You are on page 1of 16

Presented By: S.

MARIAPPAN
M.ASHOK KUMAR
SRI RAMAKRISHNA ENGINEERING COLLEGE

COIMBATORE
STEGANOGRAPHY – VECTORISED AND SCALING APPROACH

FOR DATA HIDING

ABSTRACT: The development of the Internet and the


other digital transmission channels has created
A new image data-hiding technique is the need to develop methods to protect the
proposed. The proposed approach modifies transmitted information as well as provide new
blocks of the image after projecting them services. This is due to the fact that, in the
onto certain directions. By quantizing the digital domain, perfect copying, modification,
projected blocks to even and odd values, one and redistribution of the information is
can represent the hidden information absolutely possible. Hence, it became
properly. The proposed algorithm performs necessary to find a way to protect the
the modification progressively to ensure transmitted information from being copied or
successful data extraction without the need tampered with.
for the original image at the receiver side. In this paper, images are used as an
Two techniques are also presented for example for the digital information. Data
correcting scaling and rotation attacks. The hiding provides a solution for such a problem.
first approach is an exhaustive search in Data hiding is the process of inserting invisible
nature, which is based on a training information into the image in order to perform
sequence that is inserted as part of the certain tasks. These could include proof of
hidden information. The second approach authenticity, copyright protection, detecting,
uses wavelet maxima as image semantics for and correcting possible tampering, preventing
rotation and scaling estimation. Both unlawful copying, and covert communication
algorithms have proved to be effective in systems. Several excellent data-hiding tutorials
correcting rotation and scaling distortion. have been published in the past few years.
Hiding information into the image is performed
either in the spatial or in the transform domain.
I. INTRODUCTION The idea lies behind modifying the image in an
invisible manner. Transform domain
embedding is performed by modifying attacks have the effect to desynchronize the
coefficients in the frequency domain, the DCT hidden data. This is due to the fact that
domain or in the wavelet domain. changing the image size and/or its orientation,
In the spatial domain, by modifying the even by slight amount, could dramatically
statistics of the image a spread spectrum based reduce the receiver ability to retrieve the
technique or a simple additive method is used. watermark. This can be compared to loosing
On the other hand, quantized projection-based synchronization in a communication channel.
embedding techniques for data hiding have Several attempts have been devoted to the
been exploited in the literature under different synchronization recovery problem.
contexts. In particular, the simplest form is In this paper, the proposed embedding
called the low-bit(s) modulation (LBM) where technique exploits an adaptive successive
the least significant bit(s) in the quantization of quantized projection strategy. The successive
the host signal are replaced by a binary quantized projection method performs two
representation of the embedded signal. A class operations, namely, projection and
of quantized projection approach has been quantization. The image is divided into sub
called the quantization index modulation(QIM) blocks; here, we use blocks of size 8×8. Each
which include the generalized LBM, spread block hides one bit in it by projecting it onto
LBM, spread transform dither modulation certain direction(s) and quantizing the
STDM, and distortion compensated QIM (DC- projection to satisfy certain constraints. Eigen
QIM). Transparency of the embedded data is space of the image is used as the projection
guaranteed by using the visual masking model space. Even and odd spaces are considered as
either in the spatial or the transform domain. the corresponding constraint. In particular,
In watermarking applications, the projections located in the even space
robustness of the embedded data to intentional correspond to one symbol and in the odd space
and unintentional attacks is of great correspond to the other symbol.
importance. Attacks range from simple signal The proposed approach is adaptive in
processing operations such as filtering and nature. In particular, the adaptive approach has
image coding like JPEG to severe intentional the advantage that the subspaces that we
ones such as rotation, scaling, cropping and project onto are image dependent, which means
