Robust Audio Watermarking Using Multiwavelet Transform and Genetic Algorithm

Transcription

1 Robust Audio Watermarkig Usig Multiwavelet Trasform ad Geetic Algorithm PRAYOTH KUMSAWAT 1, KITTI ATTAKITMONGCOL 2 AND ARTHIT SRIKAEW 2 1 School of Telecommuicatio Egieerig, 2 School of Electrical Egieerig Istitute of Egieerig, Suraaree Uiversity of Techology 111 Uiversity Aveue, Muag District, Nakho Ratchasima, 30000, Thailad {prayoth, kitti, ra}@sut.ac.th Abstract:- I this paper, we propose a ew approach for optimizatio i digital audio watermarkig usig artificial itelliget techique. The watermarks are embedded ito the low frequecy coefficiets i discrete multiwavelet trasform domai. The embeddig techique is based o quatizatio process which does ot require the origial audio sigal i the watermark extractio. We have developed a optimizatio techique usig the geetic algorithm to search for optimal quatizatio step i order to improve both quality of watermarked audio ad robustess of the watermark. I additio, we aalyze the performace of the proposed algorithm i terms of sigal-to-oise ratio, ormalized correlatio ad bit error rate. The experimetal results show that our proposed method ca improve the quality of the watermarked audio sigal ad give more robustess of the watermark as compared to previous works. Key-Words: - Audio watermarkig, Multiwavelet, Geetic algorithm, Artificial itelligece, Optimizatio 1 Itroductio Digital watermarkig is oe of the most popular approaches cosidered as a tool for providig the copyright protectio of digital cotets. This techique is based o direct embeddig of additioal iformatio data ito the digital cotets. Ideally, there must be o perceptible differece betwee the watermarked ad origial digital cotets, ad the watermark should be easily extractable, reliable ad robust agaist data compressio or ay sigal maipulatios [1]. The mai requiremets of digital watermarkig are ivisibility, robustess ad data capacity. These requiremets are mutually coflictig, ad thus, i the desig of a watermarkig system, the trade off has to be made. Accordig to the Iteratioal Federatio of the Phoographic Idustry (IFPI) [2], audio watermarkig should have the followig specificatios: 1) Audio watermarkig should ot degrade perceptio of origial sigal. 2) Sigal to oise ratio ( SNR ) should be greater tha 20 db ad there should be more tha 20 bits per-secod (bps) data payload for watermark. 3) Watermark should be able to resist most commo audio processig operatios ad attacks. 4) Watermark should be able to prevet uauthorized detectio, removal ad embeddig, uless the quality of audio becomes very poor. I geeral, digital audio watermarkig ca be performed i time domai ad trasform domai, where the properties of the uderlyig domai ca be exploited. Curretly, watermarkig techiques based o trasform domai are more popular tha those based o time domai sice they provide higher audio quality ad much more robust watermark. Seok ad Hog [3] itroduced direct sequece spread spectrum audio watermarkig based o the discrete Fourier trasform (DFT). The stregth of the embedded watermark sigal depeds o the huma perceptual characteristics of the audio sigal. The detectio procedure does ot require access to the origial audio sigal to detect the watermark. For may years, multi-scale represetatios ad multiresolutio aalysis have proved useful i may sigal processig applicatios. Wavelet aalysis is a example to geerate such represetatio [4, 5, 6]. I [6], Wag et al. proposed a digital audio watermarkig algorithm based o the discrete wavelet trasform (DWT). The watermark iformatio is embedded i audio low-middle frequecy coefficiets i wavelet trasform domai. A scheme of watermark detectio is preseted by usig liear predictive codig, ad it does ot use the origial sigal durig watermark extractig process. I [7], Che ad Worell proposed a class of embeddig ISSN: Issue 6, Volume 9, Jue 2010

