Image and Vision Computing 26 (2008) 16701676

Contents lists available at ScienceDirect

Image and Vision Computing

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

An improvement of EMD embedding method for large payloads by pixel

segmentation strategy
Chin-Feng Lee a,*, Chin-Chen Chang b,c, Kuo-Hua Wang b
a

Department of Information Management, Chaoyang University of Technology, 168 Jifong E. Road, Wufong Township, Taichung County 41349, Taiwan, ROC
Department of Information Engineering and Computer Science, Feng Chia University, 100 Wenhwa Rd., Seatwen, Taichung 40724, Taiwan, ROC
c
Department of Computer Science and Information Engineering, National Chung Cheng University, 160 San-Hsing, Ming-Hsiung, Chiayi 621, Taiwan, ROC
b

a r t i c l e

i n f o

Article history:
Received 22 November 2007
Received in revised form 1 April 2008
Accepted 5 May 2008

Keywords:
Covert communication
Digit steganography
Data hiding
Embedding capacity

a b s t r a c t
In this paper, a novel data hiding method by using pixel segmentation strategy is proposed. The proposed
paper keeps (16  pm) MSBs of a pixel-pair unchanged and alters pm LSBs to indicate the virtual modications on an m-dimensional pseudo-random vectors for carrying the secret data, where m 2pm 1  1.
The embedding rate of proposed method is R = (log2(2m + 1))/2, which is greater than that of the EMD
embedding method proposed by Zhang and Wang [X. Zhang, S. Wang, Efcient steganographic embedding by exploiting modication direction, IEEE Communication Letters 10 (2006) (113), pp.781783],
because the embedding rate of EMD embedding method is R = (log2(2n + 1))/n, when m > 2 and n = 2.
The experimental results show that the proposed method increases the number of embedded secret bits
more than 1.7 times compared with the EMD embedding method. Even with such high embedding capacity, the average PSNR of 44.3 dB shows that the visual quality does not decline to an unacceptable degree.
2008 Elsevier B.V. All rights reserved.

1. Introduction
Message transmission over the Internet is quite common in almost everywhere. Some problems may arise due to the security
aws of communication channel. Digital steganography, a kind of
data security technique, has been developed quickly and it receives
a great deal of attentions from both the academic and industrial
communities [111]. An original image, also called a cover image,
is used to embed the secret data. The stego-image is a version of
the cover image where secret messages are embedded. By means
of creating stego-images that are perceptually identical to the cover images with small embedding distortion, the imperceptibility
for protecting the sensitive and condential information can be
maintained without being detected or extracted. Two important issues for current data hiding techniques are to preserve the imperceptibility and the embedding capacity at the same time. However,
this is an irreconcilable conict: if we want to preserve high stegoimage quality, we usually have to scarify the embedding capacity,
and vice versa.
Nowadays, many steganographic methods have been proposed
to hide secret messages into an image. A commonly used method
is the least signicant bit (LSB) replacement, which is the simplest
hiding technique by using some least bits of cover pixels to embed
secret data [46,10]. Mielikainens LSB matching revisited method
* Corresponding author. Tel.: +886 4 23323000 4293; fax: +886 4 23742337.
E-mail addresses: [email protected] (C.-F. Lee), [email protected], m9503085@
fcu.edu.tw (C.-C. Chang).
0262-8856/$ - see front matter 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.imavis.2008.05.005

