A Digital Image Watermarking Scheme Based on Visual Cryptography

Amir Houmansadr*, Shahrokh Ghaemmaghami**

*
Electrical Engineering Department, Sharif University of Technology, Azadi St., Tehran, Iran
**
Electronic Research Center, Sharif University of Technology, Azadi St., Tehran, Iran
Emails: [email protected], [email protected]

lawbreakers, unintentional modifications caused by

Abstract coding and compression that are made to the digital
An image watermarking scheme based on the media prior to transmission and/or storage, and errors
visual cryptography is proposed in this paper. The occurred during the transmission of the media through
owner’s mark which can be a visible logo is split the networks.
into two share-images using a visual secret sharing In this paper, we propose a novel image watermarking
scheme. One of the shares is embedded into the scheme, based on the concept of visual cryptography.
host image whereas the other is kept as the A digital watermark (which can be the visible logo of
system’s secret-key to be used in the watermark the owner) is split according to a visual secret sharing
detection process. In case the inserted watermark scheme [2], [3]. One of the shares is inserted into the
is a visible logo, the mark’s existence can be image and the other serves as the detection key.
verified just by the human’s eye. Experimental
results show that the proposed scheme is robust In section 1, we describe the concept of visual
against major watermarking attacks. cryptography. Sections 2 and 3 explain the proposed
watermark insertion and detection schemes,
Keywords: digital image watermarking, visual respectively. Section 4 is dedicated to experiments
cryptography. and discusses the experimental results.

Introduction 1. Visual Cryptography

The extreme development of internet has made the In 1979, Blakley and Shamir developed the concept of
transmission, distribution and access to digital media secret sharing independently [4], [5]. A secret sharing
very convenient. So, media producers are more scheme shares a secret into a number of shares so that
frequently dealing with illegal and unauthorized usage the cooperation of a predetermined group of
of their productions. shareholders reveals the secret whereas the secret
reconstruction is impossible for any unauthorized set
Over the last two decades, digital watermarking has
of shareholders. Visual cryptography is a kind of
been addressed as an effective solution to safeguard
secret sharing in which the secret reconstruction can
copyright laws and an extensive effort has been made
be done only by the human visual system [2]. This is
to design robust watermarking algorithms [1].
why it is also called visual secret sharing (VSS).
Basically, a digital watermark is an invisible mark
that is inserted into a digital media such as audio, Many VSS schemes have been proposed from which
image, or video to identify illegal distributions of we consider the (2,2) VSS scheme proposed by Naor
copyright protected digital media and also et al. in [2], [3]. According to the algorithm, each
lawbreaking customers. A digital watermark must pixel of the binary-valued secret image is expanded
have special features to achieve functionalities into 2*2 pixels, as shown in table 1. To share a white
desired. The embedded mark should be robust enough pixel of the secret image, one row from the first 6
against various watermarking attacks, while keeping rows of table1 is chosen randomly. Similarly, the two
the perceived quality of the host signal unchanged shares of a black pixel are determined by a random
(the imperceptibility requirement). Watermarking selection from the 6 last rows of table 1. As a result,
attacks consist of deliberate attacks made maliciously an M*N pixels secret image is expanded into two
to remove or change the mark sequence by 2M*2N pixels share-images.
 Table 1 A (2x2) VSS scheme using 2x2 subpixels. In the next sections, the mentioned (2,2) VSS scheme
is utilized in the proposed watermarking algorithm.

2. Watermark Insertion Scheme

In this section, we describe the proposed watermark
insertion algorithm which exploits the mentioned
(2,2) VSS.

Figure 2 Block diagram of the proposed watermark insertion

scheme.