2 methods called quatizatio idex modulatio (QIM) that achieves probably good rate-distortiorobustess performace. Wu et al. [8] proposed a self-sychroizatio algorithm for audio watermarkig usig QIM techique. They embed the sychroizatio codes with hidde iformative data so that the hidde data has selfsychroizatio ability. Sychroizatio codes ad iformative bits are embedded ito low-frequecy subbad i DWT domai. Their simulatios suggest that the quatizatio step S (embeddig stregth) greatly depeds o types ad magitudes of the origial audio sigals. It is ot the best choice to use a fixed S. I [9], Kim ad Bae proposed a robust algorithm to estimate the modified quatizatio step size with a optimal search iterval. The equatio to determie the optimal search iterval is derived aalytically, which ca satisfy both detectio performace ad computatioal complexity. The authors coclude that the derived optimal search iterval provides the accurate estimatio of the modified quatizatio step size uder amplitude modificatio attack. I recet years, some multiwavelet-based digital watermarkig algorithms have bee proposed. Kwo ad Tewfik [10] proposed a adaptive image watermarkig scheme i the discrete multiwavelet trasform (DMT) domai usig successive subbad quatizatio ad a perceptual modelig. The watermark is Gaussia radom sequece with uit variace ad the origial image is eeded for watermark detectio. Kumsawat et al. [11] proposed a image watermarkig algorithm usig the DMT ad geetic algorithm is applied to search for optimal watermarkig parameters to improve the quality of the watermarked image ad the robustess of the watermark. Ghouti ad Bouridae [12] proposed a ovel audio figerpritig framework for robust perceptual hashig of audio cotet usig balaced multiwavelets. The extracted hash values are used for idetifyig, searchig, ad retrievig audio cotet from large audio databases. I [13], Kumsawat et al. proposed a multiwavelet-based audio watermarkig scheme by utilizig the audio statistics characteristics ad QIM techique. The watermarks are embedded ito the low frequecy coefficiets i discrete multiwavelet trasform domai to achieve robust performace agaist commo sigal processig procedures ad oise corruptios. Improvemets i performace of digital audio watermarkig schemes ca be obtaied by exploitig the characteristics of the huma auditory system (HAS) i watermarkig process. It is possible to embed perceptually iaudible watermarks with more eergy i a audio, which makes watermark more robust [14]. Dig et al. [15] proposed a audio watermarkig scheme based o wavelet packet ad psychoacoustic model. The maskig effects of the huma auditory system are calculated i each subbad by wavelet packet decompositio. The embeddig stregth is cotrolled by the maskig threshold. Thus, the watermarkig scheme has good secrecy ad high robustess. Aother way to improve the performace of watermarkig schemes is to make use of artificial itelliget (AI) techiques. The watermarkig system ca be viewed as a optimizatio problem. Therefore, it ca be solved by Geetic Algorithm (GA), support vector machie (SVM), adaptive tabu search (ATS) or eural etwork (NN) [16]. There has bee little research i applicatio of GA to digital audio watermarkig problems. Huag ad Wu [17] proposed a watermarkig method based o the discrete cosie trasform (DCT) ad Geetic Algorithm. They embed the watermark with visually recogizable patters ito the image by selectively modifyig the middle-frequecy parts of the image. The Geetic Algorithm is applied to search for the locatios to embed the watermark i the DCT coefficiet block such that the quality of the watermarked image is optimized. Sedghi et al. [18] proposed a ovel approach i spread spectrum watermarkig based o Geetic Algorithm for recoverig Pseudo-oise (PN) sequece ad detectig locatio of watermark sigal without ay iformatio from the trasmitter. Ketchem ad Vogpradhip [19] preseted audio watermarkig techique usig multiple image-based watermark scheme based o Geetic Algorithm i the DWT domai. They make use of Geetic Algorithm to fid the optimum frequecy bads for watermark embeddig which ca simultaeously improve robustess ad audio quality of the watermarked audio. I [20], Sriyigyog ad Attakitmogcol proposed a robust audio watermarkig method based o the DWT ad the adaptive tabu search. Adaptive tabu search is the artificial itelliget searchig techique desiged for the solutio of optimizatio problems. ATS is applied to search for optimal itesity of watermark such that the watermarked audio quality is optimized. Wag et al. [21] proposed a support vector machies-based audio watermarkig scheme i wavelet domai. This algorithm embeds the template iformatio ad watermark sigal ito the origial audio by adaptive quatizatio accordig to the local audio correlatio ad huma auditory maskig. I this paper, we propose a audio watermarkig method based o the discrete multiwavelet trasform for the applicatio of copyright protectio. I our algorithm, the watermark is embedded ito the multiwavelet trasform ISSN: Issue 6, Volume 9, Jue 2010

