Pattern Recognition Letters 24 (2003) 1613–1626

www.elsevier.com/locate/patrec

A steganographic method for images

by pixel-value differencing
Da-Chun Wu a, Wen-Hsiang Tsai b,*

a
Department of Information Management, National Kaohsiung First University of Science and Technology,
Kaohsiung 811, Taiwan, ROC
b
Department of Computer and Information Science, National Chiao Tung University, 1001 Ta Hsueh Rd.,
Hsinchu 30050, Taiwan, ROC

Received 29 October 2001; received in revised form 11 December 2002

Abstract

A new and efficient steganographic method for embedding secret messages into a gray-valued cover image is pro-
posed. In the process of embedding a secret message, a cover image is partitioned into non-overlapping blocks of two
consecutive pixels. A difference value is calculated from the values of the two pixels in each block. All possible difference
values are classified into a number of ranges. The selection of the range intervals is based on the characteristics of
human visionÕs sensitivity to gray value variations from smoothness to contrast. The difference value then is replaced by
a new value to embed the value of a sub-stream of the secret message. The number of bits which can be embedded in a
pixel pair is decided by the width of the range that the difference value belongs to. The method is designed in such a way
that the modification is never out of the range interval. This method provides an easy way to produce a more im-
perceptible result than those yielded by simple least-significant-bit replacement methods. The embedded secret message
can be extracted from the resulting stego-image without referencing the original cover image. Moreover, a pseudo-
random mechanism may be used to achieve secrecy protection. Experimental results show the feasibility of the proposed
method. Dual statistics attacks were also conducted to collect related data to show the security of the method.
Ó 2002 Elsevier Science B.V. All rights reserved.

Keywords: Steganography; Data hiding; Cover image; Stego-image; Security

1. Introduction applications leads people into the digital world,

and communication via digital data becomes fre-
Text, image, audio, and video can be repre- quent. However, new issues also arose and have
sented as digital data. The explosion of Internet been explored (Artz, 2001; Zhao et al., 1998), such
as data security in digital communications, copy-
right protection of digitized properties, invisible
* communication via digital media, etc.
Corresponding author. Tel.: +886-35-720631/712121; fax:
+886-35-727382/721490. The term steganography is derived from the
E-mail addresses: [email protected] (D.-C. Wu), Greek language and means covert writing. It is the
[email protected] (W.-H. Tsai). technique of encoding secret information in a

0167-8655/03/$ - see front matter Ó 2002 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0167-8655(02)00402-6
1614 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

communication channel in such a manner that the digital images may be used as the carriers of secret
very existence of the information is concealed. messages in steganographic methods to deliver or
Computer-based image steganography (Anderson hide data like secret letters, military maps, favorite
and Petitcolas, 1998) is one way of data hiding pictures, etc. In such data embedding applications
which provides data security in digital images. It is with emphasis on cheating or hiding, attackers do
considered as a technique inspired from ancient not know that the stego-image has included secret
steganography. The aim is to embed and deliver messages, so they will not disturb the stego-image.
secret messages in digital images without any sus- And so non-robust techniques may also be used
piciousness. The secret message might be a cap- here. To explore a large amount of space for effi-
tion, a plain text, another image, a control signal, ciently embedding data into an image is an em-
or anything that can be represented in bit stream phasis of such steganographic studies.
form. The secret message may be compressed and Many steganographic techniques about em-
encrypted before the embedding steps begin. bedding data in images have been proposed. A
Another way of data hiding is image water- number of data embedding techniques are based
marking (Swanson et al., 1998; Hartung and on the method of replacing the least-significant-
Kutter, 1999; Voyatzis and Pitas, 1999; Wolfgang bits (LSBs) of the pixels of the cover image
et al., 1999; Cox et al., 2001) which is a novel way (Walton, 1995) and a pseudo-random number
for embedding watermark information in host generation mechanism is often used to accomplish
images. The main purposes of image watermark- the security work (Turner, 1989). In (Ohnishi and
ing include copyright protection (Podilchuk and Matsui, 1996), an LSB-like embedding technique
Zeng, 1998; Nikolaidis and Pitas, 1998; Koch and is used in a wavelet-based method by adding or
Zhao, 1995; Cox et al., 1997; Podilchuk and Delp, subtracting one unit from the transform coeffi-
2001) and authentication (Yeung and Mintzer, cients of the image.
1997; Fridrich, 1998). In such applications, it is Some other steganographic methods based on
required that the embedded information be unal- modifying small details in images are also pub-
tered, or altered only up to an acceptable degree of lished in the literature. A series of text marking
distortion, no matter how the watermarked image methods embedding data by slightly shifting the
is attacked. Many other new applications of wa- contents in an electronic document have been
termarking are also introduced (Cox et al., 2000), conducted by Maxemchuk (1994). The texture
such as broadcast monitoring, proof of owner- block coding method proposed in (Bender et al.,
ships, transactional watermarks, copy control, and 1996) copies a small block with random texture
covert communication. into a region with similar texture. The fractal-
Computer-based image steganography mainly based steganography method proposed in (Davern
considers the requirement that the steganographic and Scott, 1996) creates a new range block, which
result, the so-called stego-image, be undetectable, is visually like the original range block, by trans-
as pointed out in (Anderson and Petitcolas, 1998). forming the selected domain block. A method that
Such steganographic techniques may be used in uses color palettes for hiding data is proposed by
various applications. In the application of image Johnson and Jajodia (1998). Non-adaptive and
database retrieval, auxiliary information, like adaptive steganographic techniques for images in
captions, time stamps, news, etc., may be embed- palette format were reported by Fridrich and Rui
ded into images for convenience of simultaneous (2000). Some more literature about computer-
handling of the images and the embedded infor- based image steganography can also be found
mation. In such a kind of application, although the (Anderson and Petitcolas, 1998; Franz and Pfitz-
existence of the embedded data may be publicly mann, 2000; Petitcolas et al., 1999; Ramkumar
known, the systems are basically closed and there and Akansu, 2001). Some steganalytic methods for
is no worry about the possibility of being attacked detecting the presence of hidden information in
from outside worlds. Non-robust data embedding images, such as visual attacks and statistical at-
methods are appropriate here. On the other hand, tacks, have also been explored in (Fridrich and
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1615

