International Journal on Information Technologies & Security, № 1 (vol.

11), 2019 107

DIGITAL WATERMARKING SCHEME FOR

COPYRIGHT PROTECTION AND TAMPERING
DETECTION
Ching-Sheng Hsu, Shu-Fen Tu

Department of Information Management

Ming Chuan University, Chinese Culture University
e-mails: [email protected], [email protected]
Taiwan

Abstract: Digital watermarking scheme is a common way for protecting

copyright or verifying integrity for digital images. The working domain of
the watermark is either spatial or frequency. Recently, some researches
tried to work on another domain which is derived from matrix factorization.
Depending on the purpose, either robust or fragile watermark is embedded
into the host image. The aim of the proposed scheme in this paper is to
embed robust and fragile watermark into the host image at the same time,
so that the proposed scheme can possess dual functions: one is copyright
protection, and the other is tampering detection. Finally, we provide some
experiments to show the performance of the proposed scheme.
Key words: Digital watermark, QR decomposition, Robustness,
Tampering detection.

1. INTRODUCTION
The development of the Internet has made the transmission of digital images
faster and more convenient. However, some related issues have followed. For
example, a person who is interested may tamper with the image content for deception
purposes. Sometimes the created images may be stolen without the permission of the
author. In view of this, the tamper detection side and copyright protection of digital
images have become important issues. Digital watermarking is a commonly used
method to protect digital images.
Digital watermark refers to a piece of information embedded in a digital image.
A digital watermarking scheme consists of two phases: one is the embedding of the
watermark; the other is the extraction of the watermark [8]. Digital watermarks can
be divided into robust watermarks and fragile watermarks depending on the purpose
of use. Robust watermarking is used to protect the copyright of digital images.
Authors of digital images can choose an image representing a personal logo as a
108 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

watermark. When the copyright of a digital image is suspected of being violated, the
author can extract the watermark from the image to prove his copyright. The basic
requirement of the robust watermarking scheme is that after the partial image is
edited, the extracted watermark must still have enough recognition to prove the
copyright of the image. Fragile watermarking is used to protect the integrity of digital
images. When digital image content is tampered with, the fragile watermark must be
able to detect the part of the image that has been tampered with. Fragile watermarks
are usually composed of features of the image so that the watermark is sufficiently
sensitive to image tampering.
Generally speaking, when someone directly uses the image without the
authorization of the original author, the person usually makes some modifications to
the image. By doing so, the person hopes to evade the copyright controversy;
furthermore, the person hopes the modification can destroy the robust watermark in
the image. If the watermark is robust enough, the author can extract the robust
watermark to prove the copyright. Let us think about a situation. Supposed that the
person who steals the image is not for copyright infringement, and that his purpose
of modifying the image is not to destroy the robust watermark in the image. but to
intentionally add some fake content to the image and use it to frame the author. This
person may add some fake content to the image to frame the original author. The
author will argue that the content of these frauds is not from himself, but that
someone has maliciously framed him. However, if the original author is asked to
extract the robust watermark, the robust watermark will make the original author
unable to prove his innocence. If the image contains a fragile watermark, the original
author can prove his innocence by fragile watermarking. Because the fragile
watermark can help the original author to point out that these images have been
tampered with, the content of the fake is not the original. Therefore, a scheme with
both robust and fragile watermarking is necessary. However, few studies have
proposed a dual-purpose digital watermarking scheme. The purpose of this paper is
to expand the previously proposed robust watermarking scheme into a dual-purpose
watermarking scheme. The previous proposed watermarking method has a good
robustness against cropping attacks. If we can combine the fragile watermark to the
previous method, we can avoid the situation mentioned above that is maliciously
framed by others.
The following sections are arranged as follows. The second section will discuss
the relevant research. The third section explains in detail the practice of the proposed
scheme. The fourth section is the experimental results and some discussions of this
study. Finally, the fifth section is the conclusion of this study.
International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 109