3 coefficiets usig quatizatio idex modulatio techique. The watermark ca be ot oly detected but also extracted to verify the ower. We apply the GA to search for optimal watermarkig parameters i order to achieve optimum performace. Fially, we have compared the experimetal results before ad after optimizatio usig GA with the results of previous works. This paper is orgaized as follows: I Subsectios 2.1 ad 2.2, the prelimiaries of multiwavelets ad GA are itroduced, respectively. Watermarkig i the DMT domai with GA optimizatio is described i Sectio 3. I Sectio 4, the experimetal results are show. The coclusios of our study ca be foud i Sectio 5. 2 Prelimiaries 2.1 Multiwavelet Trasform I recet years, multiwavelet trasformatio has gaied a lot of attetio i sigal processig applicatios. The mai motivatio of usig multiwavelet is that it is possible to costruct multiwavelets that simultaeously possess desirable properties such as orthogoality, symmetry ad compact support with a give approximatio order [22, 23]. These properties are ot possible i ay scalar wavelet (wavelet based o oe scalig fuctio). Oe of the well-kow multiwavelets was costructed by Doova, Geroimo, Hardi, ad Massopust (DGHM) [24]. DGHM multiwavelets simultaeously possess orthogoality, compact support, a approximatio order of 2 ad symmetry. A brief overview of the multiwavelet trasform is described ext. Let Φ deotes a compactly supported r orthogoal scalig vector ( 1, 2,..., ) T Φ = φ φ φ where r is the umber of scalar scalig fuctios. The Φ(t) satisfies a two-scale dilatio equatio of the form Φ( t ) = 2 h( ) Φ(2t ) (1) for some fiite sequece h of r r matrices. Furthermore, the iteger shifts of the compoets of Φ form a orthoormal system, that is l l' ( l, l', ' < φ ), φ ( ') >= δ δ. (2) Let V 0 deote the closed spa of l { φ ( ) Z, l = 1,2,..., r} ad defie V j = { f ( ) f V0}. The ( V j j ) j Z is a 2 2 R multiresolutio aalysis of L ( ). Note that the decreasig covetio V j+1 V j is chose. Let W j deotes the orthogoal complemet of V j i V j 1. The there exists a orthogoal r multiwavelet ( 1, 2,..., ) T Ψ = ψ ψ ψ such that l { ψ ( ) l = 1,2,..., r ad Z} form a orthoormal basis of W 0. Sice W 0 V 1, there exists a sequece g of r r matrices such that Ψ( t ) = 2 g( ) Φ(2t ). (3) Let f V0, the f ca be writte as a liear combiatio of the basis i V 0 : T f ( t) = c0 ( k) Φ( t k) (4) for some sequece c0 l2( Z). Sice V = V, f ca also be expressed as 0 1 W1 1 T t f ( t) = c1 ( k) Φ( k) 2 k Z T t d 1( k) ψ ( k). (5) 2 k Z 2 The coefficiets c 1 ad d 1 are related to c 0 via the followig decompositio ad recostructio algorithm: c 1 ( k) = h( ) c0 (2k + ) (6) d 1 ( k) = g( ) c0(2k + ) (7) 0 ( T 1 T c k) = h( k 2) c ( ) + g ( k 2) d1 ( ). (8) Ulike scalar wavelet, eve though the multiwavelet is desiged to have approximatio order p, the filter bak associated with the multiwavelet basis does ot iherit this property. Thus, i applicatios, oe must associate a give discrete sigal ito a sequece of legth r vectors without losig some certai properties of the uderlyig multiwavelet. Such a process is referred to as prefilterig. The block diagram of a multiwavelet with prefilter Q (z) ad postfilter P (z) r ISSN: Issue 6, Volume 9, Jue 2010

4 is show i Fig. 1 where c 1 is the approximatio subbad which maily represets the low frequecy compoet of the audio sigal, ad d 1 is the detail subbad which maily represets the high frequecy compoet of the audio sigal. H (z) ad G (z) are the z trasform of h () ad g (), respectively. Two audio subbads are obtaied from each level of decompositio; oe detail subbad ad oe approximatio subbad. For the ext level of decompositio, the multiwavelet trasform is applied to the approximatio subbad of the previous decompositio level. Thus, levels of decompositio result i + 1 subbads at the aalysis filter bak. Fig.1 Multiwavelet filter bak 2.2 Geetic Algorithm Geetic algorithm (GA) is oe of the most widely used artificial itelliget techiques belogig to the area of evolutioary computatio. Geetic algorithm based o the mechaisms of atural selectio ad geetics, has bee developed sice 1975 [25] ad has bee applied to a variety of optimizatio ad search problems [11, 17, 18, 19, 23]. GA has bee prove to be very efficiet ad stable i searchig for global optimum solutios. Usually, a simple GA is maily composed of three operatios: selectio, geetic operatio ad replacemet. A brief summary for implemetig GA ca be summarized as follows: Defiig the solutio represetatio of the system is the first task of applyig GA. GA uses a populatio, which is composed of a group of chromosomes, to represet the solutios of the system. The solutio i the problem domai ca the be ecoded ito the