Post on 27-Jul-2018
Applied Mathematical Sciences, Vol. 11, 2017, no. 54, 2681 - 2695
HIKARI Ltd, www.m-hikari.com
https://doi.org/10.12988/ams.2017.78271
A Robust Public Key Watermarking Scheme
Based on Improved Singular Value Decomposition
Cao Thi Luyen, Nguyen Hieu Cuong and Pham Van At
Faculty of Information Technology, University of Transport and Communication
Ha Noi, Viet Nam
Copyright © 2017 Cao Thi Luyen, Nguyen Hieu Cuong and Pham Van At. This article is
distributed under the Creative Commons Attribution License, which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
Abstract
Digital watermarking has been considered as an effective solution for multimedia
rightful protection and authentication. Singular value decomposition (SVD) is
used for many watermarking schemes. This method has encountered some
challenges, such as computational complexity and robustness. In this paper,
instead of using the conventional SVD, we employed a new algorithm to directly
compute the largest eigenvalue and eigenvector of segmented image blocks.
Theoretically, by using this approach, our proposed watermarking scheme has
computational complexity lower than various SVD based schemes. This
improvement is essential for watermarking systems in practice, where ones often
have to work with large-scale image datasets. Experimental results also showed
that our watermarking scheme outperforms several widely used schemes in terms
of robustness. Moreover, using the proposed algorithm, we designed a new robust
public key watermarking scheme, where watermarked images can be verified
without using the pre-defined watermark and a secret key.
Keywords: Public key watermarking, SVD, eigenvalue, eigenvector
1 Introduction
With the advent of the Internet and the wide availability of technical equipments,
most people easily copy and distribute illegally digital products. For decades,
digital watermarking is commonly used for copyright protection and
authentication of multimedia content (i.e. digital images, audio, video, etc.). With every watermarking scheme, a watermark is embedded into digital objects in order to
2682 Cao Thi Luyen et al.
present their authorship. In fragile schemes, the embedded watermarks can be
distorted when any operation is applied to the watermarked objects. Conversely,
in robust ones, the embedded watermarks should be retained, even when some
operations are deliberately applied to the watermarked objects.
Robust watermarking has been proposed as an important method for protecting
the copyright of digital images. To this end, it is required to provide trustworthy
evidence for protecting rightful ownership. A good robust watermarking
algorithm is mostly resisted due to common image manipulations, such as
Gaussian noise addition, JPEG compression, filtering, etc. Moreover, the
perceptual difference between an original image and its watermarked version
should not be distinguishable to human eyes [12].
In order to design robust watermarking schemes, many image operations and
transformations were employed. Among them, the singular value decomposition
(SVD) is widely used due to its robust properties. In the SVD based image
watermarking schemes, first each segmented image block is decomposed into
three matrices U, D and V (see 2.1). After that, watermarks can be embedded in
different positions in the matrices: (1) in D [6, 15, 17, 19, 28] (2) in the first
column U or V [4, 13, 28] or (3) in both D and U [2, 7]. In various SVD based
algorithms [2, 6, 8, 10, 13], only the first coefficient of D and the first column of
U (or V) are used. Besides, SVD can be combined with some other image
transforms, such as discrete Cosine transform (DCT) and discrete wavelet
transform (DWT) [16, 24, 25].
The most important problem of every SVD based algorithms is the step of SVD
analysis. This decomposition requires finding all eigenvalues and eigenvectors of
the analyzed matrices. Since the transform is complex and must be employed to
every block, this step is rather time consuming. Inspired by the work of [7], we
show that in watermarking schemes, one does not need to compute SVD for
image blocks in the conventional way, but only need to use the largest eigenvector
and eigenvalue of the blocks. We proposed a new algorithm to directly compute
the aforementioned values and utilize them in watermarking schemes.
Consequently, both of the procedures of embedding and extracting of
watermarking schemes are speeded up. This improvement is essential for
watermarking systems in practice, when ones often have to work with large-scale
image dataset. We also conducted experiments to evaluate the robustness the
proposed scheme and some widely used watermarking schemes. The results show
that all of the schemes are robust against common attacks and the proposed
scheme is more robust than others.
To improve the security of watermarking systems, most algorithms have
employed a secret key. Thus, for verification, ones need to know both the
pre-embedded watermarks and the secret key. That limits the applicability in
practice, where the watermarks and secret keys are not easily distributed to legal
recipients. To overcome this drawback, various public key watermarking schemes
were proposed [3, 11, 18]. The schemes borrowed some ideas from public key
cryptography, where the watermark was encrypted by the private key of the legal owner. The corresponding public key was known by everyone and in the verification
A robust public key watermarking scheme 2683
phase, it will be used to decrypt embedded watermarks. Although public key
watermarking schemes have solved the problems of distribution and verification,
robustness is also a crucial requirement of public key watermarking [1, 19, 20]. In
this work, in order to improve the robustness of public key watermarking schemes,
we employ again the proposed algorithm to extract robust features from
segmented image blocks. As a result, our scheme achieved lower computational
complexity and higher degree of robustness.