noise addition. The watermark should be able that the receiver extracts these directions from
to survive these attacks. Rotation and scaling the image once it has been received. This leads
to no storage requirement at all at the receiver performs exhaustive search to lock for the
end. Moreover, given any distortion for the correct size and orientation. In the second
image, both the projection directions and the approach, the scaling factor γ and the rotation
image blocks should undergo the same amount angle θ are estimated based on edges
of distortion which implies relative stability. information, as salient points, computed from
This is different from the non adaptive case in the wavelet maxima of the image. The scaling
which the directions do not depend on the factor γ is approximated by the edges standard
image itself but are rather random. The deviation ratio (ESDR) which measures the
proposed approach differs from the other deviation of the edges from a reference point,
quantized projection techniques, such as the before and after scaling. While the rotation
QIM and the related algorithms, e.g., the angle, θ, is approximated by the average edges
generalized LBM, STDM, and DC-QIM in that angles difference (AEAD), which measures the
the spreading vectors “projection spaces” there average angles difference of the edges
in are random which lacks the flexibility of locations, in a predetermined region, before
being self-contained algorithms. In particular, and after rotation. It should be noted that the
the receiver needs to know the projection choice of edges as features in our technique is
spaces before hand in order to complete the motivated by the fact that they reflect both size
detection process which adds to the complexity and orientation changes that occur as a result of
of the receiver. Unlike the proposed embedding geometric manipulations.
algorithm, schemes which use the low- A brief overview of the multi scale
frequency DCT vectors as the projection spaces edge detection technique and wavelet maxima
lacks the flexibility of being adaptive schemes. computation which is used in the paper is
The proposed algorithms do not assume presented in Section II. The proposed
the knowledge of the original image at the embedding algorithm is then discussed in
receiver side which allow for blind detection. Section III. Finally, simulation results are
This has the advantage of reducing the receiver presented in Section IV.
complexity and adding more practicality to the
designed system. The first method depends on
exhaustive search strategy. That is, by using a II. MULTI SCALE EDGE DETECTION
training sequence inserted in the image as part Multi scale edge detection was first
of the hidden information, the receiver proposed by Mallat, and then used by Xu et al.
for noise reduction in medical images. Edges Fig. 1. Analysis filter bank for dyadic wavelet
are detected by spatially correlating the transform.
estimated edges across the various scales after
decomposing the image by the wavelet where ψ(t) is the corresponding non
transform. In edge detection, the non orthogonal wavelet and u is the spatial/delay
orthogonal wavelet transform is used as parameter. The wavelet coefficients in (1) can
opposed to the orthogonal counterpart. This is be computed using a fast filter bank algorithm.
due to the fact that the transformed signal, in For simplicity, we will only consider the one-
the latter case, is uncorrelated across scales, dimensional case as it can be easily generalized
and, hence, detecting edges would be difficult. to the 2-D case using separable bases. Let h[n]
It is worth mentioning at this point that the non be the discrete time filter corresponding to the
orthogonal wavelet transform is also named the scaling function, g[n] be the discrete time filter
dyadic wavelet transform in the literature. The corresponding to the wavelet function, and the
use of the dyadic wavelet transform for signal discrete time signal at resolution j=0 is a0[n].
decomposition yields an over complete Then, the dyadic wavelet transform of a0[n] can
representation. The reason is that wavelet be represented as a convolutional process with
functions are only sampled across scale but not a cascade filter bank. The coarse
time sampled. Then, at each decomposition approximation in the decomposition stage at
stage, the resulting signal has the same length resolution j 0 is
as the original signal which is advantageous (2)
from edge detection point of view. The dyadic
and the details
wavelet transform of a function, f є L2(R) is
(3)
where * is the convolution process.
W f(u,2j)= (1)
is the filter obtained by inserting