2. RELATED WORKS

2.1. QR decomposition
Because a digital image can be seen as a matrix, recently, some researchers
employed matrix factorization, such as singular value decomposition (SVD) [1-2, 7]
or QR decomposition [3, 5, 6, 9-11], to design their digital watermarking schemes.
The main idea is to utilize matrix factorization to decompose the image and embed
the watermark into one of the decomposed matrices. Because the computational
complexity of SVD is higher than that of QR decomposition [10], QR decomposition
is a more common method.
The QR decomposition can factor a m  n matrix A into a m  m orthogonal
matrix Q and a m  n upper triangular matrix R, where
A= QR.
If the columns of A have correlation with each other, the absolute values of the
elements of the first row of R are probably greater than those of the other rows. Since
there usually exits correlations between adjacent pixels of an image, QR
decomposition was common in researches related to image processing [3, 8-11].
There are some algorithms for computing the QR decomposition, such as Gram–
Schmidt process, Householder transformations, or Givens rotations.
2.2. QR-based watermarking scheme
In our previous work, the host image is divided into 4  4 blocks, and each block
is decomposed via QR factorization. The robust watermark is embedded in the first
row of R matrix. In the proposed scheme, watermark bit 1 and bit 0 were represented
using f(x) ≥ 0 and f(x) < 0, respectively, and function f is defined as follows:
𝜋
𝑓 (𝑥 ) = 𝑠𝑖𝑛 (𝑥 ∙ 𝐾 ∙ 180), (1)
where K ≥ 0 is a real number. The wavelength  of the function f is (360/K), and
thus, parameter K is a wavelength controller. For any element x0 to embed the
watermark bit, if the state of f(x0) is not match the watermark bit, x0 is modified to
x0 such that |f(x0)| = 1 and the difference between x0 and x0 is as small as possible.
The R matrix with the robust watermark is merged with the original Q matrix, and
the digital image is inverted. The main concern of a QR-based digital watermarking
scheme is that the modification to the decomposed matrix may cause the pixel values
out of range. Moreover, the allowable bounds vary from pixels to pixels.
Nevertheless, few of papers gave a discussion about the allowable bounds of the
modification. To cope with this issue, our previous work built several rules to define
the allowable bound before modifying the elements of matrix R. Therefore, the
watermark can be embedded to achieve the possible minimal modification within the
allowable bounds.
110 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

3. THE PROPOSED SCHEME

3.1. Watermark embedding

Figure 1 illustrates the whole process of watermark embedding. Suppose that H
is a gray-level image and RW is a binary image, which represents the robust
watermark. At first, H is divided into nonoverlapping blocks of 4  4 sizes. Each
block A is factorized to matrix Q and R via QR decomposition. Each four bits of RW
are embedded into the four elements of the first row of R, respectively. Then,
multiply Q and the modified R to obtain the watermarked block A. Repeat the above
process to embed the watermark bits into the blocks. Because the focus of this study
is to incorporate fragile watermarks into our previous methods, we will explain the
method of embedding fragile watermark in detail. As for the method of hiding the
robust watermark, readers can refer to our previous research for the method of
embedding the robust watermark [5]. It should be noted that because the space of the
lowest bit plane of the image is used to embed the fragile watermark, the lowest bits
of each image block is not included in the QR decomposition operation.

Fig. 1. The process of watermark embedding

Suppose that H is the image with the robust watermark. At first, H is divided
into nonoverlapping blocks of 4  4 sizes. Supposed that pi denotes the pixel value
of a block B of H, where i = 0..15, and supposed that (x, y) denotes the coordinate
of B in H. For each block B, perform the following steps:
Step 1: pi = pi >> 2, where i = 0..15.
Step 2: Calculate MD5(x, y, p0, …, p15) to generate 128-bit digest (d0, d1, …, d127).
Step 3: Dj = (d16j, d16j, …, d16j+15), where j = 0..7. Perform XOR operation on D0
to D7 to get 16-bit authentication message A = (a0, a1, …, a15), which
represents the fragile watermark FW.
Step 4: Set pi = (pi >> 2) + ai to watermark the block B with authentication message,
where i = 0..15.
Repeat the above four steps to get the final watermarked image H.
International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 111