chromosome i the GA domai ad vice versa. Iitially, a populatio is radomly geerated. The fitess fuctio the uses objective values from objective fuctio to evaluate the fitess of each chromosome. The fitter chromosome has the greater chace to survive durig the evolutio process. The objective fuctio is problem-specific; its objective value ca represet the system performace idex (e.g. a error). Next, a particular group of chromosomes is chose from the populatio to be parets. The offsprig is the geerated from these parets by usig geetic operatios, which ormally are crossover ad mutatio. Similar to their parets, the fitess of the offsprig is evaluated ad used i replacemet processes i order to replace the chromosomes i the curret populatio by the selected offsprig. The GA cycle is the repeated util a desired termiatio criterio is satisfied, for example, the maximum umber of geeratios is reached or the objective value is below the threshold. There are various techiques i desigig GA that we have to take ito accout. These iclude ecodig schemes, fitess evaluatio, paret selectio, geetic operatios ad replacemet strategies. 3 Proposed Method I this sectio, we first give a brief overview of the watermark embeddig ad watermark extractig algorithms i the DMT domai based o the cocept of the quatizatio idex modulatio techique. We the describe our proposed optimizatio techique i watermarkig scheme usig geetic algorithm. 3.1 Watermark Embeddig Algorithm The watermark embeddig algorithm is described as follows: 1. Geerate a seed by mappig a sigature or text through a oe-way determiistic fuctio. The seed is used as the secret key for watermarkig. 2. To icrease security, perform a pseudoradom permutatio i order to disperse the spatial relatioship of the biary watermark patter. Therefore, it would be difficult for a pirate to detect or remove the watermark. We use W ad W to deote the origial watermark image ad the permuted watermark image, respectively. The relatioship betwee W ad W ca be expressed as W ( i, j) = W ( i, j ), where ( i, j ) is permuted to the pixel positio ( i, j) i a secret order usig the secret key. Sice the audio sigal is oedimesioal, we should trasform the permuted watermark image ito the oe-dimesioal sequece i order to embed it i the audio sigal. The, the W is trasformed ad mapped ito a biary atipodal sequece W ˆ = { wˆ i } for i = 1, 2,..., N w, where N w is the legth of watermark ad w ˆi { + 1, 1}. 3. Trasform the origial audio sigal ito fivelevel decompositio usig the DMT. Sice the approximatio coefficiets are supposed to be relatively stable ad less sesitive to slight chages of the audio sigal, they are ideal embeddig area. I order to achieve a balace betwee robustess ISSN: Issue 6, Volume 9, Jue 2010

5 ad fidelity, the coefficiets at coarsest approximatio subbad are selected for watermark embeddig based-o artificial itelliget techique. Furthermore, the coefficiets i high-frequecy subbad are ot used for watermark embeddig because of their low sigal eergy i this frequecy bad. 4. Select the sigificat coefficiets i the DMT domai which is the first N w largest coefficiets at coarsest approximatio subbad to embed the watermark bits. The positio of sigificat coefficiets will be set to the receiver as the side iformatio. To icrease the watermarkig security, we order the N w largest coefficiets i a pseudoradom maer. The radom umbers ca be geerated usig the same secret key i step (1). 5. For watermark embeddig, the sequece { w ˆ i } is embedded ito the selected coefficiets by quatizatio idex modulatio techique. The quatizatio fuctio is give as follows: ci / S S + 3S / 4 if wˆ i = + 1 c i = (9) ci / S S + S / 4 if wˆ i = 1, where x rouds to the greatest iteger smaller tha x, { c i } ad { c i } are the DMT coefficiets of the origial audio data ad the correspodig watermarked audio data respectively, ad S is quatizatio step. A large S makes the watermark robust, but it will destroy the origial quality of the audio. Thus, the value of S should be as large as possible uder the costrait of imperceptibility. I order to improve both quality of watermarked audio ad robustess of the watermark, this work employs the Geetic Algorithm to search for the optimal quatizatio step. This quatizatio step is varied to achieve the most suitable watermarked audio sigal for each give audio sigal. The details of GA optimizatio process will be described i details i Sectio Perform iverse DMT to obtai the watermarked audio sigal. The overall watermark embeddig process is show i Fig Trasform the watermarked