Professional Documents
Culture Documents
Abstract— In this work, we propose a model for an don’t provide any evidence that their model can be used
optimal detector of digital watermarks for Laplacian in practical watermarking systems.
channels, the watermark embedding algorithm is the In [2] a semi-blind multiplicative watermarking ap-
multiplicative embedding rule in time domain, the op-
timal detection equation and the threshold equation are proach for audio and speech signals has been introduced.
derived using maximum likelihood (ML) and Neyman- The detection of the watermark is accomplished by us-
Pearson criterion. The resulting system is blind and has ing the optimal ML detector aided by the channel side
very low complexity. Computer simulations are carried information for Gaussian and Laplacian signals in noisy
out applying the proposed model to audio signals and environment. They applied their scheme to speech and
results prove that the proposed system is able to de-
tect watermarks even if the watermarked object was audio signals. The algorithm was applied to low frequency
severely attacked by noise, low pass filtering among components of the host signal. In addition, the power of
other attacks. the watermark was controlled to have inaudibility using
Keywords: Digital Watermarking, Laplacian channel, perceptual evaluation of audio quality (PEAQ) and per-
Maximum likelihood, Neyman-Pearson criterion, Audio ceptual evaluation of speech quality (PESQ) algorithms.
watermarking system. However, a drawback of this proposal is that it is semi-
blind.
I. Introduction The authors of [3] propose a system designed to enforce
the integrity and security of fingerprint images using
Digital watermarking has been an active research field watermarking techniques. They apply the ML watermark
which had produced very interesting approaches, however, detection method to detect the watermark within discrete
there is a lot of work that has be done before considering wavelet transform (DWT) coefficients of fingerprint im-
watermarking as a solid discipline. ages. The ML method they use is based on Bayes’ decision
A very helpful approach is to establish analogies to theory and the Neyman-Pearson criterion, they use both
the very strong field of the theory of communication, in Generalized Gaussian (GG) and Laplacian models as PDF.
this context, we can think of a watermark as a signal In [4] a digital watermarking system in the Discrete
that propagates through some transmission channel. This Cosine Transform (DCT) domain is proposed. This wa-
channel can be modeled using a known Probability Density termarking system satisfy some perceptual constraints.
Function (PDF). Under these constraints, robust and fragile watermark-
In this work, we focus in the case of the watermark ing techniques are designed by varying the watermark
propagating through a Laplacian channel, a Laplacian strength. The optimum detection structure is constructed
channel is a channel that can be statistically modeled using based on the Laplacian model. It is claimed that it can
a Laplacian PDF. detect the tampered regions very accurately. That sys-
There are some proposals by other authors that studied tem can be used in copyright notification and protection,
the case of the Laplacian channel, for example, in [1], broadcast monitoring and tracking, and authentication
the Laplacian model is used to derive a detector for and tamper-proofing.
watermarks in the wavelet transform domain using ML An finally, a multiplicative spread spectrum watermark-
criterion, and the decision threshold equation is derived us- ing technique in curvelet domain is presented in [5] .
ing the Neyman-Pearson criterion. Their work was applied Watermarked curvelet coefficients are modeled using GG
to image watermarking in the discrete wavelet transform distribution, Laplacian distribution and Cauchy model.
domain. They compared their results to a similar approach Watermarking detectors are designed employing locally
that considered modeling the discrete wavelet transform most powerful (LMP) approach.
coefficients using a Gaussian distribution and concluded In this paper, we derive an optimal detector for the
that Laplacian model can result in a better performance Laplacian channel that unlike previous approaches, has
compared to the Gaussian model approach. However, they very low complexity. First, the derived detector variables
IV. Watermark Detection Variables for Now, finding the maximum of (12), we get:
Laplacian Channels N
∂L X ∂ yi
In this section, the optimal decision variables will be de- =− α(1 + gwi ) ,
(13)
rived considering a Laplacian channel and a multiplicative ∂g i=1
∂g
embedding rule. resulting in the following equation:
A. Watermark Embedding Rule
N
The multiplicative embedding rule is defined as: ∂L X yi yi
=− Sign wi (1 + gwi )−2 .