3.2. Watermark extraction

Since we hide two kinds of watermarks in the image, we can extract different
watermarks for different purposes. When we suspect that the image is pirated, we
can extract the robust watermark to prove the copyright. First, the image with doubts
is first decomposed by QR decomposition. Next, the watermarking bits are extracted
from each R matrix to form a watermarked image. The extracted watermark is
compared with the original watermark image. If the similarity between the two
reaches a certain level, we can claim our own copyright. Similarly, we focus on the
extraction of fragile watermark in this section. For the extraction of robust
watermarks, please refer to our previous research [5].
If we want to confirm the integrity of the image T, we can extract the fragile
watermark to verify. The method of extracting the robust watermark is described as
follows. First, split the image into blocks that do not overlap in size by 4  4.
Supposed that pi denotes the pixel value of a block B of H, where i = 0..15, and
supposed that (x, y) denotes the coordinate of B in H. Let EM denote a binary error
map of the image, which visually displays the result of tampering detection. For each
block B, perform the following steps:
Step 1: pi = pi >> 2, where i = 0..15.
Step 2: Calculate MD5(x, y, p0, …, p15) to generate 128-bit digest (d0, d1, …, d127).
Step 3: Dj = (d16j, d16j, …, d16j+15), where j = 0..7. Perform XOR operation on D0
to D7 to get 16-bit message M = (m0, m1, …, m15).
Step 4: Extract the authentication message A = (p0%2, p1%2, …, p15%2).
Step 5: If A  M, mark a black on EM; otherwise, mark a white on EM.
Repeat the above five steps to get the error map of the image T.

4. EXPERIMENTAL RESULTS AND DISCUSSIONS

In this section, we will present two experimental results, one for the
experimental results of copyright verification and the other for the experimental
results of tamper detection. The experiment of copyright verification is mainly to
show that the hiding of fragile watermark does not affect the robustness of the robust
watermark. The experiment of tamper detection is mainly to show that the fragile
watermark can successfully mark the modified part of the image.
4.1. Copyright verification
In general, a robust watermarking scheme is evaluated by its imperceptibility
and robustness [8]. The imperceptibility means that the difference between the host
image and the watermarked image should not be perceived by human eyes, and the
robustness means that the extracted watermark should be like the original watermark.
In this paper, we use PSNR to evaluate the imperceptibility of our scheme.
2552
PSNR  10 log
MSE , (2)
112 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

where
M N
1
MSE  
M  N i 1 j 1
( pi , j  pi, j ) 2 . (3)

The notation M and N in Eq. (3) denote the width and height of an image,
respectively, and pi,j and pi,j denote the pixel of the original and watermarked image,
respectively. Generally, the difference between the watermarked image and the host
image is visually imperceptible if PSNR is greater than 30. The robustness is
evaluated by the indicator NC as follows.
W H
1
NC 
W H
 (w
i 1 j 1
i, j  wi, j ) (4)

The notation W and H in Eq.(4) denote the width and height of the watermark
image, respectively, wi,j and wi,j denote the bit of the original and extracted
watermark, respectively, and ‘’ represents the logical XOR operation. The larger
the NC is, the more robust the extracted watermark is.
Figure 2 is our experimental images. When the watermark bit is embedded in
the R matrix, the K value in the Eq. (1) needs to be determined first. In our scheme,
we set different K values for the four elements of the R matrix. The set of parameter
K in the robust watermarking scheme is (25, 30, 40, 60). The PSNR of the
watermarked image showed that our scheme has good imperceptibility. In this
experiment, we simulated that image was cropped to destroy the embedded
watermark. In order to show the robustness of the proposed scheme, we simulated
different cropping ratios, as shown in figure 3. Figure 4 is the watermark images
corresponding to each of the attacked images in figure 3. Figure 4 shows that even
if the watermarked image is subjected to a cropping ratio of 60%, the NC value is
still as high as 0.8 or more. Therefore, embedding the fragile watermark after hiding
the robust watermark does not impact the effect of copyright verification.