Goljan, 2002; Westfeld and Pfitzmann, 2000). The an image can tolerate equal amounts of changes
dual statistics methods proposed by Fridrich et al. without causing notice to an observer. The largest
(2001) can be employed to detect the existence of number of LSBs whose gray values can be changed
LSB steganography in images. without producing a perceptible artifact in each
In this paper, we propose a new and efficient pixel is different. Changes of the gray values of
steganographic method to hide data in gray-valued pixels in smooth areas in images are more easily
images imperceptibly with no consideration of noticed by human eyes. In the embedding method
robustness. It was based on a simple visual effect of we propose, we simply divide the cover image into
the human visual perception capability. We use the a number of non-overlapping two-pixel blocks.
differences of the gray values in the two-pixel Each block is categorized according to the differ-
blocks of the cover image as features to cluster the ence of the gray values of the two pixels in the
blocks into a number of categories of smoothness block. A small difference value indicates that the
and contrast properties. Different amounts of data block is in a smooth area and a large one indicates
can be embedded in different categories according that it is in an edged area. The pixels in edged
to the degree of smoothness or contrast. This areas may, as mentioned previously, tolerate larger
method provides an easy way to produce more changes of pixel values than those in the smooth
imperceptible results than those yielded by simple areas. So, in the proposed method we embed more
LSB replacement methods. The method was de- data in edged areas than in the smooth areas. And
signed in such a way that there is no need of using it is in this way that we keep the changes in the
the original image in recovering the secret message resulting stego-image unnoticeable.
from the stego-image. Moreover, while hiding data A flowchart of the proposed embedding method
the cover image is traversed in an order provided is sketched in Fig. 1. The process of quantization
by a pseudo-random number generator to achieve of the differences of the gray values of two-pixel
secrecy, and so to prevent tampering access to the blocks and the process of data embedding are
embedded data from illicit users. described subsequently.
The remainder of this paper is organized as
follows. In Section 2, the proposed data embed- 2.1. Quantization of differences of gray values of
ding method is presented. The process for ex- two-pixel blocks
tracting the embedded data is described in Section
3. Several experimental results are illustrated in The cover images used in the proposed method
Section 4. Finally, concluding remarks as well as are 256 gray-valued ones. A difference value d is
some suggestions for future works are stated in computed from every non-overlapping block of
Section 5. In the Appendix A, the proof of an two consecutive pixels, say pi and piþ1 , of a given
equation used in the data embedding process is cover image. The way of partitioning the cover
included. image into two-pixel blocks runs through all the
rows of each image in a zigzag manner, as shown
in Fig. 2. Assume that the gray values of pi and piþ1
2. Proposed data embedding are gi and giþ1 , respectively, then d is computed as
giþ1  gi , which may be in the range from )255 to
Hiding data in the LSBs of the pixels of a gray- 255. A block with d close to 0 is considered to be
valued image is a common information hid- an extremely smooth block, whereas a block with
ing method that utilizes the characteristic of the d close to )255 or 255 is considered as a sharply
human visionÕs insensitivity to small changes in the edged block. By symmetry, we only consider the
image. This simple LSB embedding approach is possible absolute values of d (0 through 255) and
easy for computation, and a large amount of data classify them into a number of contiguous ranges,
can be embedded without great quality loss. The say Ri where i ¼ 1; 2; . . . ; n. These ranges are as-
more LSBs are used for embedding, the more signed indices 1 though n. The lower and upper
distorted result will be produced. Not all pixels in bound values of Ri are denoted by li and ui ,
1616 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

