 27, 31–34, 36, 37]. Recently, Zhang and Wang [37] have proposed
an exploiting modification direction (EMD) method to improve the embedding capacity of the
method introduced in [23]. In EMD, for a group of n pixels, the secret bits are divided into a
(2n + 1)-ary notational system, where only one pixel in each group is increased or decreased by
1. The maximum embedding rate of this scheme is 1.161 bpp for n = 2.
Many methods have been proposed to improve the embedding rate of the EMD scheme
[13, 16, 18–20, 24, 27, 31, 32]. In [19], by modifying values of two pixels simultaneously, the
embedding capacity is increased by 1.7 times more than that of EMD, but image quality is
degraded by more than 8 dB. Kim et al. [13] have proposed EMD-2 which modifies at most
only two pixels in a group of n pixels, where only one LSB is changed, leading to a maximum
possible error of ±1 in each pixel. EMD-2 can achieve hiding capacity of 1.58 bpp with similar
image quality of EMD for n = 2. For further improvement, EMD-3 [24] was proposed to
exploit modification direction fully. In EMD-3 a group of n pixels are modified to embed a
secret digit in 3n-ary notational system. EMD-3 can increase embedding capacity with
preserving good image quality compared to EMD and EMD-2. In EMD, EMD-2 and EMD-
3, the maximum possible error in the modified pixel is ±1 because at most only one LSB value
is changed by either increasing or decreasing it by one. In other words, these methods provide
a modification interval of [−1, 1] for a pixel to carry secret digit. This narrow interval limits the
embedding capacity of these methods.
Other variants of EMD to increase its data hiding capacity based on cn-ary notational
system were also introduced in [20]. In this method, in a group of n pixels at most n pixels are
changed at c different ways of modifications. This method can adjust the embedding capacity
by providing an average capacity of 1 ~ 4.75 bpp in an average PSNR of 51.14 ~ 30.3 dB. In
[31] a steganographic method combines pixel-value differencing (PVD) and EMD, where
PVD determines the base digits from each pair of pixels for EMD embedding process. Saha
et al. [27] introduced an improved EMD, where each secret digit in a (2kn)-ary notational
system is embedded into a group of n cover pixels in which k bits are embedded in each pixel,
which outperforms the schemes in [31, 32] in both payload and quality.
Although the above EMD methods are simple, they do not exploit the characteristics of the
human visual system (HVS) as done by non-EMD methods to hide more secret bits in edge
areas than the smooth areas. Some of non-EMD methods determine edge and smooth areas/
pixels by utilizing either pixel-value differencing (PVD) [11, 31, 34, 36] or edge detectors [3,
8, 15, 33] such as Canny and fuzzy edge detectors. The disadvantage of these schemes is the
deformation of stego-image quality, where additional information besides the secret message is
required to be inserted to ensure edge locations of the cover and stego images remain the same,
reducing the actual data hiding capacity.
Therefore, in this paper, an image steganography method based on improved EMD and
neutrosophic set is proposed. The proposed scheme avoids embedding extra edge information
in the cover image by utilizing the proposed MNSED on the adjusted cover image to estimate
the precise edge areas and guarantees the exact edge locations after embedding the secret
Multimedia Tools and Applications (2021) 80:21751–21769 21753

message. Unlike the existing variants of EMD, the proposed scheme takes into account the
HVS and hence allows embedding the secret digits in two different bases, where the secret
digits in mnedge -ary and mnnon−edge -ary notational systems are embedded into the edge and non-
edge blocks that each consists of n pixels, respectively. Exploiting HSV in reducing the
visibility of distortions at image edges can lead to more secret bits to be hidden at edges. As
a result, the proposed method has the ability to control the embedding capacity as well as the
image quality by proper selection of medge and mnon-edge.
The main contributions of the paper are: 1) Introducing an improved version of the existing
embedded modification direction, called Advanced EMD (AEMD) to facilitate embedding the
secret digits in two different -ary notational systems and hence to increase the embedding
capacity. 2) Proposing neutrosophic based edge detector, called MNSED, to increase the
number of edge pixels with lower computational complexity compared to fuzzy edge detectors
and hence to embed more secret bits at edge pixels. Thus, the proposed AEMD along with
MNSED can accommodate high payload with good image fidelity.
The organization of the paper is as follows: Section 2 presents the introduced MNSED edge
detector. The proposed embedding scheme is described in Section 3. Experimental results and the
steganalysis tests are presented in Section 4. Finally, Section 5 gives some concluding remarks.

2 Modified neutrosophic set-based edge detection

Earlier edge detection methods are based on calculation of the image gradient such as Sobel, Canny,
Robert and Prewitt. Although these gradient-based edge detectors are simple, their performance is
limited as they are sensitive to noise. Fuzzy logic-based edge detectors have also been proposed [12,
17]. Most of these edge detectors are based on a type-1 fuzzy set, whereas type-2 fuzzy set which is
characterized by fuzzy membership within fuzzy, and performs better than type-1 [9].
Neutrosophic set-based edge detector (NSED) was proposed in [3] to reduce the computational
burden of fuzzy logic-based edge detectors. In this paper, NSED is modified to deal with
indeterminacy presented in the images and hence more accurate edges can be obtained. In the
modified NSED (MNSED), an image P is transformed into the neutrosophic domain (NS) image
PNS defined by three membership sets T, I and F, meaning each pixel P(i, j) is transformed into
PNS(i, j) = {T(i, j), I(i, j), F(i, j)}, where T(i, j), I(i, j) and F(i, j) are the membership values belonging
to true, indeterminate, and false sets, respectively. Unlike NSED, T(i, j), I(i, j), and F(i, j) represent
the degree of being an object pixel, edge pixel and background pixel, respectively. The indetermi-
nacy degree of the NS image PNS is reduced using the α-mean and β-enhancement operations on T
set. The reduction continues until the entropy of the indeterminate set I becomes unchanged. More
details regarding calculation of the three sets T, I and F can be found in [3]. Consequently, two
thresholds in T and F sets are calculated iteratively to separate each set into two new groups of
pixels. The initial threshold values thT and thF in T and F, respectively, are selected and updated in
each iteration. The final values of thT and thF are derived when the error between the current and the
previous values are small. Finally, the edge image EI is determined based on the three sets T, I and F
used to evaluate whether the pixels belong to edge pixels or not with the following rule:


1; if ðT ði; jÞ < thT j F ði; jÞ < th F Þ&ðI ði; jÞ ≥γ Þ
EI ði; jÞ ¼ ð1Þ
0; otherwise
21754 Multimedia Tools and Applications (2021) 80:21751–21769

A threshold value γ in the range of (0, 1) was empirically obtained through the tested images in
the experimental section. When γ is close to 1, the number of edge pixels decreases whereas
when it is closer to 0, the number of edge pixels increases. It was found that the proper value of
γ that achieves the highest number of edge pixels is in the range of (0, 0.1] and it was therefore
set to 0.01. However, if less edge pixels are required, the value of γ in I domain can be
iteratively adjusted using the same method of calculating two threshold values thT and thF. In
this case the proposed edge detector is named MNSED_γ.

