A watermark algorithm against de-synchronization attacks
VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
A Watermark Algorithm Against De-Synchronization Attacks
Luong Viet Nguyen*, Trinh Nhat Tien, Ho Van Canh
VNU University of Engineering and Technology, Hanoi, Vietnam
Abstract
In this paper, a robust method to the ability of the watermark to resist against attacks is proposed for hiding
information into images. The proposed method is blind because the original image is not required at the decoder
to recover the embedded data. The robustness of the watermarking scheme is inspired by using a PJND
(Pyramidal Just Noticeable Difference) model and the message is inserted into these DoG (Difference of
Gaussians) [1, 2]. Our proposal takes into account three main characteristics of Human Visual System, namely:
contrast sensitivity, luminance adaptation and contrast marking. Therefore, it not only provides an invisible and
robust watermarking but also optimizes watermarking capacity. The performance of the proposed technique is
evaluated by a series of experiments with different input images. In terms of transparency, besides using the
subjective experiments, eight objective metrics are calculated in comparison with other methods such as PSNR,
MSSIM, SVDm, etc. Our approach always presents the outperform values. In terms of robustness, many kinds of
attacks from global transformation (rotation, scaling, etc) to local transformation (stirmark, checkmark
benchmarks, de-synchronization attacks) are implemented. Many image processing tools are applied to simulate
the attacks such as Print-Screen, Using Photo editing software, Camcorder, Print-Scan, etc. The experimental
results show an outstanding robustness in resisting these attacks.
Received 04 December 2015, revised 09 January 2016, accepted 14 January 2016
Keywords: Digital Watermarking, Print-Scan process, DoG, De-synchronization attacks, Camcorder.
1. Introduction*
embedding phase to encode the watermark. In
the detection phase, the same key is required to
Along with the rapid development of the
media in communication, it is important and
necessary to protect the ownership information
of digital images because illegal copying of
digital multimedia has become much easier.
Recently, many watermarking schemes
regarding copyright issue have been proposed
for digital media but few methods have been
proposed for un-digital content such as the print
and scan attack is a challenging one because it
not only alters the pixel values but also changes
the positions of original pixels. Most of the
watermarking systems use a secret key in the
decode the embedded watermark. The
watermarked content is then transmitted via a
distribution channel. In transmission process, it
may suffer some intentional as well as
unintentional manipulations (called attacks) that
try to remove or invalid the watermark. There
are two major categories of attacks:
• Unintentional Attacks: This type of
attack consists of all processes that do not
initially aim at removing or suppressing
watermark. They involve some deteriorations
due to compression (Jpeg, Mpeg, etc.), filtering,
A/D conversion, changing of coding format or
resolution, etc. that a watermarked content may
encounter through the transmission process.
________
This work is dedicated to the 20th Anniversary of the IT
Faculty of VNU-UET
49
50
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
• Malicious Attacks: These attacks aim at
error correction. In detection scheme, the
message is extracted from input image based
on autocorrelation function after filtering,
masking and adaptive line searching with
Hough transform. A Hamming coding may be
used to ensure that the message can be
decoded correctly.
Our paper is structured as follows. Section
2 introduces the related works to summarise the
existing methods. Our proposal will be
presented in section 3 and 4. In section 3, a
detailed embedding scheme is presented and the
corresponding extraction scheme is described in
section 4. In section 5, the performance of the
proposal is evaluated and discussed based on a
series of experiments. Finally, the paper ends
with a conclusion and states some potential
directions for future work.
making the watermark useless, for example,
camcorder copy, print-scan, de-synchronization
attacks or collusion attacks. However, this
group is challenging because they not only alter
the pixel intensities but also change the pixel
location. In contrast to removal attacks,
de-synchronization attacks do not actually
remove the embedded watermark itself, but
tend to make loss the synchronization between
the embedder/detector (i.e alter the location of
the watermark in the content). The watermark
still exists, but it is undetectable by the detector.
Obviously, in order to preserve robustness
of watermark, it is necessary to increase the
watermark volume but it accidentally reduces
the transparency. This raises a novel problem
which is how to tradeoff between robustness
and imperceptibility to obtain the best
watermark. Many recent physiologic researches
show that perceptual factors in HVS (Human
Visual System) could be a potential solution to
this problem. HVS modeling has become an
important issue in image and multimedia
processing such as image compression, quality
assessment as well as watermarking. Many
perceptual watermarking schemes have been
proposed [3-6].
2. Related Work
De-synchronization attacks are considered
as one of the most serious threats for any
watermarking system [8]. Therefore, many
countermeasures have been introduced in the
literature to cope with this type of attack [9]
[10],
but
a
perfect
robustness
to
de-synchronization attacks has not been
thoroughly obtained and still remains an
outstanding area of watermarking research.
De-synchronization attacks do not try to
eliminate the watermark but aim to make the
watermark undetectable although it still remains
in the content. In general, we can loosely
classify de-synchronization attacks into two
categories as below though there is no clear
distinction between them:
In this work, the proposed method deploys
a Pyramidal Just Noticeable Difference (PJND)
model in [1-2] fixed parameters are tuned as in
visual experiments by [2] to attain fidelity. The
embedding scheme is similar to the one
proposed by [7] but here, the embedding is
based on a pyramidal JND (Just-Noticeable-
Difference) model in which its strength is
controlled by a threshold and the DoG
(Difference of Gaussians) representation [1, 2].