Start

Select a two-pixel block

from C. Assume gray Cover
value difference d of the image C
two pixels is in range k

To embed data No
in this block may cause gray
value fall off of range of 0
through 255
Take bit(s) from the bit
stream of secret data for
Secret data
embedding according to
the width of range k

Replace d with a new

No Yes value d' which is based on
bit values to be embedded
and the lower bound value
of range k

Perform inverse gray

values differencing of the
two pixels by using d'

All bits Yes

Stego-
of secret data were
image
embedded

Fig. 1. The data embedding process.

respectively, where l1 is 0 and un is 255. The width the difference values of smooth blocks are chosen
of Ri is ui  li þ 1. In our proposed method, the to be smaller while those which represent the dif-
width of each range is taken to be a power of 2. ference values of edged blocks are chosen to be
This restriction of widths facilitates embedding larger. That is, we create ranges with smaller
binary data. The selected range intervals are based widths when d is close to 0 and ones with larger
on the human visual capability mentioned previ- widths when d is far away from 0 for the purpose
ously. The widths of the ranges which represent of yielding better imperceptible results. A differ-
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1617

in this block, is calculated by n ¼ log2 ðuk  lk þ

1Þ. Since the width of each range is selected to be a
power of 2, the value of n ¼ log2 ðuk  lk þ 1Þ is an
integer. A sub-stream S with n bits is selected next
from the secret message for embedding in B. A new
difference d 0 then is computed by


0 lk þ b for d P 0;
d ¼ ð1Þ
ðlk þ bÞ for d < 0;

where b is the value of the sub-stream S. Because the

value b is in the range from 0 to uk  lk , the value of
d 0 is in the range from lk to uk . According to the
previous discussions, if we replace d with d 0 , the
resulting changes are presumably unnoticeable to
the observer. We then embed b by performing an
Fig. 2. The non-overlapping two-pixel blocks are constructed
inverse calculation from d 0 described next to yield
by grouping every two consecutive pixels in a cover image in a
zigzag scanning of the image rows. the new gray values ðgi0 ; giþ1
0
Þ for the pixels in the
corresponding two-pixel block ðpi0 ; piþ1 0
Þ of the
ence value which falls in a range with index k is stego-image. The embedding process is finished
said to have index k. All the values in a certain when all the bits of the secret message are embedded.
range (i.e., all the values with an identical index) The inverse calculation for computing ðgi0 ; giþ1
0
Þ
are considered as close enough. That is, if a dif- from the original gray values ðgi ; giþ1 Þ of the pixel
ference value in a range is replaced by another in pair is based on a function f ððgi ; giþ1 Þ; mÞ which is
the same range, the change presumably cannot be defined to be
easily noticed by human eyes. In the proposed f ððgi ; giþ1 Þ; mÞ ¼ ðgi0 ; giþ1
0
Þ
method, we embed some bits of the secret message 8
>
> ðgi  ceilingm ; giþ1 þ floorm Þ
into a two-pixel block by replacing the difference <
if d is an odd number;
value of the block with one with an identical index, ¼
>
> ðg i  floorm ; giþ1 þ ceilingm Þ
i.e., we change a difference value in one range into :
if d is an even number;
any of the difference values in the same range. In
other words, in the proposed data embedding ð2Þ
process, we adjust the gray values in each two- 0
where m ¼ d  d, ceilingm ¼ dm=2e, and floorm ¼
pixel pair by two new ones whose difference value bm=2c. The above equation satisfies the require-
causes changes unnoticeable to an observer of the ment that the difference between gi0 and giþ1 0
is d 0 . It
stego-image. More details are described next. is noted that a distortion reduction policy has been
employed in designing Eq. (2) for producing gi0 and
0
2.2. Data embedding giþ1 from gi and giþ1 so that the distortion caused
by changing gi and giþ1 is nearly equally distri-
We consider the secret message as a long bit buted over the two pixels in the block. The effect
stream. We want to embed every bit in the bit is that the resulting gray value change in the block
stream into the two-pixel blocks of the cover is less perceptible.
image. The number of bits which can be embedded In the above inverse calculation, a smaller value
in each block varies and is decided by the width of of d 0 produces a smaller range interval between gi0
0
the range to which the difference value of the two and giþ1 while a larger d 0 produces a larger interval.
pixels in the block belongs. Given a two-pixel So, ðgi ; giþ1 Þ may produce invalid ðgi0 ; giþ1 0
Þ, i.e.,
block B with index k and gray value difference d, some of the calculations may cause the resulting gi0
0
the number of bits, say n, which can be embedded or giþ1 to fall off the boundaries of the range ½0; 255.
1618 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