zeros between each sample of h[n].


The analysis filter bank is illustrated in Fig. 1.
It is clear that there is no down sampling after
each decomposition stage which clarifies the
redundancy introduced by the dyadic wavelet
transform. Multi scale edges are computed by
spatially correlating the edges, maxima, for
each scale. Let |W f(u,2j)| be the modulus
wavelet transform of a function f(u) at scale j .
Then, the modulus maximum at any point (u0,
2j) is such that
|W f(u,2j)| |W f(u0,2j)| (4)
where u is the spatial/delay parameter.
For each scale, points satisfy (4) are connected
together defining the wavelet maxima contour
for this scale. Fig. 2 shows the first four levels
of the modulus maxima for the F16 image
which were computed using XWAVE. It
should be noted that low levels preserve high-
Fig. 2 Modulus maxima of the first four levels
frequency information while higher levels tend
to preserve coarse information. This is well
understood from the filter bank formulation
(2), (3), since high-frequency edge information
is not robust to low-pass filtering, which may
occur with the embedded image. The global
wavelet maxima are computed by spatially
correlating the corresponding high-levels
wavelet maxima to increase the robustness to
low-pass filtering. Fig. 3 shows the global
wavelet maxima constructed from levels 3 and
4 of the F16 image.

Fig. 3 Global wavelet Maxima


III. EMBEDDING ALGORITHM

The proposed approach is a projection


followed by quantization operation under
certain constraints. In particular, consider
v є to be a vector that corresponds to a reconstructed space will be different from Sv.
certain 8×8 block of the image and we want Thus, no perfect reconstruction is guaranteed.
this vector to represent a binary bit “1” or “0.” To overcome this problem, we use a successive
The key idea is to project v onto the subspace, data embedding scheme as explained below.
extracted from the image itself; thus, the space
is image dependent. By dividing the A. Successive Projection
corresponding space into two sets, and Sv0, The main point in the projection process is
appropriately, one can hide data in each vector. to define the direction(s)/spaces onto which the
This is accomplished by forcing the projection projection is performed. The authors found that
to be in Sv1 if we want to represent “1” or Sv0 if the space spanned by the first M=1 eigenvector
v to represent “0”. Sv1 is the set of all of the correlation matrix of a certain sub image
projections of onto such that the projection is Av is robust to JPEG coding, rotation, and
odd, and Sv0 consists of even projections. This scaling. As mentioned before, a successive
can be done by quantizing the projection of v projection strategy is used, i.e., during the
to even if we want v to represent “0” and to current iteration, the modified blocks will be
odd if we want it to represent “1.” In particular, used with previously modified ones as the new
let P be the projection operator onto Sv, then sub image for the next iteration. In particular,
the even and odd spaces Sv0 and Sv1 are defined Fig. 4 describes the ith iteration. Here, i=7 for
as, blocks of size 8×8, where is
Sv0 = {x : x є Pv,x even}, the principal sub image from which Sv is
Sv1 = {x : x є Pv,x odd}. (5) constructed. Note that Sv is the subspace for all
vs that are located in the strip i, as illustrated in
The reconstructed vector ϋ should differ Fig. 4. In the (i+1) step, the vs will be those
from v by an invisible amount to ensure vectors in the strip i+1 and Sv will be
perceptual transparency. Since Sv is image constructed from the principal sub image
dependent, it should be constructed from the , which is formed by concatenating
modified vectors, the ϋ, to guarantee perfect
with the strip i. To describe how the
reconstruction. To be more specific, if Sv was
subspace Sv is constructed from Av, let Rv
constructed from the unmodified vectors, the v
R64×64 be the empirical correlation matrix
at the transmitter, the receiver will reconstruct
estimated from Av in step i, then Sv is the space
it from the modified vectors, ϋ, and this
spanned by the first M columns of the matrix
U, where
Rv = U (6)
is the singular value decomposition of the
correlation matrix Rv.
Once the projection is performed, the
quantization should be done in such a way that
correctly represents the hidden bit. Each block
is projected onto and then quantized
appropriately. Fig. 5 shows the 64 singular
values for the corresponding spaces for the f16
image. As it is clear from the figure, the energy
is concentrated in the first eigenvector while
the remaining vectors have little energy which
justifies using only the first eigenvector as the Fig. 4. Iteration “i”
projection space for the following two reasons.
First, increased robustness due to the high-
energy concentration is used in the first
eigenvector. Second, due to the ill-conditioned
nature of the subspaces for M > 1, blocking
artifacts are emphasized if spaces of
dimensionality M > 1 are used. Finally, from
Fig. 5, we can also conclude that the proposed
algorithm is very similar to block-DC
modification except it is done in a progressive
manner.

Fig 5. Singular values for projection spaces for


F16
Choosing the Initial Block:
The choice of the upper left block as the
initial block in the iteration is arbitrary. In
principle, it can be any block in the image.
However, the drawback of choosing the upper
left block is that it is vulnerable to cropping
attack. This is due to the fact that by cropping
the borders of the image, the formed eigen
space, at the receiver, would be different than
what has been used at the transmitter and then
the error will propagate to the rest of the image. (a)
The effect of cropping attack is that it shifts the
coordinates. Choosing the initial block to be
coordinate independent would help reducing
the effect of cropping attack. Our choice is to
use the wavelet maxima locations man (MLM)
as the reference point. Simply, the location of
the initial block will be determined by the
MLM. The embedding process starts at the 8 ×
8 block containing the MLM as the initial
subspace and then projecting and quantizing
the surrounding blocks onto this initial
(b)
subspace. The modified blocks will be used as
the new subspace for the outer unmodified
surrounding blocks as before. Fig. 6(a) shows
Fig. 6. MLM for original and cropped
the MLM of the original Cameraman image
Cameraman marked by “*”.
which is marked by “*”.
(a) Original, (b) Cropped.