Experimental results have proved the performance
of our approach in terms of transparency and
robustness against severe attacks from Stirmark
and Checkmark benchmarks as well as Photo
editing software manipulations, especially to
de-synchronization, print-scan and camcorder
attacks.
Global geometric distortion is
a
parametric transformation which is applied on
the whole image. All the pixels are affected in
the same manner. An example of typical global
geometric transformations which includes
affine transforms [11, 12, 13] (including
rotation, scaling, translation (RST)) and
projective transforms is given in Figure 1.
Such attacks are quite simple to apply but
challenge for current
watermarking techniques. Indeed, there is no
perfect solution for this problem, the robustness
to global affine transformations is more or less
For embedding process, we used a
template based method where watermark is
transformed in a template pattern and the
corresponding transform coefficients are used
as input message with synchronization and
really present
a
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
51
to estimate the transformation parameters.
Furthermore, local geometric attacks are
"dangerous" in the sense that they destroy the
watermark synchronization without creating
significant visual distortion because the Human
Visual System (HVS) is less sensitive to
slightly local modifications. Hence, resistance
to local random alterations like RBAs still
remains as an open problem for most of
watermarking schemes due to the high
complexity of the attack parameter space.
handled by using some approaches such as
template-based re-synchronization [14],
self-synchronizing watermarks [15] and
embedding in invariant domains [16, 17].
Geometrical
synchronization
attacks
errors
[11,
in
12] cause
watermark
detection/extraction [19, 20]. Recently, several
better approaches dealing with this type of
attack have proposed resynchronization using
additional template.
Figure 1. Example of global geometric distortion.
Left: An Affine transform,
Right: A Projective transform
Authors embedded [21] an additional
template together with the watermark in the
DFT domain. This template contains no
information about the embedded message but
could be later used to recover the
transformation undergone by the image. During
the detection phase, the template is detected
first using inverse transformation before
extracting the watermark. However, one major
drawback of this approach is that templates can
easily be detected and erased by searching local
peaks in the transform domain. Furthermore,
template-based approach seems to be robust
only to some global geometrical transforms
such as RST rather than to local geometric
distortions. It was also discussed in [22] that
local geometrical transforms such as RBAs not
only increase the search space and computation
significantly for Exhaustive Search Detector,
but also raise a serious problem for the
template-based watermarking algorithm.
Local geometric distortion involves a set
of different geometric transforms (with different
parameters) applied to different portions of the
image so that pixels are warped in different
ways. This kind of attack mainly includes
random displacement (also known as random
jitter attack, introduced in Unzign benchmark),
Random Bending Attack (RBA), also called
Stirmark attack [18] (see Figure 2) and the two
recently reported Desynchronization Attacks
proposed by [8]. In the case of random attack, it
is almost impossible to estimate the
transformation parameters.
Another alternative approach in [11, 12]
[18] inserted a periodic matrix brand in the
DWT domain. The estimation of the geometric
attacks is evaluated based on the brand
Figure 2. Example of local geometric distortion.
Left: Original Image, Right: Local Random
Bending Transform from Stirmark
autocorrelation
function
to
obtain
autocorrelation peaks. If the image is attacked
by geometric operations, this plane will
undergo the same attack. Indeed, the correlation
or cross correlation of a signal by itself detects
repeated patterns in a signal as a periodic signal
is disturbed by a lot of noise. So, thanks to the
periodicity of the brand, the brand's
autocorrelation function locates periodic peaks.
The mark detector then estimates the
Since the parameters needed to describe the
local geometrical transformation are normally
much more than those needed for global
geometrical transformation, resynchronization
from local geometrical distortion is much more
difficult than from the global one. In the case
that the attack is random, it is almost impossible
52
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
geometrical transformation performed on the
image with reference to the plane of the peaks
of the extracted mark. The initial state of the
image can also be reconstructed.
3. Our Watermark Embedding Scheme
In our proposal, the watermark is embedded
into an original image based on a pyramid JND
model in order to improve the robustness in
countering some attacks on images as well as
preserving perceptibility. For the sake
of simplicity, we consider only the additive
embedding scheme. The detailed diagram of the
watermarking method is shows in Figure 3; the
order of operations involved is depicted in the
following section.
Among other various attacks the print and
scan attack is challenging. The printed images
are first converted to digital format (scanning)
and the information is extracted from the
detected copies; watermark image needs to
convert analog data into a digital format in
database applications. Both the printing and
capturing processes cause various attacks
including A/D or D/A conversion, compression,
quantification, dithering, filtering, blurring,
sampling, noise adding, contrast enhancement,
etc. Some of geometrical attacks and distortions
caused by devices become challenging
problems in watermark extracting. In [7], a
template based watermarking embedded multi-
bit messages into image spatial blocks using
periodic patterns, while one block embeds
synchronism information. The embedding
scheme is similar to the one proposed by [7] but
this is done as follows:
○ In contrast to the above methods, we
embed the watermark and JND mask performed
into different scales of the DoG scale space,
hence reducing the complexity of the method.
