Robust multirate lattice quantization index
modulation watermarking resilient to multiple
description transmission channel
Miin-Luen Day
National Chiao-Tung University Department of Computer Science and
Information Engineering 1001 Ta Hsueh Road Hsinchu, Taiwan 30050
and
Chunghwa Telecom Company Telecommunication Laboratory Chung-Li, Taiwan
Suh-Yin Lee
National Chiao-Tung University Department of Computer Science and
Information Engineering 1001 Ta Hsueh Road Hsinchu, Taiwan 30050
E-mail: [email protected]
I-Chang Jou
National Kaohsiung First University of Science and Technology
Department of Computer and Communication Engineering Kaohsiung, Taiwan
Abstract. Multiple-description 共MD兲 transmission results in nonlinear value-metric共value-scaling兲 distortion in the case when some of the sent descriptions are not received. An acceptable image can still be received in the above situation due to the characteristic of MD. However, it is quite damaging to the traditional quantization-based watermarking technique for payload detection. To overcome the problem, a straightforward ap-proach would be to increase the quantization step size in watermark embedding so as to keep the distortion within the tolerant range. How-ever, as a larger quantization step size in watermark embedding would result in a worsened watermarked image, it is not feasible to adopt it without further consideration. The proposed multirate lattice quantization index modulation 共MRL-QIM兲 encodes two watermark bits into each of the four co-set points of a lattice共so-called multirate兲. Compared to tra-ditional vector-based quantization encoding or combined spread spectrum-quantization encoding, it significantly increases the payload 共capacity兲 and enhances the robustness of watermark detection while preserving the fidelity of the watermarked image. Comprehensive experi-ments have confirmed that the overall performance and the effective-ness of the proposed scheme are superior to previous approaches. © 2007 Society of Photo-Optical Instrumentation Engineers. 关DOI: 10.1117/1.2715597兴 Subject terms: watermarking; multiple description; multirate lattice quantization index modulation.
Paper 060368R received May 17, 2006; revised manuscript received Aug. 15, 2006; accepted for publication Aug. 18, 2006; published online Mar. 29, 2007.
1 Introduction
Watermarking is a well-known technique used to hide data or information imperceptibly within image, audio, or video so that valuable contents can be protected. The research that developed robust watermarking techniques that can survive intentional attacks has been extensively explored in the past decade. However, few researchers have put their emphasis on dealing with incidental attacks, such as an attack caused by packet network or error-prone wireless transmission over an unreliable network. In Ref. 1 Hartung and Ramme point out that as the technology of second-generation and third-second-generation 共3G兲 mobile networks keeps advancing, digital media distribution for mobile E-commerce will eventually evolve into a huge business. Under these circumstances, watermarking applications such as media identification and copy control are getting more and more feasible for mobile E-commerce with the virtue that the identity of a user is known. Aiming at the error-prone nature of wireless communications, Checcacci et al.2 propose a robust MPEG-4 watermarking technique for video sequences corrupted with errors. Chandramouli et al.3 propose a multiple-description framework for oblivious wa-termarking, which uses one description to embed water-mark information and another for referential original image
to assist detection. However, it is not a completely blind technique, in the sense that a relationship 共defined by the watermark key兲 between the watermarked coefficients in one description and the corresponding coefficients in the referential one will be needed during the watermark detec-tion process. And since the embedded watermark cannot be extracted without receiving both descriptions, it is not suit-able for error-prone packet transmission network applica-tions.
Multiple-description coding共MDC兲5–9is different from either layered coding or simulcast coding or even error-resilient tools described in MPEG-4.4On a wireless multi-hop network or a packet-switched network, there are sev-eral parallel channels between the source and the destination and each channel may be temporarily down or suffering from long burst errors. The MDC scheme is de-signed such that the quality of the decoded signal is accept-able with receiving any individual description and can be further improved as more descriptions are received.
