Data hiding based quality access control of digital images using adaptive QIM and lifting

Contents lists available at SciVerse ScienceDirect

Signal Processing: Image Communication

Signal Processing: Image Communication 26 (2011) 646–661

0923-59

doi:10.1

n Corr

E-m

santipm

journal homepage: www.elsevier.com/locate/image

Data hiding based quality access control of digital images usingadaptive QIM and lifting

Amit Phadikar a, Santi P. Maity b,n

a Department of Information Technology, MCKV Institute of Engineering, Liluah, Howrah 711204, Indiab Department of Information Technology, Bengal Engineering and Science University, Shibpur, Howrah 711103, India

a r t i c l e i n f o

Article history:

Received 10 December 2009

Accepted 31 July 2011Available online 25 August 2011

Keywords:

Data hiding

Data modulation

Access control

Security

Lifting

QIM

65/$ - see front matter & 2011 Elsevier B.V. A

016/j.image.2011.07.008

esponding author. Tel.: þ91 33 2668 4561; fax:

ail addresses: [email protected] (

[email protected] (S.P. Maity).

a b s t r a c t

This paper proposes a joint data-hiding and data modulation scheme to serve the

purpose of quality access control of image(s) using quantization index modulation

(QIM). The combined effect of external information embedding and data modulation

cause visual degradation and may be used in access control through reversible process.

The degree of deterioration depends on the amount of external data insertion, amount

of data modulation as well as step size used in QIM. A weight function (q) and the

quantization step size (Db) of JPEG 2000 are used for defining the step size of QIM.

Lifting based discrete wavelet transform (DWT), instead of conventional DWT, is used to

decompose the original image in order to achieve advantages, namely low loss in image

quality due to QIM, better watermark decoding reliability and high embedding capacity

for a given embedding distortion. At the decoder, data are demodulated first and

watermark bits are then extracted using minimum distance decoding. Extracted

watermark is used to suppress self-noise (SN) that provides better quality of image.

Simulation results have shown that an improvement in peak-signal-to-noise-ratio

(PSNR), mean structural-similarity-index-measure (MSSIM) and Watson distance by

an amount of about 30%, 12% and 77%, respectively, are gained by authorized user when

50% coefficients of high–high (HH) coefficients are modulated.

& 2011 Elsevier B.V. All rights reserved.

1. Introduction

Today Internet brings the world to the origination.With the emergence of the communication network andWorld Wide Web (WWW), digital data can travel fromone source to an endless range of targets in any part of theworld. Simultaneously with the growth in trade andindustry, more and more organizations move intoE-Commerce and E-business, conducting transaction online[1]. The sale of high value products and services throughInternet are also on the rise. Moreover, multimedia data

ll rights reserved.

þ91 33 2668 2916.

A. Phadikar),

such as text, image, video and audio can be digitized andstored without any degradation in quality. Due to lack ofsecurity in the communication network, vendors may beunwilling to distribute their products such as digital imageover the Internet as digital media can be accessed illegally,tampered with, and copied without loss of quality for thepurpose of reselling violating the rules of copyright.Since commercial interests like to use the digitalnetworks for profit making, vendors have a keen interestin somehow protecting their digital contents. Access controltechnique may be used in providing a kind of securityeither to deny fully or to allow partial accessing of thedigital content [2,3].

The service providers and the vendors, in general, havetwo different trade-off objectives. On one hand, they wantto put their large volume of creative works in the website

www.elsevier.com/locate/image

www.elsevier.com/locate/image

dx.doi.org/10.1016/j.image.2011.07.008

mailto:[email protected]

mailto:[email protected]

dx.doi.org/10.1016/j.image.2011.07.008

A. Phadikar, S.P. Maity / Signal Processing: Image Communication 26 (2011) 646–661 647

for the large media hype. On the other hand, they normallyallow un-authorized users to enjoy a certain low level ofquality with an expectation that the un-authorized viewersmay become the potential consumers in near future. Thusan efficient access control scheme typically allows all thereceivers of the broadcast channel to display a low qualitydata, for example, image with no or little commercial value.But in the meantime, the authorized users would enjoyaccess of an image at higher quality levels depending onaccess rights that usually are determined by the subscrip-tion agreement. Such digital content, in general, are storedin their standard compressed form like JPEG and morerecently like JPEG 2000. Typical applications of accesscontrol may be seen in future generation mobile commu-nication system, video-on demand and real-time videomulticast system where royalty is regulated by the fulfill-ment of degree in quality of services (QoS).

It has been seen in recent times that significant researcheffort has been devoted for access control. Access control ofdigital data is normally accomplished through the modula-tion of transform coefficients like discrete cosine transform(DCT), discrete wavelets transform (DWT), etc., of theimage and video signals using secret key. Typically twoclasses of techniques namely, data encryption and datahiding either independently or in combined form are usedto meet the goal of access control. We classify them hereas non-data hiding and data hiding based access controlmethods, respectively.

1.1. Non-data hiding based access control

Data encryption is a popular technique to establishsecurity in commutation. Some of these concepts are nowused in access control. Grosbois et al. [2] propose a jointauthentication and access control (on resolutions andqualities) technique for images in wavelet domain thatcan be easily integrated in a JPEG-2000 codec. The schemeuses encrypted hash information for checking authentica-tion. The goal of access control in resolution level and inquality are achieved by adding pseudo-random noise inthe high frequency subbands and by scrambling of lastlayer, respectively. Imaizumi et al. [3] offer a new private-key encryption for JPEG 2000 code streams for flexibleaccess control of layers, resolution levels and color com-ponents. The above cryptographic approaches used foraccess control are usually complex in nature. Moreover,an error in a cipher text may lead to a partially or totallyfailed decryption. In a broadcast environment with largevolume of data exchange, error in transmission may occurfrom time to time. Retransmission of data is not allowedand leads to a serious problem in data communication.

To overcome the above problem, Wen et al. [4] introducea new technique of performing selective encryption andspatial/frequency shuffling of compressed MPEG-4 videocontent in the wireless channel. Wang et al. [5] also offer atransparent scrambling technique, where the image degra-dation is varied according to parameter control. Kankahalliet al. [6] propose an access control scheme on video usingjoint lightweight encryption and compression. Bertino et al.[7] suggest a novel approach to support multilevel accesscontrol in video database so that different classes of users

can access different video elements (a semantic cluster, a subcluster, a video scene, a video shot, a video frame, or even asalient object i.e. region of interest) or even the same videoelement with different qualities according to their permis-sions. Recently, Phadikar et al. [8] propose a quality accesscontrol for DCT compressed gray scale image. The secret keyis padded as a metadata at the end of the bit string.However, the padding of a secret key as a metadata maycreate a risk as the metadata may be lost especially if thecontent undergoes a variety of format changes. The aboveproblem is overcome in [9] where the secret key isembedded as a watermark in the host data for its efficientand secured transmission.

1.2. Data hiding based access control

Digital data hiding has been used widely in a numberof applications like copyright protection, ownership ver-ification, authentication, secured communication, etc.[10,1]. Recently data hiding is used for the assessmentof quality of services (QoS) and access control of multi-media data due to commercial and/or security reasons[11–14,15]. Pickering et al. [16] propose a blind data-hiding scheme in dual-tree complex wavelet transform(DT-CWT) for access control of video where compliantDVD players deny access to the pirated copies of video.The watermark length is of six bits and is embedded inthe high textured areas of the video frames. The workreports that complex wavelets can provide both shiftinvariance and good directional selectivity, with only amodest increase in signal redundancy and computationalload. Coria et al. [17] offer a similar type of work foraccess control of DVD player by developing a contentdependent watermark using dual tree complex wavelettransform (DT-CWT) and embed the same into theselected frequency components of the video sequence.Chang et al. [18] offer a structure to perform layeredaccess control for scalable media. Authors pointed outthat by taking advantages of both encryption and water-marking, copyrights of multimedia contents can be wellprotected and at the same time multiple-grade servicescan be provided.

It is found that most of the traditional DWT and DCTbased access control schemes reported in the literaturesuffer from various shortcomings. The most prominentshortcoming is high computation complexity that makesthe algorithms unsuitable for faster implementation.It is reported in the literature that compared to DCT,conventional DWT has less computational cost [19]. Inspite of this advantage of DWT, computation cost is still aproblem when DWT is applied to the whole image withlarge size [19,20]. The computation issue becomes animportant selection criterion for an efficient design, asmany algorithms in recent time demand real-time imple-mentation for which hardware realization becomes aviable alternative. Lifting scheme is an effective methodto improve the computation speed of conventionalDWT as well as for digital design [21]. Integer waveletcoefficients cause no loss in signal values during analysisand synthesis [22]. This makes lifting based DWT as a

Table 1Entropy of different subbands for normal DWT and lifting.