[7] can resist immune against the steganographic attacks because

it has not the asymmetric property of LSB replacement methods.
LSB matching revisited method also achieves the same quality of stego-images as well as LSB replacement does. However, it does not enhance the embedding capacity. For a cover image I with M  N pixels,
the maximum data hiding capacity of LSB steganography is
M  N bits. Dene the embedding rate R as the rate of the length of
embedded messages to the maximum capacity M  N. If the rate R
is more than one, it represents the steganographic embedding method with high embedding capacity [8,9,11].
Recently, Zhang and Wang [9] proposed a steganography to
transform the binary secret data into a stream of secret digits in
a (2n + 1)-ary notational system. Zhang and Wangs embedding
method uses n cover pixels to carry one secret digit in a (2n + 1)ary notational system. The embedding rate, R = (log2(2n + 1))/n, is
dened for calculating the number of secret bits that can be carried
by a cover pixel. Though Zhang and Wangs method could achieve
high embedding efciency and secrecy with low distortion, a larger
n leads to a smaller embedding rate according to the equation
R = (log2(2n + 1))/n. In other words, the EMD embedding method
has its maximum embedding rate which is approximate 1.161 bits
per pixel when n = 2.
This paper improves the EMD embedding method, especially in
embedding capacity. In our proposed method, each secret digit in a
(2m + 1)-ary notational system is carried by a pair of cover pixels.
The data hider can segment the pair of cover pixels into two parts:
one is used to embed the secret data and the other is used as an
indicator for the decoder to extract the secret data. The embedding

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

rate of our method is steadily increased as m gets larger without

losing the stego-image quality and security. Experimental results
show that our proposed method still keeps pretty good quality of
the stego-image, while the embedding capacity is exactly enlarged.
The rest of this paper is organized as follows. In Section 2, we
review previous approaches of embedding in grayscale images.
Then, in Section 3, we introduce the proposed embedding and
extracting procedures for grayscale images. We make comparisons
of embedding rate and stego-image quality between our method
and Zhang and Wangs method in Section 4. Finally, we conclude
this in Section 5.
2. The EMD embedding method
An embedding method to exploit the modication directions for
data hiding (also called the EMD embedding method for short) was
proposed by Zhang and Wang [9]. The EMD embedding method
pseudo-randomly permutes all pixels in a cover image with a
secret key to partition the pixels into a series of groups. A pixelgroup contains n grayscale pixels and is denoted as (g1, g2, . . . , gn).
The secret message to be embedded needs to be converted into a
sequence of secret digits of a (2n + 1)-ary notational system, where
each secret digit falls within [0, 2n]. Eq. (1) depicts that a secret
message of binary stream can be segmented into many pieces with
L bits, and the decimal value of each secret piece is represented by
K digits in the (2n + 1)-ary notational system, where

L bK  log2 2n 1c:

In the EMD embedding method, the data hiders can choose one of
(2n + 1)-ary notational systems and determine the length K of segmented bits to convert the secret message into a sequence of secret
digits. This feature can increase the security if the attacker does not
know the value of K and n. Zhang and Wang dened an embedding
function f as weighted sum function modulo (2n + 1) for each pixelgroup:

f g 1 ; g 2 ; . . . ; g n g 1  1 g 2  2 . . . g n  nmod2n 1:

1671

because at most one cover pixel needs to be increased or decreased

by one for a pixel-group of cover pixels. Though, the EMD embedding method gets high quality of stego-image, there is still much
room to further improve the embedding capacity, because the
embedding rate R = (log2(2n + 1))/n for the best case of n = 2 is only
1.2 secret bits per cover pixel.
3. Our proposed method
In this section, we rstly introduce the developed pixel-pair
segmentation strategy. Then, the proposed data embedding and
extracting procedures are presented to improve information hiding
ratio than that of Zhang and Wangs method.
3.1. Pixel-pair segmentation
The message embedding is performed onto a pair of cover pixels. With two grayscale pixels, 16 bits can be segmented into four
pieces (see Fig. 1(a)), which are denoted as S1, S2, S3, and S4, respectively. S1 refers to the piece which has PV1 most signicant bits of
the rst cover pixels, S2 refers to the piece which has (8  PV1) least
signicant bits of the rst cover pixels, S3 refers to the piece which
has PV2 most signicant bits of the second cover pixels, and S4
refers to the piece which has (8  PV2) least signicant bits of the
second cover pixels. Let pm is the length of S2 concatenating with
S4 i.e., pm = (8  PV1) + (8  PV2) = 16  (PV1 + PV2). After the pixel
segmentation is done, S1 and S3 are combined to be an area of a
vector of coordinates (VCA) as shown in Fig. 1(b) which can be
further processed to form a group of integer coordinates for logically embedding data. Moreover, S2 and S4 are combined to be a
coordinate vector modication area (VMA) as shown in Fig. 1(c).
Example 3.1. For a given pair of cover pixels C = (125, 101)10, we
rst transform C into its corresponding binary stream (01111101,
01100101)2. Assume a key is used to obtain PV1 = 7 and PV2 = 6, i.e.,
pm = 3, then the bits of C in VCA are (0111110 011001)2 and the bits
of C in VMA are (101)2.