Contrasted to the limited capacity inherent in spread spectrum-based watermarking techniques,10–12 the quantization-based watermarking techniques13,14 normally have a relatively high capacity. Chen and Wornell13 pre-sented a quantization index modulation 共QIM兲 scheme based on the concept of dither modulation, which uses the watermark information as an index to select a dither signal. The dither signal is then added to the host signal, and a
least distorted quantizer is selected from a set of possible quantizers. The dithered host signal is quantized using this selected quantizer, and finally the dither signal is subtracted from the quantized signal to form a watermarked value:
s共x;m兲 = Q关共x + d共m兲兴 − d共m兲, 共1兲
where x苸X is the host signal, d共m兲 is the dither signal representing watermark message bit m苸W, Q共.兲 denotes the selected quantizer, and s共x;m兲 corresponds to the host signal embedded with watermark message m. In the detec-tion process, different dither signals representing the water-mark message are added to the received signal using Eq. 共1兲, and the index 共m=0 or 1兲 of the dither signal is the extracted watermark information. The detected index m*is chosen so that it gives the minimum distance between the received signal共x
⬘
兲 and its closest quantized signal:m*= arg min
m 储x
⬘
− s共x⬘
;m兲储. 共2兲We argue that the distortion introduced by losing some of the transmitted descriptions of MD transmission can be viewed as a nonlinear value-metric attack 共described in Section 2.1兲. While in this situation an acceptable image can still be received due to the characteristic of MD, it is quite damaging to the traditional quantization-based water-marking technique 共QIM兲 for payload detection. Some other works address the problem of QIM detection in the presence of value-metric distortion,14–19yet few papers fo-cus on QIM-related watermarking scheme under MD at-tack.
In Ref. 20 the authors integrate oblivious watermarking techniques 关quantization index modulus modulation 共QIMM兲 and QIM with the multiple-description coding 共MDC兲兴 to get a multiple-description watermarking 共MDW兲 framework. In this framework, the watermark em-bedding is computed in either description and could be ex-tracted with the reception of either one or two descriptions. The main drawback of Ref. 20 is that both the watermark embedding and detection are performed on side description rather than on central description. The other problem is that both QIMM and QIM are quite limited under value-metric attack.
Stimulated by the above-mentioned issues, in this paper we attempt to find an improvement by studying the prob-lem of watermarking under multiple-description diversity transmission from a different perspective; namely, water-mark embedding is done in the central description while watermark detection can be done in either central or side description. The merit of watermark embedding done in the central description is that the embedding and detection do not interfere with the MD mechanism. Therefore, this ap-proach is more flexible than the one done in the side de-scription. Furthermore, we propose a blind multirate lattice quantization index modulation共MRL-QIM兲 watermarking technique to boost the effectiveness. As the proposed MRL-QIM encodes two watermark bits into each of the four co-set points of a lattice共multirate兲, with the above design, the payload 共capacity兲 and robustness of watermark detection will be significantly upgraded. In the meantime, the fidelity of the watermarked image is also preserved. In next sec-tion, the MD attack channel and the hexagonal lattice
quan-tization共HLQ兲 algorithm will be introduced. We will elabo-rate on the proposed MRL-QIM watermarking technique in Section 3. In Section 4, experimental results are presented. The concluding remarks together with future work are ad-dressed in Section 5.
2 Multiple-Description Coding„MDC… and
Hexagonal Lattice Quantization„HLQ…
In this section we begin by describing the multiple-description coding 共MDC兲,6,7 which introduces possible nonlinear value-metric attack for watermark detection through transmission. Then, we shall introduce a hexagonal lattice quantization 共HLQ兲 algorithm, which will be adopted in our proposed MRL-QIM watermarking tech-nique.
2.1 The MDC Approach6,7
The practical MDC approach for wavelet-based coding was proposed by Survetto et al.6The basic two-description ar-chitecture of MDC6,7 is illustrated in Fig. 1. The major contribution of the MDC scheme is its capability on receiv-ing satisfactory data quality even if part of the channel is broken. As shown in Fig. 1, the quality of a decoded signal will be acceptable if either receiver 1 or receiver 2 receives the correct signal. In addition, the quality of a received signal can be better if both receivers function well. The most crucial component of the MDC scheme is its multiple description scalar quantizer. It consists of a scalar quantizer component, which quantizes the continuous sample values to smaller countable integer values, and an index assign-ment component. The source input signal x苸X is first scalar-quantized to xQ苸XQ. The function of the index
as-signment component f : xQ→共x1, x2兲 is then to split a quan-tized coefficient xQ into two complementary and possibly
redundant smaller coefficients x1苸X1 and x2苸X2, so that each of these two small coefficients only needs a lower bit rate to describe and both could be recombined later to re-cover the original quantized coefficient. That is, with the reception of two description values, a perfect reconstructed value x= x0= xQ can be achieved by using x0= f−1共x1, x2兲. When only one of the description values is received, an acceptable estimated value x= xd共兩xd兩 Ɱ兩xQ兩兲 can be
ob-tained through xd= f−1共xd兲, where d=1,2.