Fig. 2 shows the block diagram of the watermark

embedding algorithm which is performed in the
spatial domain of the host image. We transform the
watermark (binary logo) and all the shares from
binary (0,1) to signed format (-1,+1), which leads an
Considering security of the method, presence of only approximately zero-mean pseudo-random watermark
one share image reveals nothing about the sequence, so we can proceed with the detection
corresponding secret image, i.e., each 2*2 pixels process in a correlation-based manner. Accordingly,
block of one share-image may correspond to either a white and black pixels are represented by 1 and -1,
white pixel or a black pixel of the secret image. As respectively. The watermark, which can be the visible
table 1 shows, stacking the shares of a black secret logo of the owner, is split into two pseudo-spread
pixel results in 4 black subpixels, whereas only 2 spectrum shares according to the mentioned VS
black subpixels is gained by stacking shares of a scheme. One of the gained shares is then added to the
white secret pixel. So, secret image is revealed to host image to obtain the watermarked image as below:
human eyes by stacking the shares without
performing any cryptographical computations. F = IM + µ * SH1 (1)
where SH1 is the first share of the mark , IM is the
digital host image, F is the watermarked image, and µ
is a parameter defining the mark’s power which must
be chosen so that the best trade off between
watermark robustness and watermark imperceptibility
is made. We set µ equal to two in our simulations.
Fig. 3 compares an image watermarked using the
proposed watermarking scheme and the
unwatermarked host image. As shown, the watermark
in the marked image is imperceptible.

Figure 1 Stacking two share images of a secret image. (Top-

left) first share, (top-right) second share, (bottom-left) stacked
shares without reduction and (bottom-right) reducing stacked image
which is the same as split secret image (fig. from [1]).
Figure 3 Imperceptibility of the proposed method. (Left) original
image and (right) watermarked image.
Figure 1 shows the result of superimposing the share-
images of a secret image. Original secret image can be
obtained using a simple reduction algorithm from the
superimposed image.
3. Watermark Detection Scheme L(i, j ) = FST (2i − 1, 2 j − 1) + FST (2i − 1, 2 j )
(5)
Though watermark detection can be performed in an + FST (2i , 2 j − 1) + FST (2i , 2 j )
informed structure (by using the unwatermarked host The reduction block decides whether the secret pixel
image at the detector), the blind detection scheme is corresponding to each 2*2 block of FST is a black
explained in this paper. The block diagram of the pixel or a white pixel, according to the last column of
watermark detecting algorithm is depicted in fig. 4. table 1. Due to the fact that watermarked image F
The inserted mark is a high frequency pseudo-spread contains the first share of the logo, running the stack
spectrum sequence. So, we get the received image function and the reduction function using the two
passed through a high-pass filter (HPF) in order to 2M*2N-pixels images FHL and the second share of the
weaken the low frequency components which more logo SH2 produces an M*N-pixels image which is
correspond to the unwatermarked host image. To gain highly correlated by the watermark logo.
a binary valued image, the filtered image is hard
limited by comparing every pixel to a threshold, TH, The watermark can be detected due to the correlation
as: between the gained image and the watermark. The last
block of the watermark detection scheme of fig. 4
⎧1 if FF (i, j ) > TH computes the correlation between this image and the
FHL (i, j ) = ⎨ (2)
inserted watermark, as:
⎩− 1 otherwise
U •W (6)
where FF (i,j) is the (i,j)-th pixel of the filtered image ρ=
and FHL ( i , j ) is the (i,j)-th pixel of the hard-limited E (U ) * E (W )
image. We set threshold TH equal to the mean where W is the logo, U is the stacked image passed
luminance of the input image. through the reduction function and E(X) computes the
energy of image X. By comparing this correlation
coefficient to a threshold, THR, it is decided whether
the watermark W exists in the image F or not.
However, the presence of the watermark can be
investigated just by the human eyes without any
mathematical computation. Fig. 5 illustrates the
resulted image after reduction in the detection process
Figure 4 Block Diagram of the Detection Algorithm besides the original watermark logo.