The receiver should be able to extract


the embedded data after it has been cropped by
a forger. To be able to do so, the receiver first
estimates the MLM from the cropped image then quantize it accordingly. Let Q0 and Q1 be
and uses it as its initial block. Then perform the the even and odd scalar quantizers defined as
detection process in a progressive manner as follows:
was described in the embedding process. Fig.
6(b) shows the corresponding wavelet maxima
(7)
and the MLM which is marked by “*” for the
cropped Cameraman image. The original image
size was 256 × 256 while the cropped image Fig. 7 illustrates the role of Δ in reducing the
size is 211 × 211 which amounts to 15% quantization error, where p is the projected
cropping. By comparing Fig. 6(a) and (b), it is value and the corresponding quantized levels
clear that the MLM has not been altered are as shown in Fig. 7 where it is assumed that
although the image has been cropped by 15%. Q0 quantizes to the upper level while Q1
It should be noted that cropping the image by quantizes to the lower level.
more that this value would change the location
of the MLM.

B. Quantization
Even and odd quantization are used to
represent “0” and “1,” respectively. In other
words, if the projection is quantized to an even
number then this block represents “0” and if it
is quantized to an odd number then it
represents “1”. The receiver performs the same
process to recover the hidden information. In
particular, each block is projected onto the Fig.7 Quantization step
corresponding subspace and a rounding C. Embedding
operation is performed in order to recover the Let ϋ be the embedded vector, then we
hidden bit. Let є be the rounding error, then, in are going to embed a binary bit i є {0,1}.
order to correctly retrieve the hidden data, one Here, masking is implicitly introduced
must have є < 0.5. To achieve this bound, first by Δ. In particular, there is a tradeoff between
rescale the projection by a scale factor Δ and the robustness of the embedded data and its
transparency. As the value of Δ increases, the
robustness increases but the transparency
decreases and vice versa. In principle, Δ may
take any integer value, typical values for Δ
range from 1 to 20 and up to 30 in severe
attacks. The value of Δ is chosen based on the
anticipated severity of attacks and the required
transparency. For example, in applications
where the hidden information cannot be
compromised, such as in military situations,
high is recommended, while in situations
where attacks are performed unintentionally, (a)
such as transmission over noisy channels, low
values could be chosen.
In order to compute the embedded data
rate, suppose that the image size is and each 8
× 8 block hides one bit of information. Then
the number of bits k that can be hidden is
k = (⌊N/8⌋2-1). It should be noted that the
initial block does not convey any information
as it is considered the initial subspace. Thus,
for a 256×256 image k=1023 and k=4095 for a
512 × 512 image. Fig.8 (a) shows the
embedded Pot image using Δ=10, while Fig.
8(b) illustrates a heavily watermarked image (b)
with Δ=40. The corresponding PSNR with Fig.8. Demonstrating the effect of increasing Δ.
respect to the original image are 45 and 32.9 (a) Embedded using Δ=10.
dB, respectively. In the latter case, blocking (b) Embedded using Δ=40.
artifacts are visible due to quantization effect.
IV. PERFORMANCE

A. Robustness of the Eigen space

Since the proposed embedding


algorithm assumes projection onto the
eigenspace SM, spanned by the first M
eigenvectors of the matrix U in (6), then it is
natural to examine the robustness of the
eigenspace to different attacks. In particular, (a)

we explore the robustness of SM to JPEG

coding, rotation, and scaling. Let SM and be


the eigenspaces spanned by the first M
eigenvectors of U before and after performing
the attacks, respectively. Then, the angle ψ(SM,

between SM and defined below is used

as a measure for the robustness of the


eigenspaces under attacks.

ψ(SM, = ψ(f , g) (14) (b)

where f and g are arbitrary eigenvectors Fig.9. Robustness of eigenspace to JPEG.