audio sigal ito five-level decompositio usig the DMT to obtai detail coefficiets ad approximatio coefficiets. The, we choose the first N w largest coefficiets i the coarsest approximatio subbad from positio i the side iformatio. We further order the N w largest coefficiets i a pseudoradom maer usig the secret key. 2. Let c i deote the N w largest coefficiets of the coarsest approximatio subbad. The embedded watermark ca be extracted from c i by usig the followig rule: * w + 1 if ci ci / S S S / 2 i = 1 (10) if ci ci / S S < S / 2 3. Iverse the permutatio of W * where * * W = { w, i = 1, 2,... } to obtai the extracted i N w watermark W. I our proposed method, the extracted watermark is a visually recogizable image. After extractig the watermark, we used ormalized correlatio coefficiets to quatify the correlatio betwee the origial watermark ad the extracted oe. A ormalized correlatio (NC) betwee W ad W is defied as: N w wi w i NC( W, W ) = i= 1 (11) N w 2 2 wi w i i= 1 i, where W ad W deote a origial watermark ad extracted oe, respectively ad W = { w i } for i = 1, 2,..., N w. The watermark extractig process is show i Fig. 3. Fig. 3 Watermark extractig process Fig. 2 Watermark embeddig process 3.2 Watermark Extractig Algorithm The watermark extractig algorithm is outlied as follows: 3.3 Improvig Performace usig Geetic Algorithms I the desig of digital audio watermarkig system, there are three goals that are always coflicted. These goals are imperceptibility, robustess ad data capacity. I order to miimize such coflicts, this work employs the geetic algorithm to search for a optimal watermarkig parameters. This ISSN: Issue 6, Volume 9, Jue 2010

6 allows the system to achieve optimal performace for digital audio watermarkig. For the optimizatio process, GA is applied i the watermark embeddig ad the watermark extractig processes to search for quatizatio step ( S ). The objective fuctio of searchig process is computed by usig factors that relate to both robustess ad imperceptibility of a watermark. A high quality output audio ad robust watermark ca the be achieved. The diagram of our proposed algorithm of applyig GA is show i Fig. 4 ad details of GA are described as follows: audio sigal should be greater tha 20 db. Therefore, the value of desired SNR has bee assiged to 24 db i all experimets. Durig GAbased optimizatio processes, three attacks are chose to evaluate the imperceptibility ad robustess of the embedded watermark. They are MP3 compressio at 64 kbps, Gaussia oise additio, ad re-quatizatio. Details of these attacks will be thoroughly described i Sectio 4.3. After obtaiig the SNR i the watermarked audio, the DIF value ad the average of the three ormalized correlatios ( NC ave ) after attackig, we are ready to start the objective fuctio evaluatio. A illustrative diagram is show i Fig. 4. The objective fuctio f obj ca be evaluated as follow: f obj δdif DIF+ δnc NCave = (12) Fig. 4 Optimizatio diagram for digital audio watermarkig usig geetic algorithm Chromosome Ecodig: Chromosomes i GA represet desired parameter to be searched. Number of chromosomes used i this work is 20. The ecodig scheme is biary strig with 32 bit resolutios for each chromosome. Hece, the parameter S is represeted by chromosome with legth of 32 bits. Objective Fuctio Evaluatio: The most critical step i the GA optimizatio process is the defiitio of a reliable objective fuctio. I this paper, the objective fuctio of GA uses both ormalized correlatio ( NC ) ad differece ( DIF ) betwee desired sigal-to-oise ratio ( SNR ) ad obtaied SNR from each iteratio as performace idexes. DIF is a imperceptibility measure, while NC is a robustess measure. Accordig to the Iteratioal Federatio of the Phoographic Idustry, the SNR of watermarked, where δ DIF ad δ NC are weightig factors of DIF ad, respectively. These weightig NC ave factors represet the sigificace of each idex used i GA searchig process. If both idexes are equally sigificat, the values of these factors will be 0.5 each where the relatioship δ DIF + δ NC = 1. 0 must always hold. I this work, the weightig factorsδ DIF ad δ NC are equally set to 0.5. I order to gai the optimal performace of the quatizatio-based audio watermarkig system, f obj should be optimized at GA processes. By usig objective fuctio f obj above, the parameter S ca be optimally searched to achieve the best of both output audio quality ad watermark robustess. Selectio, Geetic Operatio, ad Replacemet: After evaluatig fitess value of each chromosome based o the proposed objective fuctio, chromosomes will be selected to produce offsprigs by crossover ad mutatio operatios. I this work, a rakig selectio is chose for selectio mechaism. The crossover is uiform, with probability of 0.7. Mutatio is stadard, with probability of The chromosomes are the partially replaced by the best chromosome for each geeratio. The GA will be iteratively performed o a iput audio sigal util a desired termiatio is satisfied. I this work, the maximum umber of geeratios is set to 30 as our stoppig criterio. The the chromosome (the solutio) with the best fitess value, i.e., the quatizatio step S, is determied. ISSN: Issue 6, Volume 9, Jue 2010