The rest of this paper is organized as follows. Section 2 reviews the essential
background of SVD and some related work. In Section 3, we describe the
proposed the algorithm and design a robust public key watermarking scheme.
Sections 4 reports and discusses experimental results. Lastly, the paper is
concluded in Section 5.
2 Related word
2.1 Singular value decomposition
SVD is not only be used in linear algebra, but also in many different applications,
such as in image processing. The main benefits of SVD in image processing are that
it provides a method to decompose a large matrix to smaller and more manageable
matrices and the singular values (SVs) of an analyzed matrices have a good
stability. Thus, when a manipulation is applied to an image, its SVs do not
significantly changed. Concretely, an m×n image block (can be considered as a
matrix) is decomposed into three matrices U, D and V [9]:
𝐴 = 𝑈 × 𝐷 × 𝑉𝑇 = 𝑈1𝐷(1,1)𝑉1𝑇 + 𝑈2𝐷(2,2)𝑉2
𝑇 + ⋯ + 𝑈𝑠𝐷(𝑠, 𝑠)𝑉𝑠𝑇, (1)
where s = min(m,n), 𝑈 ∈ 𝑅𝑚×𝑚 , 𝑉 ∈ 𝑅𝑛×𝑛 are normalized orthogonal
matrices and 𝐷 ∈ 𝑅𝑚×𝑛 is a diagonal matrix. The main diagonal of D consists of
the singular values of A, where D(1,1) ≥ D(2,2) ≥ ... ≥ D(s,s) ≥ 0 (s is called the
rank of the matrix D).
2.2 Watermarking schemes based on SVD
2.2.1 The scheme of Chung et al.
In Chung et al. [13], to embed a watermark 𝑊 = {𝑤1, 𝑤2, … , 𝑤𝑠} into an image I,
first the image is divided into non-overlapped blocks. Next, in following steps,
each image block 𝐼𝑖, (𝑖 = 1, 2, … , 𝑠) can be used to insert one bit 𝑤𝑖:
Step 1: Applying SVD to every Ii:
𝐼𝑖 = 𝑈𝑖 × 𝐷𝑖 × 𝑉𝑖𝑇
Step 2: Inserting a bit wi into Ui by adjusting 𝑈𝑖(2,1) and 𝑈𝑖(3,1) following
the rule:
|𝑈𝑖(3,1)| − |𝑈𝑖(2,1)| ≥ 𝜃 if 𝑤𝑖 is 0
|𝑈𝑖(3,1)| − |𝑈𝑖(2,1)| < 𝜃 if 𝑤𝑖 is 1
2684 Cao Thi Luyen et al.
where θ is a pre-defined thereshold.
After the embedding process, 𝑈𝑖 becomes 𝑈𝑖′, which includes 𝑤𝑖.
Step 3: Computing 𝐼𝑖′
𝐼𝑖′ = 𝑈′𝑖 × 𝑉𝑖 × 𝑉𝑖
𝑇 In the processes of transmission and distribution, the watermarked image 𝐼′ can
be affected by unintentional manipulations or deliberated attacks. Subsequently,
the receiver usually obtained the attacked version 𝐼∗, which is not quite the same
as 𝐼′. Therefore, the watermark 𝑊∗ (extracted from 𝐼∗) should be compared to
the original watermark W in order to evaluate the robustness of the scheme.
In the reverse direction, the extracted watermark is recovered in following steps:
Step 1: Applying SVD to each non-overlapped segmented block 𝐼𝑖∗:
𝐼𝑖∗ = 𝑈𝑖
∗ × 𝐷𝑖∗ × 𝑉𝑖
∗𝑇
Step 2: Computing a bit 𝑤𝑖∗ from 𝑈𝑖
∗:
𝑤𝑖∗= {
0 If |𝑈𝑖(3,1)| − |𝑈𝑖(2,1)| ≥ θ1 Otherwise
Step 3: Evaluating the accuracy of the obtained watermark:
𝐸𝑅𝑅 =1
𝑡∑ 𝑤𝑖 𝑋𝑂𝑅 𝑤𝑖
∗𝑡𝑖=1 (2)
For verification, we employ a pre-define threshold τ (τ[0,1]) to conclude the
extraction results: ERR < τ means that W was legally embedded in 𝐼∗.