3 Proposed advanced EMD (AEMD) scheme

In EMD [37], only one pixel in each group of n pixels is increased or decreased by 1 (i.e. the
modification interval is [−1, 1]). In the proposed steganographic scheme to achieve higher
hiding capacity, the modification interval is increased to [rL, rU]. That is, the proposed scheme
decomposes the secret binary message into steams (segments) of mn-ary notational system
digits. The scheme then embeds each digit of L secret bits, where L = ⌊log2mn⌋ bits, into a
vector of n cover pixels g = (g1, g2, ..., gn) by increasing or/and decreasing the n pixel values at
most by rU and rL, respectively or leaving any of them unchanged. It implies that there are m
states (the number of modification intervals) to change the value of each cover pixel in a block
to embed a secret digit and hence it fully exploits modification direction. The modification
interval of the pixels in a group is [rL, rU], where rL and rU are given by:
  
−ðm−1Þ=2 þ 1; if m is even; m−1
rL ¼ and rU ¼ : ð2Þ
−ðm−1Þ=2; if m is odd; 2

Prior to embedding, an n- dimensional basis vector b is constructed. In EMD, the basis vector is
b = (1, 2, ..., n) whereas in EMD-2 [13] the used basis vector is b = (1, 2, ..., 6 + 5(n-3)) for n > 2. In
the proposed scheme, the basis vector is defined as b = (m0, m1, …, mn-1). The proposed AEMD
exploits the modifications in different directions to represent different secret digits and hence a
higher embedding capacity is achieved. The scheme allows all pixels in each group to be modified
and at most all pixels’ values would increase or decrease by rU and by rL, respectively. A group of n
cover pixels are represented as a vector g = (g1, g2, ..., gn) and are mapped into the extraction function
f which is the inner product of g and b modulo mn and is defined as:
 
n
f ðg; bÞ ¼ ∑ ðg i  bi Þ mod mn : ð3Þ
i¼1

The value of the index is used to encode secret digit d and is given by:

d− f ; if d ≥ f :
index ¼ ð4Þ
mn −jd−f j; otherwise:

Each cover pixel gi in group g is then modified by the value of element cvi(index) in the code
0
vector cv such that cvi ∈ [rL, rU]. cvi ∈ [rL, rU] The modified pixel (stego-pixel) gi in each group
in the stego image represented by vector g′ of length n is calculated as:

0 gi þ cvi ðindexÞ; if d ≥ f :
gi ¼ ð5Þ
gi þ cvi ðmn −indexÞ; otherwise:
Multimedia Tools and Applications (2021) 80:21751–21769 21755

Therefore, when n and m are given, the code vector cv for a given index can be created as
follows: the procedure for generating the elements of the code vector cv works by dividing the
size of the interval into m subintervals in which each element resides. To find the value of each
element of the code vector cv for a given index which is formulated as
 
index ¼ ∑ni¼1 ðbi  cvi Þ , where bi = mi − 1, a recursive algorithm is developed for identifying
the boundaries of the interval containing an element cvi of cv. Each cvi is linearly distributed
into m number of subintervals between its upper U(i) and lower L(i) limits of the given index.
These upper and lower limits are given by:

U ðiÞ ¼ ∑ij¼1 b j  rU and LðiÞ ¼ ∑ij¼1 b j  rL : ð6Þ

Thus, for each cvi, the distance between its upper and lower values is divided equally into m
ðiÞ
subintervals. The lower limit for each subinterval SLk can be calculated as:
i−1

ðiÞ
SLk ¼ bi  ðrU þ 1−k Þ þ ∑ b j  rL ; k ¼ 1; …; m−1: ð7Þ
j¼1

The subinterval where the index lies can be identified and the value of cvi can be easily
determined. The proposed recursive algorithm for identifying the values of the code vector is
summarized in Algorithm 1.
21756 Multimedia Tools and Applications (2021) 80:21751–21769

For example, let n and m be 4 and 5, respectively, so that b = (50, 51, 52, 53) and hence the
modification interval is [−2, 2] as shown in Fig. 1. Suppose index = 59, the value of cv4 is first
calculated according to the above algorithm by calculating upper U(4) and lower L(4) which are
ð4Þ ð4Þ ð4Þ ð4Þ
312 and − 312, respectively. The boundaries of the subintervals SL1 , SL2 , SL3 and SL4 are
also calculated and are equal to 188, 63, −62 and − 187, respectively. Since index = 59 resides
ð4Þ ð4Þ
in the subinterval boundaries SL2 and SL3 which is [63, −62], the value of cv4 is 0. The
value of index is updated to new index = 59–53 × 0 = 59. The value of cv3 is then calculated and
this process is repeated until cv1 is obtained. The elements of the code vector become cv = (−1,
2, 2, 0).
This way of embedding secret digits into a cover image uses m states to change the cover
pixels in each group. However, considering the characteristics of the human visual system
which are less sensitive to deviation at edges than on the smooth areas, more secret bits can be
embedded into the edge areas than the non-edge ones. This can be achieved by using medge and
mnon-edge states (medge > mnon-edge and mnon-edge ≥ 2) to change the cover pixels in the groups
classified as edge and non-edge groups, respectively. Therefore, in the proposed scheme the
cover image is divided into 4 × 4 non overlapping pixel blocks. Each 4 × 4 pixels block is
divided into four 2 × 2 blocks. Each 2 × 2 pixels block is defined as edge/non-edge block
according to the types of the pixels in the edge image (EI) obtained by applying MNSED or
any edge detector as follows:

0; if nnon−edge ≥ N =2;
blk type ¼ ð8Þ
1; otherwise;

where N is the number of pixels in a 2 × 2 block. In other words, if the number of non-edge
pixels nnon − edge in a 2 × 2 block in an edge image is greater than or equal to N/2, the block at
the same spatial location in the cover image is defined as non-edge block (i.e., the type of the
block (blktype) is zero). Otherwise the block is defined as edge block (i.e., blktype is one). The
secret digits are embedded into edge and non-edge blocks using medge and mnon-edge states,
respectively. Moreover, the modification intervals of edge and non-edge blocks are MIedge =
[rL_edge, rU_edge] and MInon-edge = [rL_non-edge, rU_non-edge], respectively.
There are two main problems with the existing steganography schemes based on edge
detection which are: 1) the overhead information representing the edge information should be
hidden in the cover image; and 2) after embedding the secret bits, the non-edge pixels in the
cover image may be changed into edge pixels in the stego-image and vice versa. Consequent-
ly, it is difficult to reconstruct the secret bits properly. The proposed AEMD scheme over-
comes these drawbacks, which are explained in three stages of edge detection, embedding, and
extraction in the following subsections.