Low–low(LL)

High–low(HL)

Low–high(LH)

High–high(HH)

Lifting 7.26 5.24 5.42 5.40

DWT 10.41 8.30 6.78 8.62

A. Phadikar, S.P. Maity / Signal Processing: Image Communication 26 (2011) 646–661648

potential tool to develop cost effective access control forcompressed image and video signals.

Quantization index modulation (QIM) has recentlybecome a popular form of data hiding based on the frame-work of communications with side information [23,24]. ForQIM scheme, a common practice for ensuring low distor-tion and reducing the interference between the watermarkand the host media is the so-called informed embedding inwhich the information about the host media is exploited bythe embedder [25]. Chen and Wornell [26] have proposed awatermarking technique using dither modulation (DM),which is a special case of QIM for full self-noise (SN)suppression. In [27], authors proposed a hidden QIM water-marking for compressed digital image using channel codingand lifting. The combined use of channel coding and liftingrecreates a virtual redundancy space that was lost due tocompression and offers flexibility in watermark casting. Thelong run goal is to design optimized watermarking oncompressed data. The implementation is not intended forthe application to quality access control, instead to protectcopyright of the document.

The objective of quality access control is a compromisebetween full denial and complete acceptance. In thispaper, we propose a quantization based data-hidingscheme together with data modulation to achieve thegoal. The proposed scheme embeds the watermark usingDM in the DWT coefficients that are obtained using Lifting.Integration of QIM and lifting offers advantages of (1) lowbit error rate (Pe) in binary watermark decoding comparedto conventional DWT, as the former shows low variance ofthe cover coefficients compared to the latter and (2) lowloss in image information due to data embedding. The stepsize for QIM watermarking is determined by exploring therelationship among the wavelet coefficients, human visualsystem (HVS) and quantization step size (Db) used instandard JPEG 2000. This combined effect leads to achievebetter invisibility, robustness and gain in access control.Moreover, the scheme only modulate the sign bit of theselected coefficients that has the following advantages, (1)change in the sign of DWT coefficients may reduce theunderstandability of the images, (2) modulation of sign bitfor the selected coefficients keep compression ratiounchanged and also makes the encryption process be oflow computation cost. Simulation results show that theuser having full knowledge of the key can get the bestcopy of the image, while all other users can only access theimage quality to a certain level.

The rest of the paper is organized as follows: basicprinciples and key features of lifting based DWT areoutlined in Section 2. Section 3 describes proposed water-mark encoding and decoding scheme. In Section 4, wefocus on the error performance of adaptive QIM basedwatermarks, while the empirical performance of theproposed access control scheme is demonstrated inSection 5. Finally conclusions are drawn in Section 6along with the scope of future works.

2. Basic principles and key features of lifting based DWT

The lifting scheme is a technique both designing fastwavelets and performing the discrete integer wavelet

transform. The technique was introduced by Wim Sweldens.Lifting-based filtering consists of a sequence of very simplefiltering operations for which alternately odd sample valuesof the signal are updated with a weighted sum of evensample values and even sample values are updated with aweighted sum of odd sample values [28–30] as described byEqs. (1) and (2) [31,32]

yð2nþ1Þ ¼ xextð2nþ1Þ�xextð2nÞþxextð2nþ2Þ

2

� �ð1Þ

yð2nÞ ¼ xextð2nÞþyð2n�1Þþyð2nþ1Þþ2

4

� �ð2Þ

where xext is the extended input, y is the output signal andab c indicates the largest integer not extending ‘a’ [31,32].

This mathematical form of lifting allows creation of correla-tion among the sample values. This form of correlation maybe beneficial for data hiding on the coefficients and leads tobetter image visual quality of the watermarked data. Tosupport the above argument we calculate the entropy ofdifferent subbands for both traditional DWT and liftingbased coefficients. The average result is shown in Table 1.The lower values of entropy for lifting based decompositionindicate that there exits large correlation between thecoefficients in lifting method than the traditional DWT. Inother words, the large values of correlation indicate lowrandomness and high redundancy in the coefficients. Datahiding based access control exploits the advantages of thisredundancy to embed more bits of watermark for a givenembedding distortion that leads to better access controlthrough the reversible process. Interested readers may con-sult [33] for further information regarding various benefits oflifting operation over conventional DWT.

3. Proposed watermark encoding and decoding scheme

In this section, we present the encoding and thedecoding schemes of the proposed method. The objectiveof the encoding scheme is to hide a watermark and tomodulate some coefficients in the cover image. On theother hand, the decoding scheme removes the self-noise(SN) due to data embedding from the watermarked imageand also demodulates the selected coefficients, thus offersbetter image quality.

3.1. Watermark encoding

The block schematic of the proposed encoding schemeis shown in Fig. 1. The encoding process is performed indifferent steps as follows:

Step 1: creation of binary watermark: in the presentscheme, some coefficients of HH subbands are selected

Fig. 1. Block diagram of watermarking process: encoder.

Fig. 2. (a) Original watermark; (b) permuted watermark.

Table 2Variance values for different subbands due to lifting and

traditional DWT.

Low–low(LL)

High–low(HL2)

Low–high(LH2)

High–high(HH2)

Lifting 2165.9 91.62 219.46 91.62

DWT 34,656.0 364.51 877.36 147.19


pseudo randomly depending on secret key (P) supplied byowner of the media. The key P is generated through theselection of an integer number. In other words, by choos-ing this integer number, the secret key stores the posi-tional information of HH coefficients to be selected fordata modulation. For efficient transmission of secret key,the potential information i.e. the integer value is nowmapped to a binary image, which would be used aswatermark (W). Fig. 2(a) shows an example of binarywatermark (where the pixel value is either 0 or 1) withsize (32�32). This image indeed represents an integer‘‘2008’’, for example as a binary image, which would beused to generate the pseudo-random sequence.

Step 2: permuted watermark calculation: in order toconstruct a good watermark for embedding, the originalwatermark is spatially dispersed using a pseudo-randomsequence and is shown in Fig. 2(b). The pseudo-randombits are generated using a secret key. Let the originalwatermark (W) and a 2-D pseudorandom binary sequenceK (which is used as a secret key), each with size ðn� nÞ, bedescribed as follows:

W ¼ fwði,jÞ,1r i,jrn,wði,jÞ 2 ð0,1Þg ð3Þ

K ¼ fkði,jÞ,1r i,jrn,kði,jÞ 2 ð0,1Þg ð4Þ

The permuted watermark is calculated as follows:

W 0 ¼W � K

where � denotes the X-OR operation.Step 3: selection of proper transform coefficient: the

present scheme uses lifting based DWT (LWT) for decom-position of cover signals. It has been shown that the biterror rate (Pe) in binary watermark decoding is related tothe standard deviation of the cover image coefficients [34]

as follows:

Pe ¼2ðM�1Þ

MU

ffiffiffiffiffiffiffiffiffiNd2

0

4s2x

s0@

1A ð5Þ

where d0 indicates step size (D), M indicates the numberof different level of step sizes, sx

2 is the variance of animage block, Uð:Þ indicates the complimentary error func-tion and N is the number of cover signal points over whicha single watermark bit is embedded. It is seen from thesimulation results over large number of images that thevariance of lifting based DWT coefficients is much lowerthan DWT coefficients. The result is shown in Table 2 andsupports the use of lifting based method for cover imagedecomposition in order to design robust watermarking.

Step 4: image transformation: lifting-based n-level 2DDWT is performed on the original image to decompose itinto high and low subbands as shown in Fig. 3(a). Thenumber of levels of wavelet decomposition is implemen-tation dependent; however, four (4) levels are demon-strated in the present experimentation. This is due to thefact that the test image is of size (512�512) and thewatermark is of size (32�32). In the present scheme, onewatermark bit is embedded by modulating a group ofcoefficients, one coefficient is selected from LL subband.As a matter of fact, the test image is decomposed into fourlevels so that it generates a LL subband of size (32�32).