2.2.2 The scheme of Lai
The main steps of [6] are similar to [13], but there are several differences:
1) A watermarking bit is not embedded into every block, but only in the blocks
satisfy some criteria of visual entropy,
2) [6] employs DCT to the selected blocks before using SVD, and
3) Watermarks are embedded into 𝑈𝑖(3,1) and 𝑈𝑖(4,1) instead of using
𝑈𝑖(3,1) and 𝑈𝑖(4,1) as in [13].
2.2.3 Comments on the SVD based schemes
In numerous SVD based schemes [5, 6, 13, 17, 19, 23], each segmented image
block Ii is usually divided into matrices 𝑈𝑖 , 𝑉𝑖 and 𝐷𝑖. Although each bit wi can
be embedded into 𝑈𝑖(3,1) and 𝑈𝑖(4,1) [6,13] or only in 𝐷𝑖(1,1) [5,17,19,23],
in the processes of embedding and extraction, we have to analyze all of the
matrices. Next, in watermarking schemes, we show that, it does not need to
analyze all matrices, but only need to compute the first column of 𝑈𝑖 and 𝑉𝑖, or
the first coefficient of 𝐷𝑖 . Obviously, we have:
𝐼𝑖 = 𝑋𝑖 + 𝐶𝑖 where
Xi= Di(1,1)Ui(1)Vi(1)T
A robust public key watermarking scheme 2685
Ci= ∑ Di(j,j)Ui(j)Vi(j)Tsj=2
s = min {m, n}
After embedding a watermarking bit into Ui(1), Vi(1) or Di(1,1), then 𝑋𝑖
become 𝑋𝑖′ , while 𝐶𝑖 is not changed. Therefore:
𝐼𝑖′ = 𝑋𝑖′ + 𝐶𝑖 where
𝐶𝑖 = 𝐼𝑖 − 𝑋𝑖
Consequently, in order to insert one bit into 𝐼𝑖, we only need to compute
𝑋𝑖. That means we only determine 𝐷𝑖(1,1), 𝑈𝑖(1) and 𝑉𝑖(1), instead of finding
𝐷𝑖(𝑗, 𝑗), 𝑈𝑖(𝑗) and 𝑉𝑖(𝑗), where 𝑖 = 1 … 𝑠, 𝑗 = 1. . . 𝑛 .
Based on this idea, next section describes the new watermarking scheme,
which achieves low computational complexity. For hiding a watermarking bit in a
segmented image block, we employ a pair {𝑈𝑖(1,1), 𝑈𝑖(2,1)} instead of using
{𝑈𝑖(2,1), 𝑈𝑖(3,1)} in order to increase the robustness of watermarking schemes.
2.3 Using SVD for feature extraction
In order to design robust public key watermarking schemes, extracting robust
features from images is crucial. In embedding process, the features are utilized to
create watermarks. There are two main directions to generate watermarks in
public key watermarking schemes:
1) using standard random distribution [18],
2) using robust image features [26, 27].
In this paper, the second approach is considered. Next, we review some SVD
based algorithms for extracting robust features from images and employ them to
watermarking schemes.
2.3.1 The algorithm of Zhou and Jin
In the algorithm of Zhou and Jin [27], the features B are extracted from an image I
by following steps:
Step 1: Using the L-level DWT transform (in the paper, the authors use 2-level
DWT) to I in order to obtain low frequency part, which is denoted by 𝐹𝐿.
Step 2: Dividing 𝐹𝐿 into non-overlapped blocks, denoted 𝐹𝐿𝑖, 𝑖 = 1. . 𝑘.
Step 3: Applying SVD to every block 𝐹𝐿𝑖:
𝐹𝐿𝑖= 𝑈𝑖𝐷𝑖𝑉𝑖
𝑇 , 𝑖 = 1. . 𝑘
Step 4: Determining the special features B as following rule:
If 𝐷2𝑖−1(1,1) > 𝐷2𝑖(1,1) then B(i) = 1 else B(i) = 0, where 𝑖 = 1. . ⌊𝑘
2⌋.
2.3.2 The algorithm of Ye
According to Ye [26], the particular features B are extracted by the following
steps:
Step 1: Dividing the host image I into non-overlapped blocks denoted by 𝐼𝑖 , 𝑖 =1. . 𝑘.
Step 2: Applying SVD to every block 𝐼𝑖:
𝐼𝑖 = 𝑈𝑖𝐷𝑖𝑉𝑖𝑇 , 𝑖 = 1. . 𝑘
2686 Cao Thi Luyen et al.
Step 3: Employing DCT for 𝐷𝑖 and using its DC coefficient, denoted 𝐷𝐶𝑖.
Step 4: Determining the special features B as the following rule:
If 𝐷𝐶2𝑖−1(1,1) > 𝐷𝐶2𝑖(1,1) then B(i) = 1 else B(i) = 0, where 𝑖 = 1. . ⌊𝑘
2⌋.