To better explain the concept, we use an example to discuss the approach. A quantized coefficient xQvalued at
392 is split into the ordered pair共x1, x2兲=共130,131兲, where 130 and 131 are the values assigned to descriptions 1 and 2, respectively. Similarly, a quantized coefficient xQvalued at
813 is split into the ordered pair共x1, x2兲=共271,270兲, where 271 and 270 are the values assigned to descriptions 1 and 2, respectively. On receiving only description 1 for the trans-mitted value 392, the estimated x= x1using x1= 130 will be
391; while receiving only description 2, the estimated x = x2using x2= 131 will be 394. Similarly, on receiving only description 1 for the transmitted value 813, the estimated
x
= x1 using x1= 271 will be 814; while receiving only de-scription 2, the estimated x= x2 using x2= 270 will be 811. As can be seen from this example, this leads to nonlinear value-metric distortion. The detailed algorithm for index assignment can be found in Refs. 6 and 7.
2.2 HLQ (Hexagonal Lattice Quantization)
It is well known that quantization error includes overload error and granular error.21 The overload error can be re-duced by vector quantization, and the granular error is af-fected by the size and shape of a quantization region. Since the best shape for a quantization region in two dimensions is a hexagon,22 we adopt an efficient lattice quantization algorithm23 for hexagonal lattice quantizer. In order to make this algorithm better fit our watermarking scheme, we employ the concept from nested lattice24 to implement the software.
2.2.1 Lattice quantization
An N-dimensional lattice is a set of points ⌳=兵x其=兵u1a1
+¯ +uNaN其, where x is an N-dimensional row vector
共point兲 in RN,兵a
1, . . . , aN其 is a set of basis vectors in RN,
and u1, . . . , uNrange through all integers. That is, x = uA; where x =关x1,x2, . . . ,xN兴, u =关u1,u2, . . . ,uN兴, and A =
冤
a1 a2 . . . aN冥
.The Voronoi region, or the nearest-neighbor region for the lattice⌳ with respect to x
⬘
, is defined asV共x
⬘
兲 = 兵x 苸 RN:储x⬘
− x储 ⱕ 储y − x储, for all y 苸 ∧ 其. Let Q be the quantization function mapping x苸Rn to thenearest point x
⬘
in⌳. To quantize x to x⬘
, we need to find nearest lattice points x⬘
to x, that is, Q共x兲=x⬘
. Some of the fast algorithms for lattice quantizer have been proposed in the literature.23–25We adopt the method proposed in Ref. 23 for its simplicity and then make some minor modifications on its format. This minor change makes the lattice quanti-zation feasible for watermark embedding and detection. Our modified quantizer algorithm for two-dimensional hex-agonal lattice A2in R2can be implemented as follows:Step 1: x = uA.
Step 2: u*= uA−1共as u*might not be integers, x*= u*A
might not be a lattice point, but a close point to a lattice point x
⬘
defined below兲.Step 3: u*=共u 1, u2兲.
Step 4: Round u* to integer point 共u
1, u2兲=关round共u1兲, round共u2兲兴.
Step 5: Find seven neighbor integer pairs for u*.
u共1兲=共u1, u2兲, u共2兲=共u1+ 1 , u2兲, u共3兲=共u1, u2+ 1兲, u共4兲=共u1+ 1 , u2+ 1兲, u共5兲=共u1− 1 , u2兲, u共6兲=共u1, u2− 1兲, u共7兲=共u1− 1 , u2− 1兲.
Step 6: Choose i*, where i* is the index i that gives the minimum u共i兲A−x
⬘
;i*= arg min
i 储u共i兲A − x储.
Step 7: Get the closest lattice point x
⬘
to x:x
⬘
= u共i*兲A.After the process of lattice quantization, the next step is to further partition the constructed lattice into co-sets. With the nested lattice structure, one is able to correctly deal with the payload issue.
2.2.2 Nested lattice
For the purpose of watermarking, the constructed lattice should be further partitioned into several co-sets, where points belonging to a different co-set represents a different watermark payload. Figures 2共a兲 and 2共b兲 depict the con-cept of nested lattice pair 共⌳f,⌳c兲,26 where ⌳f is a fine
lattice and⌳cis a coarse lattice. The generating matrix Af
of⌳f and Acof ⌳care related by Ac= JAf.