The resulted image gets Stacked to the second share

of the watermark, which was kept private as the
detection key of the watermarking system. Stack
function acts as the reverse of (2,2)-VSS scheme used
in the watermark insertion algorithm and resembles
stacking images printed on two transparent sheets.
The stack function returns a 2M*2N pixels image
whose (i,j)-th pixel is: Figure 5 (Left) watermark and (right) detected mark (logo from
FST (i, j ) = min( FHL (i, j ), SH 2(i, j )) (3) [1])

Pixels of FHL and SH2 only take two values 1 and -1, The detection algorithm discussed above is a blind
representing white and black pixels respectively. As a detection scheme. Substituting the high-pass filter in
result, only the two values 1 and -1 are allocated to fig.4 by an image subtractor and using the host image,
the pixels’ values of the gained image FST. If a black IM, transforms the detection scheme to informed
pixel is stacked to another pixel, a black pixel is structure. This leads better robustness to various
obtained, whereas a white pixel is gained only if two watermarking attacks, however the original
white pixels are stacked. unwatermarked image is needed at the detector.
The last block before evaluating the correlation is the
reduction function, which returns an M*N-pixels
image whose (i,j)-th pixel is given as: 4. Experimental Results
⎧1 if L(i, j ) > 0 We simulated the proposed watermark insertion and
U (i, j ) = ⎨ (4) detection schemes and investigated the effect of
⎩− 1 otherwise
various attacks on the watermarking system. We
where: examined different high pass filters of order 21 to find
the optimum filter that is to maximize the ratio of
correct detection (correlation coefficient ρ, in case of the watermarked image, which more correspond to the
genuine watermark logo) to false detection high-frequency inserted watermark.
(correlation coefficient ρ for an irrelevant watermark
Cropping can be taken as the dual of the JPEG
logo). The maximum ratio was about 8.3 for different
compression, which zeros out spatial components of
watermark logos, which gained by using the HPF
the image. Because of its structure, the proposed
whose frequency response is shown in fig. 6.
scheme is quite robust to the cropping attack, if spatial
synchronization is made. Table 3 shows decrement of
the correlation coefficient of the cropped watermarked
image to the typical watermarked correlation
coefficient, for both cases with and without
performing synchronization.

Table 3 Decrement of the correlation coefficient and the detection

status after image cropping, with respect to the availability of
synchronization.

Cropping
10 20 30 40 50
Percentage
Figure 6 Frequency response of the HPF used in the detection Decrement
scheme.
of ρ (%) 6 13 8 11 13
Synch.
The threshold THR in the detection scheme should be Detection
Y Y Y Y Y
properly selected to minimize the false alarm ratio Status
(detecting a watermark in an unwatermarked image or Decrement
denying the watermark in a watermarked image). We No 36 43 57 55 75
of ρ (%)
set THR equal to 0.12 based on certain experiments. Synch.
We also surveyed the effect of different attacks on the Detection
watermark detection. Y Y Y Y N
Status
JPEG Compression acts as a low-pass filtering, which The effect of Image Rotating on the proposed
zeros out frequency components. We marked different watermarked image is also taken as a watermarking
images using the proposed scheme, and then attack. By performing spatial synchronization
performed JPEG compression with various quality (rotating the detection key by a value equal to what is
factors, Q, on the marked images. Subsequently, we applied to the marked image), the correlation value
verified the existence of the watermark in the shows up to 4% decrement for various rotation angels.
compressed images by evaluating the correlation However, the scheme is fragile to rotation if no
coefficient as in (6). Table 2 depicts decrement of the synchronization is made.
correlation coefficient of the JPEG compressed
watermarked image over the typical uncompressed Changing the aspect ratio (AR) of the watermarked
correlation coefficient. A “Y” in the last row of the image, which is the image’s height to width ratio, has
table indicates that the watermark’s existence is a little impact on the correlation coefficient. Table 4
successfully detected, whereas an “N” shows that illustrates the effect of changing the aspect ratio of a
watermark is not detected due to the performed 256*256 pixels watermarked image on the detection
watermarking attack. process.

