MINIMIZING DISTORTION IN STEGANOG-RAPHY BASED ON IMAGE FEATURE

International Journal of Computer Science & Information Technology (IJCSIT) Vol 11, No 1, February 2019
DOI: 10.5121/ijcsit.2019.11106 77
MINIMIZING DISTORTION IN STEGANOG-RAPHY
BASED ON IMAGE FEATURE
Dong Wu
Department of Information Engineering, Lingnan Normal University,
ZhanJiang, China
ABSTRACT
There are two defects in WOW. One is image feature is not considered when hiding information through
minimal distortion path and it leads to high total distortion. Another is total distortion grows too rapidly
with hidden capacity increasing and it leads to poor anti-detection when hidden capacity is large. To solve
these two problems, a new algorithm named MDIS was proposed. MDIS is also based on the minimizing
additive distortion framework of STC and has the same distortion function with WOW. The feature that
there are a large number of pixels, having the same value with one of their eight neighbour pixels and the
mechanism of secret sharing are used in MDIS, which can reduce the total distortion, improve the anti-
detection and increase the value of PNSR. Experimental results showed that MDIS has better invisibility,
smaller distortion and stronger anti-detection than WOW.
KEYWORDS
Information Hiding; Minimal Distortion; Secret Sharing; Eight Neighbour Pixels; PSNR; Anti-Detection
1. INTRODUCTION
With the rapid development of information technology, people enjoy the convenience of the
information society but also suffer the threats of information disclosure and attacks. How to
ensure the security of information has become the focus of information research fields [1]. In
addition to the traditional information encryption technology, more and more scholars began to
focus on the research of steganography. In recent years, image-based steganography has
developed rapidly. In general it can be divided into two types of methods in which one is based
on spatial domain and another based on frequency domain. The methods based on spatial domain
have received much attention because of its large hidden capacity and simple implementation.
The more classic spatial information hiding algorithms are LSB (Least Significant Bit) and
Patchwork [2] early. These two algorithms are simple to implement, but have small hidden
capacity and poor anti-detection. In 2011, Gul et al. proposed an algorithm named HUGO(Highly
Undetectable steGO), which has larger capacity and higher anti-detection compared to LSB and
Patchwork. At the same year, Filler et al. proposed a complete practical methodology for
minimizing additive distortion in steganography with general (non-binary) embedding operation
using STC (Syndrome-Trellis Codes) [3]. Based on this framework, most of exist algorithms
including HUGO can be implemented by designing different distorting functions. In 2012, in
their later work, Holub and Fridrich proposed a new spatial algorithm named WOW which has
also large capacity and high anti-detection based on the former framework.
Experimental results in [4] show that the overall performance of the WOW is superior to HUGO,
LSB and Patchwork. After deep research, we found there are two defects in WOW. The total
distortion grows rapidly with the increase of hidden capacity, and the anti-detection of the