Fig. 3(b) shows 4-level DWT of the 8-bits/pixel Lenaimage tile, while Fig. 3(c) shows the same due to liftingbased DWT decomposition. A subjective visualizationshows that lifting based operation offers relatively bettermultiresolution information processing capability com-pared to the normal DWT operation. We have performedour experimentation using the lifting based ‘haar’ waveletfor image transformation due to its simplicity and ease of

Fig. 3. (a) The lifting based DWT structure; (b) 4-level DWT (haar) of the

8-bit Lena image tile; (c) 4-level lifting based DWT (haar) of the 8-bit

Lena image tile.

Fig. 4. Selected coefficients for one bit of watermark embedding.


implementation, although other lifting based waveletscan also be used.

Step 5: coefficients selection criteria for one bit of water-

mark embedding: one bit of spatially dispersed watermark(W0) is inserted into different energy levels. The choice ofembedding the watermark information into the HH sub-bands of all levels and all subbands of the rth level wasmotivated by the experimental tests, as this one offers thebest compromise between robustness and invisibility.Moreover, in the case of scalable decoding if only thehigh-energy subbands are sent to the decoder, watermarkcan be detected efficiently from those subbands withoutwaiting for the others. Fig. 4 shows the group of selectedcoefficients (G) for one bit of watermark insertionwhen the image is decomposed using 4-level DWT. The

subbands LL, HL, LH and HH denote low–low, high–low,low–high and high–high coefficients, respectively.

Step 6: categorization of coefficients: the selected coeffi-cients (G) are divided into ‘b’ number of different cate-gories (C) depending on the subbands in which they fall

C ¼ ðC1,C2,C3,. . .. . .CbÞ ð6Þ

In present scheme, the numbers of categories (b) are of 7types as the host image is decomposed into 4 levels andone watermark bit is embedded into ‘b’ number ofdifferent subbands.

Step 7: selection of steps size for QIM based dither

modulation: the step sizes ðDbÞ for a category/subbands(b) are calculated using a weight function (q) and thequantization step size (Db) of JPEG-2000. The weightfunction is constructed based on the relation betweenthe wavelet coefficients and HVS. The weight function (q)is adapted for better imperceptibility, while the step sizeDb is adapted to increase robustness against compression.

(1)
Calculation of weight function (q): Lewis and Knowles[35] propose a quantization method of DWT coeffi-cients for image compression, where the quantizationmatrix is modified in accordance with the local esti-mated noise sensitivity of HVS. This model is slightlymodified by Barni et al. [36] to fit it into the water-marking problem. They have used the model as aweighing function that allows exploiting the maskingcharacteristics of the HVS. However, this is used here tocalculate the step size for QIM to achieve betterimperceptibility due to data embedding. The quantiza-tion step size for each DWT coefficient is calculated bymeans of the weighted product of three (3) terms:
qsrðx,yÞ ¼ q0 � Fðr,sÞ � Iðr,x,yÞ � Tðr,x,yÞ0:2 ð7Þ

where q0 is a normalization constant. Let qsrðx,yÞ be the

coefficient value of the decomposition q at level r,orientation s 2 fLL,LH,HL,HHg and position x, y withinthat subband. The first term takes into account of noisesensitivity depending on the subband and can beexpressed as

Fðr,sÞ ¼

ffiffiffiffiffi2,p

if s¼HH

1, otherwise

( )�

1:00 if r¼ 0

0:32 if r¼ 1

0:16 if r¼ 2

0:10 if r¼ 3

8>>><>>>:

9>>>=>>>;ð8Þ

The second term takes into account the local bright-ness based on the gray level values of the lowpassversion of the image. Since Lewis and Knowles [35]assumed that the eye is less sensitive in the regionswith high brightness, they proposed to compute thisfactor in the following way:

Iðr,x,yÞ ¼ 1þ lðr,x,yÞ ð9Þ

where

lðr,x,yÞ ¼ 3þ1

256X3,LL 1þ

x

23�r

� �,1þ

y

23�r

� �� ð10Þ

It is reported in the work [37] that the majorproblems in a low intensity level (dark region) which


any image processing system has to face are (i)detection of changes occurring in a low steady butvisible illumination (i.e. minimum detectable change)and detection of the mere presence or absence of thelight under a dark adapted condition (i.e. absolutethreshold value). Accordingly, based on the consid-eration that the human eye is less sensitive to thechanges in very dark regions also [37,38], the factor ismodified as follows [36] to:

l0ðr,x,yÞ ¼

1�lðr,x,yÞ, if lðr,x,yÞo0:5

lðr,x,yÞ otherwise

( )ð11Þ

This modification of brightness component in Eq. (11)becomes effective and contributes in Eq. (7) to makethe quantization step size adaptive with HVS. Finally,the third term Tðr,x,yÞ gives a measure of textureactivity in the neighborhood of the pixel. In particu-lar, this term consists of the product of two contribu-tions. The first contribution represents the distancefrom the edges, whereas the second one correspondsto the texture. Those two terms are multiplied, as theeye is less sensitive in textured areas, but moresensitive near edges

Tðr,x,yÞ ¼X3�r

k ¼ 1

16�kXHH,HL,LH

s

X1

i ¼ 0

X1

j ¼ 0

ðXkþ r,sðiþx=2k,jþy=2kÞÞ

2

þ163�rvarðX3,LLðf1,2gþx=23-r ,f1,2gþy=23�rÞÞ ð12Þ

where varð:Þ is the variance of the four coefficientblocks, and the summation is zero when its lowerlimit exceeds its upper limit. Then the minimumvalue of ‘q’ is used as a final quantization step (qb)to quantize the entire band.

(2)
Calculation of Db: the quantization step size Db for theJPEG 2000 is calculated using Eq. (13) [39]
Db ¼ 2Rb�eb 1þmb

211

� �ð13Þ

where Rb is the nominal dynamic range of subband b.

The symbols eb and mb are the number of bits allottedto the exponent and mantissa of the subband’scoefficients, respectively.

(3)
Calculation of final step size ðDbÞ : to compromisebetween watermark robustness and watermarkobtrusiveness, a weighted sum of qb and Db isconsidered that is represented by Eq. (14)
Db ¼ @Db� ð1�DbÞþ@qb

� qb ð14Þ

where @Dband @qb

are the positive weight factors ofDb and qb, respectively. Each weighting factor repre-sents how important each index is during the adap-tive data embedding. For example, if both indices areequally important, they should be of 0.5 (the rela-tionship @Db

þ@qb¼ 1:0 must hold). In our simulation,

we provide equal importance on the both factors.

Step 8: generation of binary dither for each subband: letthe number of DWT coefficient(s) for a particular category/band be L. For coefficient(s) in category ‘b’, two dithersequences, each one of length L, are generated pseudo

randomly using a key as follows:

db,qð0Þ ¼ fRðkeyÞ � Dbg�Db=2 0rqrL�1 ð15Þ

db,qð1Þ ¼db,qð0ÞþDb=2 if db,qð0Þo0

db,qð0Þ�Db=2 if db,qð0ÞZ0

8<: ð16Þ

where RðkeyÞ is a random number generator. Note that theelements of db,q are real-valued.

Step 9: watermark insertion: in traditional QIM [26]based data hiding scheme, the qth watermarked DWTcoefficient Sq corresponding to a category ‘b’ is obtainedas follows:

Sq ¼QfXqþdb,qð0Þ,Dbg�db,qð0Þ if W 0ði,jÞ ¼ 0

QfXqþdb,qð1Þ,Dbg�db,qð1Þ if W 0ði,jÞ ¼ 1

8<: ð17Þ

where Q is a uniform quantizer (and dequantizer) withstep Db for category ‘b’. We have modified Eq. (17) to fitinto quality access control problem and is described as

Sq ¼QfXq�k0 � db,qð0Þ,Dbgþðk

0�nÞ � db,qð0Þ if W 0ði,jÞ ¼ 0

QfXqþk0 � db,qð1Þ,Dbg�ðk0�nÞ � db,qð1Þ if W 0ði,jÞ ¼ 1

8<:

ð18Þ

where Xq is the original qth DWT coefficient. The second partof Eq. (18) plays the key role in the access control process.The symbols k0 and n are positive integers and the relation-ship k04n must hold. The factor k0 represents the degree ofquality degradations for the image. The values of k0 and n areapplication dependent and would be high and low, respec-tively, if content owner wants to provide low quality imageto general users. In the present set of simulations, the value ofk0 and n are set to 2 and 1, respectively. One may usedifferent k0 and n values for efficient access control.