3.1 Edge detection

At this stage, the proposed steganography scheme avoids the burden of embedding
overhead edge information to fully assign the embedding space to only the secret bits.

Fig. 1 The modification interval for m = 5 and the corresponding boundaries of the subintervals
Multimedia Tools and Applications (2021) 80:21751–21769 21757

Moreover, it maintains edge pixels without being changed after embedding the secret
message. Therefore, the MNSED edge detector is performed on the adjusted image
(AIMSB) instead of the cover image to obtain an edge image (EI). For a given 8-bit
gray-level cover image, the adjusted image is generated by setting seven LSBs of each
pixel in the cover image to zero and preserving the most significant bit (MSB). It
means that all the pixels with grey levels greater than or equal to 128 will contribute in
identifying the edges.
Unlike the existing edge-based detection steganography schemes, the proposed
embedding scheme guarantees that the embedding space is allocated fully to embed
only the secret message and hence the embedding capacity increases without affecting
the quality of stego-image. Moreover, at the receiver side, the edge image of the
stego-image is the same as that of the cover image. Figure 2 shows the adjusted
images and their edge images which are generated by the proposed MNSED, Canny,
type-1 and type-2 fuzzy edge detectors. It is clear from this figure that the introduced
MNSED edge detector provides more edge pixels than the other edge detectors. This
leads to increasing the embedding capacity as will be demonstrated in the experimen-
tal results section.

3.2 Embedding procedure

The embedding process is as follows: firstly, the edge image (EI) is obtained as
described in Section 3.1. Secondly, Let SM be a series of binary secret messages and
C be the cover image of size W × H pixels. The cover image is divided into 2 × 2
blocks which are classified into edge and non-edge blocks as described in Section 3.
SM is then divided into NSedge and NSnon-edge segments equal to the number of 2 × 2
blocks defined as edge and non-edge blocks, respectively. Similarly, each segment of
Ledge and Lnon-edge bits are embedded into a 2 × 2 edge and non-edge blocks, respec-
tively. It means that each secret segment is converted into a secret digit d based on
mnedge -ary or mnnon−edge -ary notational system according to the type of each block (i.e.,
the binary segment is converted into decimal and the decimal value is converted into d
in the mnedge -ary or mnnon−edge -ary base number). The values of Ledge and Lnon-edge are
given by:
j k j k
Ledge ¼ log2 mnedge and Lnon−edge log2 mnnon−edge : ð9Þ

45237 73366 77254 57090 122878


(a) (b) (c) (d) (e) (f)
Fig. 2 Comparison among various edge detectors and the proposed MNSED edge detector. (a) Adjusted image
(AIMSB) and the generated edge images using (b) Canny, (c) type-1 fuzzy, (d) type-2 fuzzy (e) MNSED_γ (at
obtained γ =0.2336) and (f) MNSED (at fixed γ =0.01) edge detectors and their corresponding number of edge
pixels for Baboon image of size 512 × 512 pixels
21758 Multimedia Tools and Applications (2021) 80:21751–21769

The theoretical embedding rate Rth of the proposed method is given by:


NS edge  log2 mnedge þ NS non−edge  log 2 mnnon−edge
Rth ¼ : ð10Þ
W H

Accordingly, the embedding capacity of the proposed scheme depends on the values of
the medge and mnon-edge and the method of edge detector that specifies the number of
edge and non-edge blocks (NSedge and NSnon-edge). While the theoretical embedding rates
Rth of the Zhang and Wang’s EMD method [37] and EMD-2 [13] are Rth = log2(2n + 1)/
n = 1.16 bpp and Rth = log2(2w + 1)/n = 1.5850 bpp, respectively, which are their max-
imum rates at n = 2 and w = 4.
After calculating the secret digit d to be embedded in a block, the embedding
function f and the index are calculated from (3) and (4), respectively. If the block is
defined as edge/non-edge, then medge/mnon-edge is used in (3) and (4) (i.e., m ∈ {medge,
0
mnon − edge}). Consequently, by using (5) each stego-pixel gi is obtained by modifying
each cover pixel gi in a 2 × 2 block by the i-th value of the code vector cv generated
according to the algorithm described in Section 3. After embedding, overflow or
0
underflow problem may occur for 8-bit image pixels when the stego-pixel gi is greater
than 255 or less than zero, respectively. This can be rectified by inspecting the cover
pixels in a 2 × 2 block according to:

8 0
< g i −Max; if g i > 255;
mg i ¼ g i −Min; ifg i < 0;
0
ð11Þ
:
gi ; otherwise;

where Max and Min are the maximum and minimum values in the modification
interval, respectively. This means that if the blktype is edge, Max and Min are equal
 
to rU ∈ rU edge ; rU non−edge and rL ∈ rLedge ; rLnon−edge , respectively. The secret digit d is
embedded again but using the modified cover pixel vector mg.
Moreover, after embedding, mismatch edge problem may occur when the cover
0
pixel gi in the range of [127- rU + 1, 127] is modified into stego pixel g i in the range
of [128, 128+ rU -1] or the cover pixel gi in the range of [128, 128+ rU -1] is
0
modified into g i in the range of [127- rU + 1, 127]. This mismatch edge problem can
also be rectified according to:

8 0
< g i −Max; if gi ≤ 127&g i ≥ 128:
0
mgi ¼ g i þ Max; if gi ≥ 128 & gi ≤ 127: ð12Þ
:
gi ; otherwise:

The secret digit d is embedded again but setting g equal to mg. The embedding procedure is
summarised in Algorithm 2.
Multimedia Tools and Applications (2021) 80:21751–21769 21759

3.3 Extraction procedure

At the receiver, the stego-image S and parameter mnon-edge are received from the sender. The
secret message is recovered by firstly applying the MNSED edge detector on adjusted image
(AIMSB) generated from S to obtain the edge image EI. Similar to the embedding algorithm,
the stego-image is divided into 2 × 2 blocks which their types are identified by EI. The secret
digit d is extracted from each block g′ by calculating the extracted function f using (3) (i.e., d =
f). Then, the secret digit d is converted into Lnon-edge /Ledge bits of the original secret message
SM. This process is repeated until all the secret digits and their corresponding secret bits are
extracted to retrieve the original secret message SM.
21760 Multimedia Tools and Applications (2021) 80:21751–21769