This model takes into account three main
characteristics of the Human Visual System
(HVS), namely: contrast sensitivity, luminance
adaptation and contrast masking. In [7] only
with luminance does not consider neither
contrast sensitivity, contrast masking and
informed coding or the color channel
properties. Furthermore, to ensure transparency,
the embedding strength is determined using the
pyramidal JND model proposed in [1] [2] and
adapted here for the scale space transform.
○ Both embedding and extraction are
adaptive, with no need to change parameters
settings for different images. It is shown that
3.1. Computing the Pyramidal JND map
The JND model takes into account only the
luminance channel. Hence, in order to apply for
color images, the image is first transformed into
YCbCr color space, and then only the Y
component is watermarked. The JND model in
our model is similar to one in [1] [2]. The input
image is first decomposed into DoG scale
images; the Gaussian image is replaced by the
original image and the JND map is then
computed for each DoG level. In each level of
the scale space, a JND map is computed by
incorporating the most relevant HVS’s
properties such as contrast sensitivity function
(CSF), luminance adaptation and contrast
masking. Fixed parameters are tuned as in
visual experiments by [2] to attain fidelity.
The image and the JND image are then
divided into blocks, and several bits are
embedded in each block. The following section
explains the message encoding and embedding.
3.2. Preparing the message
3.2.1 Hamming encoding message
The watermark is read and processed block
by block, and the watermark capacity depends
on the number of bits embedded in each block.
In our experiments, the image was divided into
sixteen blocks and four bits in each block are
used for watermarking. Thus, the watermark
capacity is of 16x4=64 bits. In this application,
we used a (64, 57) extended Hamming code.
This is an extension of the original (63, 57)
Hamming code by adding an additional
redundant bit. A Hamming code [23] is used to
background variation, or
a
change of
printer/print material (two printers, two
materials) has no significant effect on the
performance of the method.
○ Completing the features balance between
the three categories for a watermarking system
namely transparency, robustness and capacity.
○ The message is protected with extended
Hamming (64, 57) error correction coding that
is capable of correcting three bits.
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
53
add redundancy to the bits so that the errors can
be detected or corrected to a certain extent.
Hamming code is a linear block code. The main
advantage of linear block codes is the simplicity
in implementation and low computational
complexity. A linear block code is usually
composed of two parts. The first part contains
the information bits, the original bits to be
transmitted. The second part contains the parity
checking bits, which are obtained by summing
over a subset of the information bits. A linear
block code with length n and k information bits
is denoted as a (n; k) code.
codeword to get even parity error correction
coding that is capable of correcting one bit or
detecting three erroneous bits. Each Hamming
coded sequence of message is transformed into
rotation angles by assigning each sequence a
value between 0 to 180 degrees. The value is
chosen by quantizing the rotation angles and
assigning each of the values a number of bits,
as illustrated in Figure 4. The quantization
angle is determined by equation (1):
180
2m
α
=
(1)
The embedded message is protected with an
extended Hamming (64,57) which is
where m is the number of bits embedded in
constructed by a parity bit at the end of the
each of the blocks.
H
Create message
10101011001001010…
Original Image
Hamming encoding message
(64,57)
Convert a block 4-bit message into degree Eq. (1)
Generate
pseudorandom
pattern
Divide Original Image and PJND
image into 16 blocks
Rotate patterns according to the rotation
angles. Eq. (2)
…
Embed watermark block
Eq. (4)
Figure 3. Illustration of watermark embedding process.
54
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
- Three matrices are all shear matrices.
- The first and the last matrices are the
same.
- The determinant of each matrix is 1.0
(each stage is conformal and keeps the area
the same).
Figure 4. Creating encoding table.
As the shear happens in just one plane at a
3.2.2. Generating pseudorandomtemplate pattern
The patterns are formed for each block.
time, and each stage is conformal in area, no
aliasing gaps appear in any stage.
Each pattern is formed by first repeating a small
rectangular pseudorandom sequence until the
sequence covers an area the size of the block. A
3.3. Embedding rule
We evaluate the smoothness of the image
area in each block instead of 16×16 sub blocks
[7] using average gradient magnitude on an
image sharpened with an un-sharp mask. We
use linear relationship between β and the
average gradient magnitude to place more
watermark strength on textured blocks
according to equation:
bipolar random sequence W
∈ {-1,1} with mean
zero and variance one is generated size 64x64.
Note: in order to get good results, the size
of W and the size of one block original image
(in this original photo 1/16) should be close
together, the closer together the better. The
reason is interpolation image will not affect
many temples pattern.
MY
β =
(3)
When the direction of the template pattern
is detected and the direction is erroneously
interpreted to the adjacent quantization step, the
Hamming coding ensures that the message can
be decoded correctly because only one bit
changes between adjacent quantization steps.
Each pattern is then rotated according to the
message and cut to the size of the block. The
process does not affect periodicity. The pattern
is embedded in the image block.
M JND * MW
In the equation (3) MY, MJND and MW,
respectively stand in average gradient
magnitude on original image, Compute
Pyramidal JND and interpolate watermark
blocks. The watermark is directly embedded in
the Pyramidal JND, in which its strength is
controlled by using the JND threshold in [2].