Lattice⌳f may be decomposed into兩det J兩 co-sets, and ⌳f
is the union of co-sets.⌳fand⌳care related by
⌳f= 艛 k=0 兩det J兩−1
⌳k.
For example, let
Af=
冋
1 0 − 1/2冑
3/2册
, J = 2I2=冋
2 0 0 2册
. By mapping 共1,2兲苸Z2 to共
1− 1 22, 冑3 22兲
苸A2, we can therefore use共1, 2兲苸兵共0,0兲, 共0,1兲, 共1,0兲, 共1,1兲其 to gen-erate 4 co-set leaders:C =兵d0,d1,d2,d3其 =
再
冉
1−1 22,冑
3 2 2冊
冎
=再
共0,0兲,冉
− 1 2 ,冑
3 2冊
,共1,0兲,冉
1 2,冑
3 2冊
冎
. The four co-sets are then defined as follows: ⌳0=⌳c+ d0,⌳2=⌳c+ d2, ⌳3=⌳c+ d3,
where∧1共marked with “o”兲, ∧2共marked with “䊐”兲 and ∧3 共marked with “䉭”兲 are the translated versions of ∧0 共marked with “.”兲 shifted by d1, d2, and d3, respectively 关see Fig. 2共c兲兴. Note that the generating matrix Afof⌳fand Ac of ⌳c共∧0兲 are related by Ac= 2Af, where the lattice
points of∧f are shown in Fig. 2共a兲, the lattice points of ∧c
are shown in Fig. 2共b兲, and the lattice points of ∧k 共k
= 0 , . . . , 3兲 are shown in Fig. 2共c兲.
3 The Proposed Multirate Lattice Quantization
Index Modulation„MRL-QIM… Watermarking
In this section, a multirate lattice quantization index modu-lation共MRL-QIM兲 watermarking scheme is described. The design goal of the MRL-QIM scheme is to use multirate watermark encoding to increase payload hiding and, at the same time, to increase the robustness of watermark detec-tion. The advantage of the proposed scheme is two-fold: First, it can increase the detection robustness for error-prone transmission over unreliable networks. Second, it is able to increase the watermark capacity while preserving the perceiving transparency. This is achieved by modulat-ing the selected coefficients pair appropriately so that two bits of information can be embedded.
Figure 4共a兲 shows the flow of MRL-QIM, which is com-posed of a watermark embedding process and a transmis-sion process. The original image is first transformed into the discrete cosine domain. The transformed coefficients are then grouped into 64 feature bands X共i, j兲
=共x1, x2, . . . , xc兲, where i=1, ... ,8; j=1, ... ,8, and c is the
total coefficient count of each feature band. Next, each of two bits共m兲 of the watermark message W is embedded in the selected two-coefficient pair 共where the selection de-pends on key K兲 using MRL-QIM quantizer as described in Section 3.1. The perturbed coefficients Y are then pro-cessed by multiple description scalar quantizer共MDSQ兲 to generate two independent descriptions, Y1and Y2, and sent through two independent channels. The watermarked im-ages共Iw1
⬘
from side decoder 1, Iw2⬘
from side decoder 2, orIw0
⬘
from central decoder兲 could then be obtained byreceiv-ing either one description 关decoder 1 共Y1兲 or decoder 2 共Y2兲兴 or two descriptions 关decoder 0 共Y0兲兴 and inversing the discrete cosine transforms. In the detection process 关Fig. 4共b兲兴, the received image Iwr
⬘
共r=0,1 or 2兲 first goesthrough a discrete cosine transform. The DCT coefficients are then grouped into 64 feature bands, Yr共i, j兲
=共x1, x2, . . . , xc兲, where i=1, ... ,8; j=1, ... ,8 and c is the
total number of coefficients of each feature band. Finally, apply the detection process of MRL-QIM on Yrdepending
on key K to extract the embedded watermark message W*. 3.1 MRL-QIM quantizer
The proposed MRL-QIM quantizer takes two DCT coeffi-cients each time to encode two watermark bits into each of the four co-set points of a lattice共so called multirate兲. For the host signal X, watermark message W, and key K, the Fig. 2 Nested lattice.共a兲 Fine lattice ∧f,共b兲 coarse lattice ∧c= 2∧f,
embedding function is defined as Y = f共X,W,K兲. For a chosen DCT coefficient pair x =共x1, x2兲苸X depending on key K, if Q共x兲 is used for finding the nearest point of a lattice⌳, then Q共x−dm兲+dm can be used for finding the
nearest point of a co-set ⌳+dm. To embed message m = 00 or 01 or 10 or 11苸W into the host signal x,
we calculate y = Q0共x−dm兲+dm to replace x with the
watermarked coefficient pair y =共y1, y2兲苸Y. For a received signal Yr共r苸兵0,1,2其兲 and key K, the detection function
is defined as W = f共Yr, K兲. To detect watermark
message from watermarked signal y苸Yr, we calculate m*= arg min
m
储y−Q0共y−dm兲−dm储 to get the watermark
message m*.