Although we may re-adjust the two new values into only examining the values of ð^ gi ; g^iþ1 Þ which are
the valid range of ½0; 255 by forcing a falling-off- produced by the case of using uk . If either g^iþ1 or g^i
boundary value to be one of the boundary values of falls off the boundary of 0 or 255, we regard the
0 and 255, and adjusting the other to a proper value block to have the possibility of falling-off, and
to satisfy the difference d 0 , yet this might produce abandon the block for embedding data.
abnormal spots in contrast with the surrounding In addition, the inverse calculation in Eq. (2) is
region in some cases. To solve this problem, we designed in such a way that it satisfies the fol-
employ a checking process to detect such falling- lowing property:
off-boundary cases, and abandon the pixel blocks
which yield such cases for data embedding. The f ððgi ; giþ1 Þ; mÞ ¼ f ðf ððgi ; giþ1 Þ; m0 Þ; m00 Þ
gray values of the abandoned blocks are left intact for m ¼ m0 þ m00 : ð3Þ
in the stego-image. This strategy helps us to dis-
tinguish easily blocks with embedded data from The proof can be found in Appendix A.
abandoned blocks in the process of recovering data This equation means that the inverse calcula-
from a stego-image, which will be discussed in the tion can proceed directly or progressively. This
next section. It is noted that such abandoned pixel property is useful for judging the existence of
blocks are very few in real applications according embedded data in each block in the data recover-
to our experiments. ing process.
The proposed falling-off-boundary checking An illustration of the data embedding process
proceeds by producing a pair of ð^ gi ; g^iþ1 Þ from the is shown in Fig. 3. In the figure, the gray values
inverse calculation of the value of the function of a sample two-pixel block are assumed to be
f ððgi ; giþ1 Þ; uk  dÞ. Since uk is the maximum value ð50; 65Þ. The difference value is 15, which is in the
in the range from lk to uk , the resulting pair of range of 8 through 23. The width of the range is
ð^
gi ; g^iþ1 Þ produced by the use of uk will yield the 16 ¼ 24 , which means that a difference value in the
maximum difference. That is, this maximum range range can be used to embed four bits of secret data.
interval g^iþ1  g^i covers all of the ranges yielded Assume that the four leading bits of the secret data
by the other ð^ gi ; g^iþ1 Þ pairs. So, the falling-off- are 1010. The value of this bit stream is 10. It is
boundary checking for the block can proceed by added to the lower bound value 8 of the range

New gray values

Gray values of a pixel pair after inverse differencing

50 65 48 66

- +
8 + New gray
18 value
difference
15 23
+
Gray value
difference 1 0 1 0 0 1 1 ...

Bit stream of secret message

Quantization ranges from 0

through 255
Fig. 3. An illustration of the data embedding process.
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1619

to yield the new difference value 18. Finally, by By Eq. (3), the above result can be transformed
Eq. (2) the values ð48; 66Þ are computed for use as further to be
the gray values in the stego-image. Note that 66
48 ¼ 18. f ððgi ; giþ1 Þ; d  d þ uk  d Þ
¼ f ðf ððgi ; giþ1 Þ; d  dÞ; uk  d Þ
3. Process of recovering embedded data from stego-
images ¼ f ððgi ; giþ1

Þ; uk  d Þ ¼ ð^
gi ; g^iþ1

Þ: ð5Þ

The process of extracting the embedded mes- This completes the proof.
sage proceeds by using the seed of the pseudo- The above property shows that the results of
random scheme to produce the same traversing both of the falling-off-boundary checking pro-
order for visiting the two-pixel blocks as in the cesses, one in data embedding and the other in
embedding process. Each time we visit a two-pixel data recovery, are identical. This in turn implies
block in the stego-image, we apply the same fall- that if either of the gray values of the computed
ing-off-boundary checking as mentioned previ- values ð^gi ; g^iþ1