78
algorithm also decreases drastically. In this paper, we first introduce the basic idea and the defects
of WOW, then proposed a new information hiding algorithm named MDIS. MDIS and WOW use
the same distortion function. However, in the process of embedding the secret information using
the minimal distortion path, the information hiding rule is redesigned based on the facts that many
pixels always have the same value with one of its eight neighbour pixels. At the same time, the
mechanism of secret sharing [5] is also used in MDIS. There are two advantages in MDIS
compared to WOW. MDIS get smaller total distortion and higher anti-detection after using
feature mining of over image. It is also ensured that the anti-detection does not decrease
drastically when the hiding capacity increases after using the mechanism of secret sharing.
Experimental results show that MDIS is superior to WOW in terms of PNSR, total distortion and
anti-detection.
2. BASIC IDEA AND DEFECTS OF WOW
2.1. BASIC IDEA OF WOW
In 2012, WOW was proposed by Holub and Fridrich based on the STC framework. For a given
cover image(X) and secrete information (M), the stego image(Y) is got through the following
steps:
(1)Calculating the hiding capacity according to the cover image and secrete information. If
secrete information is m bits, the number of rows and columns is n1 and n2 in cover image, then
the hiding capacity is as follows:
Payl oad m n n1 2
/= × （1）
(2)Calculating the minimal distortion path through distortion function according to STC
framework. The distortion function is as follows:
1 2
1 1
( , ) ( , )
n n
i j i j i j i j
i j
D X Y X Y X Yρ
= =
= − （2）
In formula (2), Xi,j and Yi,j respectively represent the value of pixel(i,j) in cover image and stego
image. ρij is the distortion parameter when Xij becomes Yij . In WOW, WDFB-D Filter is used to
calculate the value of ρij . The specific calculation method is shown in formulas (3) and (4).
( ) ( ) ( ) ( )
[ , ]* *k k k k k k
ij i jR R R R Kε = − ≈ （3）
1
( ) ( )
1
( )ρ ε
−
=
= 
pF
p k p
ij ij
k
（4）
In formula (3), for given filter { }(1 ) ( )
, ...,= F
nB K K , the value in the kth direction is
( ) ( )
*k k
R K X= . If the value of pixel (i,j) changes, then the new value in the kth direction is
[ , ]
k
i j
R
. In formula (4), the value of P is generally -1 and F represents the filtering direction.
(3)Information hiding is performed through change the value (plus or minus one) of pixels in
minimal distortion path.

79
2.2. DEFECTS OF WOW
WOW is best in existing spatial algorithms. However, we found that there are two defects in
WOW. One is that only the change of the value of pixels is considered in total distortion, but the
feature of cover image is not utilized to server for hiding information. Another is that the total
distortion grows obviously with the increase of hidden capacity, and the anti-detection of the
algorithm also decreases drastically. Here we analyze these two defects in detail.
(1)Analyzing the feature of cover image
Spatial redundancy is ubiquitous in image data and it is usually caused by the spatial coherence
between the colors of sampling points in a same scene surface. Especially the value of some pixel
is always the same with one of its eight neighbor pixels. In experiment 1, we selected randomly
200 images with different sizes and types, and then respectively counted the number of pixels,
having the same value with one of their eight neighbor pixels. The way of counting is shown in
formula (5) and the results are seen in table 1.
1 2
1 1
( X( i , j ) ( i , j ) ) 1
n n
i j
i f X Num Num
= =
′ ′== = +  （5）
In formula (5), n1 and n2 is the number of rows and columns. X(i’,j’) is one of the eight neighbor
pixels which has the same value with pixel X(i,j).
In order to compare and analyze intuitively, the results of multiple images are averaged. The
specific statistical results are shown in Table 1:
Table 1. The number of pixels and percentage who have the same value with one of
their eight neighbor pixels
Image type Image size Number Percentage(Number/n1*n2)
landscape 600*800 226704 47.23%
people 680*1024 478094 68.66%
plants 780*1024 532946 66.73%
fruits 640*780 299970 60.09%
As can be seen from Table 1, there are nearly 50% pixels in the all randomly selected types of
images that have the same values with one of their eight neighbor pixels. If we can utilize this
feature to hide information, the change to cover image will be reduced. Then the total distortion
also can be reduced and the anti-detection ability of the algorithm can be improved too.
To illustrate this feature of image can be applied to the algorithm, we designed the experiment 2.
We selected randomly 50 images as cover images and use them to hide information with WOW.
Then we counted the number of pixels whose value changed in minimal distortion path and
among these pixels also counted the number of pixels who have the same values with one of their
eight neighbor pixels. The statistical results are shown in Table 2.