(a)F16. (b) Tea pot
in SM and , respectively. The motivation

for using the angle ψ as a measure for


Figs. 9–11 show ψ(SM, for different
robustness is that it gives us a sense of how
M values for JPEG coded with different Qs,
much the space has been rotated due to the
scaled and rotated images. It is clear that the
performed attacks. This is acceptable since a
first eigenvector is highly robust to the
rotated space means different projection values
performed attacks, this justifies the use of the
and, hence, more errors in the embedding
first eigenvector as the projection space in the
results.
proposed embedding algorithm. Unfortunately,
the eigenspace is not robust to blurring attack,
and, hence, the embedding algorithm does not
survive blurring attacks.

(a)

(a)

(b) (b)

Fig.10. Robustness of eigenspace to scaling. (a) Fig.11. Robustness of eigenspace to rotation.


Barbara. (b) Cameraman. (a) F16 (b) Cameraman
B. Robustness to Attacks Fig. 12(a) and (b) illustrates the percentage
The robustness of the proposed embedding BER versus Δ for different PSNR for f16 and
algorithm to AWGN, JPEG coding, combined pot images, respectively. As can be seen from
scaling, and rotation has been tested. the figures, the BER decreases with the
increase of Δ and PSNR, as expected. Fig.
13(a) and (b) illustrates the performance of the
embedding algorithm to JPEG coding for
different Q factors for the F16 and Mandrill
images, respectively. Robustness to combined
scaling and rotation attacks is illustrated by
Fig. 14(a) and (b) for Cameraman and Lena
images with Δ=20 and Δ=30, respectively.
The figures show the percentage BER as a
function of different scales and rotation angles
after regaining synchronization. As it is clear
(a)
from the figures, the BER increases
dramatically with downscaling the image. This
can be understood from the inherent data loss
accompanied with the down sampling process.
On the other hand, enlarging the image has
little effect on the error due to the redundancy
introduced by the interpolation process. The
performance of the algorithm as a function of
angle and scaling factor estimation error is
illustrated in Fig. 15. It is clear from the figure

(b) that the algorithm is sensitive to estimation


error which can be understood from the

Fig.12. BER versus Δ for different PSNR. desynchronization effect produced by the

(a)F16. (b)Tea pot estimation error. Hence, the fine tuning step
introduced in Section IV-B is important to
maintain low BER. Finally, Fig. 16 shows the
contour plot of the 2-D autocorrelation function
of the original pot image and the corresponding
watermark embedded with Δ=40. It is clear that
the watermark power spectrum1 follows that of
the original image, i.e., the power spectrum
condition is satisfied.

(a)
Fig. 14. BER after combined scaling and rotation
attacks after correction. (a)Cameraman, Δ=20.
(b)Lena, Δ=30.

(b)
Fig.13. BER versus Δ for different JPEG-Q.
(a) F16 (b) Mandrill.
Fig.15. BER versus angle and scaling factor estimation
error for cameraman with Δ = 30.
exhaustive search technique uses pre embedded
training sequence for locking to the right scale
and orientation. On the other hand, scaling
factors and rotation angles are estimated with
the aid of wavelet maxima. Performance
analysis and robustness tests were also
presented.
The drawback of this algorithm is, it
(a) can not overcome the blurring attack and also
very sensitive to estimate errors.

VI. REFERENCES

1. M. Swanson, M.
Kobayashi, and A. H. Tewfik, “Multimedia
data embedding and watermarking
(b) technologies,” Proc. IEEE, vol. 86, no. 6,
Fig.16. Contour plot of the autocorrelation of pp. 1064–1087, Jun. 1998.
the watermark and the original pot. 2. P. Bas, J.-M. Chassery, and B. Macq,
“Geometrically invariant watermarking
using feature points,” IEEE Trans. Image
V. CONCLUSION Processing, vol. 11, no. 9, pp. 1014–1028,
Sep. 2002.
A new data-hiding technique has been 3. Masoud Alghoniemy and Ahmed H.
presented in this paper. Data is embedded by Tewfik, Fellow, IEEE, “Projection
quantizing the projection of the 8 × 8 blocks approach to Data hiding”, VOL. 15, NO.
onto the eigen subspaces extracted from the 2, FEBRUARY 2006.
image. The proposed data embedding
algorithm assumes blind detection where no
overhead is required for detection. The

You might also like