Þ falls off the boundaries of the
ously to the block to find out whether the block range ½0; 255, it means that the current block was
was used or not in the embedding process. Assume not used for embedding data, or that the block was
that the block in the stego-image has the gray abandoned in the embedding process. On the
values ðgi ; giþ1

Þ, and that the difference d of the contrary, if both of the values ð^ gi ; g^iþ1

Þ do not fall
two gray values is with index k. We apply the off the range, it means that some data was em-
falling-off-boundary checking process to ðgi ; giþ1
Þ bedded in the block. The value b, which was em-

by using f ððgi ; giþ1 Þ; uk  d Þ. bedded in this two-pixel block, is then extracted
We now want to prove that the resulting out using the equation


gi ; g^iþ1
ð^
Þ computed from f ððgi ; giþ1
Þ; uk  d Þ are d  lk for d P 0;
identical to the gray values ð^ gi ; g^iþ1 Þ which were b¼ ð6Þ
d  lk for d < 0:

computed by f ððgi ; giþ1 Þ; uk  dÞ in the embedding
process. The proof is as follows. First, Note that in the recovery of the secret message
from the stego-image using the previously de-
ð^
gi ; g^iþ1 Þ ¼ f ððgi ; giþ1 Þ; uk  dÞ
scribed extraction process, there is no need of
¼ f ððgi ; giþ1 Þ; d  d þ uk  d Þ: ð4Þ referencing the cover image.

Fig. 4. The cover images ð512  512Þ (a) ‘‘Lena’’, and (b) ‘‘Baboon’’.
1620 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

4. Experimental results and discussions Table 1

Values of the capacities for embedding data by using the cover
image
4.1. Experimental results
Cover image Maximum capacity
In our experiments, four cover images ‘‘Lena’’, Embedding using Embedding using
‘‘Jet’’, ‘‘Peppers’’, and ‘‘Baboon’’ were used, each the range widths the range widths
of 8, 8, 16, 32, 64, of 2, 2, 4, 4, 4, 8, 8,
with size 512  512. Two of them are shown in and 128 16, 16, 32, 32, 64,
Fig. 4. Two sets of widths of ranges of gray value and 64
differences were used in the experiments. The first Lena 50960 25940
experiment was based on selecting the range Jet 51243 24177
widths of 8, 8, 16, 32, 64, and 128, which parti- Peppers 50685 27269
tion the total range of ½0; 255 into ½0; 7; ½8; 15; Baboon 56291 36061
½16; 31; . . . ; ½128; 255. The second experiment was
based on the use of the range widths of 2, 2, 4, 4, 4, capacities for embedding data by using the cover
8, 8, 16, 16, 32, 32, 64, and 64. The values of the image and the two sets of range widths are given in

Fig. 5. Two of the resulting images and their enhanced difference images after embedding a Word-format file consisting of the text of
this article by using a set of range widths of 8, 8, 16, 32, 64, and 128. (a) The stego-images, (b) the enhanced difference images between
the cover images and the stego-images.
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1621

Fig. 6. Two of the resulting images and their enhanced difference images after embedding a Word-format file consisting of the text of
this article by using a set of range widths of 2, 2, 4, 4, 4, 8, 8, 16, 16, 32, 32, 64, and 64. (a) The stego-images, (b) the enhanced difference
images between the cover images and the stego-images.

Table 1. We used a Word-formatted file which on the edges in the images. This means that such
consists of the text of this article as the secret distortions will be less noticeable because changes
message in the experiments. Some results of the in edge parts of images are generally less obvious
first experiment are shown Fig. 5. In Fig. 5(a), we to human eyes.
show the two of the stego-images resulting from In contrast, for the purpose of comparison, two
embedding the given secret data using the first set stego-images resulting from embedding random
of range widths, and in Fig. 5(b) we show the data into the three LSBs of the pixel values using
corresponding enhanced difference images between the conventional LSB-embedding steganographic
the stego-images of Fig. 5(a) and the cover images technique, and the corresponding enhanced differ-
of Fig. 4 (with the differences of gray values being ence images are shown in Fig. 7. It is seen from this
scaled five times). Similarly results of the second figure that distortions are spread all over the image
experiment are shown in Fig. 6. The difference which are more obvious to observers than distor-
images are shown here to indicate the distortions tions resulting from our method which are lim-
resulting from the data embedding process. From ited essentially at edge areas, as shown in Figs. 5
them, we see that most of the distortions are found and 6.
1622 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

Fig. 7. Two of the resulting images and their enhanced difference images after embedding random data in the three LSBs. (a) The
stego-images, (b) the enhanced difference images between the cover images and the stego-images.