For example, as depicted in Fig. 3共a兲, if one wants to embed watermark bits m = 00, the original point marked with “+” is quantized to the point marked with “.”, with “x” superimposed on the latter. And similarly, as depicted in Fig. 3共b兲, if one wants to embed watermark bits m=01, the original point marked with “+” is quantized to the point
Fig. 4 共a兲 Flow of proposed MRL-QIM watermark embedding
scheme for error-prone transmission over unreliable network.共b兲 Flow of proposed MRL-QIM watermark detection scheme.
Fig. 3 Examples of multirate lattice watermark embedding.共a兲 For embedding watermark bits 00, the
original point marked with “+” is quantized to the point marked with “.”, with “x” superimposed on the latter;共b兲 for embedding watermark bits 01, the original point marked with “+” is quantized to the point marked with “o”, with “x” superimposed on the latter;共c兲 for embedding watermark bits 10, the original point marked with “+” is quantized to the point marked with “䊐”, with “x” superimposed on the latter; 共d兲 for embedding watermark bits 11, the original point marked with “+” is quantized to the point marked with “䉭”, with “x” superimposed on the latter.
marked with “o”, with “x” superimposed on the latter. Fig-ure 3共c兲 illustrates the case of embedding bits 10, the origi-nal point marked with “+” is quantized to that marked with “䊐”, with “x” superimposed on the latter. The case of em-bedding bits 11 is illustrated in Fig. 3共d兲.
3.2 The Embedding and Transmission Process
To embed n bits of watermark message W, the algorithm is described as follows:
1. The original image I is transformed using an 8-by-8-block DCT transform.
2. The DCT coefficients are then grouped into 64 fea-ture bands X共i, j兲=共x1, x2, . . . , xc兲, where i=1, ... ,8,
j = 1 , . . . , 8, and c is the total number of coefficients
of each feature band.
3. Apply embedding process of MRL-QIM on X de-pending on key K to embed watermark message W to obtain perturbed coefficients Y.
4. Each of the perturbed coefficients Y is quantized by a uniform scalar quantizer.
5. Two descriptions 共Y1, Y2兲 of the quantized cient are created by mapping each quantized coeffi-cient of YQto a pair of numbers by the index
assign-ment component.
6. Transmit these two watermarked descriptions over network via two different channels.
7. Apply an inverse transform to obtain a watermarked image Iwr
⬘
共r=0,1, or 2兲 depending on receivedde-scriptions Yr共r=0,1, or 2兲. 3.3 The Detection Process
To extract n bits of watermark message W*, the algorithm is described as follows:
1. The received image Iwr
⬘
is transformed using an8-by-8, block DCT transform.
2. The DCT coefficients are then grouped into 64 fea-ture bands Yr共i, j兲=共x1, x2, . . . , xc兲, r=0,1, or 2, i
= 1 , . . . , 8; j = 1 , . . . , 8, and c is the total number of coefficients of each feature band.
3. Apply detection process of MRL-QIM on Yr 共r
= 0 , 1, or 2兲 depending on key K to get the extracted watermark message W*.
4 Experimental Results
To evaluate the effectiveness of the proposed method, ex-perimental simulations on both of the Monte Carlo simu-lated Gaussian images and several real images共Lena, Bar-bara, House, and boat兲 were performed. To save space, only “Lena” 关Fig. 5共a兲兴 and “Barbara” 关. 5共b兲兴 as well as the average of detection ratios of Gaussian images are given here. For each run of Monte Carlo simulation, host signals
X drawn from 256⫻256 samples of a Gaussian zero-mean
random variable were generated, each having standard de-viationX ranging from 10 to 100 with step size 10. All
these Gaussian simulated data were then normalized to the range of 0⬃255 to simulate Gaussian gray-level images. We performed 100 times on each of the above simulations
using different seeds, so that 1000共10⫻100兲 Gaussian im-ages were employed to obtain the average detection ratios. In order to further demonstrate the effectiveness of our pro-posed MRL-QIM, the state-of-the-art watermark technique QIM13 and a combined spread spectrum and QIM 共SS-QIM兲18,27
were simulated for comparison. The SS-QIM scheme utilizes spread spectrum approach, in which a wa-termark strength weighting parameter␣is needed, to obtain a correlation value. This correlation value will then be quantized based on specified embedding quantization step size␦ and watermark bit共0 or 1兲 to produce a watermark value.