7 4 Experimetal Results ad Discussios I order to demostrate the performace of the proposed algorithm, some umerical experimets are carried out to measure the audio quality of the watermarked audio ad evaluate the robustess of the watermark uder typical attacks. A set of te audio sigals have bee used as host sigals, represetig five geeral classes of music: classical, coutry, jazz, rock, ad pop. This delieatio has bee chose because each class has differet spectral properties. Each audio sigal has duratio of 30 secods i the WAV format ad is moo, 16 bits/sample, with samplig rate of 44.1 khz. A biary logo SIP SUT of size pixels ( N w = 1,024) is used as the visually recogizable watermark. Cosequetly, the total watermark data rate is bps which satisfies the IFPI requiremet described i Sectio 1. Figures 5(a) ad (b) show the origial watermark ad permuted watermark, respectively. We use SNR (Sigal-to-oise ratio), NC (Normalized correlatio) ad BER (Bit error rate) to aalyze the performace of the proposed algorithm. The BER ad SNR are defied as: Number of error bits BER = 100% (13) Number of total bits 2 ( fi ) SNR = 10 log i 10 (14) fi fi 2 ( ) i, where f i ad f i deote the origial ad modified audio, respectively. Fig. 5 (a) Origial watermark ad (b) permuted watermark 4.1 Results of Geetic Algorithm Optimizatio Figure 6 shows the covergece of GA optimizatio at 30 geeratios of the Pop2 audio sigal. It is obvious that as the umber of geeratio icreases, the improvemet of audio quality ( SNR ) gradually approaches to a saturatio value. The resultig parameters, which are quatizatio steps S from GA optimizatio of 10 test audios, are show i the Table 1. These parameters are optimally varied to achieve the most desirable oes for origial audios with differet characteristics. 4.2 Imperceptibility Test Results The watermarked audio quality is examied by watermarkig the origial audio sigals with the resultig parameters from GA. The, the SNR test is coducted, which serves as a objective measuremet of audio sigal quality. The SNR is measured by comparig the watermarked sigal with the origial oe. Fig. 6 SNR, DIF, NC ave ad fobj from GA optimizatio process Table 1 Parameter S from GA process of each audio sigal Host sigal S Classical Classical Coutry Coutry Jazz Jazz Rock Rock Pop Pop Figure 7 shows the origial Classical1 audio sigal waveform ad the correspodig watermarked audio sigal waveform. Note that SNR as high as db for watermarked audio sigal. However, there is o obvious differece betwee origial sigal ad watermarked sigal by usig iformal listeig test, ad from Fig. 7(a) ad (b). It demostrates that the proposed algorithm has perfect isesibility i the sese of hearig. The results of watermarked audio quality are show i Table 2. The results obtaied from our proposed method which is called With-GA (After optimizatio) are compared with the method without usig Geetic Algorithm which is referred to as Without-GA (Before optimizatio). I Without-GA method, the quatizatio step is fixed at 0.4. We ca see that the proposed method ca ISSN: Issue 6, Volume 9, Jue 2010

8 improve the SNR of the watermarked audio about 2 db. Fig. 7 (a) Origial audio sigal (Classical1), (b) watermarked audio sigal Table 2 Sigal-to-oise ratio of watermarked audio sigals SNR (db) Host sigals Without-GA With-GA Classical Classical Coutry Coutry Jazz Jazz Rock Rock Pop Pop Average Robustess Test Results We first tested the robustess of the proposed algorithm to 10 audio samples uder o attacks. If the BER of the recovered watermark sequece is 0, it meas that the embedded bit ca be recovered exactly. The effects of the followig six types of attacks are the ivestigated. 1. Re-samplig: The audio sigal is first dowsampled at khz, ad the up-sampled at 44.1 khz. 2. Re-quatizatio: The 16-bit watermarked audio sigals have bee re-quatized dow to 8 bits/sample ad back to 16 bits/sample. 