All of the results were produced by embedding Table 2

the secret data in the two-pixel blocks of the cover Values of RMSEs and PSNRs of stego-images in which a file
consisting of the text of this article is embedded using two sets
image in a random traversing order generated by a of range widths in the embedding process
pseudo-random scheme, which walks through the
Cover Embedding using Embedding using
cover image and visits each two-pixel block only image the range widths of the range widths of
once. 8, 8, 16, 32, 64, and 2, 2, 4, 4, 4, 8, 8, 16,
Finally, the values of the peaks of the signal- 128 16, 32, 32, 64, and
to-noise (PSNR) and the root-mean-square error 64
(RMSE) of the embedding results are shown in RMSE PSNR RMSE PSNR
Table 2. The quality is still good even in a stego- Lena 2.07 41.79 0.97 48.43
image with a lower PSNR. Jet 2.28 40.97 1.09 45.67
Peppers 2.09 41.73 1.20 47.19
Baboon 3.25 37.90 1.59 44.10
4.2. Discussions

The proposed method may be used to embed It does not replace the LSBs of pixel values di-
variable numbers of bits into blocks of two pixels. rectly; instead, it changes the differences of the two
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1623

pixel values in a block. We cannot find obvious Such diagrams are referred to as RS-diagrams.
suspicious artifacts on the resulting images or bit- According to Fridrich et al. (2001), if more and
planes by simple visual inspection. more LSBs are replaced with random data, then
To check whether the function of the proposed the percentages Rm and Sm of the two pixel groups
embedding method can be detected with some (regular and singular, respectively) in the diagram
newly announced related statistical steganalytic will become equal gradually when the mask of m is
techniques, we tested the stego-images yielded by adopted in the statistics analysis process, or the
our method with the dual statistics method pro- percentages Rm and Sm will become more and
posed by Fridrich et al. (2001) as a demonstration. more unequal when the mask of m is adopted.
By the method, an image is divided into disjoint From the RS-diagram of Fig. 8(a), we can see that
pixel groups. The regularity of each group is the function of conventional LSB-embedding ste-
computed by a discrimination function. By com- ganographic techniques in images indeed can be
bining the function and an invertible operation, detected because when the percentage of pixels
three types of pixel groups are defined: regular, embedded with data into their LSBs approaches
singular, and unusable. Two complemental masks 100%, the percentages of the regular and singular
are used for simulating the act of different noise pixel groups will become more and more equal or
adding. Some test results using this method were unequal. But the results of our method as shown by
transformed into the two diagrams shown in Fig. 8. the RS-diagram of Fig. 8(b) indicate that the stego-
In the diagrams, the x-axes depict the percentage of images seemingly do not contain any embedded
image pixels into which message data are embed- data in their LSBs, because the expected value of
ded, and the y-axes depict the relative numbers (in Rm is seen close to that of Rm and so are the case of
percentage) of regular and singular pixel groups Sm and Sm , i.e., Rm ffi Rm and Sm ffi Sm . This
with masks m ¼ ½0 1 1 0 and m ¼ ½0  1  1 0. proves that our steganographic method is secure
from the viewpoint of the dual statistics method.

60%
50% 5. Conclusions and suggestions
Rm
40%
Sm A new and efficient computer-based stegano-
30%
R-m graphic method for embedding secret messages
20% into images without producing noticeable changes
10% has been proposed. There is no need of referencing
0% the original image when extracting the embedded
(a) 0% 20% 40% 60% 80% 100% data from a stego-image. The method utilizes the
characteristic of the human visionÕs sensitivity to
60% gray value variations. Secret data are embedded
50% into a cover image by replacing the difference
Rm values of the two-pixel blocks of the cover image
40%
Sm with similar ones in which bits of embedded data
30%
R-m are included. The method not only provides a
20%
better way for embedding large amounts of data
10% into cover images with imperception, but also of-
0% fers an easy way to accomplish secrecy. This em-
(b) 0% 20% 40% 60% 80% 100% bedding method can be easily extended to
efficiently carry content-related messages such as
Fig. 8. RS-diagrams yielded by the dual statistics method by
Fridrich et al. (2001) for stego-images produced by conven- captions or annotations in audios and videos by
tional LSB-embedding steganographic technique and our embedding data in each adjacent pair of signals of
method. the data-streams.
1624 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

Acknowledgements II(I) m is even, m0 is odd, and m00 is odd;