∂g i=1
α(1 + gw i ) α
yi = xi (1 + gwi ), (7) (14)
Since wi ∈ {−1, 1} and |gwi | ≪ 1, then 1+gwi is always
where yi is watermarked sample, wi is i -th watermark positive, we can get rid of the absolute value operation,
bit and g is watermark embedding gain which controls and by using the Taylor series for (1 + x)m ≈ 1 + mx for
the watermark energy and robustness. Multiplicative rules |x| < 1, we have:
exhibit several desirable properties, the most important is
the inherent masking effect that allows greater embedding N
strength while watermark imperceptibility holds. ∂L X y y
i i
=− Sign wi (1 − 2ĝwi ). (15)
∂g i=1
α α
B. Optimal Detection of Watermarks
Solving (15) for the embedding gain ĝ we get:
We can proceed to derive the optimal detection vari-
ables: the detection variable d, which is a measure of the N
yi w i
P
presence of the watermark within a given signal; and the α
i=1
threshold T h, which provides reference to decide if the ĝ = N
. (16)
y
signal is watermarked.
P i
α
A handy approach to watermark detection is performing i=1
an estimation of the gain (ĝ), analyzing (7), clearly, we And finally, since α > 0 the optimal detection variable
have two possible outcomes for an estimator: is given by:
N
(
0 if no watermark was embedded 1 X
ĝ ≈ . (8) d = ĝ = |yi |wi , (17)
g otherwise N αy i=1
where,
N
1 X
αy = |yi |. (18)
N i=1
In order to detect the watermark, the decision variable d
must be compared to a threshold, a watermark is present
if d ≥ T h, the general threshold equation derived from the
Neyman-Pearson criterion proposed in [6] is given as:
N N
1 X µ|yi | X In next section we will evaluate the model as an optimal
E(d) = E(|yi |)wi = wi = 0, (20) estimator of the embedding gain, and then, we will present
N αy i=1 N αy i=1
an evaluation of performance of the model in the presence
1 X
N of attacks.
V(d) = V(|yi |)wi2 , (21)
N 2 αy2 i=1 A. Detector Performance
Since the PDF of a distribution of a random variable |x| One of the most important evaluation parameters for
distributed accordingly f (x) is f (|x|) = f (x) + f (−x) for a watermarking system is the detection behavior of a
x > 0, we have for the L(αy ) that g(x) = 2L(αy ) (Because given model for an arbitrary set of different watermarks;
its symmetry), clearly it becomes an exponential and thus ideally, the detector variable from (17) should have a zero
its variance is αy2 , using (21), we get: response for any watermark different from the embedded
watermark. In practice, it is not possible due the fact that
1 the cover signal is correlated to the watermark. In practical
V(d) = . (22)
N situations, the detector variable outputs a very low value
Finally, the threshold equation is: for any watermark different to the embedded one.
r In a good detection model, only the system’s response
2 to the watermark that was embedded in the cover should
T h = erfc−1 (1 − 2Pf p ) . (23)
N cross the threshold value, furthermore, this response
This completes the derivation of the detection variables should be much larger than the response for any other
for the detection model proposed in this paper. In next watermark, and the smaller the response of watermarks
section, computer simulations will be carried out. different to the embedded one, the better the model is.