2
If a secret digit to be embedded into a given cover pixel-group
(g1, g2, . . . , gn) is not equal to the value calculated from the embedding function, only one of the cover pixels has to be modied by
either increasing or decreasing one; otherwise, no modication
needs to be done. That is why the EMD embedding methods distortion induced in the stego-image is not great.
In the extracting procedure, if the stego pixel-group is
g 01 ; g 02 ; . . . ; g 0n , then the secret digit can be extracted by the following extraction function.

f g 01 ; g 02 ; . . . ; g 0n g 01  1 g 02  2 . . . g 0n  nmod2n 1:
3
The EMD embedding method provides a pretty high stego-image
quality with the PSNR value greater than 51 dB in average. This is

3.2. Embedding procedure

As we know, the distortion of grayscale image increases greatly
if we change the most signicant bits (MSBs) instead of the least
signicant bits (LSBs). In addition, the modication of the most signicant bits is much easier to be detected. Therefore, in the proposed method, only pm least signicant bits distributed over the
VMA of the given pair of cover pixels are possibly altered. In other
words, we do not physically modify the (16  pm) most signicant
bits to achieve the embedding of secret messages. In this way, good
stego-image quality can be preserved and the embedding ratio is
high.
In fact, the (16  pm) bits of VCA are merely used as a seed of
random function fr which can be predetermined and known by
both data hiders and extractors. The outcome generated by the

Fig. 1. Pixel-pair segmentation. (a) A grayscale pixel-pair; (b) VCA; (c) VMA.

1672

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

random function is a vector of coordinates with m positive integers, which can be denoted as (g1, g2, . . . , gm) and can be expressed
as a unit of hyper-cube in an m-dimensional space. Here, m is
determined by the value of pm such that m 2pm 1  1. The hyper-cube related to the vector of coordinates (g1, g2, . . . , gm) is inspired from the hyper-cube of a pixel-group introduced by Zhang
and Wang. The difference between the vector of coordinates (g1,
g2, . . . , gm) proposed here and the pixel-group proposed by Zhang
and Wang is that the vector of coordinates is projected from the
16  pm bits of VCA for imitating the embedding way proposed
by Zhang et al. That is, m integers of vector coordinates are not cover pixels. Therefore, no physical modication is performed on the
coordinate values. Also, no bits of VCA are modied during the
embedding procedure. The embedded secret message has to be
converted into a sequence of secret digits of the (2m + 1)-ary notational system, where each secret digit falls within [0, 2m]. We modify the pm bits of VMA to indicate which integer of g1, g2, . . . , and gm
will be logically increased/decreased one when a secret digit is
going to be embedded.
The embedding steps are presented as follows.
Step 1:
Step 2:

Step 3:

Step 4:

Step 5:
Step 6:

Determine the pm value as introduced in the procedure

of pixel-pair segmentation.
Pseudo-randomly permute all grayscale cover pixels
with a secret key, and divide these cover pixels into a
series of pixel-pairs.
For each pair of cover pixels, say C, we further proceed to
divide the pixel-pair C into two areas, VCA and VMA as
shown in Fig. 1.
Use the (16  pm) bits in VCA as a seed of random function to generate a vector of coordinates denoted as
(g1, g2, . . . , gm), where gi can be any positive integer, for
i = 1, 2, . . . , m, where m 2pm 1  1.
Convert a binary secret message into a sequence of
secret digits in the (2m + 1)-ary notational system.
Apply the embedding function which can be calculated
in Eq. (4) to embed every secret digit in the (2m + 1)ary notational system.