2.3.3 Comments on extracting features
In the mentioned algorithms [26,27], specific features were created by using SVD
and the inequalities 𝐷2𝑖−1(1,1) > 𝐷2𝑖(1,1) or 𝐷𝐶2𝑖−1(1,1) > 𝐷𝐶2𝑖(1,1). When
an image has been attacked, these relationships can be ruined and the extracted
features from attacked images become different from the features of the original
image. Therefore, if we choose a threshold T > 0, the features satisfy the
condition |𝐷1(1,1) − 𝐷2(1,1)| < 𝑇.
3 The proposed algorithm
3.1 Determining Di (1,1), Ui (1) and Ui (1)
In this section, we revise the algorithm for computing 𝐷𝑖(1,1), 𝑈𝑖(1) and 𝑉𝑖(1)
directly without using the conventional SVD. To achieve that goal, we employ the
method suggested in [14] to find the largest eigenvalue and eigenvector of a
non-negative matrix. The algorithm is described as below:
For this algorithm, the input is a matrix Ii and the output are the first coefficient of
Di and the first columns of 𝑈𝑖 and 𝑉𝑖. To this end, we set:
Ai = Ii IiT, Ai ≥ 0, 𝐴𝑖 ∈ 𝑅𝑚×𝑚
Because Ui and Vi are normalized orthogonal matrices, we have:
𝐴𝑖 = 𝑈𝑖𝐷𝑖𝑉𝑖𝑇(𝑈𝑖𝐷𝑖𝑉𝑖
𝑇)𝑇=UiDi2Ui
T ,
𝐴𝑖𝑈𝑖 = 𝑈𝑖𝐷𝑖2 and 𝐴𝑖𝑈𝑖(1) = 𝐷𝑖
2(1,1)𝑈𝑖(1).
Following the definition of eigenvalues and eigenvectors, 𝐷𝑖2(1,1) and 𝑈𝑖(1)
will be the largest eigenvalue and eigenvector of the matrix 𝐴𝑖 , respectively.
Similarly, if we set: Bi = IiTIi, Bi ≥ 0, 𝐵𝑖 ∈ 𝑅𝑛×𝑛 then Di
2(1,1) becomes the largest
eigenvalue and 𝑉𝑖(1) is the corresponding eigenvector of the non-negative matrix
𝐵𝑖.
As a result, the values of 𝐷𝑖2(1,1) , 𝑈𝑖(1) and 𝑉𝑖(1) can be computed by
evaluating the maximum eigenvalue and corresponding eigenvector of the
non-negative matrix 𝐴𝑖. In other words, by directly applying this technique to the
proposed algorithm, we can bypass the conventional SVD. Therefore, the
computational complexity of feature extracting and watermarking schemes will be
decreased.
3.2 The robust features extracting
To reduce the computational complexity as well as further improve the robustness
of extracted features, we employ a threshold T along with the proposed algorithm
(Section 3.1). The extracting algorithm includes following steps:
Step 1: Dividing an original image I into non-overlapped m×n blocks 𝐼𝑖 , 𝑖 =1. . 𝑘.
A robust public key watermarking scheme 2687
Step 2: Applying the mentioned algorithm (in Section 3.1) to each block 𝐼𝑖 in
order to compute 𝐷𝑖(1,1).
Step 3: Determining robust features of the cover image by the formulas as below:
If |𝐷2𝑖−1(1,1) − 𝐷2𝑖(1,1)| > 𝑇 then B(i) = 1 else B(i) = 0, where 𝑖 = 1. . ⌊𝑘
2⌋.
Experimental results (in Section 4) show that the proposed algorithm is more
robust than some other related schemes [26, 27].
3.3 The watermark embedding procedure
The input of the embedding algorithm is an image I, a watermark 𝑊 =(𝑤1, 𝑤2, … , 𝑤𝑡), and pre-defined thresholds (𝜃, 𝜏). Its output is the watermarked
image 𝐼′.
The embedding algorithm consists of following steps:
Step 1: Dividing the original image I into non-overlapped m×n blocks 𝐼𝑖.
Step 2: Using the proposed algorithm to directly compute the first column of 𝑈𝑖.
We obtained 𝑈𝑖(1) as the eigenvector corresponding to the largest eigenvalue of
the non-negative matrix 𝐼𝑖 × 𝐼𝑖𝑇.
Step 3: Embedding 𝑤𝑖 into the the image block 𝐼𝑖 using the pair
{𝑈𝑖(1,1), 𝑈𝑖(2,1)}.