3.4 Examples

The following examples illustrate how the embedding and extraction procedures work.
Assume mnon-edge is set to 4 and medge is equal to mnon-edge+1. A 2 × 2 block g is given as
(130, 1 92; 130, 184) with a t ype identified as b l k t y p e = edge . T h er e for e

b= m0edge ; m1edge ; m2edge ; m3edge =(1, 5, 25, 125). The segment of the secret bit-stream to be
embedded with length Ledge calculated using (9), SM = 110,110,011, is converted into a
secret digit d = 411 in a 625-ary (54). According to (3), the value of f is 465. Since d is less
than f, using (4) the value of index is 571 and the code vector is then generated according
to the proposed recursive algorithm cv = (−1, 1, 2, 0). The stego pixels are obtained using
(5) and are equal to S=g′=(131, 191; 128, 184). Consequently, at the receiver end, the
stego-pixels are g′=(131, 191; 128, 184). By analyzing EI, it is easy to determine the type
of that block which is blktype = edge. By knowing the value of mnon-edge, the values of medge
and Ledge can be determined which are 5 and 9, respectively. The extracted secret digit d is
calculated as d = f = (131 × 1 + 191 × 5 + 128 × 25 + 184 × 125) mod 54 becomes 411 and
the recovered message in its binary form is 110,110,011.
Another example is given to illustrate the mismatch edge situation. In this example,
mnon-edge, medge and blktype are also 4, 5 and edge. g is given as (123, 127; 126, 126) and
SM = 011000001 is changed into digit d = 262 in a 625–ary. Since d is less than f
which is 283, the value of index is 604 and the code vector is then generated and is
equal to cv = (1, −1, 1, 0). The stego pixels are g-cv = g′ = (122, 128; 125, 126). Since
0
only g2 is in mismatch edge situation, it is modified using (12) into mg2 = 127-Max,
whereMax ¼ rU ¼ rU edge ¼ 2. After that, g is set to mg = (123, 125; 126, 126). The
value of f is recalculated using the new g and according to (3) it is equal to 273. The
value of index is 614 and the corresponding code vector is cv = (1, 2, 0, 0). Therefore,
the stego-pixels are g′=(122, 123; 126, 126). The extraction process is the same as the
previous example. At the receiver, the stego-image S and parameter mnon-edge are
received from the sender. The secret message is recovered by firstly applying the
MNSED edge detector on adjusted image (AIMSB) generated from S to obtain the edge
image EI. Similar to the embedding algorithm, the stego-image is divided into 2 × 2
blocks which their types are identified by EI. The secret digit d is extracted from each
block g′ by calculating the extracted function f using (3) (i.e., d = f). Then, the secret
digit d is converted into Lnon-edge /Ledge bits of the original secret message SM. This
process is repeated until all the secret digits and their corresponding secret bits are
extracted to retrieve the original secret message SM.

3.5 Content-selective blocks at low payloads

Recently content-adaptive steganography methods have been suggested for testing the
detectability and security of steganography methods. Some of them are designed for
gray-scale images such as S-UNIWARD [10] and others for true color images. In the
case of true color images data are hidden either by independently embedding messages
into color channels [1] or by exploiting inter-channel dependencies to adaptively
partition payload for color channels [21]. The single image steganographic schemes
have been extended to multiple images steganography schemes with introducing adap-
tive payload distribution strategies [22].
Multimedia Tools and Applications (2021) 80:21751–21769 21761

For single gray-scale images steganography methods, they embed a given message
while minimizing a suitably defined distortion function for distributing the message
among pixels to reduce detectability measure using blind steganalyzers and then
improve the security. Since the detectability of the steganographic algorithms is tested
at an arbitrarily relative payload α, 0 < α ≤1, with the main focus on small relative
payloads, α ≤1/2, the proposed method, which is designed for embedding at high
payloads, should be adapted to perform embedding at low payloads by selecting the
blocks that minimize the embedding cost.
The additive approximation of distortion function D(C,S) is used for embedding with all
steganographic methods proposed in [10] to compute the cost ρij of modifying each cover pixel
cij to stego pixel sij or leaving all other cover pixels unchanged. In the proposed method, this
principle is extended to select the blocks adaptively for embedding the secret message. To
perform embedding at low payloads, mnon-edge and medge are set to 2. That means the
modification interval of the pixels in each block is [−1, 1] and the pixels in each block are
changed by 1 or − 1. So like the algorithms in [21], the embedding pattern can be defined as
I = S − C = {−1, 0,+1}W × H, that means the cover pixel values are changed by adding +1 or − 1
or are not changed.
Having defined the distortion function and a fixed relative payload, the constrained
optimization problem is solved to find the probability distribution p(I) that minimizes the
embedding impact and it can be written as:
arg min EðDðC; SÞÞ ¼ arg min ∑ DðC; SÞ⋅pðI Þ Subject to ð13Þ
p p I

l ¼ HðpÞ≜− ∑ pðI Þ  logðpðI ÞÞ; ð14Þ


I

where l is the length of the message in bits=α × W × H. The distortion function can be written
as:
W H  
DðC; SÞ ¼ ∑ ∑ ρij ⋅cij −sij ; ð15Þ
i¼1 j¼1

where ρij is the embedding impact of changing a pixel by 1 or − 1 which is called the
embedding cost. The optimization problem is then solved using Lagrangian multipliers
method resulting in probability distributions for embedding changes given by:
 −λρ


e ij = 1 þ 2e−λρij ; if I ij ¼ 1;
p I ij ¼
ð16Þ
1= 1 þ 2e−λρij ; if I ij ¼ 0;

where λ is the Lagrangian multiplier calculated using the payload constraint in (14). The
probability of changing each 2 × 2 block by 1 or − 1 is then defined as:

p blk ð1Þ ¼ max pij ð1Þ; piþ1; j ð1Þ; pi; jþ1 ð1Þ; piþ1; jþ1 ð1Þ
ð17Þ
p blk ð−1Þ ¼ max pij ð−1Þ; piþ1; j ð−1Þ; pi; jþ1 ð−1Þ; piþ1; jþ1 ð−1Þ

That means the cover image is classified into two disjoint classes of blocks with probability
p_blk(I) – those that are likely to be changed are marked as embedded blocks and those that are
21762 Multimedia Tools and Applications (2021) 80:21751–21769

less likely to be changed are marked as unembedded blocks. The proposed embedding
algorithm (Algorithm 2) is then applied into the blocks that are marked as embedded blocks.