"
#
m
X
F f g 1 ; g 2 ; . . . ; g m
g i  i mod2m 1:

to Eq. (5), the F value of hyper-cube corresponding to (7, 13, 9)10 is

4. Meanwhile, the values of hyper-cubes corresponding to the six
neighbors of (7, 13, 9)10 are all within [0, 6] as shown in Fig. 2.
Example 3.3. Lets continue Example 3.2. We have gotten the
hyper-cube value (say F = 4) by given the vector of coordinates
(7, 13, 9)10, where m = 3. If a secret digit falling in [0, 7] is going to
be embedded, the coordinates should be logically modied to
be one of the (2m + 1) outcomes, totally depending on the
value of the embedded secret digit s. The seven outcomes are
(g1, g2, g3), (g1  1, g2, g3), (g1 + 1, g2, g3), (g1, g2  1, g3), (g1, g2 + 1, g3),
(g1, g2, g3  1), and (g1, g2, g3 + 1), respectively, discussed as follows.

Case 1:

Case 2:

Case 3:

Case 4:

Case 5:

Case 6:

Case 7:

if we want to embed (0)7, we rst calculate d = s  F mod

(2m + 1) = 0  4 mod 7 = 3, which is not greater than m,
where m = 3. Therefore, we add one to g3 such that the
vector of coordinates (7, 13, 9)10 can be replaced by
(7, 13, 10)10.
if we want to embed (1)7, we calculate d = s  F mod
(2m + 1) = 1  4 mod 7 = 4, and get d is greater than m.
Therefore, we decrease g2m+1d by 1 to obtain (7, 13, 8)10.
if we want to embed (2)7, we do the calculation as
d = s  F mod (2m + 1) = 2  4 mod 7 = 5, and see d is
greater than m. Therefore, we decrease g2m+1d by 1 such
that g2 = 12 and the vector of coordinates is (7, 12, 9)10.
if we want to embed (3)7, d is calculated as d = s  F mod
(2m + 1) = 3  4 mod 7 = 6, which is greater than m.
Therefore, we decrease g2m+1d by 1 such that g1 = 6
and the vector of coordinates is (6, 13, 9)10.
To embed (4)7 in (7, 13, 9)10, according to Fig. 2, if we
want to embed (3)7 in (7, 13, 9)10, no modication is
needed.
if we want to embed (5)7, we have d = s  F mod
(2m + 1) = 5  4 mod 7 = 1, which is not greater than
m. Therefore, we add one to g1 such that the vector of
coordinates (7, 13, 9)10 can be replaced by (8, 13, 9)10.
if we want to embed (6)7, we calculate d = s  F mod
(2m  1) = 0  4 mod 7 = 2, which is not greater than
m. Therefore, we add one to g2 such that the vector of
coordinates (7, 13, 9)10 can be replaced by (7, 14, 9)10.

i1

Since the vector of coordinates (g1, g2, . . . , gm) corresponds to a unit

hyper-cube in an m-dimensional integer space, the hyper-cube
can be labeled with its F value. Obviously, the F value of any
hyper-cube and its 2m neighbors are integers within [0, 2m], and
they are mutually different.
There are (2m + 1) possible cases when embedding a secret digit
s. If s is equal to F, no modication for the vector of coordinates is
needed. Otherwise, one of the m coordinates, i.e., gi needs to be
logically modied. From Eq. (5), d can be obtained rst.

d s  F mod 2m 1:

As mentioned in Section 3.1, the most (16  pm) signicant bits

of VCA can be used as a seed of random function and promise no bit
modication will be occur during the whole process. In the last
step of embedding procedure, we physically modify the pm bits
of VMA to indicate how the vector of coordinates are changed
according to Eq. (4). We use the most (pm  1) bits of VMA to which
gi of the coordinate vector (g1, g2, . . . , gm) should be logically modied. The last one bit of VMA is used to be a sign indicator. If the bit