By this way, salient regions tend to mask non-
salient regions. JND threshold is hence
modulated by two masking mechanisms: the
contrast masking and the "saliency masking".
Recent JND models [7] do not take into account
this phenomenon and therefore do not
completely exploit HVS limitation.
3.2.3. Rotating template pattern
It is called the three shear rotation
method. The heart of this method is the
expansion of the single 2D rotation matrix
The embedding of the message in the host
image is realized in spatial domain utilizing the
equation:
into three matrices [24]:
x*
(2)
Wο
=
y *
i
*Wθ (x, y))
1
− tan(
θ
/ 2)
1
0
1
1
0
− tan(
θ
/ 2)
x
Ii (x, y) = Y (x, y) + JND (x, y).*(
β
i
i
i
=
0
1
sin
θ
1
y
(4)
where Ii is the ith watermarked block of the
image, Yi is the original image, JNDi is Multi-
scale JND Map and the Wiθ represents the
Here, bilinear interpolation is used to resize
image. Parameter θ defines the rotation angle as
in equation (1) of the Hamming coded bit
sequence through the results in Figure 4.
- There are some very interesting properties
of these three matrices:
coded watermark information in the form of
directed template pattern.
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
55
printing industries for crispening edges [25]. A
signal proportional to the un-sharp or low pass
filtered version of the original noisy image is
subtracted from the image so that the resulting
4. Watermark Detection Process
Print and capture involve several
distortions, because of the user interaction, the
devices and air interface, which are taken into
account in designing watermark extraction
algorithm. Our detection scheme is shows in
Figure 5.
First, the captured image is downsampled
by using bilinear interpolation. It was necessary
to compromise between the high processing
time and the amount of data processed. After
capturing a picture with a scan, camcorder the
captured image is divided into blocks. The
existence of a watermark is processed. The
message is read by processing the blocks.
image is a crisp high-contrast image [26].
~
For each block, a W i
estimation of the template watermark structure
is calculated
(x, y) un-sharp filter
:
~
W i (x, y) = Y * (x, y) −Y *smooth (x, y)
(5)
is the ith watermarked block
is a low pass filtered version
i
i
where Y*
(
,
x
,
y
)
i
andY*
(
x
y
)
i
smooth
of represents the adaptive Wiener filtering
Y*
(
x
,
y
)
. Bilinear interpolation is used to
i
~
resize watermark W i (x, y) of template pattern.
The autocorrelation function (ACF) is
calculated of the Sharpening estimate, and this
operation doubles the size of the processing
block. Autocorrelation function utilized in order
to reveal the periodicity in the extracted
watermark estimate can be calculated as:
~
~
R~ (a,b) = (W i (x, y)W i (x + a, y + b))dx
~
∫
W i W i
(6)
The autocorrelation is scaled to the range of
[0,1]
R* (a,b)
~
~
W
W
i
i
R* (a,b) =
(7)
max(max(R* (a,b))
~
W
~
W
i
i
~
W
~
W
i
i
The enhanced autocorrelation peaks are
then thresholded, and a binary grid is formed
with equation:
*
1, when M (a,b)xR (a,b) ≥ T
~
W
~
W
i
i
G* (a,b) =
0, when M (a,b)xR* (a,b) < T
~
W
~
W
i
Figure 5. The Detection Scheme.
i
(8)
The un-sharp filter is a simple sharpening
operator which derives its name from the fact
that it enhances edges (and other high
frequency components in an image) via a
procedure which subtracts an un-sharp, or
smoothed, version of an image from the
original image. The un-sharp filtering technique
is commonly used in the photographic and
where M(a, b) is a circular masking
operation. The central area of the
autocorrelation image contains noise, which
causes errors in line detection. Therefore, the
center of the grid is masked out.
Finding edges in intensity image: edge
takes an intensity or a binary image G*
(a,b)
56
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
as its input, and returns a binary image BW of
the same size as G*
, with 1's where the
function finds edges in G*
and 0's
5. Experimental results and discussion
(a,b)
To validate the performance of our method
in terms of robustness and imperceptibility,
some experiments are carried out on 512x512
scale images. We test the robustness of
watermark with some common attacks namely:
JPEG compression, Gaussian noise, cropping,
and low-pass filtering.
(a,b)
elsewhere. The Sobel method finds edges using
the Sobel approximation to the derivative. It
returns edges at those points where the gradient
of G*
(a,b) is maximum.
The peaks are aligned according to the
direction of the pseudorandom sequence pattern
and this alignment is detected with Hough
transform and line detection. These detected
lines then give the angle of the pattern and thus
the message.
The JND model takes into account only the
pixel luminance. Hence, in order to apply for
color images, the image is first transformed into
YCbCr color space and then, only the Y
component
is
watermarked.
Further
experiments are also carried out on a variety of
natural images to validate the performance of
our method in terms of robustness and
imperceptibility. Due to the limited space and
in order to facilitate the comparison, we only
report the results for a set of 10 images, each of
which contained 57 bits and error coding.