4 Experimental results

The performance of the proposed scheme was analyzed in terms of data hiding capacity and
invisibility benchmarks. The embedding capacity (EC) or payload is the number of secret bits
that can be embedded into the pixels of the cover image. The invisibility of a stego-image was
measured on 512 × 512 pixels greyscale images through Peak Signal to Noise Ratio (PSNR)
and Structural Similarity Index measurement (SSIM) metrics.
First, various edge detection methods were utilized on the proposed method as shown in
Table 1. For edge detectors based on type-1 and type-2 fuzzy systems, the same inputs, output
and rules defined in [9] are used in this paper except the output MR for type-2 which is set
constant. The results in this table were obtained when mnon-edge was set to 2, 4, or 9 and medge
was equal to mnon-edge+1. It is clear from Table 1 that the proposed steganography algorithm
with MNSED achieves higher embedding capacity than that with other edge detection
methods. This is because as mentioned in Section 2, when γ is set to 0.01, MNSED always
produces the highest number of edge pixels than MNSED_γ and other edge detection methods
and that is clear in Fig. 2. Therefore, the performance of AEMD with MNSED in terms of
averaged PSNR and averaged SSIM is slightly lower than that with other edge detectors but at
higher EC.
Canny and the proposed MNSED edge detectors have computational complexity of the
order of O(W H(1 + 2 k)) and O (W H (9k2 + nT × nF + 23)), respectively, with nT and nF
iterations and k is the filter length. However, the CPU time consumed by MNSED edge detector
is less than the other edge detectors based on type-1 and type-2 fuzzy systems as shown in
Table 2. Although the computational complexity of Canny is less than MNSED edge detector,
in the next experiments MNSED edge detector is utilized in the proposed method since
MNSED provides more edge pixels than the other edge detectors as demonstrated in Fig. 2
and hence it leads to increased embedding capacity by the proposed steganography method.
Second, the proposed steganography method, AEMD, was compared with the other
methods based on the variants of EMD, including EMD [37], EMD-2 [13], and the state of
the art methods such as Shen et al. [32], Saha et al. [27], and Lee et al. [20] methods. Table 3
shows the results for n = 2 in both EMD and EMD-2 to achieve maximum embedding rates.
Although theoretical embedding rates of EMD and EMD-2 are 1.161 and 1.58 bpp, practically,
these methods achieve maximum embedding rates of 1 and 1.5 bpp, respectively, which are not
high. The corresponding averaged PSNR values for the EMD and EMD-2 are 52.1 and 49.8 dB,
respectively. The proposed method was implemented with MNSED at mnon-edge equal to 2 and
3. As can be seen from Table 3, the proposed method provides better embedding rate but with
lower image quality compared to EMD and EMD-2 because the proposed method is exploited
at its full capacity. Unlike EMD and EMD-2, the proposed method has the flexibility to adjust
the embedding capacity as well as the image quality by selecting the value of mnon-edge.
The computational complexity of proposed embedding algorithm shown in Algorithm 2 is
approximately equal to the computational complexities of Algorithm 1 and MNSED edge
detector algorithm. These two algorithms have computational complexity the order of
O(nm(WH/n)) and O (W H (9k2 + nT × nF + 23)), respectively. In other words, the proposed
embedding algorithm (Algorithm 2) has computational complexity of the order of O(WH(m +
Table 1 Performance of the proposed steganography algorithm utilizing various edge detection methods

Cover images AEMD with Canny AEMD with type-1 fuzzy AEMD with type-2 fuzzy AEMD with MNSED_γ AEMD with MNSED

EC (bpp) PSNR (dB) SSIM EC (bpp) PSNR (dB) SSIM EC (bpp) PSNR (dB) SSIM EC (bpp) PSNR (dB) SSIM EC (bpp) PSNR (dB) SSIM

Lena 1.0148 52.07 0.9969 1.0561 51.81 0.9968 1.0547 51.82 0.9968 1.0235 52.01 0.9969 1.0615 51.75 0.9967
2.0074 46.45 0.9885 2.0281 46.29 0.9883 2.0274 46.28 0.9882 2.0117 46.42 0.9885 2.0307 46.21 0.9880
3.0074 39.43 0.9454 3.0281 39.39 0.9455 3.0274 39.41 0.9458 3.0117 39.43 0.9456 3.0307 39.42 0.9460
Baboon 1.0350 51.94 0.9990 1.1069 51.53 0.9989 1.1209 51.44 0.9989 1.0557 51.81 0.9990 1.2142 50.96 0.9988
2.0175 46.39 0.9963 2.0535 46.12 0.9961 2.0604 46.07 0.9961 2.0278 46.33 0.9963 2.1071 45.73 0.9959
3.0175 39.42 0.9815 3.0535 39.30 0.9813 3.0604 39.26 0.9815 3.0278 39.39 0.9815 3.1071 39.09 0.9812
Peppers 1.0098 52.12 0.9971 1.0435 51.91 0.9970 1.0403 51.92 0.9970 1.0281 51.99 0.9970 1.0454 51.88 0.9970
Multimedia Tools and Applications (2021) 80:21751–21769

2.0049 46.52 0.9894 2.0217 46.39 0.9893 2.0202 46.39 0.9893 2.0140 46.44 0.9894 2.0227 46.37 0.9892
3.0049 39.68 0.9509 3.0217 39.61 0.9507 3.0202 39.63 0.9508 3.0140 39.66 0.9509 3.0227 39.64 0.9509
Jet 1.0095 52.15 0.9966 1.0460 51.93 0.9966 1.0493 51.90 0.9966 1.0190 52.09 0.9966 1.0443 51.97 0.9966
2.0048 46.55 0.9878 2.0230 46.42 0.9878 2.0247 46.39 0.9877 2.0095 46.52 0.9878 2.0222 46.41 0.9877
3.0048 39.72 0.9444 3.0230 39.66 0.9442 3.0247 39.65 0.9441 3.0095 39.72 0.9444 3.0222 39.67 0.9444
Lake 1.0121 52.13 0.9978 1.0581 51.85 0.9978 1.0581 51.86 0.9978 1.0208 52.07 0.9978 1.0587 51.89 0.9978
2.0061 46.56 0.9920 2.0291 46.39 0.9919 2.0290 46.38 0.9919 2.0104 46.53 0.9919 2.0293 46.38 0.9919
3.0061 39.77 0.9630 3.0291 39.66 0.9629 3.0290 39.67 0.9631 3.0104 39.75 0.9630 3.0293 39.65 0.9627
Boat 1.0123 52.16 0.9972 1.0604 51.87 0.9971 1.0653 51.84 0.9971 1.0267 52.07 0.9972 1.0622 51.89 0.9971
2.0061 46.53 0.9896 2.0302 46.35 0.9895 2.0326 46.33 0.9895 2.0134 46.46 0.9895 2.0311 46.33 0.9895
3.0061 39.74 0.9545 3.0302 39.63 0.9541 3.0326 39.61 0.9533 3.0134 39.71 0.9545 3.0311 39.65 0.9545
Average 2.0104 46.07 0.9816 2.0412 45.90 0.9814 2.0432 45.88 0.9814 2.0193 46.02 0.9815 2.0540 45.83 0.9814
21763
21764 Multimedia Tools and Applications (2021) 80:21751–21769