For the MRL-QIM scheme, two DCT transformed coef-ficients were simultaneously used to embed two bits of wa-termark information, and 1024 coefficients in total were used to embed 1024 bits of watermark information. For the traditional vector QIM scheme, two DCT transformed co-efficients were formed as a vector to embed one bit of watermark information, and 2048 coefficients in total were used to embed 1024 bits of watermark information. As for SS-QIM, an algorithm adopted from Ref. 27 was imple-mented, and the trellis error correction coding module was removed to make fair comparisons. Note that the detection rate can be further enhanced by employing an error correc-tion coding module for all these three schemes.
From our experiments, the degree of PSNR dropped de-pending on the embedding quantization step size. A larger quantization step size brought more robustness, but it also introduced more distortion. We follow the common practice by fixing two requirements, namely watermark capacity and the transparency 共distortion兲 of watermarked image, and then comparing the robustness. To make the compari-son fair, the parameters that defined the embedding quanti-zation step size or watermark strength weighting parameter were adjusted so that similar PSNR values 共about 41 dB兲 for these three schemes could be obtained. In our setting, the embedding quantization step size was set to 28, 44, and 7 for MRL-QIM, vector-QIM and SS-QIM, respectively. And for SS-QIM, the other watermark strength weighting parameter␣was set to 0.9. The original and watermarked images for MRL-QIM were shown in Figs. 5共a兲 and 5共b兲 共one sample of Gaussian images兲, Figs. 5共c兲 and 5共d兲 共Lena兲 and Figs. 5共e兲 and 5共f兲 共Barbara兲, respectively.
To evaluate the reliability of watermark detection, the detection ratio was defined as
=total number of correctly detected bits
total number of embedded bits . 共3兲 A higher value ofindicated a more reliable detection. The perfect recognition rate could be achieved when the value of= 1.
In addition to the degree of robustness against packet loss, a desirable and fundamental property for a watermark-ing algorithm is to survive compression attack. In real-world applications, compression is frequently used to facili-tate efficient storage and transmission. Here, we used images compressed by JPEG 共low-quality factor ranging from 60 to 80兲 as test images.
The performance of the detection on receiving only de-scription 1共similar results can be obtained via description 2兲 against the MD attack over various transmission rates is
Fig. 5 共a兲 Original Lena 共256⫻256 with a gray-scaled level兲. 共b兲 Watermarked Lena 共PSNR 41.10兲. 共c兲
Original Barbara共256⫻256 with a gray-scaled level兲. 共d兲 Watermarked Barbara 共PSNR 41.11兲. 共e兲 One sample of original Gaussian images共256⫻256 with a gray-scaled level兲. 共f兲 Watermarked Gauss-ian image共PSNR 41.68兲.
evaluated first. The detection ratios of “Lena,” “Barbara,” and 1000 Gaussian images are depicted in Figs. 6共a兲–6共c兲, respectively. It is clear that the proposed MRL-QIM outper-formed traditional vector QIM for all testing images over all transmission rates, and MRAL-QIM performed better than SS-QIM except in the case of high-rate transmission 共MD quantization step size⫽18兲 for the Gaussian and “Bar-bara” images.
As for JPEG compression attack, the detection ratios of
“Lena,” “Barbara,” and 1000 Gaussian images are shown in Figs. 7共a兲–7共c兲, respectively. The performance of proposed MRL-QIM is still better than traditional vector QIM and superior to SS-QIM for most of the compression rates, ex-cept at JPEG quality⫽60 for “Lena” and “Barbara.”
Note that, even though the detection capability of SS-QIM approach is not as good as the other two schemes when under weaker attacks, it has a smoother decay in detection ratios than pure quantization-based 共vector QIM and MRL-QIM兲 ones when the attacks become stronger. Fig. 6 The comparison in terms of detection ratios among QIM,
SS-QIM, and proposed MRL-QIM against various MD transmission rates.共a兲 “Lena”; 共b兲 “Barbara”; 共c兲 the average of 1000 Gaussian images.