Lines are searched from the binary grid
using Hough Transform. The dominating
direction is found by evaluating the number of
peaks allocated to the same bin in the Hough
transform matrix. Due to the properties of the
Hough transform, the possible false peaks in the
autocorrelation function have little or no effect.
However, it is important to locate as many of
the correct peaks as possible for reliable
determination of the angle. These peaks are
then presented to Hough transform as input to
find the dominating direction formed by these
peaks and thus giving an angle. Obtained angles
are decoded to message nibble using the
encoding equation (1).
5.1. Transparency Evaluation
The results in Figure 6 show that the
proposed algorithm provides
a
good
imperceptibility the results in Table 2, at the
same perceptual quality, the better the model
[7] by subjective test. The proposed JND
estimator has been compared with JND models
[7] in Figure 7.
The order of operations, to extract the
message from a captured image, is presented in
Figure 5. The message is extracted by analyzing
the autocorrelation peaks of the embedded
template pattern. Hough transform is used to
detect the angle of lines made by these
autocorrelation peaks. It helps in correctly
identifying the aligned peaks as the small errors
in the detection, due to the misleading peaks
that appear due to thresholding, are minimized
because of the robust properties of Hough
transform at the watermark detection side.
The process is repeated for each block and
the angle information is quantized. The same
quantization step size and encoding table as
during embedding is used. Each quantized
angle value represents a coded bit sequence
which is interpreted using a coding table and
decoding. Finally, Hamming (64, 57) error
decoding is used to decode the message.
Although subjective assessment approach is
the appropriate and accurate solution to
watermark transparency evaluation, it is usually
inconvenient, expensive and time-consuming,
and not always easy to use. These drawbacks
have led to the use of objective assessment as
an alternative method. The goal of objective
quality evaluation is to assess the quality of
image/video by means of an automatic tool
(objective metric) without performing any
subjective test. In this paper, we have
investigated the performance of objective
assessment is also done HVS inspired quality
metrics, the SSIM (Structural Similarity Index
Measure) [27], the so-called Watson metric [28]
which can measures the Total Perceptual Error
[29] (TPE) between the original and the
watermarked image and a wavelet-based metric
(PSNR_wav1 and PSNR_wav2) [30], Peak
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
57
Signal-to-Noise Ratio (PSNR) comparision for
Images Gray , weighted peak signal-to-noise
ratio (wPSNR)[28][29] using weighted signal-
to-noise ratio (wSNR)[31], Singular value
decomposition (SVD)[32]. These metrics have
been designed to general image quality
assessment, it is therefore necessary to study
their performances to the specific purpose of
watermark transparency assessment. As
mentioned in the first section, there is no
objective metric specifically designed for
watermarking purpose. It would be then very
useful if we can determine, amongst the
existing metrics, the one which is the most
appropriate
for
watermark
transparency
assessment. The results in Table 2 show that the
algorithm provides an excellent imperceptibility.
However, the objective measures do not fully
correlate with the subjective evaluation. It is
mainly due to the variation of visual image
content. Hence, a specific quality metric for
watermarking is still missing.
G
T
(a)
Original image
(b)
Figure 6. Original image middle and Watermarked image: Proposal left (a), Model [7] right (b).
Proposal
Model [7]
Figure 7. JND maps of different models.
Experimental results are reported in Table
2. It can be seen that the proposed model yields
slightly lower as PSNR, PSNR_wav1,
PSNR_wav2, SVD, wPSNR, wSNR scores than
[7] for most of test sequences. However, we can
see that MSSIM scores of the proposed model
are higher than [7], which means that the
proposed JND model not only conceals more
distortions but also achieves a better quality. To
demonstrate the invisibility of the watermark
and the advantage of our method with method
[7], we use the Structural Similarity Index
Measure (SSIM) metric, [27] proposed a multi-
scale version of SSIM (MSSIM) where the
images are low-passed filtered and down
sampled by a factor of two and the contrast and
the structure are computed for each sub-
sampled level for evaluating the quality of
watermarked image of different 512x512 of 10
images. It works under the assumption that
human visual perception is highly adapted for
extracting structural information from a scene.
The SSIM index is based on a combination of
luminance, contrast and sensitivity of the
58
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
watermarked image with the original. The
comparisons are performed on local windows;
the overall image quality is averaged on these
local windows. SSIM has become a very well-
known metric for perceptual image quality
assessment and has been extended in various
directions. Our results are reported in the
following figures and in Table 2. As shows in
Figure 6, the watermarked image and the
original are visually undistinguishable.
As for the watermark invisibility, quality
results of watermarked image of different
methods for different images are reported in
Table 2. We can observe that the quality of the
watermarked images of our method is
equivalent or even better than that of other
method [7].
C-LPCD (Constraint LPCD), MF (Markov
Random field). Watermark detection results are
shown in Table 3.
Table 1. Robustness Evaluation (Stirmark [33] and
Checkmark [34] benchmarks attack)
Explicit
scheme
Attack Type
Method [7]
Random Cropping
Jpeg compression
1%
0.8%
QF=9%
0.1 bpp
σ= 67%
Ok
QF=3%
Jpeg 2000 compression 0.08 bpp
Gaussian Noise
σ= 64%
Wiener filtering
Median filtering
Sharpening
Ok
5x5
Ok
Ok
Ok
Ok
Ok
Ok
3x3
Failed
Failed
Failed
Ok
Blurring
Bit plan reduction
Histogram Equalization
Rescale (45%)
Affine Transform
5.2. Evaluation of robustness
Ok
The robustness of our algorithm is tested
via a wide range of attacks including “signal
processing” and de-synchronization types (DA).