Table 2 Averaged CPU time consumed by the three edge detection algorithms applied on various images of
512 × 512 pixels

Canny Type-1 fuzzy Type-2 fuzzy MNSED

CPU Time 0.0235 (sec) 3.9315 (min) 9.5267 (min) 1.5750 (sec)

9k2 + nT × nF + 23)). It should be mentioned that the computational complexity of the stega-
nography methods that utilize edge detection to facilitate embedding more secret bits in the
edge areas than the smooth ones are higher than those which do not utilize edge detection.
Therefore, AEMD is more complex than EMD and EMD-2 due to the computational
complexity of MNSED edge detector.
The performance comparison of the proposed method to the state of the art methods based
on the EMD is presented in Table 4. As can be seen, the proposed method with MNSED at
mnon-edge = 4 provides higher embedding capacity than Shen et al. [32], Saha et al. [27] and Lee
et al. [20] methods with achieving higher PSNR values than Shen et al. [32], Saha et al. [27]
and comparable PSNR values to Lee et al. [20] for most cover images. However, when the
number of edge pixels are reduced by using MNSED_γ, the proposed AEMD with
MNSED_γ, as demonstrated in Table 1 for mnon-edge = 4, has a better EC and PSNR than that
of Lee et al. [20].
The proposed AEMD with MNSED method was also compared with methods based
on PVD including Wu et al’s method [36] and PBPVD [11] and the state of the art
method based on wavelet packet transform, WPT [3]. The results summarized in Table 5
indicate that the proposed method at mnon-edge = 9 achieves approximately 50%, 38%
and 19% on average higher embedding capacity than Wu et al. [36], PBPVD [11] and
WPT [3] methods, respectively. It also provides an averaged PSNR gain of approxi-
mately 2.1, 1.6 and 0.8 dB over those of Wu et al. [36], PBPVD [11] and WPT [3]
methods, respectively. For Baboon image, the PSNR of AEMD with MNSED is lower
than that with WPT but with approximately 20% increase on the EC. However, when
the proposed AEMD with MNSED_γ is used, the obtained PSNR and EC are
39.3924 dB and 3.0278 bpp equal to 99,215 bytes (as indicated in Table 1), respec-
tively which are higher than those obtained by WPT. Furthermore, the superiority of

Table 3 Performance comparison of the proposed steganography algorithm with the conventional EMD and
EMD-2 methods

Cover images EMD [37], EMD-2 [13], Proposed AEMD with MNSED
Rth=1.16 Rth=1.58
mnon-edge =2 mnon-edge =3

EC PSNR EC PSNR EC PSNR Rth EC PSNR Rth


(bpp) (dB) (bpp) (dB) (bpp) (dB) (bpp) (dB)

Lena 1.00 52.12 1.50 49.89 1.0615 51.80 1.0719 1.5615 49.29 1.6360
Baboon 1.00 52.11 1.50 49.89 1.2142 51.01 1.2506 1.7142 48.16 1.7627
Peppers 1.00 52.11 1.50 49.89 1.0454 51.93 1.0531 1.5454 49.49 1.6226
Jet 1.00 52.10 1.50 49.89 1.0443 51.97 1.0519 1.5443 49.54 1.6218
Lake 1.00 52.11 1.50 49.89 1.0587 51.89 1.0686 1.5587 49.40 1.6337
Boat 1.00 52.11 1.50 49.88 1.0622 51.89 1.0728 1.5622 49.35 1.6366
Couple 1.00 52.11 1.50 49.89 1.0971 51.61 1.1136 1.5971 49.01 1.6656
Average 1.00 52.11 1.50 49.89 1.0833 51.73 1.0975 1.5833 49.18 1.6541
Multimedia Tools and Applications (2021) 80:21751–21769 21765

Table 4 Performance comparison of the proposed AEMD steganography algorithm (mnon-edge = 4) with some
state-of-the-art methods based on EMD. The best result is shown in boldface

Cover images Shen et. al [32] Saha et. al [27] Lee et. al [20] AEMD with MNSED

EC PSNR EC PSNR EC PSNR EC PSNR


(bpp) (dB) (bpp) (dB) (bpp) (dB) (bpp) (dB)

Lena 1.685 40.89 2.00 44.03 2.00 46.38 2.031 46.21


Baboon 1.906 36.13 2.00 44.03 2.00 46.35 2.107 45.73
Peppers 1.688 39.99 2.00 44.22 2.00 46.36 2.023 46.37
Jet 1.700 40.10 2.00 43.94 2.00 46.36 2.022 46.41
Lake 1.723 39.65 2.00 43.95 2.00 46.35 2.029 46.38
Boat 1.718 40.02 2.00 44.03 2.00 46.36 2.031 46.33
Couple 1.732 39.63 2.00 44.04 2.00 46.37 2.049 46.30

the proposed AEMD at various values of mnon-edge over other state of the art methods
including DT-CWT [2], Sahu et al. [30], Sahu et. Al [29] and PVDMF [1] is shown in
Table 6.
Figure 3 shows the visual quality of the stego-images generated by the proposed scheme
and some of the mentioned steganography methods. It is clear that the distortions resulted from
embedding using the proposed method are imperceptible to human vision.
It is worth noting that the performance of the proposed method is not just limited to a
certain resolution of the cover image. To justify this claim, the performance of AEMD method
was evaluated using 1000 grayscale 512 × 512 pixel images selected randomly from
BOSSbase1.01 database [4]. The selected images were down sampled to 256 × 256 pixels
resolution. The grayscale 256 × 256 pixel images are also selected from USC-SIPI Image
database (http: //sipi.usc.edu/services/database/Database.html) to further demonstrate the
performance of AEMD. Table 7 illustrates the averaged PSNR and payload obtained by the
proposed method, WPT, EMD and EMD-2 using these two datasets. It is clear that the
performance of the AEMD method is resolution independent of the cover image and confirms
the results reported in Tables 3 and 5. Thus, the performance of the proposed method is
superior to the other algorithms in terms of both the embedding capacity and quality.