3. Low-pass filterig: Low-pass filterig usig a secod order Butterworth filter with cut-off frequecy of 6 khz, 12 db/octave roll-off, is performed to the watermarked audio sigals. 4. Additio of oise: White Gaussia oise with 1% of the power of the audio sigal is added. 5. Croppig: Two thousad samples of each testig sigal are cropped out at 5 radom positios. 6. Low bit-rate codec: The robustess agaist the low-rate codec was tested by usig MPEG 1 Layer III compressio (MP3) with compressio rates of 56, 64, 96, ad 128 kbps. Detectio results for the various attacks described above are show i Table 3 which displayed the NC ad BER from watermark extractio. The experimetal results give i Table 3 show that the watermark is ot affected by resamplig, re-quatizatio, additive oise, ad MP3 compressio at 64, 96, ad 128 kbps. This idicated that the watermark is very robust to these attacks. For low-pass filterig, croppig ad MP3 compressio at 56 kbps attacks, the BER values of the recovered watermark sequece are %, % ad % for the Without-GA method ad %, % ad % for the With- GA method, respectively. Although a lot of loss occurred i the audio sigal, the bit error rates are still acceptable. The results show that our proposed method yields better results tha the method without GA. Because GA search guaratees the global optimum solutio, the proposed method ca thus improve the quality of the watermarked audio ad give almost the same robustess of the watermark. Table 3 Robustess compariso of our algorithm Without-GA With-GA Attack type NC BER (%) NC BER (%) Attack free Re-samplig Re-quatizatio Low-pass filterig Additive oise Croppig MP3-56kpbs MP3-64kpbs MP3-96kpbs MP3-128kpbs Fially, results obtaied from our proposed method which is called With-GA are compared i fier details with the method based o wavelet trasform ad ATS i [20]. I order to compare robustess betwee the two techiques i a fair maer, parameters for each scheme should be adjusted so that watermarked audio sigals of approximately close imperceptibility are produced. I these experimets, the SNR of watermarked audio i each scheme has bee set to 24 db. Accordig to the experimetal results, the value of the embeddig capacity has bee assiged to bps i all experimets. The compariso results are listed i Table 4. Table 4 shows test results of Rock1 audio sigal with o attack, re-samplig, re-quatizatio, lowpass filterig, additio of oise, croppig ad MPEG 1 Layer III compressio with compressio rates of 56, 64, 96 kbps ad 128 kbps, respectively. The BER of watermark image ad the SNR of digital audio sigal are also displayed. ISSN: Issue 6, Volume 9, Jue 2010

9 Table 4 Robustess compariso of our algorithm With-GA [20] Attack type BER (%) SNR (db) BER (%) SNR (db) Attack free Re-samplig Re-quatizatio Low-pass filterig Additive oise Croppig MP3-56kpbs MP3-64kpbs MP3-96kpbs MP3-128kpbs Accordig to these results, the extracted watermark images from our proposed method have some distortio for low-pass filterig ad croppig attacks but they are still visually recogizable. I additio, the bit error rates of the extracted watermarks usig our proposed method are always lower tha the oes usig method i [20]. The results demostrate that our proposed method yields sigificatly more robust watermark tha the method i [20] does. 5 Coclusio This paper proposes a digital audio watermarkig algorithm i the multiwavelet trasform domai. I order to make the watermarked sigal iaudible, the watermark is embedded ito low frequecy part of the highest eergy of audio sigal by takig advatage of multi-resolutio characteristic of multiwavelet trasform. The watermark isertio ad watermark extractio are based o the quatizatio idex modulatio techique ad the watermark extractio algorithm does ot eed the origial audio i the extractio process. We have developed a optimizatio techique usig the geetic algorithm. I our optimizatio process, we use geetic algorithm searchig for optimal parameter which is the quatizatio step. This parameter is optimally varied to achieve the most suitable for origial audios with differet characteristics. The testig results of the watermarked audio quality ad watermark robustess with various watermark attacks show that our proposed method ca improve the performace of the watermarkig process such that the better watermarked audio quality ad watermark robustess are achieved. Further research ca be cocetrated o the developmet of our proposed method by usig the characteristics of the huma auditory system. Ackowledgemet This work was supported by a grat from Suraaree Uiversity