Some of them are very severe attacks like Print-
Scan, Print Screen, Camcorder attack, Using
Photo editing software and new de-
synchronization attacks developed in [8] values
are shown in Table 3. To facilitate the task,
there are various tools that can test and evaluate
Ok
Print-scan attack: this attack consists of
printing image on a classical laser printer: HP
LaserJet 4250 PCL6, EPSON Stylus. Scanning:
Epson Perfection 4490. The image is printed in
color, grayscale level on an A4 paper at 300 dpi
resolution (tests were done on image printed on
a white paper) and scanned, witch is shows in
Figure 8. Watermark detection results are
shown in Table 3.
watermarking
algorithms
systematically.
Among them, the following two tools are most
known to Stirmark [33] and Checkmark [34]
benchmarks. However, when regarding the
results in [7], we can see that the proposed
method has a nearly equivalent robustness for
geometric attacks values shown in Table 1.
Furthermore, it resists other specific attacks
(Camcorder, Print Screen, Using Photo
editing software, new DAs…) that the method
[7] cannot.
Camcorder attack: we get the picture of
image on the computer screen with the Nikon
D90 Digital SLR Camera with 18-105mm VR
Lens Kit (12.3MP) 3inch LCD. The
watermarked test images were captured 10
times with each of the camera setting and each
image contained 57 bits and error coding when
images were captured by tilting the camera
randomly is shows in Figure 9 and values are
shown in Table 3.
New DAs: these de-synchronization attacks
are an extension of classical geometric attacks
proposed by [8]. They are proved to be more
powerful and less intrusive than the Stirmark
attack. We tested three types of these with
default parameters as in [8]: the LPCD (Local
Permutation with Cancellation and Duplication,
Using Photo editing software: Do you still
use Microsoft Paint, or some other under-
powered paint packages that allow you to rotate
an image by an arbitrary angle (Figure 9).
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
59
Table 2. Imperceptibility Evaluation
Image
Objective Method
AVG
Baboon Barbara
Boat
Car
Clown
Fruit
Isabe
Lena
Peppers
Plane
Keskinarkaus
25,79
26,10
9,36
28,41
27,88
11,64
10,48
12,43
10,87
34,82
29,64
0,10
33,01
32,80
33,66
36,68
36,18
35,45
35,51
35,82
33,33
32,39
16,36
14,71
17,20
15,44
16,79
14,98
0,09
PSNR
Proposed
32,46
15,96
14,66
16,33
14,87
16,47
13,94
0,08
31,98
16,93
15,35
17,69
15,95
17,35
15,20
0,08
32,98
17,13
15,54
18,13
16,45
12,51
11,18
0,08
34,84
19,25
17,02
19,68
17,53
9,74
35,25
16,78
15,15
18,56
16,90
9,28
34,47
17,72
15,92
19,11
17,14
10,70
9,88
34,35
19,16
16,97
19,99
17,75
8,07
33,61
19,65
16,69
20,03
17,08
11,10
10,82
0,06
Keskinarkaus
Proposed
PSNR
wav1
9,31
Keskinarkaus
Proposed
10,08
9,88
PSNR
wav2
Keskinarkaus
Proposed
37,82
33,36
0,19
SVDm
TPE
9,65
8,27
7,87
Keskinarkaus
Proposed
0,07
0,07
0,07
0,07
0,19
0,13
0,08
0,09
0,10
0,08
0,07
0,07
0,08
0,07
0,10
Keskinarkaus
Proposed
0,82
0,88
0,93
0,94
0,93
0,97
0,94
0,94
0,91
0,96
0,92
mssim
wPSNR
wsnr
0,85
0,89
0,94
0,94
0,94
0,97
0,95
0,94
0,93
0,96
0,93
Keskinarkaus
Proposed
35,56
35,99
33,18
32,16
37,28
36,93
34,85
32,08
39,25
38,97
39,28
37,92
38,34
37,80
38,09
36,41
38,62
37,90
34,50
33,71
40,46
39,60
37,17
35,39
39,98
39,24
40,58
38,82
39,76
39,31
38,69
37,05
38,95
38,80
39,77
37,17
40,93
39,46
41,39
38,08
38,91
38,40
37,75
35,88
Keskinarkaus
Proposed
Table 3. Robustness Evaluation. For some type of attacks, the results showed: X/Y (bit error/bit encoded
message) the parameters demonstrate the break-down limit of the method
(the strongest attack to which the watermark still survives)
Attack
Method
Baboon
Fruit
Ok
-
Isabe
2/64
-
Lena
1/64
-
Peppers
Ours
Ok
-
3/64
-
Camcorder attack
Keskinarkaus
Ours
Ok
Ok
Ok
-
3/64
2/48
Ok
-
Ok
3/48
2/64
-
2/64
Ok
3/64
-
Ok
Ok
Ok
-
Print scan Attack
Photo editing software
Print screen Attack
DA New
Keskinarkaus
Ours
Keskinarkaus
Ours
Ok
-
Ok
-
Ok
-
Ok
-
Ok
-
Keskinarkaus
Ours
Ok
-
Ok
-
Ok
-
Ok
-
Ok
-
Keskinarkaus
60
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
f
Table 3 shows the average robustness tested
for five images (Baboon, Fruit, Isabe, Lena and
Peppers). These values denote the breakdown
limit of the tested methods, i.e. the strongest
level of attacks to which the watermark still
survives. Table 1 shows that the watermark
survives many severe attacks in both schemes
but there are no significant differences in
robustness between these two schemes (except
for Jpeg compression). Furthermore, robustness
against some attacks "like Jpeg" (Jpeg2000) is
Watermark Image Insert
Document and Print
Scanner at 300dpi
resolution
even slightly improved.