In order to obtain 𝑈𝑖′, we adjust 𝑈𝑖(1,1) and 𝑈𝑖(2,1) according to 𝑤𝑖 such that
they satify the following conditions:
|𝑈𝑖(2,1)| − |𝑈𝑖(1,1)| ≥ θ if 𝑤𝑖 is 0
|𝑈𝑖(2,1)| − |𝑈𝑖(1,1)| < θ if 𝑤𝑖 is 1 Step 4: Figuring out
𝐼𝑖′ = 𝐷𝑖(1,1) × 𝑈𝑖
′(1) × 𝑉𝑖𝑇(1) + 𝐼𝑖 − 𝐷𝑖(1,1) × 𝑈𝑖(1) × 𝑉𝑖
𝑇(1)
As a result, 𝐼𝑖′ can be computed without using conventional SVD analysis. The
correctness of the calculation of 𝐼𝑖′ is described as below.
According to the formula (1) in [9]:
𝐼𝑖 = 𝐷𝑖(1,1) × 𝑈𝑖(1) × 𝑉𝑖𝑇(1) + ∑ 𝐷𝑖(k,k) × 𝑈i(k) × 𝑉𝑖(𝑘),𝑠
𝑘=2
where 𝑠 = 𝑚𝑖𝑛 (𝑚, 𝑛).
Therefore:
∑ 𝐷𝑖(k,k) × 𝑈i(k) × 𝑉𝑖(𝑘)𝑠𝑘=2 = 𝐼𝑖 − 𝐷𝑖(1,1) × 𝑈𝑖(1) × 𝑉𝑖
𝑇(1) (3)
Besides, we have:
𝐼𝑖′ = 𝐷𝑖
′(1,1) × 𝑈𝑖′(1) × 𝑉𝑖
′𝑇(1) + ∑ 𝐷𝑖
′(k,k) × 𝑈𝑖′(k) × 𝑉𝑖
′(𝑘)𝑠𝑘=2 (4)
Since the values of 𝐷𝑖(k,k), 𝑈i(k) and 𝑉𝑖(𝑘) are retained after embedding, we
have:
𝐷𝑖′(𝑘,k) = 𝐷𝑖(𝑘,k), 𝑈𝑖
′(k) = 𝑈i(k) and 𝑉𝑖′(𝑘) = 𝑉𝑖(𝑘), 𝑘 ≥ 2.
Thus,
𝐼𝑖′ = 𝐷𝑖(1,1) × 𝑈𝑖
′(1) × 𝑉𝑖𝑇(1) + ∑ 𝐷𝑖(k,k) × 𝑈i(k) × 𝑉𝑖(𝑘)𝑠
𝑘=2 (5)
Substitute (3) for (5), we obtain the next equation:
Ii′ = 𝐷𝑖(1,1) × 𝑈𝑖
′(1) × 𝑉𝑖𝑇(1) + 𝐼𝑖 − 𝐷𝑖(1,1) × 𝑈𝑖(1) × 𝑉𝑖
𝑇(1)
2688 Cao Thi Luyen et al.
3.4 The watermark extracting procedure
The extracting procedures of the proposed algorithm and the algorithm of of [13]
are alike, except the first and the second steps. In the first step, we solve the
largest eigenvalue and the correspond eigenvector of the non-negative matrix as
below:
𝐴𝑖∗ = 𝐼𝑖
∗ × 𝐼𝑖∗𝑇
Instead of applying SVD on each blocks 𝐼𝑖
∗ to determine 𝐷𝑖∗(1,1) and 𝑈𝑖
∗(1). In
the second step, 𝑤1∗ is extracted by following way:
If |𝑈𝑖∗(2,1)| − |𝑈𝑖
∗(1,1)| ≥ θ if 𝑤𝑖 is 0 then 𝑤1∗ = 0, otherwise w1
∗ = 1.
3.5 The proposed public key watermarking scheme
To improve the usability of watermarking in practice, various watermarking
schemes borrowed some ideas from public key cryptography [11]. In this section,
we utilize the proposed algorithm and public key in order to design a robust public
key watermarking scheme.
3.5.1 The embedding watermark algorithm
The input of the algorithm is an original image I. This private key belongs to the
legal owner of I. The output is the watermarked image 𝐼′.
The embedding process is performed as below:
Step 1: Dividing the original image 𝐼 into two separated part 𝐼1 and 𝐼2.
Step 2: Creating the watermark W by using the proposed algorithm (Section 3.2) to
extract robust features (denoted by W) from 𝐼2. Assuming that, the length of W is k
(bits).
Step 3: Embedding the watermark W into the part 𝐼1following the proposed
algorithm. Consequently, we get 𝐼1′ .
Step 4: Combining 𝐼1′ and 𝐼2 to receive the watermarked image 𝐼′.
3.5.2 The extracting and detecting algorithms
The input of the algorithm is an attacked image 𝐼∗ and a pre-defined
threshold 𝜏 (𝜏[0,1]) . The output is the conclusion that the image I∗ legally
belongs to the owners of 𝐼′.