Step 10: modulation of coefficients: in our previous work[22], it is seen that QIM watermarking alone may not besufficient to provide an efficient access control aftercommon image processing operations. As a matter of fact,the sign bits of some selected HH coefficients are furthermodulated by the secret key (P) governed by the choice ofinteger number by the content owner. The coefficients inHH bands are selected for modulation as it containsinformation about both of horizontal and vertical edges.After modulation, inverse lifting based DWT is appliedand the watermarked image is formed.

3.2. Watermark decoding

The block schematic for watermark decoding is shownin Fig. 5 and the different steps are described as follows:

Step 1: steps 4, 5, 6 and 8 of watermark encodingprocess are performed on the watermarked or possiblydistorted watermarked image.

Step 2: watermark bit extraction: in traditional QIM[26] based data hiding, a watermark bit ~W ði,jÞ is decodedby examining the group of coefficients (G) of differentsubbands (Fig. 4) using the following rule:

A1 ¼XL�1

q ¼ 0

ð9Q ðYqþdbqð0Þ,DbÞ�dbqð0Þ�Yq9Þ ð19aÞ

Fig. 5. Block diagram of watermarking process: decoder.


B1 ¼XL�1

q ¼ 0

ð9Q ðYqþdbqð1Þ,DbÞ�dbqð1Þ�Yq9Þ ð19bÞ

We have modified Eq. (19a) and (19b) to fit into ourquality access control problem and is written as follows:

A2 ¼XL�1

q ¼ 0

ð9Q ðYq�ðk0�nÞ � dbqð0Þ,DbÞþðk

0�nÞ � dbqð0Þ�Yq9Þ

ð20aÞ

B2 ¼XL�1

q ¼ 0

ð9Q ðYqþðk0�nÞ � dbqð1Þ,DbÞ�ðk

0�nÞ � dbqð1Þ�Yq9Þ

ð20bÞ

where Yq is the qth DWT coefficient (possibly distorted) ofthe received signal for category/band ‘b’. The watermarkbit ~W ði,jÞ corresponding to a group of selected coefficients(G) is now decoded using the following rule:

~W ði,jÞ ¼0 if AioBi

1 otherwise

�ð21Þ

The watermark bits ~W ði,jÞ are then extracted bycomparing the value of ‘Ai’ and ‘Bi’ as shown in therespective cases of Eqs. (19a), (19b), (20a), (20b), (23a)and (23b) governed by the rules as stated in Eq. (21). Notethat ‘Ai’ and ‘Bi’ indicate the distance corresponding to thetwo different watermark bits, and the distance is obtainedby calculating the correlation between the received signaland dither vectors used. The relative distance between ‘Ai’and ‘Bi’ indicates decoding reliability of the binary water-mark where the large distance indicates low probability ofdecoding error.

Step 3: decoding of watermark bit: the extracted water-mark ð ~W Þ bits are reversely permuted i.e. spatially rear-ranged and are X-ORed with random bits to get thedecoded watermark ðWÞ. The random bits are generatedusing the same secret key that was used during the timeof watermark permutation at encoder.

Step 4: demodulation of coefficients: the decoded water-mark ðWÞ is now reverse mapped to obtain the positionalinformation (P) for the coefficients undergone modulation.The information is then used for data demodulation. TheHH coefficients that are used at the time of modulation aredemodulated first to undo the affect of modulation.

Step 5: self-noise (SN) cancellation for access control: theremaining self-noise (SN) due to watermark embedding issuppressed to get better quality of image using following

equation:

Y 0q ¼Yqþn� db,qð0Þ if ~W ði,jÞ ¼ 0

Yq�n� db,qð1Þ if ~W ði,jÞ ¼ 1

8<: ð22Þ

where Y 0q is the watermarked signal after self-noise (SN)elimination. Then inverse lifting based DWT is applied toget relatively high quality of the watermarked image. Oneof the interesting characteristics of the proposed accesscontrol scheme is that even after the operation is per-formed using Eq. (22), the signals Y 0q carry the watermark(W) signal. The watermark from Y 0q can be extracted usingthe following equations:

A3 ¼XL�1

q ¼ 0

ð9Q ðY 0q�k0 � dbqð0Þ,DbÞþk0 � dbqð0Þ�Y 0q9Þ ð23aÞ

B3 ¼XL�1

q ¼ 0

ð9Q ðY 0qþk0 � dbqð1Þ,DbÞ�k0 � dbqð1Þ�Y 0q9Þ ð23bÞ

The watermark bits ~W ði,jÞ are then extracted bycomparing the value of ‘A3’ and ‘B3’ in Eq. (23) withrespect to using Eq. (21). It is seen that robustnessperformance for the extracted watermark from Y 0q is muchhigher than the watermark extraction from Yq. This isquite reasonable as robustness of QIM based data hidingscheme is due to the contribution from the 2nd part ofEq. (18). Since after self-noise (SN) elimination usingEq. (22), the 2nd part is increased by a factor of ‘n’ thatultimately increases the robustness of the extractedwatermark. This is the other advantage for the proposedQIM based access control scheme over spread spectrum(SS) based watermarking method like [40]. In otherwords, after self-noise (SN) cancellation, the SS basedscheme [40] will fail to detect the watermark, while theproposed scheme can detect it with high confidence evenafter various common image and signal processing opera-tions. The above argument is supported by simulationresults shown in Fig. 17 in the next section.

4. Performance analysis of adaptive QIM watermarks

In this section, we first calculate the probability oferror ðreÞ in watermark decoding. Our primary focus is tohighlight how the performance of watermark decodingdepends on the steps size which is calculated based onboth Db and q. More correct decoding of watermark bitsleads to better quality of decoded host image throughreversible process. Since the value of decoded watermarkbit in (21) follows the principle of minimum distance

Table 3Analytical value of re for different subbands using Eq. (25). A: analytical, E: experimental.

Wavelets band DbRange of M M re Wavelets band Db

Range of M M re

A E A E

LWT DWT LWT DWT

HH1 62 M 2 ½62,186� 63 0.01 0.02 0.1 HH2 30 M 2 ½30,90� 31 0.03 0.05 0.09

124 0.50 0.5 0.54 60 0.50 0.5 0.65

185 0.66 0.61 0.68 89 0.66 0.61 0.68

M 2 ½186,310� 187 0.66 0.61 0.68 M 2 ½90,150� 91 0.66 0.61 0.68

248 0.50 0.5 0.54 120 0.50 0.45 0.62

309 0.40 0.42 0.47 149 0.40 0.42 0.47

Fig. 6. Test images. (a) Lena, (b) Pepper, (c) Baboon, (d) Boat, (e) Opera, (f) Cameraman, (g) Pueblo bonito and (h) F16.


decoding, the distortion within ½�Db=2,þDb=2� does notcause error in detection. The mean squared error (MSE)for QIM based embedding is given by [41]:

MSE¼1

2

D2

b

12þ

1

2

1

Db

Z 0

�Db=2

ðxþDbÞ2dxþ

1

Db

Z Db=2

0ðx�DbÞ

2dx

" #

¼D

2

b

3ð24Þ

In general, the error would be introduced providedthat the noise (channel noise like Gaussian) falls in½ð4jþ1ÞDb=2,þð4jþ3ÞDb=2� for some positive integer j

[41]. If we assume that the channel noise is uniformlydistributed between �M=2 and M=2 with M4Db, theprobability of error in watermark bit decoding can beexpressed on interval basis as follows:

re ¼1� ð2K�1ÞDb

M if M 2½ð4K�3ÞDb,ð4K�1ÞDb�

2KDbM if M 2½ð4K�1ÞDb,ð4Kþ1ÞDb�

8<: ð25Þ

where K is a positive integer and re fluctuates around 1/2and converges to 1/2, as the value of M goes to infinite.The expression of re shown in Eq. (25) is supportedthrough analytical form and experimental results arereported in Table 3. The result is obtained when the valueof K in Eq. (25) is set to one (1). The low value of re

denotes better decoding of watermark bits. This suggeststhat one may get better image quality through thereversible process.