Table 5 Performance comparison of the proposed AEMD steganography algorithm (mnon-edge = 9) with some
state-of-the-art methods based on PVD and Wavelet packet transform. The best result is shown in boldface

Cover Wu et. al [36] PBPVD [11] WPT [3] AEMD with MNSED
images k=3

EC PSNR EC PSNR EC PSNR EC PSNR


(bytes) (dB) (bytes) (dB) (bytes) (dB) (bytes) (dB)

Lena 66,064 38.80 70,403 39.09 82,960 38.77 99,311 39.42


Baboon 68,007 33.33 78,973 35.06 84,902 39.15 101,813 39.08
Peppers 66,032 37.50 70,322 39.30 82,615 39.09 99,048 39.63
Jet 66,256 37.63 69,241 38.95 83,709 38.02 99,031 39.66
Lake 66,912 36.92 73,422 37.59 84,619 38.68 99,265 39.65
Boat 66,622 35.01 73,130 37.49 84,058 38.31 99,324 39.64
Couple 66,167 39.07 72,214 37.75 83,831 38.81 99,895 39.22
21766 Multimedia Tools and Applications (2021) 80:21751–21769

Table 6 Performance comparison of the proposed AEMD steganography algorithm (mnon-edge = 5 or 6) with
other state-of-the-art methods

Cover images DE-DWT [2] Sahu et. al [30] PVDMF [28] Sahu et. al [29] AEMD WITH MNSED

EC PSNR EC PSNR EC PSNR EC PSNR EC PSNR

Lena 0.5 49.10 1.5 47.38 2.1 42.98 2.0 44.08 2.2807 44.59
Baboon 0.5 49.08 1.5 47.36 2.5 38.83 2.0 44.07 2.6071 42.42
Peppers 0.5 49.10 1.5 47.39 2.1 43.23 2.0 44.09 2.2727 44.80
Jet 0.5 49.09 1.5 47.34 2.2 42.85 2.0 44.08 2.2722 44.83
Lake 0.5 49.10 1.5 47.30 2.4 41.48 2.0 44.07 2.2793 44.81
Boat 0.5 49.10 1.5 47.20 2.5 41.75 2.0 44.08 2.5311 42.96
Couple 0.5 49.11 1.5 47.38 2.2 41.53 2.0 44.09 2.2986 44.44

Finally, the security analysis of the proposed method was carried out through universal
steganalysis. The universal steganalysis is a blind steganalysis to detect the existence of the
secret message in stego-images without any prior knowledge about the type of the

WPT [23] EMD [7] EMD-2 [12] Proposed AEMD

Peppers

PSNR (dB) 48.2059 52.11 49.89 51.93


Payload (bpp) 1.0615 1.0 1.5 1.0454

Tiffany

PSNR (dB) 48.1412 52.0404 49.9034 52.1307


Payload (bpp) 1.0957 1.0 1.5 1.0167
Shen et. al [17] Sahu et. al [33] Sahu et. al [34] Proposed AEMD

Lena

PSNR (dB) 40.89 47.38 44.08 46.21


Payload (bpp) 1.685 1.5 2.0 2.031

Baboon

PSNR (dB) 36.13 47.36 44.07 45.73


Payload (bpp) 1.906 1.5 2.0 2.107
Fig. 3 Visual comparison among various steganography methods and the proposed AEMD method
Multimedia Tools and Applications (2021) 80:21751–21769 21767

Table 7 The average PSNR and payload for the proposed, PWT, EMD, and EMD-2 steganographic methods
applied on BOSSbase1.01 and SIPI datasets

Dataset PWT EMD EMD-2 AEMD with MNSED

BOSSbase PSNR 48.33 44.28 52.06 49.90 51.899 49.53 46.497


1.01 EC 1.046 1.796 1.00 1.50 1.042 1.542 2.0211
SIPI PSNR 48.32 44.48 52.07 49.94 51.65 49.12 46.298
EC 1.056 1.806 1.00 1.50 1.075 1.575 2.038

steganography algorithms. The blind steganalysis is built based on machine learning, which
consists of feature extraction and classification. A set of features is first extracted from the
cover and stego-images. A classifier is then trained using these extracted features to classify
the images into two classes: cover images and stego images.
In this paper, the security of the proposed method against the universal analysis was
evaluated and compared to the steganographic algorithms of WPT [3], S-UNIWARD [10]
and EMD [37]. Steganalysis was implemented using the version of the spatial rich model
(SRM) [7] with a single quantization (SRMQ1) with 12,753 features. The extracted features
were trained using the ensemble with Fisher linear discriminant (FLD) classifier [14]. Exper-
iments were performed on 1000 images selected randomly from BOSSbase1.01 database [4].
The proposed method with selecting the blocks adaptively as described in Section 3.5 was
tested at various low payloads ranging from 0.2 to 0.8 bits per pixel (bpp) at mnon-edge = 2. The
blocks of EMD were also selected adaptively using the proposed algorithm in Section 3.5. The
performance of the steganalyzer is always measured on the testing set using the minimal
average decision error PE under equal probability of cover and stego images which is given by:
1
PE ¼ min ðPFA þ PFN Þ; ð18Þ
PFA 2

where PFA and PFN are the probabilities of false alarm or false positive (detecting cover as stego) and
missed detection (false negative), respectively. It should be mentioned that the security of a
steganographic method is judged by the difficulty of detecting the secret message. This difficulty
is achieved when the detection error rate is high as shown in Table 8. It is clear from this table that
the proposed method has superior security performance than the other algorithms.

5 Conclusion

This paper has proposed an image steganography method based on advanced EMD and
neutrosophic set. In the proposed scheme, the cover image is divided into blocks which are

Table 8 Detection error PE for the proposed and various steganographic methods

Steganography SRMQ1 Feature [7]


method
0.2 0.4 0.6 0.8

PWT [3] 0.0297 0.0250 0.0223 0.0220


S-UNIWARD [10] 0.4027 0.2952 0.2182 0.1412
EMD [37] 0.3965 0.2977 0.2083 0.1398
Proposed 0.4183 0.3172 0.2475 0.1752
21768 Multimedia Tools and Applications (2021) 80:21751–21769

classified into edge and non-edge blocks based on the modified neutrosophic set edge detector
(MNSED). Unlike the existing steganography methods based on EMD, the proposed method
facilitates embedding the secret digits in two different -ary notational systems into the edge and
non-edge blocks. Moreover, it guarantees estimating the exact edge locations after embedding
the secret message and hence avoids the burden of embedding of overhead edge information in
the cover image. Experimental results showed the superiority of the proposed method over
state-of-the-art schemes in terms of both embedding capacity and stego-image quality. The
proposed method is also robust against the universal analysis. It is observed that the possibility
to further improve the security of the proposed method at low payloads can be investigated by
selecting the blocks adaptively for embedding the secret message through selecting a more
proper model.