of Techology, Nakho Ratchasima, Thailad. The authors would like to thak the Thailad Research Fud ad Commissio o Higher Educatio for all great supports. Refereces: [1] S. J. Lee ad S. H. Jug, A Survey of Watermarkig Techiques Applied to Multimedia, Proc. IEEE It. Symp. Idustrial Electroics, Pusa, South Korea, vol. 1, pp , Jue [2] S. Katzebeisser ad F. A. P. Petitcolas, Iformatio Hidig Techiques for Stegaography ad Digital Watermarkig, Artech House, Massachusetts, [3] J. W. Seok ad J. W. Hog, Audio Watermarkig for Copyright Protectio of Digital Audio Data, IEE Electroics Letters, vol. 37, pp , [4] M. Shiohara, F. Motoyoshi, O. Uchida ad S. Nakaishi, Wavelet-Based Robust Digital Watermarkig Cosiderig Huma Visual System, WSEAS Tras. o Sigal Processig, vol. 3, pp , Feb [5] X. Wag, Y. O. Yag, ad H. M. Gu, A Remote Sesig Image Self-Adaptive Blid Watermarkig Algorithm Based o Wavelet Trasformatio, Proc. 7 th WSEAS It. Cof. o Sigal, Speech ad Image Processig, Beijig, Chia, vol.1, pp , Sept [6] R. Wag, D. Xu, J. Che ad C. Du, Digital Audio Watermarkig Algorithm Based o Liear Predictive Codig i Wavelet Domai, IEEE It. Cof. Sigal Processig, Beijig, Chia, vol. 3, pp , August [7] B. Che ad G. Worell, Quatizatio Idex Modulatio: A Class of Provably Good Methods for Digital Watermarkig ad Iformatio Embeddig, IEEE Tras. o Iformatio Theory, vol. 47 o. 4, pp , May [8] S. Wu, J. Huag, D. Huag ad Y. Q. Shi, Efficietly Self-Sychroized Audio Watermarkig for Assured Audio Data Trasmissio, IEEE Tras. o Broadcastig, vol. 51. o. 1, pp , March [9] S. Kim ad K. Bae, Aalysis of Optimal Search Iterval for Estimatio of Modified Quatizatio Step Size i Quatizatio-Based Audio Watermark Detectio, Lecture Notes i Computer Sciece, 4139, pp , [10] K. R. Kwo ad A. H. Tewfik, Adaptive Watermarkig Usig Successive Subbad Quatizatio ad Perceptual Model Based o Multiwavelet Trasform, Proc. SPIE It. ISSN: Issue 6, Volume 9, Jue 2010

10 Cof. Security ad Watermarkig of Multimedia Cotets IV, Sa Jose, CA, USA., vol. 4675, pp , April [11] P. Kumsawat, K. Attakitmogcol ad A. Srikaew, A New Approach for Optimizatio i Image Watermarkig by Usig Geetic Algorithms, IEEE Tras. Sigal Processig, vol. 53, pp , [12] L. Ghouti ad A. Bouridae, A Robust Perceptual Audio Hashig Usig Balaced Multiwavelets, Proc. IEEE It. Cof. Acoustics, Speech, ad Sigal Processig, Toulouse, Frace, vol. 1, pp , May [13] P. Kumsawat, K. Attakitmogcol ad A. Srikaew, Digital Audio Watermarkig for Copyright Protectio Based o Multiwavelet Trasform, Lecture Notes i Computer Sciece, vol. 5376, pp , [14] F. Hartug ad M. Kutter, Multimedia Watermarkig Techiques, Proc. IEEE, vol. 87, o. 7, pp , July [15] W. R. Dig, X. D. We ad L. Qia, Audio Watermarkig Algorithm Based o Wavelet Packet ad Psychoacoustic Model, Proc. Sixth It. Cof. Parallel ad Distributed Computig, Applicatios ad Techologies, Dalia, Chia, vol. 1, pp , Dec [16] Y. Zhag, Blid Watermark Algorithm Based o HVS ad RBF Neural Network i DWT Domai, WSEAS Tras. o Computers, vol. 8, pp , Ja [17] C. H. Huag ad J. L. Wu, A Watermark Optimizatio Techique Based o Geetic Algorithms, Proc. SPIE It. Cof. Visual Commuicatios ad Image Processig, Sa Jose, CA, USA., vol. 3971, pp Feb [18] S. Sedghi, H. R. Mashhadi ad M. Khademi, Detectig Hidde Iformatio from a Spread Spectrum Watermarked Sigal by Geetic Algorithm, Proc. IEEE Cof. Evolutioary Computatio, Vacouver, BC, Caada, vol. 1, pp , July [19] M. Ketcham ad S. Vogpradhip, Geetic Algorithm Audio Watermarkig Usig Multiple Image-Based Watermarks, Proc. It. Symp. Commuicatios ad Iformatio Techologies, Sydey, Australia, vol. 1, pp , Oct [20] N. Sriyigyog ad K. Attakitmogcol, Wavelet-Based Audio Watermarkig Usig Adaptive Tabu Search, Proc. IEEE It. Symp. Wireless Pervasive Computig, Phuket, Thailad, vol.1, pp. 1-5 Ja [21] X. Y. Wag, P. P. Niu ad H. Y. Yag, A Robust, Digital-Audio Watermarkig Method, IEEE Multimedia, vol. 16, pp , [22] K. Attakitmogcol, D. P. Hardi ad D.M. Wilkes, Multiwavelet Prefilters II: Optimal Orthogoal Prefilters, IEEE Tras. Image Processig, vol. 10, o. 10, pp , [23] P. Kumsawat, K. Attakitmogcol ad A. Srikaew, A Optimal Robust Digital Image Watermarkig Based o Geetic Algorithms i Multiwavelet Domai, WSEAS Tras. Sigal Processig, Issue 1, vol. 5, pp , Ja [24] S. J. Geroimo, D. P. Hardi ad P. R. Massopust, Fractal Fuctios ad Wavelet Expasios Based o Several Scalig Fuctios, Approximatio Theory Joural, vol. 78, pp , [25] J. H. Hollad, Adaptatio i Natural ad Artificial Systems, A Arbor, The Uiversity of Michiga Press, ISSN: Issue 6, Volume 9, Jue 2010