80
Table 2. The number of pixels whose values changed in minimal distortion path
Hiding Capacity Nchanged Nsame Percentage(Nsame/ Nchanged)
0.1 9040 1179 13.04%
0.2 20283 3074 15.16%
0.3 32970 5581 16.93%
0.4 46888 8666 18.48%
0.5 62107 12396 19.96%
0.6 78546 16829 21.43%
0.7 96299 22145 23.00%
0.8 114750 28375 24.73%
We can know that there are nearly 20% pixels whose values are same with one of their eight
neighbor pixels in the minimal distortion path in WOW from table 2. And with the hidden
capacity increasing, this image feature becomes more obvious. We can use this feature for
information hiding algorithm. For each pixel in minimal distortion path has this feature, it
represents a bit of hide information. In other word, without changing any value of pixels,
embedding hiding information is done. Then the total distortion will be reduced and anti-
detection will be increased.
(2)The total distortion growing rapidly with the increasing of hidden capacity
In wow, with the increasing of hiding capacity, the total distortion grows rapidly and caused a
decrease in anti-detection property. We have done experiments to compare the total distortion
with different hidden capacities and the results are shown in figure 1 and figure 2. Although the
total distortion of WOW is smallest compared to other algorithms (seen from figure 1), the total
distortion in WOW grows very fast with the increasing of hidden capacity (seen from figure 2).
Especially when the hiding capacity is greater than 0.5, the value of total distortion almost
becomes twice for hidden capacity increased by 0.1. We introduce the mechanism of secret
information sharing in MDIS to solve this problem.
Figure 1. Total distortion of all algorithms Figure 2. Total distortion of WOW
3. THE IDEA OF MDIS
We proposed a new algorithm named MDIS based on the above analysis of two defects in WOW.
In MDIS, we use the same distortion function with WOW, but we redesign the hiding rule
through the minimal distortion path. We combine the feature that there are lots of pixels whose

81
values are same with one of their eight neighbor pixels and hiding information together. In
addition we also join the mechanism of secret information sharing in MDIS. The basic idea of
MDIS is as follows:
(1) Combining Hiding Information and the Feature of Image Together
In MDIS, the minimal distortion path is found according to distortion function and hiding
capacity. For each pixel in the minimal distortion path, we first check if there exists one of eight
neighbor pixels whose value is same with this pixel. If the condition is true, it means there is a bit
(0 or 1) of hidden information and we need not do anything to hide information. If it is false, then
we select its adjacent pixel having the smallest distance and not in the main direction [6]
of the
image to do the operation (plus 1 or minus 1) to hide information.
We do this for the following reasons. First we avoid changing in the main direction to reduce the
impact of modification in cover image. Second, we choose to modify the adjacent pixels is to not
follow the traditional thinking in order to improve the security and reduce the probability of being
detected. Experimental results show that this hiding strategy provides MDIS better anti-detection
than WOW.
In the minimal distortion path, the algorithm for locating the position of the pixel where the secret
information is embedded is as follows:
For each pixel P(x,y) in the minimal distortion path
iSame = 0; //by default, no neighbor pixels whose value is same with this pixel
iDiff = 255; //by default, the distance is 255
changePos = (x,y); //By default, the pixel where the secret information is embedded
Foreach P(x1,y1) in sets of eight neighbor pixels
If (x1,y1) is not valid continue;
If P(x1,y1)== P(x,y) iSame=1；break;
iTempDiff = abs(P(x,y), P(x1,y1); //get the distance of these two pixels
if (iTempDiff < iDiff) && isNotMainDirection(x,y,x1,y1)
iDiff = iTempDiff;
changePos = (x1,y1); //the inner loop ends here
if (0 ==iSame)
p(changePos) = p(changePos) + 1;
or p(changePos) = p(changePos) - 1； //hiding information
(2) With the mechanism of secret information sharing
In MDIS, there is a parameter named MaxPayload. When the hidden capacity exceeds
MaxPayload, the secret information is divided into multiple parts using modulo operation. Each
part will be hidden in a same cover image. Of course, the hiding capacity in each cover image is
not exceeds MaxPayload. The design of dividing information includes three steps. Step one is
getting the value of N (round up) and N is equal to hidden capacity/MaxPayload. Step two is
choosing a random number M (M>=N) between 1to 9. Step three is using M, N and P (the
location of every bit of hide information) to divided hidden information into N parts according to
the value of P%M. Details are shown in figure 3:

82
Figure 3. N=4，M=9；the first part information(P%M =0,4,8)，the second part
information(P%M =1,5) ，the third part information(P%M =2,6)，the forth part
information(P%M =3,7)
After using secret sharing, there will be multiple stego images. So receiver must get complete
hidden information through the inverse of modulo operation from these stego images.
4. EXPERIMENTAL RESULTS AND ANALYSIS
In order to compare the performance of two hidden algorithms, MDIS and WOW, we focus on the
PNSR, total distortion and anti-detection in different hidden capacity.
4.1. COMPARING AND ANALYZING THE PNSR IN DIFFERENT HIDDEN CAPACITY
The hidden capacity of MDIS, WOW and HUTO is 0-0.8, but the hiding capacity of LSB is 0-
0.125. Here we choose PSNR to compare and analyze. PSNR is an important technical indicator
which can be used to determine whether the visual effect is good. PSNR also is an effective
parameter standard to judge the invisibility of the algorithm. The larger the value of PSNR, the
better the hiding effect is. We choose four standard images as cover images to do the test, four
standard test images are shown in figure 4.
（a）Lena (b) Barbara (c) Cameraman (d) baboon
Figure 4. Test images
We calculate the value of PSNR in different hidden capacity. The experimental results are shown
in table 3 and figure 6. From the results used Lena as cover image in table 3, we know that the
value of PSNR in MDIS is highest in the same hiding capacity and followed by WOW, HUGO
and LSB.
Table 3. The value of PSNR in different hidden capacity using Lena as cover image
Hiding Capacity PSNR（MDIS） PSNR(WOW) PSNR(HUGO) PSNR(LSB)
0.05
73.2421 72.5481 71.7548
55.1286
0.125
68.9979 67.9929 67.3914
51.1288
0.4
63.4265 62.0965 61.6407
——
0.8 60.1731 58.463 58.1732 ——

83
There is a graphical representation of average PSNR of four standard images in different hidden
capacity and different hiding algorithms in figure 5. We know that the trend of PSNR is consistent
using all algorithms. The larger the hidden capacity, the smaller the value of PSNR is. But under
the same cover image and the same hidden capacity, the performance of MDIS is best. That
means the invisibility of MDIS is best.
Figure 5. The analysis of PSNR in MDIS, WOW and HUGO
4.2. THE ANALYSIS OF TOTAL DISTORTION IN STEGO IMAGE
Because there is a same distortion function in WOW and MDIS, so it is very meaningful to
compare the total distortion in the same hidden capacity. In the same hidden capacity, the one
which has larger distortion has worse anti-detection. Using the tool of MATLAB and choosing
randomly 50 images, we design two experiments. The first experiment is comparing and
analyzing the total distortion without secret sharing. The second experiment is comparing and
analyzing the total distortion with secret sharing. In the second experiment, there are multiply
stego images, so we get the average total distortion. For MDIS, if it is not using secret sharing, we
named it as MDIS-NS, else named it as MDIS. And we set MaxPayLoad = 0.3. The experimental
results are shown in figure 6.
Figure 6. The total distortion of WOW, MDIS-NS and MDIS