5. Simulation results

This section briefly describes performance of theproposed access control scheme. Simulation is done overlarge number of benchmark images [42,43] includingeight popular test images: Lena, Pepper, Baboon, Boat,Opera, Cameraman, Pueblo bonito and F16 and are shownin Fig. 6. All test images are of size (512�512), 8-bit/pixelgray scale images. The length of the dither is 88, as DWTdecomposition is performed up to level 4. The presentstudy uses peak-signal-to-noise-ratio (PSNR) [44], mean-structural SIMilarity-index-measure (MSSIM) [45] andWatson distance [24] as a distortion measure for thewatermarked image under inspection with respect tothe original image. In [24], a perceptual model basedDCT domain QIM watermarking scheme robust againstvalumetric scaling is proposed. Watson distance isdefined as, given the original content c0, its transformdomain coefficients are denoted by C0 and Cw denotes thewatermarked transform coefficients. The Watson dis-tance, according to the formula given in [24], betweenthe original and watermarked image is developed here for


our lifting based method as follows:

DwatðC0,CwÞ ¼X

x,y,r,s

Csw,r ½x,y��Cs

o,r ½x,y�

qsrðx,yÞ

� �4( )1=4

ð26Þ

where the term in the denominator is obtained fromEq. (7), and r,s indicate the level and orientation, (x,y)indicate position information at coefficient domain. Therelatively high PSNR and MSSIM values and low Watsondistance shown in Table 4 for different test images high-light the imperceptibility of the proposed data-hidingscheme. As the step size (delta) is calculated based onthe relation between the wavelet coefficients and HVS,low values of WD are achieved.

5.1. Visibility measure

Table 4 illustrates PSNR and MSSIM values before andafter decoding process for eight test images along withnormalized cross correlation (NCC) used to quantifyrobustness measure. The NCC value between the originalwatermark image (W) and the decoded watermark ðWÞ iscalculated using Eq. (26)

NCC¼

Pi

PjWijWijP

i

PjðWijÞ

2ð27Þ

A normal user without the knowledge of the key canview lower quality images with PSNR values as shown incolumn 2 of Table 4 when only watermarking is used asan access control tool. Numerical values shown in column3rd–6th represent the lower quality images when 25%,50%, 75% and 100% of coefficients in HH subbands aremodulated along with watermarking. It is notable tohighlight that modulation from 25% to 50% coefficientscause much degradation in visual quality of �2.85 dB,while the similar for the 50–75%, 75–100% coefficientscause relatively small change in visual quality of the�2.0–1.0 dB. However, the user with a valid key candecode a better quality image with PSNR shown inTable 4. Relatively low values of Watson distance (WD)for different decoded test images indicate the novelty ofthe proposed HVS characteristics based adaptive QIMhiding scheme. These low values of Watson distance areachieved due to the proper step size obtained in Eq. (7)incorporating the noise sensitivity, modified local bright-ness component, and texture measures in different levels

Table 4PSNR(dB) and MSSIM before and after decoding process. The value of step size

(payload) is 1024 bits. P and M in the heading of the table resents the PSNR (i

Name of image Before dec. 25% 50%

P M P M P M

Lena 36.00 0.93 32.17 0.89 29.43 0.86

Pepper 36.09 0.92 32.02 0.88 29.08 0.85

Babbon 36.54 0.95 32.12 0.91 29.14 0.89

Boat 36.04 0.95 32.14 0.91 29.25 0.88

Opera 36.06 0.94 32.08 0.90 29.64 0.87

Cameraman 36.36 0.93 32.61 0.89 29.74 0.86

Pueblo bonito 36.23 0.94 32.25 0.90 29.32 0.87

F16 36.02 0.93 32.31 0.89 29.41 0.86

and orientations. Fig. 7(a) shows one of the watermarkedimage Lena after embedding of a watermark of size(32�32). Figs. 7(b)–(e) show the watermarked imageswhen 25%, 50%, 75% and 100% of coefficients in HHsubbands are modulated along with watermarking.Fig. 7(f) is the decoded images from all encoded imagesi.e. in Figs. 7(a)–(e). Similar results for three other testimages Pepper, Opera and Boat are shown in (g, i, k) after100% modulation of coefficients. Respective decodedimages are shown in (h, j, l). To highlight subjectivequality improvement supporting the respective objectivemeasures, edge information of decoded images are high-lighted along with showing the same for the modulatedimages. It is seen form P1 and P2 of (g, i, k) (modulatedimages) and (h, j, l) (decoded images) that the visualinformation (edges) for the decoded images are signifi-cantly improved and are enjoyed by the authorized users.The rectangular regions are shown here in order to high-light how the corresponding visual qualities are differentin two types of images accessed by common and author-ized users. Fig. 8 shows the extracted watermark from alltest images having NCC values one (1). Figs. 9 and 10show the improvement in image quality in terms of PSNRand MSSIM if watermark of different lengths areembedded in normal DWT and lifting based DWT, respec-tively. It is evident from the graphs in Figs. 9 and 10 thatimage quality after watermark embedding for liftingbased DWT is better than the normal DWT.

5.2. Robustness test

To test robustness performance for the proposed accesscontrol scheme, some typical image processing operationsare performed. The term robustness implies here theability of quality access control through joint self-noise(SN) cancellation and demodulation of HH coefficientsafter various image processing operations. Robustnessagainst such operations is essential as it is expected thatan un-authorized user may distort the watermarkedimages with an objective that the users having validcommercial agreement would also not be able to enjoybetter quality of the images. The average experimentalresults are shown in Table 5. During robustness check,we consider 50% of the coefficients in all HH band aremodulated along with watermarking. It is clear fromnumerical values shown in Table 5 that the quality of the

is 9.32 (i.e. watermark power (WP) [46] is 8.59 dB). Length of watermark

n dB) and MSSIM values of the image. WD: Watson distance.

75% 100% NCC After decoding

P M P M P M WD

27.95 0.82 26.90 0.79 1 39.26 0.96 12.15

27.51 0.81 26.56 0.78 1 39.23 0.96 13.82

27.32 0.84 26.24 0.81 1 39.14 0.93 12.05

27.42 0.84 26.23 0.81 1 39.26 0.97 12.62

27.31 0.83 26.51 0.80 1 39.12 0.97 12.32

27.32 0.82 26.81 0.79 1 39.85 0.96 12.32

27.51 0.83 26.32 0.80 1 39.19 0.98 12.15

27.32 0.82 26.41 0.79 1 39.24 0.97 12.28

Fig. 8. Extracted watermark image from all watermarked images

(Fig. 7(a–e)).

5 10 15 20 25 30 35 4020

25

30

35

40

45

50

55

60

Watermark Size(nxn)

PS

NR

(dB

)

Before Deco.(Lifting)After Deco.(Lifting)Before Deco.(Normal DWT)After Deco.(Normal DWT)

Fig. 9. Shows the improvement of image quality in term of PSNR (dB) for

different watermark sizes (when 50% coefficients are modulated).

(P=26.51, M=0.80, WD=66.12) (P=26.23, M=0.81, WD=66.92)

(P=36.00, M=0.93, WD=12.81) (P=27.95, M=0.82, WD=54.31)(P=29.43, M=0.86, WD=45.12)(P=32.17, M=0.89, WD=35.11)

(P=39.23, M=0.96, WD=13.82) (P=26.56, M=0.78, WD=67.15)(P=39.26, M=0.96, WD=12.15)(P=26.90, M=0.79, WD=65.32)

(P=39.26, M=0.97, WD=12.62)(P=39.12, M=0.97, WD=12.32)

Fig. 7. (a) Watermarked Lena image, (b–e, g, i, k) different watermarked images for different percent coefficients modulation, (f) decoded image Lena

from all watermarked images (a–e). (h, j, l) show the respective decoded images of (g, i, k). P, M and WD above in each image represents the PSNR (in dB),

MSSIM and Watson distance values of the image. Rectangular boxes indicate quality improvement as edges are got back in decoded images and are

enjoyed by the authorized users through access control.


decoded image is improved even after various commonimage-processing operations done on the encoded image.Fig. 11 shows various distorted images along with thedecoded good quality images from the distorted images.Figs. 12 and 13 show the improvement of quality in termsof PSNR and MSSIM, respectively, with the variation ofquality factors for JPEG compression operation. It is seenfrom Figs. 12 and 13 that even when the JPEG quality factorgoes down to 50, the scheme still can decode the imageeffectively. This leads to the superiority of the proposed


access control method. Fig. 14 shows comparative resultsof robustness for DWT and lifting based implementation.The NCC values for the extracted watermark after JPEGcompression with different quality factors show that liftingbased implementation is much effective for access controlthan DWT based implementation for the proposedalgorithm.