Table 2 Decrement of the correlation coefficient and the detection Table 4 Decrement of the correlation coefficient and the detection
status after JPEG compression with various quality factors (Q). status after changing the AR of a 256*256 pixels watermarked
image.
Quality
90 80 70 60 New Size
factor 256*128 256*64 512*256
(pixels)
Decrement
29 61 71 75
of ρ (%) Decrement
14 6 26
of ρ (%)
Detection
Y Y Y N
Status Detection
Y Y Y
Status
The high impact of the JPEG compression on the
watermark detection is due to the fact that JPEG The high robustness of the proposed scheme to
compressor eliminates high frequency components of geometrical attacks, in the presence of
synchronization, is due the VSS scheme’s robustness
to this kind of distortions. Fig. 7 illustrates the schemes are fast and do not need a high amount of
reconstructed secret image by the VSS scheme, if the computational capabilities.
shares are geometrically distorted.
It must be again mentioned that selection of the
parameter µ in (1) makes a tradeoff between
imperceptibility and robustness of the watermarking
system to different attacks. We chose µ equal to 2 in
the simulations, where a higher value for µ (up to 5)
could result in a higher robustness to the surveyed
attacks, especially the JPEG compression process.
Also, if the host image is present at the detector, an
informed watermark detection scheme leads higher
robustness to different watermarking attacks.

Conclusions
In this paper, we have proposed a new watermarking
algorithm for digital images. The proposed scheme is
simple and is based on the concept of visual
cryptography, which is performed in the spatial
domain of the host image. Digital mark to be
embedded in the host image is one of the shares of the
Figure 7 Retrieving the secret image (watermark) from its
owner’s mark (that can be a visible logo), whereas the
geometrical distorted shares. (Top) image rotating, (bottom-left): other share is utilized as a detection key in the
image cropping and (bottom-right): changing the aspect ratio. detection process. Watermark detection can be
performed in both informed and blind structures,
We also investigated the robustness of the scheme to depending on the host image’s presence at the
image scaling. We low-pass filtered the watermarked detector. The detection algorithm relies on the fact
image using a four-tap filter, prior to downsampling that stacking images, containing corresponding shares
by R, in each direction. The resulting image is of the logo, make a high correlation to the logo that is
upsampled before calculation of the correlation also detectable by the human eyes.
coefficient. Table 5 shows the decrement of the
correlation coefficient of scaled image compared to Acknowledgment
that of the original watermarked image. In case no This project is partly supported by the Iran
filtering is performed, prior to downsampling, the Telecommunication Research Center (ITRC).
scheme shows a higher degree of robustness.
References
Table 5 Decrement of the correlation coefficient and the detection
status after image scaling, with and without filtering prior to [1] J. S. Pan, H. C. Huang and L. C. Jain, Intelligent
downsampling. Watermarking Techniques, World Scientific
Publishing Co. Pte. Ltd., Singapore, 2004.
Scaling Ratio (R) 2 4 8 16
[2] M. Naor, and A. Shamir, “Visual Cryptography”,
Decrement Advances in Cryptology – Eurocrypt’94
of ρ (%) 39 37 20 43 Proceeding, LNCS Vol. 950, Springer-Verlag,
No
1995, pp. 1-12.
Filtering Detection [3] M. Naor and A. Shamir, “Visual Cryptography II:
Y Y Y Y
Status Improving the Contrast Via the Cover Base”,
Decrement Cambridge Workshop on Protocols, 1996.
of ρ (%) 65 64 65 77 [4] A. Shamir, “How to Share a secret”,
With
Communications of the ACM, vol. 22, 1996,
Filtering Detection pp.612-613.
Y Y Y N
Status [5] G. Blakley, “Safeguarding cryptographic keys”,
in: Proceedings of National Computer
The utilized VSS scheme is simple, if the
Conference, 48, AFIPS Press, New York, 1979
computational complexity is considered. As a result,
the proposed watermark insertion and detection