(a) Host image (b) Watermarked image (c) Watermark

(512  512 pixels) PSNR = 42.2096 (128  128 pixels)

Fig. 2. The experimental images

International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 113

CR = 0.05, CR = 0.10, CR = 0.15, CR = 0.20,

PSNR = 22.6674 PSNR = 19.5439 PSNR = 16.9982 PSNR = 16.9982

CR = 0.25, CR = 0.30, CR = 0.35, CR = 0.40,

PSNR = 13.9781 PSNR = 13.4861 PSNR = 13.0086 PSNR = 12.5354

CR = 0.45, CR = 0.50, CR = 0.55, CR = 0.60,

PSNR = 12.0567 PSNR = 11.717 PSNR = 11.3661 PSNR = 11.0263

CR = 0.65, CR = 0.70, CR = 0.75, CR = 0.80,

PSNR = 10.7134 PSNR = 10.4365 PSNR = 10.1525 PSNR = 9.8252

CR = 0.85, CR = 0.90, CR = 0.95, CR = 1.00,

PSNR = 9.3816 PSNR = 8.8425 PSNR = 8.1972 PSNR = 7.6432
Fig. 3. The PSNR values for cropped images with different cropping ratio (C R) values
114 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

CR = 0.05, CR = 0.10, CR = 0.15, CR = 0.20,

NC = 0.9993 NC = 0.9976 NC = 0.9930 NC = 0.9859

CR = 0.25, CR = 0.30, CR = 0.35, CR = 0.40,

NC = 0.9749 NC = 0.9647 NC = 0.9489 NC = 0.9287

CR = 0.45, CR = 0.50, CR = 0.55, CR = 0.60,

NC = 0.9061 NC = 0.8794 NC = 0.8536 NC = 0.8231

CR = 0.65, CR = 0.70, CR = 0.75, CR = 0.80,

NC = 0.7925 NC = 0.7598 NC = 0.7335 NC = 0.7057

CR = 0.85, CR = 0.90, CR = 0.95, CR = 1.00,

NC = 0.6793 NC = 0.6599 NC = 0.647 NC = 0.6415

Fig. 4. The NC values of extracted watermarks for cropped images with different cropping
ratio values

Figure 5 is a line graph of the correlation between CR and NC. It can be seen
from this figure that NC did not drop sharply with the increase of CR value, which
means that the method proposed in this study has good robustness.
International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 115

Fig. 5. The relationship between CR and NC

4.2. Tampering detection

This paper uses false negative rate (FNR) (Eq. (5)) and false positive rate (FPR)
(Eq. (6)) to evaluate the efficiency of tampering detection and peak signal to noise
ratio (PSNR) (Eq. (2)) to evaluate the quality of the watermarked image. The
notations in Eq. (5) and Eq. (6) are explained as below [4, 8]:
FN: the number of false negative pixels, that is, the number of pixels that are
tampered but are judged as untampered;
TP: the number of true positive pixels, that is, the number of pixels that are
untampered but judged as tampered with;
FP: the number of true positive pixels, that is, the number of pixels that are tampered
and judged as tampered;
TN: the number of true negative pixels, that is, the number of pixels that are
untampered and judged as untampered.
FNR = FN / (FN + TP), and (5)
FPR = FP / (FP + TN). (6)
We use figure 2 as experimental images under malicious cropping attacks. Using
the proposed tampering detection scheme, we can successfully identify areas that
have been tampered with. We visually display the detected area with the Error Map,
as shown in figure 6.
 116 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