The bit error rate (Pe) of the proposed access controlscheme is also computed and compared with other workproposed in [24] for a fixed watermark energy and pay-load. The result is shown in Table 6 for setting watermarkpower at 8.59 dB and payload size at 1024 for both themethods. It is already mentioned that method [24] isbased on DCT based perceptual model, while the presentone is lifting based method incorporating HVS

5 10 15 20 25 30 35 400.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

Watermark Size(nxn)

MSS

IM

Before Deco.(Lifting)After Deco.(Lifting)Before Deco.(Normal DWT)After Deco.(Normal DWT)

Fig. 10. Shows the improvement of image quality in term of MSSIM for

different watermark sizes (when 50% coefficients are modulated).

Table 5Performance results of the proposed access control method after different ima

Stepsize

PSNR beforedeco.

MSSIM beforedeco.

PSNR afterdeco.

MSSIMAfterdeco.

Stepsize

Median filtering (3�3)

6.21 29.80 0.86 31.73 0.87 6.21

9.32 29.63 0.85 31.17 0.83 9.32

18.64 29.04 0.80 29.44 0.72 12.43

Mean filtering (3�3)

6.21 28.99 0.87 30.15 0.86 6.21

9.32 28.91 0.86 29.92 0.83 9.32

18.64 28.71 0.84 28.92 0.74 18.64

Highpass filtering

6.21 20.11 0.82 24.55 0.88 6.21

9.32 20.27 0.76 23.81 0.85 9.32

18.64 20.74 0.58 22.70 0.72 18.64

Down and up sampling (0.75)

6.21 30.14 0.89 33.2 0.90 6.21

9.32 30.02 0.88 32.72 0.87 9.32

18.64 18.11 0.44 18.44 0.54 18.64

Histogram equalization

6.21 18.60 0.68 19.06 0.76 6.21

9.32 18.55 0.61 19.04 0.70 9.32

18.64 29.60 0.84 30.93 0.77 18.64

characteristics. The third and fourth row of Table 6 indi-cate the respective numerical values for [24] with dithermodulation based on Watson (DM-W) and modifiedWatson model (DM-MW), respectively. Authors in [24]modified adaptive DM algorithm in order to make itrobust to volumetric scaling by linearly scaling the slacksi.e. estimated step size is multiplied with the factor bywhich the amplitude of the signal is scaled. It is seen thatthe bit error rate ðPeÞ of the proposed scheme is compar-able to [24] which, in one way, claims better robustness ofthe former compared to the latter. In other words,although numerical values shown in comparison do notreflect much improvement compared to [24], still theabove claim is justified as the proposed work applies bothwatermarking and data modulation unlike watermarkingmethod applied only in the latter. This in other wayindicates that the depth of signal degradation to thewatermarked image is much higher in this work comparedto the other work [24]. The same range of numericalvalues (for fixed watermark energy and payload) in termof low Watson distance support the novelty of the pro-posed data hiding scheme through adaptive step size thattakes into consideration the relationship among the liftingbased wavelet coefficients, HVS and quantization step-size(Db) of JPEG 2000. Moreover, the proposed scheme servesthe purpose of quality access control for which it isdesigned, along with the normal application of copyrightprotection done by watermarking method in [24].

Table 7 illustrates PSNR and MSSIM values before andafter image decoding (without applying any attack) if thefinal step size ðDbÞ of QIM is determined based on onlyqb (case 1), only Db (case 2) and both qb and Db (case 3)achieved. Shown in Figs. 15 and 16 are the family of curvesfor gain in quality after image decoding as a function of JPEGquality factor. From Figs. 15 and 16, one can see that with

ge processing operations.

PSNR beforedeco.

MSSIM beforedeco.

PSNR afterdeco.

MSSIM afterdeco.

Salt and pepper noise (0.005)

25.89 0.75 27.87 0.83

25.59 0.70 27.65 0.80

24.29 0.53 26.89 0.71

Speckle noise (0.005)

26.10 0.59 28.14 0.64

25.79 0.57 27.81 0.63

24.43 0.47 26.92 0.58

Gaussian noise (0.005)

19.60 0.24 19.98 0.29

19.57 0.24 19.96 0.26

19.21 0.23 19.69 0.25

Dynamic range change [50–200]

21.78 0.84 22.17 0.89

21.79 0.82 22.20 0.88

21.55 0.69 22.10 0.83

Lossy JPEG 70

29.39 0.86 34.76 0.90

28.70 0.79 33.15 0.85

25.78 0.55 29.92 0.71

(P=30.15, M=0.86, WD=43.02)(P=28.99, M= 0.87, WD=53.12 )(P=31.73, M= 0.87, WD=37.11)(P=29.80, M=0.86, WD=45.01)

(P=20.11, M= 0.82, WD=85.11) (P=24.55, M=0.88, WD=75.12) (P= 30.14, M=0.89, WD=43.07) (P=33.2, M=0.90, WD=34.14)

(P=27.87, M=0.83, WD=54.86)(P=25.89, M=0.75, WD=68.35)(P=19.06, M= 0.76, WD=86.32)(P=18.60, M=0.68, WD=88.17 )

(P=19.98, M=0.29, WD=85.29)(P= 19.60, M=0.24, WD=86.05)(P=28.14, M=0.64, WD=53.78)(P=26.10 , M=0.59, WD=65.82)

(P=22.17, M=0.89, WD=80.12)(P=21.78, M=0.84, WD=83.12)(P= 34.76, M=0.90, WD=28.17)(P=29.39, M= 0.86, WD=45.36)

Fig. 11. (a) Median filtered image using spatial mask (3�3); (b) decoded image of (a), (c) mean filtered image using spatial mask (3�3), (d) decoded image of

(c); (e) highpass filtered image; (f) decoded image of (e); (g) down and up sampling (0.75); (h) decoded image of (g); (i) histogram equalization; (j) decoded

image of (i); (k) salt and pepper noise (variance 0.005); (l) decoded image of (k); (m) speckle noise (variance 0.005); (n) decoded image of (m); (o) Gaussian

noise (variance 0.005); (p) decoded image of (o); (q) lossy JPEG at quality factor70; (r) decoded image of (q); (s) dynamic range change to 50–200; (t) decoded

image of (s); P, M and WD above in each image represents the PSNR (in dB), MSSIM and Watson distance values of the respective image.


Fig. 12. Improvement of quality in PSNR (dB) due to access control for

lossy JPEG compression operation (for 50% coefficient modulation).

Fig. 13. Improvement of quality in MSSIM due to access control for lossy

JPEG compression operation (for 50% coefficient modulation).

Fig. 14. Results of robustness for lossy JPEG compression operation

(for 50% coefficient modulation).


the increase in JPEG quality factor, the gain increasesmonotonically. An important trend is that the variousparameters (i.e. qb and Db) of step size ðDbÞ influencesignificantly on the gain. In particular, case 3 gives bettergain over case 1 and case 2. This result shows that case 3(adaptive QIM) is preferable in designing adaptive datahiding based quality access control.

To support the validity of Eqs. (23a) and (23b), thewatermark is extracted from the decoded watermarkedimage (Fig. 7(f)). Fig. 17(a) shows the extracted watermarkfrom the decoded watermarked image Fig. 7(f) withoutapplying any common or deliberate image processingoperations. Fig. 17(b) shows the comparative robustnessperformance of the extracted watermark from the water-marked images before and after decoding for lossy JPEGcompression. Here the word ‘‘decoded image’’, means theimage accessed by the authorized users. Table 8 lists thecomparative robustness performance for other image-pro-cessing operations. It is seen from the results in Table 8 andFig. 17(b) that there is significant improvement in robust-ness performance when watermark is extracted from thedecoded watermarked image Fig. 7(f). The watermarkextracted from the decoded watermarked image Fig. 7(f)may be used as ‘‘copyright mark’’ or ‘‘tracking information’’to track illicit distribution of decoded watermarked image(image obtained after self-noise cancellation) by the author-ized user. This is the other benefits offered by the proposedQIM based access control scheme over spread spectrum (SS)modulation based watermarking method [40].

We examine the time requirement to run the wholeprocedure of encoding and decoding for quality accesscontrol of image as a measure of computational load.Table 9 illustrates the encoding and decoding time inseconds. This result evidently shows that proposed algo-rithm is much faster than the existing methods. Theexperimentations are conducted in Pentium IV, 2.80 GHzprocessor, with 512 MB RAM using MATLAB 7 version.