I(II) m is even, m0 is even, and m00 is even;
This work is supported partially by National (III) m is odd, m0 is odd, and m00 is even;
Science Council under Project NSC 89-2218-E- (IV) m is odd, m0 is even, and m00 is odd.
130-002 and partially by the MOE Program for
Promoting Academic Excellence of Universities The proof of Case (I) is conducted in the fol-
under the grant number 89-1-FA04-1-4. lowing by considering two possible situations,
namely, when the value giþ1  gi is even and when
Appendix A. Proof of Eq. (3) it is odd.
Firstly, if giþ1  gi is even, then by Eq. (A.2) we
Another way to represent Eq. (2) is as follows: have

8 
>
> mþ1 m1
>
> gi  ; giþ1 þ when giþ1  gi is odd and m is odd; ðA:1Þ
>
>
2 2

< m1 mþ1
ðgi0 ; giþ1
0
Þ¼ gi  ; giþ1 þ when giþ1  gi is even and m is odd; ðA:2Þ
> 2 2
>
> m
>
> m
gi  ; giþ1 þ ðA:3Þ
>
: 2 2
when m is even;

 
where m ¼ d 0  d represents the total changes of 0 m0  1 m0 þ 1
f ððgi ; giþ1 Þ; m Þ ¼ gi  ; giþ1  :
the gray values of gi and giþ1 to produce gi0 and 2 2
0 It is easy to see that the difference between
giþ1 . It is easy to verify that the range interval of
the resulting pair of ðgi0 ; giþ1
0
Þ produced by a posi- gi  ðm0  1=2Þ and giþ1 þ ðm0 þ 1=2Þ is odd.
tive m will cover that of ðgi ; giþ1 Þ, i.e., the produced
Therefore,

  
0 00 m0  1 m0 þ 1 00
f ðf ððgi ; giþ1 Þ; m Þ; m Þ ¼ f gi  ; giþ1 þ ;m by Eq: ðA:2Þ
2 2
 
m0  1 m00 þ 1 m0 þ 1 m00  1
¼ gi   ; giþ1 þ þ by Eq: ðA:1Þ
2 2 2 2
 
ðm0 þ m00 Þ ðm0 þ m00 Þ
¼ gi  ; giþ1 þ
2 2
¼ f ððgi ; giþ1 Þ; m0 þ m00 Þ by Eq: ðA:3Þ
¼ f ððgi ; giþ1 Þ; mÞ

range will be enlarged. On the contrary, a negative which is just the desired Eq. (3).
m will produce a reduced range. In the following, Secondly, if giþ1  gi is odd, then by Eq. (A.1)
we do not use Eq. (2) but use Eqs. (A.1)–(A.3) we have
instead.  
0 m0 þ 1 m0  1
The proof of the correctness of Eq. (3) proceeds f ððgi ; giþ1 Þ; m Þ ¼ gi  ; giþ1 þ :
2 2
by considering different combinations of m, m0 ,
and m00 which meet the condition of m ¼ m0 þ m00 . It is easy to see that the difference between
Possible combinations include: gi  ðm0 þ 1Þ=2 and giþ1 þ ðm0  1Þ=2 is even.
 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626 1625

Therefore,

  
0 00 m0 þ 1 m0  1 00
f ðf ððgi ; giþ1 Þ; m Þ; m Þ ¼ f gi  ; giþ1 þ ;m by Eq: ðA:1Þ
2 2
 
m0 þ 1 m00  1 m0  1 m00 þ 1
¼ gi   ; giþ1 þ þ by Eq: ðA:2Þ
2 2 2 2
 
ðm0 þ m00 Þ ðm0 þ m00 Þ
¼ gi  ; giþ1 þ
2 2
¼ f ððgi ; giþ1 Þ; m0 þ m00 Þ by Eq: ðA:3Þ
¼ f ððgi ; giþ1 Þ; mÞ