The process of determining the copyright of 𝐼∗ is conducted in the following
steps:
Step 1: Dividing the attacked image 𝐼∗ into two separated part 𝐼1∗ and 𝐼2
∗ follows
the same way in Step 1 of the embedding watermarking process.
Step 2: Recovering 𝐵1∗ from 𝐼1
∗ by applying the proposed extracting procedure (in
Section 3.4).
Step 3: Extracting the features of 𝐼2∗ by employing the extracting algorithm
(presented in Section 3.2) to obtain 𝐵2∗.
Step 4: Inspired by the formula to calculate ERR (formula (2), Section 2.2.1), we
will verify the copyright of attacked images by the following sub-steps:
- Computing 𝐶 = 𝑆𝑈𝑀(𝐵1∗ 𝑋𝑂𝑅 𝐵2
∗)/𝑘 (6)
A robust public key watermarking scheme 2689
- If 𝐶 < τ then we can conclude that the image 𝐼∗ was copyrighted.
4 Experimental Results
We use some standard images with size of 512×512 (Figure 1) to evaluate the
robustness of our feature extraction algorithm and related work (Section 4.1) and
the robust watermarking schemes (Section 4.2). The experimental images are
downloaded from the standard image database USC-SIPI.
(A)
(B)
(C)
(D)
Figure 1. Images used in experiments.
To this end, first we applied attacks to watermarked images. Different operations
of Gaussian noise (15%), JPEG compression (50%), Gaussian blurring (2%),
median filtering (with the window size of 2×2) and resizing (up to 20% and down
to the original size of the host image) and were independently employed to
watermarked images. In the following sections, we used the original images and
attacked images to evaluate the robustness of feature extraction algorithm and
watermarking schemes.
4.1 Robustness of feature extraction algorithms
In order to measure the robustness of the proposed feature extraction algorithm,
we obtain the features B and 𝐵∗ by applying the analyzed algorithms to the
extracted features from the original images (Figure 1, A, B, C or D) and its
corresponding attacked versions.
We considered the algorithms of [26, 27] and our proposed algorithm. We used
the threshold T = 100 in the proposed algorithm.
2690 Cao Thi Luyen et al.
Table 1. Robust feature extraction against “adding Gaussian noise”.
Algorithms A B C D
Ye 0.2852 0.1978 0.2998 0.2944
Zhou & Jin 0.2402 0.1309 0.1553 0.2231
Proposed 0.0313 0.0610 0.0713 0.0449
Table 2. Robust feature extraction against “JPEG compression”.
Algorithms A B C D
Ye 0.0752 0.0449 0.0630 0.1001
Zhou & Jin 0.0557 0.0269 0.0293 0.0669
Proposed 0.0044 0.0049 0.0093 0.0059
Table 3. Robust feature extraction against “Gaussian blurring”.
Algorithms A B C D
Ye 0.1050 0.0859 0.1060 0.2739
Zhou & Jin 0.0532 0.0410 0.0361 0.0737
Proposed 0.0161 0.0376 0.0425 0.0562
Table 4. Robust feature extraction against “median filtering”.
Algorithms A B C D
Ye 0.1401 0.1401 0.0854 0.1563
Zhou & Jin 0.0879 0.0879 0.0283 0.1230
Proposed 0.0078 0.0078 0.0303 0.0811
Table 5. Robust feature extraction against “resizing”.
Algorithms A B C D
Ye 0.0630 0.0400 0.0498 0.0728
Zhou & Jin 0.0278 0.0181 0.0205 0.0298
Proposed 0.0083 0.0034 0.0190 0.0151
Next, in each case we employed the formula (6) in the previous section to
compute the difference C between B and 𝐵∗. The experimental results are shown
in Table 1 – Table 5, where the lower the difference, the higher the robustness.
The results show that, the proposed algorithm is robust.
4.2 Robustness of watermarking schemes
As the extracting and detecting processes of watermarking algorithm described in
Section 3.5.2, in the watermark validation process, comparing B1∗ and B2
∗ shows
the robustness of the watermarking algorithm. We apply the proposed algorithm
and other algorithms of [6, 13] on the attacked version of the images in Figure 1 (A
– D). The experimental outcomes (Table 6 – Table 10) exhibit the robustness of
A robust public key watermarking scheme 2691
[13] and the proposed algorithms is essentially the same and they are much more
robust than [6]. However, due to the application of the improved SVD algorithm
(presented in Section 3), our algorithms have much lower computational
complexity.
Table 6. The robustness of watermarking algorithms against “Gaussian noise”.
Algorithms A B C D
Chung 0.2609 0.2271 0.1925 0.1481
Lai 0.2937 0.2583 0.2289 0.1920
Proposed 0.2629 0.2281 0.1948 0.1526
Table 7. The robustness of watermarking algorithms against “JPEG compression”.