Watermark detection results are shown in
Table 1 and Table 3; our method outperformed
the method [7] for most attacks. Furthermore,
the message protected with Hamming (64, 57)
error correction coding that is capable of
correcting three bits ensures that the message
can be decoded correctly. Especially, in contrast
to [7], it survives many severe attacks such as
"camcorder", "print-scan" and Stirmark,
Checkmark and new DA. However, our method
as well as the method [7] are not very robust to
"signal processing" attacks such as noise, jpeg
compression, etc. Throughout these results, it is
clear that using perceptual models helps
improve not only transparency but also
robustness of a watermarking system. The
explicit scheme, once again provides the best
robustness amongst the compared methods. The
detector outputs for some severe attacks are
also displayed in Figure 8, 9 and 11.
Figure 8. The watermark image
is printed and scanned.
Watermark Image screen with
the Nikon
Crop image and
Detection
Figure 9. The watermark screen with the Nikon.
We user vertical shear and skew the image
a negative number of degrees (1-6 degrees) in
the vertical plane which is shows in Figure 10
and values are shown in Table 3.
6. Conclusion
In this paper, we have presented a novel
content based image watermarking operating in
the DoG scale space with enhancing robustness
against de-synchronization attacks. Such
watermarking methods present additional
advantages over the published watermarking
schemes in terms of detection and recovery
from geometric attacks, and with better security
characteristics. The experimental results show
Vertical shear 5 degrees
Vertical shear 6 degrees
Figure 10. The watermark attack Using Photo
editing software.
Print screen Keyboard: When you press it,
an image of your screen is copied to the
Clipboard. This is called a screen capture or
screen shot. You will then need to further edit
using some image editing programs values
shown in Table 3.
that the proposed method has
performance in terms of robustness and
imperceptibility. In the future, this method
a
good
digital watermarking will be exten
mobile phones.
đe to used on
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
61
robust to geometric distortions,” IEEE Trans. on
Image Processing, vol. 14 (2003) 2140.
References
[12] M. Alghoniemy and A. H. Tewfik, “Geometric
invariance in image watermarking,” IEEE
Trans. on Image Processing, vol. 13, no. 2
(2004) 145.
[1] Luong Viet Nguyen, Trinh Nhat Tien, Ho Van
Canh. “Pyramidal JND Model for Grayscale
Image and its Application to Watermarking”.
Proceedings of IEEE International Conference
on Computer Science and Automation
Engineering Vol.01[C], 2013.
[13] Zhao, Yao, RongRong Ni, and ZhenFeng Zhu.
“RST transforms resistant image watermarking
based on centroid and sector-shaped
[2] Luong Viet Nguyen, Trinh Nhat Tien, and Ho
Van Canh. “A watermarking method robust for
copyright protection of images against Print-
scan.” Information Science and Control
Engineering (ICISCE), IET International
Conference on. IET, 2012.
partition.” Science China
Sciences 55.3, 650-662, 2012.
Information
[14] Wang, Caiyin, and Chao Li. “A Steganography
Approach for Printed Image Based on Image
Complexity and Template Matching.” Open
Automation and Control Systems Journal 6
(2014) 84.
[3] Ensaf Hussein, Mohamed A. Belal, “Digital
Watermarking Techniques, Applications and
Attacks Applied to Digital Media: A Survey,”
IJERT, ISSN: 2278-0118, Vol. 1 Issue 7, 2012.
[15] Schlauweg, Mathias, et al. “Self-synchronizing
robust Texel watermarking in Gaussian scale-
space.” Proceedings of the 10th ACM workshop
on Multimedia and security. ACM, 2008.
[4] S.
Voloshynovskiy,
A.
Herrigel,
N.
Baumgartner, and T. Pun, “A stochastic
approach to content adaptive digital image
watermarking,” in Proc. of the 3rd International
Workshop on Information Hiding, 1999.
[16] Joseph J. K., O’ Ruanaidh, J. J. K. et Pun, T.
“Rotation, Scale and Translation Invariant
Digital
Image
Watermarking”,
Image
Processing. International Conference, Issue, 26-
29 (1997) 536.