which is just the desired Eq. (3). Fridrich, J., 1998. Image watermarking for tamper detection.
The proofs of the other cases can proceed In: Proc. IEEE Internat. Conf. on Image Processing, Vol. II,
pp. 404–408.
similarly, and are omitted. Fridrich, J., Rui, D., 2000. Secure steganographic methods for
palette images. In: Proc. Third Internat. Workshop Infor-
mation Hiding. In: Lecture Notes in Computer Science, Vol.
References 1768. Springer-Verlag, Berlin, pp. 61–76.
Hartung, F., Kutter, M., 1999. Multimedia watermarking
Anderson, R.J., Petitcolas, F.A.P., 1998. On the limits of techniques. Proc. IEEE 87, pp. 1079–1107.
steganography. IEEE J. Selected Areas Commun. 16, 474– Johnson, N.F., Jajodia, S., 1998. Exploring steganography:
481. Seeing the unseen. IEEE Comput. (February), 26–64.
Artz, D., 2001. Digital steganography: Hiding data within data. Koch, E., Zhao, J., 1995. Towards robust and hidden image
IEEE Internet Comput. 5 (May/June), 75–80. copyright labeling. In: Proc. IEEE Nonlinear Signal and Image
Bender, W., Gruhl, D., Morimoto, N., Lu, A., 1996. Tech- Processing Workshop, Thessaloniki, Greece, pp. 452–455.
niques for data hiding. IBM Syst. J. 35 (3/4), 313–336. Maxemchuk, N.F., 1994. Electronic document distribution. AT
Cox, I.J., Miller, M.L., Boom, J.A., 2000. Watermarking & T Tech. J. 73, 73–80.
applications and their properties. In: Proc. Internat. Conf. Nikolaidis, N., Pitas, I., 1998. Robust image watermarking in
on Information Technology: Coding and Computing, pp. 6– the spatial domain. Signal Process. 66, 385–403.
10. Ohnishi, J., Matsui, K., 1996. Embedding a seal into a picture
Cox, I.J., Bloom, J., Miller, M., 2001. Digital Watermarking: under orthogonal wavelet transform. In: Proc. Multime-
Principles & Practice. Morgan Kaufmann Publishers, San diaÕ96. IEEE Press, Piscataway, NJ, pp. 514–521.
Francisco. Petitcolas, F.A.P., Anderson, R.J., Kuhn, M.G., 1999. Infor-
Cox, I.J., Kilian, J., Leighton, F.T., Shamoon, T., 1997. Secure mation hiding––a survey. Proc. IEEE 87, pp. 1062–1078.
spread spectrum watermarking for multimedia. IEEE Podilchuk, C.I., Delp, E.J., 2001. Digital watermarking: Algo-
Trans. Image Process. 6, 1673–1687. rithms and applications. IEEE Signal Process. Mag. 18 (4),
Davern, P., Scott, M., 1996. Fractal based image steganogra- 33–46.
phy. In: Proc. First Internat. Workshop Information Podilchuk, C.I., Zeng, W., 1998. Image-adaptive watermarking
Hiding. In: Lecture Notes in Computer Science, Vol. 1174. using visual models. IEEE J. Selected Areas Commun. 16,
Springer-Verlag, Berlin, pp. 279–294. 525–539.
Franz, E., Pfitzmann, A., 2000. Steganography secure against Ramkumar, M., Akansu, A.N., 2001. Capacity estimates for
cover-stego-attacks. In: Proc. Third Internat. Workshop data hiding in compressed images. IEEE Trans. Image
Information Hiding. In: Lecture Notes in Computer Process. 10, 1252–1263.
Science, Vol. 1768. Springer-Verlag, Berlin, pp. 29–46. Swanson, M.D., Kobayashi, M., Tewfik, A.H., 1998. Multi-
Fridrich, J., Goljan, M., 2002. Practical steganalysis of digital media data-embedding and watermarking technologies.
images-state of the art. In: Proc. SPIE Photonics West, Vol. Proc. IEEE 86, 1064–1087.
4675, Conference on Security and Watermarking of Mul- Turner, L.F., 1989. Digital data security system. Patent IPN,
timedia Contents, pp. 1–13. WO 89/08915.
Fridrich, J., Goljan, M., Du, R., 2001. Reliable detection Voyatzis, G., Pitas, I., 1999. The use of watermarks in the protec-
of LSB steganography in grayscale and color images. In: tion of digital multimedia products. Proc. IEEE 87, 1197–1207.
Proc. ACM Workshop on Multimedia and Security, pp. 27– Walton, S., 1995. Image authentication for a slippery new age.
30. Dr. DobbÕs J.: Software Tools Profess. Program. 20, 18–26.
1626 D.-C. Wu, W.-H. Tsai / Pattern Recognition Letters 24 (2003) 1613–1626

Westfeld, A., Pfitzmann, A., 2000. Attacks on steganography Yeung, M.M., Mintzer, F., 1997. An invisible watermark-
systems. In: Proc. Third Internat. Workshop Information ing technique for image verification. In: Proc. IEEE
Hiding. In: Lecture Notes in Computer Science, Vol. 1768. Internat. Conf. on Image Processing, Vol. II, pp. 680–
Springer-Verlag, Berlin, pp. 61–75. 683.
Wolfgang, R.B., Podilchuk, C.I., Delp, E.J., 1999. Perceptual Zhao, J., Koch, E., Luo, C., 1998. Digital watermarking
watermarks for digital images and video. Proc. IEEE 87, in business today and tomorrow. Commun. ACM 41, 67–
1108–1126. 74.