Fig. 7 The comparison in terms of detection ratios among QIM,
SS-QIM, and proposed MRL-QIM against JPEG compression.共a兲 “Lena”;共b兲 “Barbara”; 共c兲 the average of 1000 Gaussian images.
Based on the experimental result, we speculate that the pos-sible combination SS-MRL-QIM would be a very interest-ing topic worth further investigation.
5 Conclusion
We have presented in this paper an MRL-QIM watermark-ing scheme that is robust to nonlinear value-metric distor-tion introduced by MD transmission. For a tradidistor-tional bal-anced two-description case in a packet transmission network, the embedded watermark can be extracted with the reception of either one or two descriptions. The experi-mental result shows that the proposed MRL-QIM outper-forms traditional vector QIM overall and peroutper-forms better than SS-QIM in the case of high-rate transmission. Further-more, in the case of compression attack, the performance of proposed MRL-QIM still performs better than traditional vector QIM and superior to SS-QIM for most of the com-pression rates. In the future, we expect to seek other trans-forms and statistic models to further enhance the robustness and increase the watermark payload while preserving the visual quality of the transmitted image. Furthermore, the steganography security against statistical steganalysis should also be addressed to enhance the security for reli-able transmission.
References
1. F. Hartung and F. Ramme, “Digital rights management and water-marking of multimedia content for M-commerce applications,” IEEE
Commun. Mag. 38, 78–84共2000兲.
2. N. Checcacci, M. Barni, F. Bartolini, and S. Basagni, “Robust video watermarking for wireless multimedia communications,” IEEE
Wire-less Commun. Networking Conf. 3, pp. 1530-1535, Chicago共2000兲.
3. B. Graubard, R. Chandramouli, and C. Richmond, “A multiple de-scription framework for oblivious watermarking,” Proc. SPIE 4314, 585-593共2001兲.
4. Y. Wang, S. Wenger, J. Wen, and A. K. Katsaggelos, “Error resilient video coding techniques,” IEEE Signal Process. Mag. 17共4兲, 61–82 共2000兲.
5. Y. Wang, M. T. Orchard, V. A. Vaishampayan, and A. R. Reibman, “Multiple description coding using pairwise correlating transforms,”
IEEE Trans. Image Process. 10共3兲, 351-366 共2001兲.
6. S. D. Servetto, K. Ramchandran, V. A. Vaishampayan, and K. Nahr-stedt, “Multiple description wavelet based image coding,” IEEE
Trans. Image Process. 9共5兲, 813-826 共2000兲.
7. V. A. Vaishampayan, “Design of multiple description scalar quantiz-ers,” IEEE Trans. Inf. Theory 39共3兲, 821-834 共1993兲.
8. V. K. Goyal, “Multiple description coding: Compression meets the network,” IEEE Signal Process. Mag. 18共5兲, 74–93 共2001兲. 9. Y. Wang, A. R. Reibman, and S. Lin, “Multiple description coding
for video delivery,” Proc. IEEE 93共1兲, 57-70 共2005兲.
10. I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “Secure spread spectrum watermarking for multimedia,” IEEE Trans. Image Process.
6共12兲, 1673-1687 共1997兲.
11. C. S. Lu, S. K. Huang, C. J. Sze, and H. Y. M. Liao, “Cocktail watermarking for digital image protection,” IEEE Trans. Multimedia
2共4兲, 209-224 共2000兲.
12. H. S. Malvar and D. A. F. Florêncio, “Improved spread spectrum: A new modulation technique for robust watermarking,” IEEE Trans.
Signal Process. 51共4兲, 898-905 共2003兲.
13. B. Chen and G. W. Wornell, “Quantization index modulation: A class of provably good methods for digital watermarking and information embedding,” IEEE Trans. Inf. Theory 47共4兲, 1423-1443 共2001兲. 14. J. J. Eggers, R. Bäuml, R. Tzschoppe, and B. Girod, “Scalar costa
scheme for information embedding,” IEEE Trans. Image Process.
51共4兲, 1003-1019 共2003兲.
15. J. Oostveen, T. Kalker, and M. Staring, “Adaptive quantization wa-termarking,” in Security, Steganography, and Watermarking of
Mul-timedia Contents VI, Proc. SPIE 5306, 296-303共2004兲.