Then, when d is greater than m, the value of g(2m1)d has to be

logically decreased by one; otherwise, the value of gd has to be
logically increased by one. That is, when we embed a secret
digit in the (2m + 1)-ary notational system, one logical outcome
of (g1, g2, . . . , gm), (g1  1, g2, . . . , gm), (g1 + 1, g2, . . . , gm), (g1, g2  1,
. . . , gm), (g1, g2 + 1, . . . , gm), . . . , (g1, g2, . . . , gm  1), and (g1, g2, . . . ,
gm + 1) is obtained after applying the embedding function in
Eq. (5).
Example 3.2. According to Example 3.1, when pm = 3, m is three.
Assume that the bit stream (0111110 011001)2 is an input
argument of a predetermined random function fr and the random
function fr will return a vector of coordinates, (7, 13, 9)10. According

Fig. 2. the hyper-cube values of the vector of coordinates (7, 13, 9)10 and its six
neighbors.

1673

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

of sign indicator is 0, then the ith g value should be decreased by

one; otherwise, the value gi should be increased by one.
Example 3.4. Let us continue Example 3.3 with pm = 3, m = 3.
There are (2*3 + 1) cases for physically modifying the three bits of
VMA to express the logical modication of the vector of
coordinates (7, 13, 9)10. The seven cases are described in detail as
follows.

Case 1:
Case 2:
Case 3:
Case 4:
Case 5:
Case 6:
Case 7:

If the modication we have to do is g1 + 1, it means that

the three bits in VMA is (011)2.
If the modication we have to do is g1  1, it means that
the three bits in VMA is (010)2.
If the modication we have to do is g2 + 1, it means that
the three bits in VMA is (101)2.
If the modication we have to do is g2  1, it means that
the three bits in VMA is (100)2.
If the modication we have to do is g3 + 1, it means that
the three bits in VMA is (111)2.
If the modication we have to do is g3  1, it means that
the three bits in VMA is (110)2.
If we do no modication on the vector of coordinates
(7, 13, 9)10, it means that the three bits in VMA can be
either (000)2 or (001)2.

3.3. Data extraction

The same as the embedding procedure, the extracting
procedure has to segment a given pair of stego-pixels into
two parts: the vector coordinate area (VCA) composed of
(16  pm) bits, and the vector modication area (VMA) composed of pm bits.
Use the (16  pm) bits in VCA as a seed of a predetermined random function fr to generate a vector of coordinates denoted as
g 01 ; g 02 ; . . . ; g 0m , where g 0i can be any positive integer, for i = 1,
2, . . . , m and m 2pm 1  1. Apply the following extraction function
as shown in Eq. (6) to the vector of coordinates g 01 ; g 02 ; . . . ; g 0m so as
to extract the embedded secret digit s in the (2m + 1)-ary notational system.

s f g 01 ; g 02 ; . . . ; g 0m
g 01  1 g 02  2 . . . g 0m  m mod 2m 1:

In order to compare the performance of the EMD embedding

and the proposed methods, we use the embedding rate dened
as R = S/H  W to calculate the number of secret bits carried by
one cover pixel, i.e., bits per pixel (bpp), where S refers to the number of bits of the secret data, and H  W is the total number of cover pixels.
By Zhang and Wangs method, S = (H  W  log2(2n + 1))/n.
Therefore, the embedding rate R of the EMD embedding method
is (log2(2n + 1))/n, where n = 2, which has the best embedding rate
R of (log2 5)/2, when n is set to be 2. In our proposed method,
S = (H  W  log2(2m + 1))/2
that
implies
R = log2(2m + 1)/
2 = log2(pm  1)/2 and the best embedding rate is log2 15/2, which
is almost twice than that of Zhang and Wangs method. Table 1
compares the embedding rate for the EMD embedding and the proposed methods.
The proposed paper keeps (16  pm) MSBs of a pixel-pair unchanged and alters pm LSBs to indicate the virtual modications
on m-dimensional pseudo-random vectors for carrying the secret
data. Compared with the LSB replacement, the proposed method
uses pm LSBs to theoretically carry log2(pm  1)/2 secret bits, which
is slightly less than the number of secret bits embedded in the LSB
replacement. Given a cover pixel-pair, Table 2 compares the
embedding capacity for exploiting two and three LSB planes,
respectively, to carry the number of secret bits per cover pixel.
The last row of Table 2 shows the average secret bits that are exactly embedded into a cover pixel.
Peak signal-to-noise rate (denoted as PSNR) is a common measurement for evaluating the performance of embedding methods.
In the experiment, we use PSNR value as a criterion to estimate
the stego-image quality. We test the proposed algorithm and
Zhang and Wangs method on various grayscale images named
Lena, Baboon, Tiffany, F16, Pepper, Goldhill, Boat,
Babara, and Zelda. These test images have the same size, that
is 512  512.
The EMD embedding method has the maximum embedding
rate about 1.16 bits per pixel, meaning that R = (log 5)/2 bits/pixel,
and the average PSNR value of stego-images is about 51 dB with
262,144 embedded bits. At the embedding rate R = (log 15)/2 with
the embedding capacity of 441,913463,528 bits, we present the
PSNR values of our proposed method for these nine stego-images
in Fig. 3. The average PSNR value is about 44.3 dB. As
m 2pm 1  1, m is determined by the value of pm. When the