Algorithms A B C D
Chung 0.2662 0.0015 0.0621 0.0071
Lai 0.2789 0.2660 0.2184 0.2284
Proposed 0.1592 0.0020 0.0708 0.0075
Table 8. The robustness of watermarking algorithms against “Gaussian blurring”.
Algorithms A B C D
Chung 0.3123 0.2787 0.3127 0.2626
Lai 0.3995 0.4133 0.4175 0.4260
Proposed 0.3058 0.2665 0.3146 0.2609
Table 9. The robustness of watermarking algorithms against “filtering”.
Algorithms A B C D
Chung 0.1007 0.0638 0.0907 0.0760
Lai 0.2896 0.2879 0.2990 0.2969
Proposed 0.0994 0.0903 0.0898 0.0782
Table 10. The robustness of watermarking algorithms against “resizing”.
Algorithms A B C D
Chung 0.0297 0.0046 0.0259 0.0228
Lai 0.2917 0.3234 0.3181 0.3208
Proposed 0.0289 0.0047 0.0237 0.0244
This is very important when we often have to work with a large-scale image dataset
in practice. In addition, in order to increase the applicability in practice, our
algorithms use the public key model, so the watermark and secret key do not need
to be distributed to the recipients.
2692 Cao Thi Luyen et al.
5 Conclusion
In this paper, a fast and robust public key image watermarking scheme based on
improved SVD was presented. To achieve this goal, we employed a method to
directly compute the largest eigenvector and eigenvalue of a matrix, instead of
using the conventional SVD. Since the computational complexity of
watermarking schemes was decreased, the processes of watermarking embedding
and extracting were speed up. This is crucial for working with large-scale image
dataset in real-life applications. The robustness of feature extraction and
watermarking schemes can be further improved by using a threshold.
Experimental results showed that our watermarking scheme is more robust against
various attacks in comparison with several widely used SVD based schemes.
Besides, based on the mentioned algorithms, we suggested a new robust public
key watermarking scheme, where watermarked images can be verified without
using the pre-defined watermark and secret key. As a result, with the proposed
improvements, our scheme can be applied in practice.
References
[1] A. Ray, S. Padhihary, P. Kumar, P. Mihir and N. Mohanty, Development of a
New Algorithm Based on SVD for Image Watermarking, Chapter in
Computational Vision and Robotics, Springer, 2015, 79-87.
https://doi.org/10.1007/978-81-322-2196-8_10
[2] B.C. Mohan and S.S. Kumar, A Robust Image Watermarking Scheme using
Singular Value Decomposition, Journal of Multimedia, 3 (2008), no. 1, 7-15.
https://doi.org/10.4304/jmm.3.1.7-15
[3] B. Natarajan, Robust Public Key Watermarking for Digital Images, Imaging
Tech. Dept., Hewlett-Packard Company HP Labs, 1997.
[4] C.C. Chang, T.C. Tsai and C. Lin, SVD-based digital image watermarking
scheme, Pattern Recognition Letters, 26 (2005), no. 10, 1577-1586.
https://doi.org/10.1016/j.patrec.2005.01.004
[5] C. Fei, R.K. Wong and D. Kundur, Secure semi-fragilewatermarking for
image authentication, 2009 First IEEE International Workshop on
Information Forensics and Security WIFS, (2009), 141-145.
https://doi.org/10.1109/wifs.2009.5386465
[6] C.C. Lai, An improved SVD-based watermarking scheme using human
visual characteristics, Optics Communication, 284 (2011), no. 4, 938-944.
https://doi.org/10.1016/j.optcom.2010.10.047
A robust public key watermarking scheme 2693
[7] C.T. Luyen, N.H. Cuong and P.V. At, A Fast and Robust Image
Watermarking Scheme Using Improved Singular Value Decomposition,
Chapter in Intelligent Information and Database Systems, Lecture Notes in
Computer Science, Vol 9621, Springer, Berlin, Heidelberg, 2016, 780-789.
https://doi.org/10.1007/978-3-662-49381-6_75
[8] G. Bhatnagar and B. Raman, A new robust reference watermarking scheme
based on DWT-SVD, Computer Standards & Interfaces, 31 (2009), no. 5,
1002-1013. https://doi.org/10.1016/j.csi.2008.09.031
[9] G.H. Golub and C.F. Loan, Matrix Computations, Johns Hopkins University,
Press, Baltimore, 1996.