16. F. P. González, C. Mosquera, M. Barni, and A. Abrardo, “Rational dither modulation: A high-rate data-hiding method robust to gain at-tacks,” IEEE Trans. Signal Process. 53共10兲, 3960-3975 共2005兲. 17. P. Bas, “A quantization watermarking technique robust to linear and
non-linear valumetric distortion using a fractal set of floating quan-tizers,” Inform. Hiding Workshop, pp. 106-117, Barcelona, Spain 共2005兲.
18. M. L. Miller, G. J. Doerr, and I. J. Cox, “Applying informed coding and informed embedding to design a robust, high capacity water-mark,” IEEE Trans. Image Process. 13共6兲, 792-807 共2004兲. 19. A. Abrardo and M. Barni, “Informed watermarking by means of
or-thogonal and quasi-oror-thogonal dirty paper coding,” IEEE Trans.
Sig-nal Process. 53共2兲, 824-833 共2005兲.
20. M. L. Day, S. Y. Lee, and I. C. Jou, “Multiple description watermark-ing based on quantization index modulus modulation,” J. Inform. Sci.
Eng.共in press兲.
21. K. Sayood, “Vector quantization,” in Introduction to Data
Compres-sion, Chap. 9, pp. 246–249, Morgan Kaufmann Publishers, San
Fran-cisco共1996兲.
22. D. J. Newman, “The hexagon theorem,” IEEE Trans. Inf. Theory
28共4兲, 137-139 共1982兲.
23. K. Sayood and S. J. Blankenau, “A fast quantization algorithm for lattice quantizer design,” 2nd IEEE Int. Conf. on Acoustics, Speech,
and Signal Processing (ICASSP), pp. 1168-1171, New York共1988兲.
24. J. H. Conway and N. J. A. Sloane, “Fast quantizing and decoding algorithms for lattice quantizers and codes,” IEEE Trans. Inf. Theory
28共2兲, 227-232 共1982兲.
25. E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, “Closest point search in lattices,” IEEE Trans. Inf. Theory 48共8兲, 2201-2214 共2002兲. 26. R. Zamir, S. Shamai, and U. Erez, “Nested linear/lattice codes for
structured multiterminal binning,” IEEE Trans. Inf. Theory 48共6兲, 1250-1276共2002兲.
27. I. J. Cox, M. L. Miller, and J. A. Bloom, “Watermarking with side information,” in Watermarking: Principles & Practice, Chap. 5, pp. 144–151, Morgan Kaufmann Publishers, San Francisco共2001兲.
Miin-Luen Day received his BS degree in
computer science and information engineer-ing and MS degree in electronic engineerengineer-ing from Chung-Yuan Christian University, Tai-wan, in 1986 and 1990. He joined the Tele-communication Laboratories of Chunghwa Telecom Co., Ltd., in 1990 and is now pur-suing his PhD degree in the Department of Computer Science and Information Engi-neering at Chiao Tung University, Taiwan. His current research interests include multi-media security, multimulti-media communication, image processing, and pattern recognition.
Suh-Yin Lee received her BSEE degree
from the National Chiao Tung University, Taiwan, in 1972, and her MS degree in com-puter science from the University of Wash-ington, Seattle, in 1975. She joined the fac-ulty of the Department of Computer Engineering at Chiao Tung University in 1976 and received the PhD degree in elec-tronic engineering there in 1982. Dr. Lee is now a professor in the Department of Com-puter Science and Information Engineering at Chiao Tung University. Her current research interests include mul-timedia information systems, mobile computing, and data mining. Dr. Lee is a member of Phi Tau Phi, the ACM, and the IEEE Com-puter Society.
I-Chang Jou received his BS degree in
electrical engineering from National Taiwan University, Taiwan, in 1969, his MS degree in geophysics and in computer science from National Central University, Taiwan, in 1972 and 1983, respectively, and his PhD degree in electrical engineering from National Tai-wan University, TaiTai-wan, in 1986. He was with Telecommunication Labs., Ministry of Communications, Taiwan, from 1972 to 1997. Currently, he is the president of Na-tional Kaohsiung First University of Science and Technology. His major research fields are VLSI for DSP, digital signal processing, image processing, speech processing, and neural networks. He has published over 131 papers in the areas of parallel computing, image processing, speech processing, and neural networks. He is the se-nior member of IEEE.