In Fig. 3, we can see the computed detection variable, a
V. Computer Simulations gain value of g = 0.02 was used to embed a watermark and
In this section we present simulation results that val- 1000 different watermarks were tested, only the watermark
idate (17) and (23), firstly, we will show the detection that was actually embedded, in this case, watermark
capabilities of proposed model, and then, we will show number 500, crosses the computed threshold (in dotted
that (17) is an accurate estimator of the gain and in line), whilst, the other watermarks produce a very low
consequence, is an accurate watermark detection model. response from the detector variable, which confirms that
All test were carried out under the following scenario: the derived detector is optimal under the ML criterion.
the watermark was embedded in non overlapping blocks Fig. 3 shows the system performance for our worst case,
with length of 2 times the sampling frequency of the however, we can see that the performance of the model is
song using (7). Detection is made in the same block wise remarkable.
approach, d and T h are computed for each block using (17)
and (23) and the responses for each block are accumulated B. Estimator Performance
and averaged. We let Pf p = 10−6 . All sound files used for Since (17) is claimed to be the optimal estimator of the
our tests were uncompressed 16-bit stereo WAV files with embedding gain, in this section, we evaluate the model as
48000 Hz sampling rate. an estimator of g.
Fig. 4: Estimation of gain value for different covers, actual Fig. 6: Detection variable d versus decay value in an echo
embedding values are shown for comparison. addition attack.
VI. Conclusions
We performed two experiments, in the first one, the song
was watermarked with a watermark length equal to the
number of samples of the song (a full length approach), we
found that this approach has some difficulties, for example,
the memory needed for allocating both the song and the
watermark could be a drawback for some applications,
in addition, a full length detection system based on the
proposed model fails to detect watermarks for many attack
classes, making it a fragile watermark system. For that
reason, we decided to focus in a block wise approach.
Block wise approach has many advantages, first, the
song doesn’t need to be complete and stored in the
receiver, the detection can be started as soon the song
arrives and within a few seconds, the system would be
able to detect the watermark instead of waiting for the
Fig. 7: Detection variable d versus noise amplitude in a white
noise addition attack. whole song to arrive to the detector, a second advantage,
memory requirement is fairy reduced compared to the full
length approach.
Even when many approaches prefer the use of some
transform domain claiming superior performance, it is
done at the expense of many arithmetic operations, how-
ever in our approach, the number of such operations
is reduced due the low complexity of the detector, in
addition, there is not need of applying any transform
on the data since all process is performed in temporal
domain. Those properties would help for developing real
time applications.
The proposed system achieves a remarkable perfor-
mance, it was able to detect watermarks even under heavy
attacks, so model’s reliability has been proved.
Acknowledgements
Authors wish to thank the FIEC Poza Rica of The
University of Veracruz for the support for this work.
Fig. 8: Detection variable d versus cutoff frequency in an low
pass filtering attack. References
[1] T.M. Ng and H.K. Garg. Maximum-likelihood detection in dwt
domain image watermarking using laplacian modeling. Signal
Processing Letters, IEEE, 12(4):285 – 288, april 2005.
[2] M.A. Akhaee, N.K. Kalantari, and F. Marvasti. Robust multi-
plicative audio and speech watermarking using statistical mod-
eling. In Communications, 2009. ICC ’09. IEEE International
Conference on, pages 1 –5, june 2009.
[3] K. Zebbiche, F. Khelifi, and A. Bouridane. Maximum-likelihood
watermarking detection on fingerprint images. In Bio-inspired,
Learning, and Intelligent Systems for Security, 2007. BLISS
2007. ECSIS Symposium on, pages 15 –18, aug. 2007.
[4] Qiang Cheng and T.S. Huang. A dct-domain blind watermarking
system using optimum detection on laplacian model. In Image
Processing, 2000. Proceedings. 2000 International Conference on,
volume 1, pages 454 –457 vol.1, 2000.
[5] Chengzhi Deng, Shengqian Wang, Hui Sun, and Hanqiang Cao.
Multiplicative spread spectrum watermarks detection perfor-
mance analysis in curvelet domain. In E-Business and Informa-
tion System Security, 2009. EBISS ’09. International Conference
on, pages 1 –4, may 2009.
[6] Mario Gonzalez-Lee. Marcas de Agua Digitales y sus Aplicaciones
Practicas. PhD thesis, Instituto Politecnico Nacional.