Performance of the proposed scheme is also comparedwith the other access control methods reported in[2,3,17,47] in terms of PSNR, and bit error rate (BER). Thisaccess control performance comparison is not onlyrestricted with data hiding based method [17] but alsowith other scrambling and encryption based methods[2,3,47]. The word BER here implies the error in thereceived signal when transmitted through the noisychannel or is manipulated by some malicious users. BERperformance is computed in this experimentation byrandomly manipulating the watermarked image data.This affects PSNR values of the received image withrespect to the original version. The results shown inTable 10 highlight the superiority in performance for theproposed method compared to both type of access controlmethods. From Table 10, it is clear that methodsin [2,3,47] offer poor performance, compared to thepresent one, to the authorized user. This little poorperformance is due to the fact that the access control in[2,3,47] use cryptographic technique. An error in a ciphertext may lead to a partial failure in decryption andultimately results in poor image quality after decoding.Another advantage offered by the proposed method is thepossibility of partial accessing of the content by the

Table 6Results of bit error rate (Pe) when step size ðDÞ is 9.32 (i.e. watermark power (WP) [46] is 8.59 dB) and number of watermark bits (payload) is 1024.

A: JPEG 85; B: median filtering (3�3); C: mean filtering (3�3); D: highpass filtering; E: down and up sampling (0.9); F: histogram equalization;

G: dynamic range change; H: salt and pepper noise; I: speckle noise; J: Gaussian noise. DM-W: dither modulation with Watson model. DM-MW: dither

modulation with modified Watson model. WD: Watson distance.

Embeddingdomain

WD (for watermarked image) A B C D E F G H I J

Proposed Lifting 12.81 0 0.15 0.39 0 0.01 0.33 0 0 0.04 0.35

DM-W[24] DCT 11.75 0.09 0.18 0.45 0.002 0.15 0.35 0.01 0.01 0.05 0.38

DM-

MW[24]

DCT 12.95 0.02 0.19 0.42 0.001 0.01 0.36 0.01 0.02 0.09 0.40

Table 7

Performance if step size ðDbÞ is determined using only qb , only Db and

using both i.e. qb and Db . Watermarking along with 50% modulation of

HH subbands are considered.

Parameter (for step size) Before PSNR(dB)

After PSNR(dB)

Only qb (case 1) 23.32 40.82

Only Db (case 2) 19.50 37.40

Db and qb (proposed method)

(case 3)

21.84 39.98

50 55 60 65 70 75 80 85 90 95 1001

2

3

4

5

6

7

8

9

10

JPEG Quality Factor

Gai

n In

PS

NR

(dB

)

Case 1Case 2Case 3

Fig. 15. Gain in PSNR (dB) when step size is calculated by: case 1: using

only qb; case 2: using only Db; case 3: using both Db and qb.

50 55 60 65 70 75 80 85 90 95 100

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

JPEG Quality Factor

Gai

n in

MS

SIM

Case 1Case 2Case 3

Fig. 16. Gain in MSSIM when step size is calculated by: case 1: using

only qb; case 2: using only Db; case 3:using both Db and qb.

Fig. 17. (a) Extracted watermark image from all watermarked images

after self-noise (SN) elimination (i.e. from Fig. 7(f)), (b) comparative

robustness performance of the extracted watermarks before SN cancel-

lation, Fig. 7(a–e) and after SN cancellation, Fig. 7(f) of decoded image

for various JPEG index.


unauthorized/common users, while the methods in[2,3,47] offer full denial of the content. While comparingwith other watermarking based access control method[17], improvement in the proposed method is also seendue to the combined use of data hiding and data modula-tion. Robust extraction of watermark information leads tocorrect data demodulation and in turn helps to availbetter quality image to the authorized users. Numericalresults in table indicate watermark energy, watermarkpayload, different intensity of signal modulation and therespective quality improvement.

6. Conclusions and scope of future works

In this paper, a lifting based DWT domain joint datahiding and data modulation scheme for quality access

control of image using adaptive dither QIM is proposed.Simulation results show that the determination of stepsize using the relationship among the wavelet coeffi-cients, HVS and quantization step size (Db) of JPEG 2000

Table 8Comparative robustness performance (NCC) of the extracted watermark from before and after decoded watermarked image; A: median filtering (3�3),

B: mean filtering (3�3), C: highpass filtering, D: down and up sampling (0.9), E: down and up sampling (0.75), F: histogram equalization, G: dynamic

range change, H: salt and pepper noise (0.001), I: salt and pepper noise (0.005), J: salt and pepper noise (0.009), K: speckle noise (0.001), L: speckle noise

(0.005), M: speckle noise (0.009), N: Gaussian noise (0.001), O: Gaussian noise (0.005), P: Gaussian noise (0.009).

A B C D E F G H I J K L M N O P

Before decoded image 0.69 0.61 1 0.81 0.74 0.67 1 1 1 1 1 0.86 0.74 0.65 0.60 0.57

After decoded image 0.98 1 1 0.99 0.97 0.68 1 1 1 1 1 0.88 0.73 0.59 0.59 0.59

Table 9Encoding and decoding time in seconds for access control.

Domain Encodingprocess (s)

Decodingprocess (s)

Totaltime (s)

DCT [9] 8.18 11.59 19.77

DWT [3] 4.68 8.43 13.11

Lifting based DWT [proposed] 3.96 7.87 11.83

Table 10Comparison of results: PSNR (dB) and BER. WP: watermark power [46].

Operatingdomain

Techniqueused

Watermarkenergies inWP (dB)

Length ofwatermark(payload in bits)

BER: 10�4

PSNR (dB)BER: 10�5

PSNR (dB)BER: 10�6

PSNR (dB)

Grosbois et al. [2] DWT Scrambling N/A N/A 28.75 30.10 31.11

Imaizumi et al. [3] DWT Encryption N/A N/A 29.11 31.11 31.19

Zaidee et al. [47] DCT Scrambling N/A N/A 28.84 29.25 29.89

Coria et al. [17] WT-CWT Data hiding 13.52 1024 28.12 29.32 29.78

Proposed Lifting Data hiding

and data

modulation

8.59 1024 31.21 32.52 32.73


leads to greater improvement in image quality throughreversible process even after signal processing operationslike lossy JPEG. The simulation results also show thatour scheme is robust against various image processingoperations such as filtering, lossy JPEG compression,noise addition and histogram equalization, dynamic rangechange, down and up sampling and leads to betterquality access control of the image by the authorizedusers. The proposed scheme highlights the benefits ofQIM over SS based access control scheme. Simulationresults show that the use of lifting in QIM offers betterrobustness performance against common signal proces-sing operations compared to the DWT based implementa-tion. Future work may be done to extend the proposedaccess control method for video and audio signal, aswell as development of hardware implementation of theproposed access control scheme through applicationspecific integrated circuit (ASIC) or field programmablegate array (FPGA).

Acknowledgment

Authors gratefully acknowledge the valuable and criticalcomments and suggestions made by Dr. Frank Hartung,Associate Editor and all the anonymous reviewers for tech-nical quality improvement of the article.

References

[1] F. Hartung, F. Ramme, Digital rights management and watermark-

ing of multimedia content for m-commerce applications, IEEE

Communications Magazine 38 (2000) 78–84.

[2] R. Grosbois, P. Gerbelot, T. Ebrahimi, Authentication and access

control in the JPEG 2000 compressed domain, in: Proceedings of

46th SPIE Annual Meeting, Applications of Digital Image Processing,

San Diego, 2001, pp. 95–104.

[3] S. Imaizumi, O. Watanabe, M. Fujiyoshi, H. Kiya, Generalized

hierarchical encryption of JPEG 2000 code streams for access

control, in: Proceedings of IEEE International Conference on Image

Processing, Genoa, Italy, 2005, pp. 1094–1097.

[4] J.G. Wen, M. Severa, W. Zeng, M.H. Luttrell, W. Jin, A format-

compliant configurable encryption framework for access control of

video, IEEE Transactions on Circuits and Systems for Video Tech-

nology 12 (2002) 545–557.

[5] C. Wang, H.B. Yu, M. Zheng, A DCT-based MPEG-2 transparent

scrambling algorithm, IEEE Transaction on Consumer Electronics 49

(2003) 1208–1213.

[6] M.S. Kankahalli, T. Guan, Compressed domain scrambler/descram-

bler for digital video, IEEE Transactions on Consumer Electronics 48

(2002) 356–365.

[7] E. Bertino, J. Fan, E. Ferrari, M.S. Hacid, A.K. Elmagarmid, X. Zhu, A

hierarchical access control model for video database systems, ACM

Transactions on Information System 21 (2003) 155–191.