Example 3.5. For a given pair of stego-pixels which is

(126, 101)10 = (01111110, 01100101)2, pm = 3, m = 3, and the
vector of coordinates g 01 ; g 02 ; g 03 (7, 13, 9)10, which is the
outcome of a predetermined random function fr by using the
binary stream (011111 0110010)2 as the input argument of fr.
For the three bits, i.e., (10 1)2, of VMA which obviously comes
from the given pair of stego-pixels, the most two signicant
bits, i.e., (10)2, indicate that g 02 6 g 2 , and g 0i g i for i = 1, 3.
Similarly, since the last signicant bit of VMA is (1)2, it indicates
that g 02 g 2 1.The outcome means that the vector of coordinates has been changed from (7, 13, 9)10 to (7, 14, 9)10. After
applying Eq. (6) to the vector (7, 14, 9)10, we can extract the
secret digit (s = 6).

4. Experimental results
For evaluating the embedding capacity as well as visual stegoimage quality, we implemented the most common LSB replacement, Zhan and Wangs method and the proposed method by using
C language on the Intel Core 2 Duo 5600 CPU computer with 1GB
RAM.

Table 1
The embedding rate (bpp) comparison between the EMD and the proposed methods
The number of cover pixels
for carrying one secret digit

EMD embedding
Proposed method with 2LSBs
Proposed method with 3LSBs

log2(5)/2
log2(15)/2
log2(63)/2

log2(7)/3
log2(63)/3
log2(511)/3

log2(9)4
log2(255)/4
log2(4095)/4

Table 2
Given a cover pixel-pair, the embedding secret bits per cover pixel (bpp) for
exploiting two and three LSB planes, respectively
The number of LSB planes used
to carry secret bits

2 LSB planes

3 LSB planes

LSB replacement
Proposed method (theoretical
results)
Proposed method (the average
secret bits actually
embedded in a cover pixel
according to the experimental
results)

2
log2(15)/2 = 1.953

3
log2(63)/2 = 2.9886

1.733

2.746

1674

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

Fig. 3. Results of stego-images with R = (log 15)/2.

60
50
40

Zheng et.al
R=(log5)/2

30

Proposed Scheme
R=(log15)/2
Proposed Scheme
R=(log31)/2

20
10

er
dh
Ba ill
rb
ar
a
Bo
at
Ze
ld
a
G

ol

6
F1

pp
Pe

ny

on

ffi

Ti

bo
Ba

Le

na

Fig. 4. The PSNR values of proposed methods and Zhang et al.s method by embedding secret data with 262,144 bits and full embedding.

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

embedding rate R is (log 15)/2, the PSNR value of stego-image is

about 44.3 dB in average; when the embedding rate R is raised to
(log 31)/2, the PSNR value will reduce to the average of 38 dB
(Fig. 4). Although the PSNR values of stego-image are lower than
those of the EMD embedding method, our proposed method has
higher embedding rate and more embedded bits as well. In other
words, our proposed method can tune down the pm value to improve the embedding rate. In the proposed method, a trade-off between the embedding capacity and the stego-image quality may be
realized by adjusting the pm value.