84
From figure 6, although there is a same distortion function, even not using secret sharing, the total
distortion in MDIS is smaller than in WOW in a same hidden capacity. As the hidden capacity
increases, the gap between the two algorithms becomes more and more obvious. When the hidden
capacity is greater than 0.5, the total distortion in WOW is almost twice as in MDIS. That means
our operation for hiding information play a great role and the effect is very obvious. After using
secret sharing, the average total distortion in multiply stego images in MDIS is much smaller than
in WOW, and the effect is very prominent. Obviously, MDIS has much better performance in
total distortion than WOW.
4.3. COMPARING AND ANALYZING ANTI-DETECTION
In our experiments, we choose two features: SPAM and SRM to analyze the anti-detection of
algorithms [7, 8]
. We design two experiments to test the anti-detection in different hidden capacity
of three algorithms (MDIS-NS, MIDS and WOW). We set MaxPayload = 0.3 and use EOOB [9]
as
the measurement of anti-detection. The higher the value of EOOB, the higher the probability of
detection error, the better anti-detection performance of the algorithm has.
Figure 7. The anti-detection based on SPAM Figure 8. The anti-detection based on SRM
From figure 7 and figure 8, we can find that the trend of EOOB is very similar based on SPAM
and SRM. With the increasing of hidden capacity, the EOOB in MDIS-NS and WOW has been
falling too. But there is higher EOOB in MDIS-NS. Another point, the EOOB in MDIS is
basically kept parallel after the hidden capacity exceeds 0.3. That means the anti-detection does
not decrease with the increase of the hidden capacity. This is a very big improvement compared to
WOW. So, we can conclude that the performance of MDIS-NS and MDIS are better than WOW.
5. CONCLUSIONS
We proposed a new information hiding algorithm name MDIS, combining minimal distortion and
image feature together, based on two defects of WOW. At the same time, we joint a secret
information sharing mechanism in MDIS. The experimental results show that MDIS has better
performance than WOW. MDIS has higher PSNR, smaller total distortion and higher EOOB than
WOW. That means MDIS has better invisibility, smaller distortion and stronger anti-detection
than WOW. There is still room for improvement in MDIS. The cover images must be used when
the secret information is restored.

85
ACKNOWLEDGEMENTS
The authors are grateful to the anonymous reviewers and the helpful suggestion given by the
partners. The research was supported by the Natural Science Foundation of Lingnan Normal
University (no.QL1307).
REFERENCES
[1] Slaughter & Jason & Rahman & Syed Shawon M, (2011) “Information Security Plan for Flight
Simulator Applications”, International Journal of Computer Science and Information Technology, Vol.
3, No. 3, pp1-15.
[2] Bender W & Gruhi D & Morimoto N, (1996) “Techniques for Data Hiding”, IBM System Journal,
Vol. 35, No. 3, pp313-335.
[3] T. Filler & J. Judas & J. Fridrich. (2011) “Minimizing additive distortion in steganography using
syndrome-trellis codes”. IEEE TIFS, Vol. 6, No. 3, pp920-935.
[4] Li-qiong Lu & WuDong, (2015) “Research of the Digital Image Steganography based on spatial
domain”, Joural of lingnan normal university, Vol. 1, No. 3, pp105-111.
[5] ChenGouxi & ShenHonglei & Wuyuliang & ChenJunjie, (2012) “Research on Sharing and
Steganographic Algorithm for Batch Cover Image”, Computer Engineering, Vol. 1, No. 4, pp116-118.
[6] B.M. Mehtre, (1993) “Fingerprint image analysis for automatic identification”, Machine Vision and
Application, Vol. 1, No. 6, pp124-139.
[7] T.Pevny & P.Bas & J.Fridrich, (2010) “Steganalysis by Subtractive Pixel Adjacency Matrix”, IEEE
Trans. on Info. Forensics and Security, Vol. 5, No. 2, pp215-224.
[8] C.Chen & Y.Q.Shi, (2008) “JPEG image steganalysis utilizing both intrablock and interblock
correlations”, International Symposium on Circuits and Systems, Vol. 7, No. 1, pp3029-3032.
[9] Jan Kodovský & Jessica Fridrich, (2010) “Ensemble Classifiers for Steganalysis of Digital Media”,
Forensics and Security, Vol. 7, No. 2, pp432-444.
AUTHORS
Dong Wu is currently a teacher in the Department of Information Engineering,
Lingnan Normal University, ZhanJiang, China. His research interests include image
processing, information hiding, and data mining.

MINIMIZING DISTORTION IN STEGANOG-RAPHY BASED ON IMAGE FEATURE

More Related Content

What's hot (13)

Similar to MINIMIZING DISTORTION IN STEGANOG-RAPHY BASED ON IMAGE FEATURE (20)

Recently uploaded (20)

MINIMIZING DISTORTION IN STEGANOG-RAPHY BASED ON IMAGE FEATURE