[8] A. Phadikar, M.K. Kundu, S.P. Maity, Quality access control of a

compressed gray scale image, in: Proceedings of National Confer-

ence on Computer Vision, Pattern Recognition, Image Processing

and Graphics, Gandhinagar, India, 2008, pp. 13–19.

[9] A. Phadikar, S.P. Maity, Quality access control of compressed color

images, AEU—International Journal of Electronics and Communica-

tions, Elsevier 64 (2010) 833–843.


[10] M. Kutter, F. Hartung, Introduction to watermarking techniques, in:S. Katzenbeisser, F. Petitcolas (Eds.), Information Hiding Techniquesfor Steganography and Digital Watermarking, Artech House, 2000.

[11] P. Campisi, M. Carli, G. Giunta, A. Neri, Blind quality assessmentsystem for multimedia communications using tracking watermark-ing, IEEE Transactions on Signal Processing 51 (2003) 996–1002.

[12] S.P. Maity, M.K. Kundu, S. Maity, Dual purpose FWT domain spreadspectrum image watermarking in real-time, International Journalof Computer and Electrical Engineering in the Special Issue on Real-Time Security and Copyright Protection of Multimedia, ElsevierScience 35 (2009) 415–433.

[13] P.S. Huang, C.S. Chiang, C.P. Chang, T.M. Tu, Robust spatial water-marking technique for colour images via direct saturation adjust-ment, IEE Proceedings on Vision, Image and Signal Processing 152(2005) 561–574.

[14] C.T. Li, Reversible watermarking scheme with image-independentembedding capacity, IEE Proceedings on Vision, Image and SignalProcessing 152 (2005) 779–786.

[15] M. Wu, H.H. Yu, Video access control via multi-level data hiding, inProceedings of IEEE International Conference on Multimedia andExpo, 2000, pp. 381–384.

[16] M. Pickering, L.E. Coria, P. Nasiopoulos, A novel blind video water-marking scheme for access control using complex wavelets, in:Proceedings of IEEE International Conference on Consumer Electro-nics, Las Vegas, 2007, pp. 1–2.

[17] L. Coria, P. Nasiopoulos, R. Ward, M. Pickering, An access controlvideo watermarking method that is robust to geometric distortions,Journal of Information Assurance and Security 2 (2007) 266–274.

[18] F.C. Chang, H.C. Huang, H.M. Hang, Layered access control schemeson watermarked scalable media, Journal of VLSI Signal Processing(Springer Netherlands) 49 (2007) 443–455.

[19] X. Wu, J. Hu, Z. Gu, J. Huang, A secure semi-fragile watermarking forimage authentication based on integer wavelet transform withparameters, in: Proceedings of Australian Workshop on GridComputing and E-Research, New South Wales, 2005, pp. 75–80.

[20] P. Meerwald, A. Uhl, A survey of wavelet-domain watermarkingalgorithms, in: Proceedings of SPIE, Electronic Imaging, Security andWatermarking of Multimedia Contents, San Jose, USA, pp. 505–516.

[21] Y.H. Seo, D.W. Kim, A new VLSI architecture of lifting-based DWT,Lecture Notes in Computer Science 3985 (2006) 46–151.

[22] A. Phadikar, S.P. Maity, M.K. Kundu, Quantization based data hidingscheme for efficient quality access control of images using DWT vialifting, in: Proceedings of Sixth IEEE Indian Conference on Compu-ter Vision, Graphics and Image Processing (ICVGIP’08), Bhubanes-war, India, pp. 265–275.

[23] M. Costa, Writing on dirty paper, IEEE Transactions on InformationTheory 29 (1983) 439–441.

[24] Q. Li, I.J. Cox, Using perceptual models to improve fidelity andprovide resistance to volumetric scaling for quantization indexmodulation watermarking, IEEE Transactions on Information For-ensics and Security 2 (2007) 127–139.

[25] I. Cox, M. Miller, B. Jeffrey, Digital Watermarking: Principles andPractice, Morgan Kaufmann, 2002.

[26] B. Chen, G.W. Wornell, Quantization index modulation: a class ofprovably good methods for digital watermarking and informationembedding, IEEE Transactions on Information Theory 47 (2001)1423–1443.

[27] S.P. Maity, C. Delpha, S. Braci, R. Boyer, Hidden QIM watermarkingon compressed data using channel coding and lifting, in: Proceed-ings of the Third International Conference on Pattern Recognitionand Machine Intelligence, LNCS, India, 2009, pp. 414–419.

[28] W. Sweldens, The lifting scheme: a custom-design construction ofbiorthogonal wavelets, Applied and Computational Harmonic Ana-lysis 3 (1996) 186–200.

[29] W. Sweldens, The lifting scheme: construction of second generationwavelets, SIAM Journal on Mathematical Analysis 29 (1997) 511–546.

[30] J. Kovacevic, W. Sweldens, Wavelet families of increasing order inarbitrary dimensions, IEEE Transaction on Image Processing 9(2000) 480–496.

[31] M. Boliek, C. Christopoulos, E. Majani, JPEG 2000 Part I Final DraftInternational Standard, ISO/IEC FDIS15444-1, ISO/IEC JTC1/SC29/WG1 N1855, 2000.

[32] M.D. Adams, F. Kossentini, Reversible integer-to-integer wavelet trans-forms for image compression: performance evaluation and analysis,IEEE Transactions on Image Processing 9 (2000) 1010–1024.

[33] G. Uytterhoeven, D. Roose, A. Bultheel, Wavelet Transforms UsingLifting Scheme, Technical Report ITA-Wavelets Report WP 1.1,Department of Computer Science, Katholieke Universiteit Leuven,Belgium, 1997.

[34] S. Voloshynovskiy, T. Pun, Capacity security analysis of data hidingtechnologies, in: Proceedings of IEEE International Conference onMultimedia and Expo, Lausanne, Switzerland, 2002, pp. 477–480.

[35] A.S. Lewis, G. Knowles, Image compression using the 2-D wavelettransform, IEEE Transactions on Image Processing 1 (1992)244–250.

[36] M. Barni, F. Bartolini, A. Piva, Improved wavelet-based watermark-ing through pixel-wise masking, IEEE Transactions on ImageProcessing 10 (2001) 783–791.

[37] M.K. Kundu, S.K. Pal, Thresholding for edge detection using humanpyschovisual phenomena, Pattern Recgnition Letters 4 (1986) 433–441.

[38] D. Taskovski, S. Bogdanova, M. Bogdanov, Blind low frequencywatermarking method, International Journal of Information andCommunication Engineering 2 (2006) 146–150.

[39] C. Christopoulos, A. Skodras, T. Ebrahimi, The JPEG 2000 still imagecoding system: an overview, IEEE Transactions on ConsumerElectronics 46 (2000) 1103–1127.

[40] A. Phadikar, S.P. Maity, A cost effective scheme for contentverification and access control of quality of an image, in: Proceed-ings of the IEEE Colloquium and the Third IEEE InternationalConference on Industrial and Information Systems, IEEE Explore,2008, pp. 1–6.

[41] M. Wu, Multimedia Data Hiding, Ph.D. Thesis, Princeton University,2000.

[42] Image Database Online Sources. Available from: http://www.cl.cam.ac.uk/fapp2/watermarking

[43] Image Database Online Sources. Available from: FabienPetitcolashttp://www.petitcolas.net/fabien/watermarking/image_database/index.html

[44] R.C. Gonzalez, R.E. Woods, S.L. Eddins, Digital Image ProcessingUsing Matlab, Pearson Education, 2005.

[45] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image qualityassessment: from error measurement to structural similarity, IEEETransactions on Image Processing 13 (2004) 1–14.

[46] J.P. Boyer, P. Duhamel, J.B. Talon, Performance analysis of scalarDC-QIM for watermark detection, in: Proceedings of IEEE Interna-tional Conference on Acoustics, Speech and Signal Processing,Honolulu, Toulouse, 2006, pp. 283–289.

[47] A.A. Zaidee, S. Fazdliana, B.J. Adznan, Content access control forJPEG images using CRND zigzag scanning and QBP, in: Proceedingsof the Sixth IEEE International Conference on Computer andInformation Technology, 2006, pp. 146–146.

http://www.cl.cam.ac.uk/fapp2/watermarking

http://www.cl.cam.ac.uk/fapp2/watermarking

http://www.petitcolas.net/fabien/watermarking/image_database/index.html

http://www.petitcolas.net/fabien/watermarking/image_database/index.html

Data hiding based quality access control of digital images using adaptive QIM and lifting

Documents