CR = 0.05, FNR = CR = 0.10, FNR = CR = 0.15, FNR = CR = 0.20, FNR =

0.0000, FPR = 0.0018 0.0000, FPR = 0.0041 0.0000, FPR = 0.0036 0.0000, FPR = 0.0000

CR = 0.25, FNR = CR = 0.30, FNR = CR = 0.35, FNR = CR = 0.40, FNR =

0.0000, FPR = 0.0000 0.0000, FPR = 0.0000 0.0000, FPR = 0.0071 0.0000, FPR = 0.0041

CR = 0.45, FNR = CR = 0.50, FNR = CR = 0.55, FNR = CR = 0.60, FNR =

0.0000, FPR = 0.0048 0.0000, FPR = 0.0111 0.0000, FPR = 0.0064 0.0000, FPR = 0.0000

CR = 0.65, FNR = CR = 0.70, FNR = CR = 0.75, FNR = CR = 0.80, FNR =

0.0000, FPR = 0.0000 0.0000, FPR = 0.0000 0.0000, FPR = 0.0135 0.0000, FPR = 0.0516

CR = 0.85, FNR = CR = 0.90, FNR = CR = 0.95, FNR = CR = 1.0, FNR =

0.0001, FPR = 0.0000 0.0001, FPR = 0.1084 0.0001, FPR = 0.0760 0.0001, FPR = 0.0000
Fig. 6. The authentication result for cropping attacks
International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 117

Since the cropping attack in figure 4 is less natural, we additionally simulate a

malicious tampering that may occur. We add three objects to the watermarked image
in figure 7(a): one is a puppy on the ground, one is an eagle in the sky, and the other
is a chimney on the roof of a house. The image synthesized with the above three
objects is shown in Figure 7(b). It can be observed from the Error Map of figure 7(c)
that the objects can be detected. One can notice that the FNR of this experiment is 0,
which means that the system can be successfully detected as long as it is tampered.
The FPR of this experiment was 0.0088, indicating that the area that was not
subjected to malicious tampering, about 0.88% of the pixels were misjudged as being
tampered with. Although the place that has not been tampered with may be
misjudged as being tampered with, the false positive rate is extremely low.
Compared to false positive misjudgment, false negative is more serious because it
means that the place that was maliciously tampered with may not be noticeable.

(c) The authentication

(b) Watermarked image
(a) Watermarked image result:
with malicious attack:
PSNR = 42.2096 FNR = 0.0000
TR = 0.0389
FPR = 0.0088

Fig. 7. The authentication result for a malicious attack

4.3. Security analysis

In terms of security, this study uses a secret key to scramble the image block
when embedding the robust watermark. Therefore, even if the attacker knows the
algorithm of this study, but does not have this secret key, he cannot restore the order
of the block and learn about our watermark image. Therefore, it is possible to prevent
an attacker from modifying the embedded watermark content to claim the copyright
of the image. In addition, this study also adds a secret key when generating the image
authentication message. Therefore, the attacker cannot deduce the authentication
message according to the algorithm of the study and deliberately change the content
of the verification message to make the tampering detection invalid. In summary,
this method meets the Kirchhoff’s principle: all algorithms must be public, and only
the key is secret. Kirchhoff’s criterion is often used to evaluate the usefulness of
cryptographic methods [9, 11]. If we need to pass the secret key over a public
118 International Journal on Information Technologies & Security, № 1 (vol. 11), 2019

network to a specific person, we can use asymmetric cryptography to protect the

security of the secret key during transmission. We can encrypt the secret key with
the public key of a specific person before transmitting the secret key. After receiving
the specific key, that person can use his private key to decrypt the content of the
secret key.

5. CONCLUSIONS
This study proposes a dual function watermarking scheme: one is copyright
verification, and the other is tampering detection. Suppose we need to verify the
copyright, we take the robust watermark representing the author's logo from the
image. In order to avoid the verification of copyright, an attacker may try to destroy
the image to remove the tough watermark. The experimental results of this study
show that even if half of the image is cropped, the extracted tough watermark can
clearly show the author's logo. Assuming we need to verify that the image has been
tampered with, we can take out the fragile watermark. With the fragile watermark,
we can not only know whether the image has been tampered with, but also the part
that has been tampered with. In this research experiment, it can be clearly seen from
the error map that this research method can mark the part that has been tampered
with. In the future research, we will make the fragile watermark not only have the
function of tamper detection, but also serve as the basis for enhancing the strength
of the robust watermark.