[5] Mundher, Myasar, et al. “Digital watermarking
for images security using discrete slantlet
[17] Wang, Xiang-Yang, et al. “A new robust digital
watermarking based on exponent moments
invariants in nonsubsampled contourlet
transform domain.” Computers & Electrical
Engineering 40.3 (2014) 942.
transform.” Applied Mathematics
Information Sciences 8.6: 2823-2830, 2014.
and
[6] Kekre, H. B., Tanuja Sarode, and Shachi Natu.
“Performance of watermarking system using
wavelet column transform under various
attacks.” International Journal of Computer
Science and Information Security 12.2 (2014).
[18] M. Mitrea, F. Preteux, M. Petrescu, and A.
Vlad, “The Stirmark watermarking attack in the
dwt domain,” in Proceedings of the 12th IEEE
International Conference on Systems, Signals
and Image Processing (IWSSIP’05), Halkida,
Greece, pp. 5–9, 2005.
[7] A. Keskinarkaus, A. Pramila, T. Seppänen,
“Image watermarking with a directed periodic
pattern to embed multibit messages resilient to
print scan and compound attacks.” the Journal
of Systems and Software v. 83 (2010) 1715.
[19] Wang, Xiang-yang, et al. “Robust image
watermarking approach using polar harmonic
transforms based geometric correction.”
Neurocomputing, 2015.
[8] Barni, Mauro, Angela D'Angelo, and Neri
Merhav. “Expanding the class of watermark de-
synchronization attacks.” Proceedings of the 9th
workshop on Multimedia & security, ACM,
2007.
[20] N. Merhav, “An information-theoretic view of
watermark embedding-detection and geometric
attacks,” in Proceedings of WaCha’05,
Barcelona, Spain, 2005.
[9] Wang, Xiang-Yang, and Chang-Ying Cui. “A
novel image watermarking scheme against
[21] D. Zheng, J. Zhao, and A. Saddik, “Rst invariant
digital imagewatermarking based on log-polar
mapping and phase correlation,” IEEE Trans. Circ.
Syst. Video Tech., vol. 13 (2003) 753.
desynchronization
revision.” Journal of Visual Communication
and Image Representation 19.5(2008) 334.
attacks
by
SVR
[10] Schmitz, Roland, et al. “Towards Robust
Commutative
Based
[22] S. Pereira and T. Pun, “Robust template matching
for affine resistant image watermarks,” IEEE
Transaction on Image Processing, vol. 9, no. 6,
June (2000) 1123.
Invariant
Encryption
Watermarking-
on
Journal
Image
of
and
Histograms.” International
Multimedia Data
Management 5.4 (2014) 36.
Engineering
[23] W. Peterson., “Error-correcting codes,” 2nd ed.,
Cambridge: The MIT Press, 1980. 560p.
[11] P. Dong, J. G. Brankov, N. P. Galatsanos, Y.
Yang, and F. Davoine, “Digital watermarking
62
L.V. Nguyen et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 32, No. 2 (2016) 49-62
[24] Toffoli, Tommaso, and Jason Quick. “Three-
Workshop on Information Hiding, pp. 219-239,
Pittsburgh, PA, USA, 2001.
dimensional rotations
shears.” Graphical Models
Processing 59.2, 89-95, 1997.
by
and
three
Image
[29] Pereira, S.: “Robust digital image watermarking”,
PhD thesis, Genève, Swiss, 2000.
[25] Jain A., “Fundamentals of Digital Image
Processing,” Prentice-Hall, France, 1989.
[30] Beghdadi, A. and Pesquet-Popescu, B.: “A New
Image Distortion Measure Based on Wavelet
Decomposition.” In Proc. of 7th IEEE ISSPA,
vol. 2, pp. 485-488, Paris, France 2003.
[26] Amirgholipour S. and Naghsh-Nilchi A., “A
New Robust Digital Image Watermarking
Technique Based on Joint DWT-DCT
Transformation,” in Proceedings of the 3rd
International Conference on Convergence and
Hybrid Information Technology, Busan, vol. 2
(2008) 539.
[31] Niranjan, D. V., Thomas, D. K., Wilson, S. G.,
Brian, L. E., Bovik, A. C.: “Image quality
assessment based on a degradation model”, In
IEEE Trans Image Processing, 9(4) (2000) 636.
[32] Shnayderman, R., Gusev, E., Eskicioglu, A. M.:
“An SVD-based gray-scale image quality
measure for local and global assessment.” In
IEEE Transactions on Image Processing, 15(2)
(2006) 422.
[27] Wang, Z., Bovik, A. C., Sheikh H. R.,
Simoncelli, P. E.: “Image quality assessment:
from error visibility to structural similarity.”
IEEE Transactions on Image Processing, 13(4),
pp. 600-612, 2004.
tirmark/
[28] Pereira, S., Voloshynovskiy, S., Madueño, M.,
Marchand-Maillet, S., Pun, T.: “Second
generation benchmarking and application
oriented evaluation.” In Procs of 3rd Int.
html
14 trang
13080
Free
Bạn đang xem tài liệu "A watermark algorithm against de-synchronization attacks", để tải tài liệu gốc về máy hãy click vào nút Download ở trên
File đính kèm:
- a_watermark_algorithm_against_de_synchronization_attacks.pdf