1675

The difference image histograms for the detection of LSB

replacement and proposed methods are shown in Fig. 5. Fig. 5(a)
is the image histogram for the original Lena image of size
512  512. Fig. 5(b) and (c) represent the stego-image histograms
produced using the LSB replacement and the proposed methods
by exploiting two LSB planes to carry secret data. Similarly, Fig.
5(d) and (e) represent the stego-image histograms produced using
the LSB replacement and the proposed methods by exploiting three
LSB planes to carry secret messages. In terms of the statistical analysis, Fig. 5(b) through (e) illustrate the comparable behaviors for

Fig. 5. The difference image histograms for the detection of LSB replacement and proposed methods.

1676

C.-F. Lee et al. / Image and Vision Computing 26 (2008) 16701676

both the LSB replacement and the proposed methods; their stegoimage histograms are different from the original one and can be
noticeable when the data hiding are performed on three LSB
planes. However, the LSB replacement is less secure, because it is
easy to sequentially collect secret message in terms of pm LSB bits,
thereby easily leaving recognizable ngerprints. In contrast, the
proposed method uses a random generator fr to project an mdimensional pseudo-random vector onto a virtual hyperspace
rstly and then deals with the virtual modications to hide secret
messages by exploiting the EMD embedding. The random generator fr as well as the value of m act as secure keys. Without knowing
the keys, the attacker has no awareness of the secret message.
Therefore, the proposed method can achieve some degree of
security.

5. Conclusions
The embedding capacity and the quality of stego-image are two
important issues for data hiding methods. Our proposed method
has the embedding rate R = (log2(2m + 1))/2 which is greater than
that R = (log2 (2n + 1))/n of the EMD embedding method proposed
by Zhang et al., when m > 2 and n=_ 2. The experimental results
show that the proposed method is able to embed more information

than the EMD embedding method can do and still keeps good stego-image quality at an acceptable level.
References
[1] Hide and Seek. Available from: <http://www.jjtc.com/Security/stegtools.htm>.
[2] S. Katzenbeisser, F.A.P. Petitcolas, Information Hiding Techniques for
Steganography and Digital Watermarking, Artech House, Boston, 2000.
[3] H. Wang, S. Wang, Cyber warfare: steganography vs. steganalysis,
Communications of the ACM 47 (10) (2004) 7682.
[4] S. Dumitrescu, X. Wu, Z. Wang, Detection of LSB steganography via sample pair
analysis, IEEE Transactions on Signal Processing 51 (7) (2003) 19952007.
[5] A. Ker, Improved detection of LSB steganography in grayscale images, in: J.
Fridrich (Ed.), Proceedings of the Sixth International Workshop on Information
Hiding, vol. 3200, Springer, Toronto, Canada, May 2004, pp. 97115.
[6] A. Ker, Quantitative evaluation of pairs and RS steganalysis, in: Edward J. Delp
III, Ping W. Wong (Eds.), Proceedings of SPIE Security, Steganography, and
Watermarking of Multimedia Contents VI, SPIE, vol. 5306, California, USA,
January 2004, pp. 8397.
[7] J. Mielikainen, LSB matching revisited, IEEE Signal Processing Letters 13 (5)
(2006) 285287.
[8] C.C. Thien, J.C. Lin, A simple and high-hiding capacity method for hiding digitby-digit data in images based on modulus function, Pattern Recognition 36
(12) (2003) 28752881.
[9] X. Zhang, S. Wang, Efcient steganographic embedding by exploiting
modication direction, IEEE Communications Letters 10 (113) (2006) 781783.
[10] C.K. Chan, L.M. Cheng, Hiding data in images by simple LSB substitution,
Pattern Recognition 37 (3) (2004) 469474.
[11] T. Zhang, X. Ping, A new approach to reliable detection of LSB steganography in
natural images, Signal Processing 83 (10) (2003) 20852093.