References

1. Ahani S, Ghaemmaghami S (2015) Colour image steganography method based on sparse representation.
IET Image Process 9(6):496–505
2. Atawneh S, Almomani A, Al Bazar H, Sumari P, Gupta B (2017) Secure and imperceptible digital image
steganographic algorithm based on diamond encoding in DWT domain. Multimed Tools Appl 76(18):
18451–18472
3. Atta R, Ghanbari M (2018) A high payload steganography mechanism based on wavelet packet transfor-
mation and neutrosophic set. J Vis Commun Image Represent 53:42–54
4. Bas P, Filler T, Pevný T (2011) Break our steganographic system: The ins and outs of organizing BOSS, in
International Workshop on Information Hiding Springer, pp. 59–70
5. Chan CK, Cheng LM (2004) Hiding data in images by simple LSB substitution. Pattern Recogn 37(3):469–
474
6. Cheddad A, Condell J, Curran K, Mc Kevitt P (2010) Digital image steganography: survey and analysis of
current methods. Signal Process 90:727–752
7. Fridrich J, Kodovský J (2012) Rich models for steganalysis of digital images. IEEE Trans Inform Forensics
Secur 7(3):868–882
8. Ghosal SK, Mandal JK, Sarkar R (2018) High payload image steganography based on Laplacian of
Gaussian (LoG) edge detector. Multimed Tools Appl 77(23):30403–30418
9. Gubbi A, Azeem MF (May 2012) Fuzzy inference system - theory and applications, 1st ed.. InTech, ch. 13,
pp. 297–298
10. Holub V, Fridrich J, Denemark T (2014) Universal distortion function for steganography in an arbitrary
domain. EURASIP J Inf Secur 1:1–13
11. Hussain M, Wahab AWA, Ho AT, Javed N, Jung KH (2017) A data hiding scheme using parity-bit pixel
value differencing and improved rightmost digit replacement. Signal Process Image Commun 50:44–57
12. Kaur EK, Mutenja EV, Gill ES (2010) Fuzzy logic based image edge detection algorithm in Matlab. Int J
Comput Appl 1(22):55–58
13. Kim HJ, Kim C, Choi Y, Wang S, Zhang X (2010) Improved modification direction methods. Comput
Math Appl 60(2):319–352
14. Kodovský J, Fridrich J, Holub V (2012) Ensemble classifiers for steganalysis of digital media. IEEE Trans
Inform Forensics Secur 7(2):432–444
15. Kumar S, Singh A, Kumar M (2019) Information hiding with adaptive steganography based on novel fuzzy
edge identification. Defence Technol 15(2):162–169
16. Kuo WC, Wuu LC, Kuo SH (Aug. 2012) The high embedding steganographic method based on general multi-
EMD, in International Conference on Information Security and Intelligent Control (ISIC’12), pp.288–291
17. Kushwaha N (2012) Edge detection using fuzzy logic in Matlab. International Journal of Advanced
Research in Computer Science and Software Engineering 2(4):38–40
18. Lee CF, Wang YR, Chang CC (2007) A steganographic method with high embedding capacity by
improving exploiting modification direction, in Proc International Conference on Intelligent Information
Hiding and Multimedia Signal Processing, pp. 497–500
Multimedia Tools and Applications (2021) 80:21751–21769 21769

19. Lee CF, Chang CC, Wang KH (2008) An improvement of EMD embedding method for large payloads by
pixel segmentation strategy. Image Vis Comput 26(12):1670–1676
20. Lee CF, Chang CC, Pai PY, Liu CM (2015) Adjustment hiding method based on exploiting modification
direction. J Netw Secur 17(5):607–618
21. Liao X, Yu Y, Li B, Li Z, Qin Z (2020) A new payload partition strategy in color image steganography.
IEEE Trans Circuits Syst Video Technol 30(3):685–696
22. Liao X, Yin J, Chen M, Qin Z (2020) Adaptive payload distribution in multiple images steganography
based on image texture features. IEEE Transactions on Dependable and Secure Computing
23. Mielikainen J (2006) LSB matching revisited. IEEE Signal Process Lett 13(5):285–287
24. Niu X, Ma M, Tang R, Yin Z (2015) Image steganography via fully exploiting modification direction. Int J
Secur Appl 9(5):243–254
25. Petitcolas FA, Anderson RJ, Kuhn MG (1999) Information hiding-a survey. Proc IEEE 87(7):1062–1078
26. Pevný T, Bas P, Fridrich J (2010) Steganalysis by subtractive pixel adjacency matrix. IEEE Trans Inform
Forensics Secur 5(2):215–224
27. Saha S, Ghosal SK, Chakraborty A, Dhargupta S, Sarkar R, Mandal JK (2018) Improved exploiting
modification direction-based steganography using dynamic weightage array. Electron Lett 54(8):498–500
28. Sahu A, Swain G (2019) An optimal information hiding approach based on pixel value diferencing and
modulus function. Wirel Pers Commun 108(1):159–174
29. Sahu AK, Swain G (2019) A novel n-rightmost bit replacement image steganography technique. 3D
Research 10(1)
30. Sahu AK, Swain G, Babu ES (2018) Digital image steganography using bit flipping. Cybernetics Inform
Technol 18(1):69–80
31. Shen SY, Huang LH (2015) A data hiding scheme using pixel value differencing and improving exploiting
modification directions. Comput Secur 48:131–141
32. Shen S, Huang L, Yu S (2017) A novel adaptive data hiding based on improved EMD and interpolation.
Multimedia Tools Appl 14:1–17
33. Tseng HW, Leng HS (2014) A high-payload block-based data hiding scheme using hybrid edge detector
with minimal distortion. IET Image Process 8(11):647–654
34. Wang CM, Wu NI, Tsai CS, Hwang MS (2008) A high quality steganographic method with pixel-value
differencing and modulus function. J Syst Softw 81(1):150–158
35. Westfeld A, Pfitzmann A (2000) Attacks on steganographic systems. Lecture Notes in Computer Science,
Springer:1768
36. Wu HC, Wu NI, Tsai CS, Hwang MS (2005) Image steganographic scheme based on pixel-value
differencing and LSB replacement methods. IEE Vision, Image and Signal Processing 152(5):611–615
37. Zhang X, Wang S (2006) Efficient steganographic embedding by exploiting modification direction. IEEE
Commun Lett 10(113):781–783

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

You might also like