[10] H. Chen and Y. Zhu, A robust watermarking algorithm based on QR
factorization and DCT using quantization index modulation technique,
Journal of Zhejiang University SCIENCE C, 13 (2012), 573-584.
https://doi.org/10.1631/jzus.c1100338
[11] H.Y. Kim, A New Public Key Authentication Watermarking for Binary
Document Images resistant to parity attacks, IEEE International Conference
on Image Processing 2005, (2005).
https://doi.org/10.1109/icip.2005.1530245
[12] I. Cox, M. Miller and J. Bloom, Digital Watermarking, Morgan Kaufmann,
2002.
[13] K.L. Chung, W.N. Yang, Y.H. Huang, S.T. Wu and Y.C. Hsu, On
SVD-based watermarking algorithm, Applied Mathematics and Computation,
188 (2007), 54-57. https://doi.org/10.1016/j.amc.2006.09.117
[14] P.V. At, Determination of the greatest eigenvalue and the corresponding
eigenvector of a non-negative matrix, USSR Comput. Math. Math. Phys., 21
(1981), no. 4, 1–19. https://doi.org/10.1016/0041-5553(81)90030-6
[15] P. Bao and X. Ma, Image adaptive watermarking using wavelet domain
singular value decomposition SVD, IEEE Transactions on Circuits and
Systems for Video Technology, 15 (2005), 96-102.
https://doi.org/10.1109/tcsvt.2004.836745
[16] N.R. Sawant and Pravin S. Patil, Comparative study of SWT-SVD and
DWT-SVD digital image watermarking technique, International Journal of
Computer Applications., 166 (2017), no. 12, 9-17.
https://doi.org/10.5120/ijca2017913817
2694 Cao Thi Luyen et al.
[17] R. Sun, H. Sun and T. Yao, A SVD and quantization based semi-fragile
watermarking technique for image authentication, 6th Proc. Intl. Conf. on
Signal Processing, (2002), 1952–1955.
https://doi.org/10.1109/icosp.2002.1180102
[18] R. Munir, B. Riyanto, S. Sutikno and W.P. Agung, A public key
watermarking method for still image based on random permutation, In Proc
Iternational Joint Conference in Engineering, (2008).
[19] R. Liu and T. Tan, An SVD-based watermarking scheme for protecting
rightful ownership, IEEE Trans. Multimedia, 4 (2002), no. 1, 121–128.
https://doi.org/10.1109/6046.985560
[20] S. Lee, D. Jang and C. Yoo, An SVD-based watermarking method for image
content authentication with improved security, In Proc. of IEEE
International Conference on Acoustics, Speech, and Signal Processing, IEEE
Press, 2005. https://doi.org/10.1109/icassp.2005.1415457
[21] S. Byun, S. Lee, A. Tewfik and B. Ahn, A SVD-Based Fragile
Watermarking Scheme for Image Authentication, Chapter in International
Workshop on Digital Watermarking, Springer, 2002, 17-178.
https://doi.org/10.1007/3-540-36617-2_14
[22] M. Rahman, S. Ahammed, R. Ahmed and M.N. Izhar, A Semi Blind
Watermarking Technique for Copyright Protection of Image Based on DCT
and SVD Domain, Global Journal of Research In Engineering, 16 (2016),
9-16.
[23] X. Sun, J. Liu, J. Sun, Q. Zhang and W. Ji, A robust image watermarking
scheme based on the relationship of SVD, 2008 Int. Conf. Intelligent
Information Hiding and Multimedia Signal Processing, Harbin, (2008),
731-734. https://doi.org/10.1109/iih-msp.2008.190
[24] U.M. Gokhale and Y.V. Joshi, A Semi Fragile Watermarking Algorithm
Based on SVD-IWT for Image Authentication, International Journal of
Advanced Research in Computer and Communication Engineering, 1 (2012),
217-222.
[25] T.H. Rassem, N.M. Makbol and B.E. Khoo, Performance evaluation of
RDWT-SVD and DWT-SVD watermarking schemes, AIP Conference
Proceedings, 1774 (2016), 050021. https://doi.org/10.1063/1.4965108
[26] T. Ye, A Robust Zero-Watermark Algorithm Based on Singular Value
Decomposition and Discreet Cosine Transform, Chapter in Parallel and
Distributed Computing and Networks, Springer, 2011, 1–8.
A robust public key watermarking scheme 2695
https://doi.org/10.1007/978-3-642-22706-6_1
[27] Y. Zhou and W. Jin, A novel image zero-watermarking scheme based on
DWT-SVD, 2011 Int. Conf. Multimed. Technol. ICMT, (2011), 2873–2876.
https://doi.org/10.1109/icmt.2011.6002066
[28] X. Zhu, J. Zhao and H. Xu, A digital watermarking algorithm and
implementation based on improved SVD, Proc. 18th Int. Conf. Pattern
Recognition, Hong Kong, (2006), 651-656.
https://doi.org/10.1109/icpr.2006.59
Received: September 12, 2017; Published: October 29, 2017