REFERENCES
[1] I.A. Ansari, M. Pant, and C.W. Ahn, “SVD based fragile watermarking scheme
for tamper localization and self-recovery,” International Journal of Machine
Learning and Cybernetics, 10.1007/s13042-015-0455-1, October 2015.
[2] S.C. Byun, S.K. Lee, A.H. Tewfik, and B.H. Ahn, “A SVD-Based Fragile
Watermarking Scheme for Image Authentication,” Lecture Notes in Computer
Science, Vol. 2613, pp. 170-178, April 2003.
[3] J. Gao, L. Fan, and L. Xu, “Solving the face recognition problem using QR
factorization,” WSEAS Transl. Mathematics, Vol. 11, Iss. 8, pp. 712-721,
August 2012.
[4] C.S. Hsu, S.F. Tu, “Probability-based Tampering Detection Scheme for Digital
Images,” Opt. Commun., Vol. 283, No. 9, pp. 1737-1743, 2010.
[5] C.S. Hsu and S.F. Tu, “Digital watermarking scheme enhancing the robustness
against cropping attack,” Proceedings of The 6th International Conference on
Frontier Computing (FC2017), pp. 143- 152, Osaka, Japan.
International Journal on Information Technologies & Security, № 1 (vol. 11), 2019 119

[6] H.F. Huang “A Fragile Watermarking Algorithm based on Chaos and QR

Decomposition,” International Journal of Advancements in Computing
Technology, Vol. 5 Issue 4, pp. 117-124, February 2013.
[7] Q. Kang, K. Li, and H. Chen, “An SVD-based Fragile Watermarking Scheme
With Grouped Blocks,” Proceedings of the 2nd International Conference on
Information Technology and Electronic Commerce, Dalian, China, December
2014.
[8] S. Katzenbeisser, F.A.P. Petitcolas, Information Hiding: Techniques for
Steganography and Digital Watermarking, Artech House, Inc., MA, USA,
2000.
[9] O. Su, Y. Niu, H. Zou, Y. Zhao, and T. Yao, “A blind double color image
watermarking algorithm based on QR decomposition,” Multimed. Tools Appl.,
Vol. 72, Iss. 1, pp 987-1009, September 2014a.
[10] Q. Su, Y. Niu, G. Wang, S. Jia, and J. Yue, “Color image blind watermarking
scheme based on QR decomposition,” Signal Processing, Vol. 94, pp. 219-235,
2014b.
[11] Q. Su, G. Wang, X. Zhang, G. Lv, and B. Chen, “An improved color image
watermarking algorithm based on QR decomposition,” Multimed. Tools Appl.,
vol. 76, no. 1, pp.707-729, 2017.

Information about the author(s):

Ching-Sheng Hsu got his BA degree from the Department of Information Management,
National Cheng-Chi University, Taiwan, in 1994, MA degree from the Institute of
Information Management, National Chi-Nan University, Taiwan, in 1998, and PhD degree
from the Institute of Information Management, National Central University, Taiwan, in 2005.
From 1998 to 1999, he was a software engineer at the Syscom Group Co, Taiwan, where his
work focused on the Web-based stock trading systems. Currently, he is an associate professor
of Department of Information Management, Ming Chuan University. His current research
interests include blockchain technology, digital watermarking and information hiding, and
intelligent computer-assisted learning and testing systems.
Shu-Fen Tu is a professor in Department of Information Management at Chinese Culture
University in Taiwan. She received the BS degree in management information system from
National Cheng-Chi University, Taiwan in 1996, the MS degree in information management
from National Chi-Nan University, Taiwan in 1998, and the PhD degree from the Institute of
Information Management, National Central University, Taiwan in 2005. From 1998 to 1999,
she was a software engineer of the Syscom Group Co., Taiwan. Her current research interests
include digital watermarking, secret sharing, and applications of blockchain.
Manuscript received on 2 January 2019