Host cancelation-based spread spectrum watermarking for audio anti-piracy over Internet

Transcription

1 SECURITY AND COMMUNICATION NETWORKS Security Comm. Networks 2016; 9: Published online 20 October 2016 in Wiley Online Library (wileyonlinelibrary.com) RESEARCH ARTICLE Host cancelation-based spread spectrum watermarking for audio anti-piracy over Internet Rangkun Li 1 *, Shuzheng Xu 1,BoRong 2 and Huazhong Yang 1 1 Department of Electronic Engineering, Tsinghua University, Beijing 10084, China 2 Communications Research Centre, Ottawa, Ontario, Canada ABSTRACT This paper addresses the audio piracy problem over Internet by tweaking data redundancy of host media signal for optimized embedding and extraction of spread spectrum watermarking. In particular, we take into account a special feature that host audio signals have the short-time stationary property and their major power can be removed using linear prediction filter. By combining the redundancy removing and the improved spread spectrum modulation, host interference is canceled with minimized embedding distortion to the host audio. The data redundancy removing is also applied at the receiver to achieve matched filtering and improved performance. Experiments based on real audio signals show that our proposed scheme performs robustly against various kinds of channel attacks while maintaining high extraction performance. Copyright 2016 John Wiley & Sons, Ltd. KEYWORDS spread spectrum watermarking, data redundancy, AR model, linear prediction filtering *Correspondence Rangkun Li, Department of Electronic Engineering, Tsinghua University, Beijing 10084, China. lrk06@mails.tsinghua.edu.cn 1. INTRODUCTION Digital watermarking embeds additional security information into digital materials or media streams and becomes a popular solution to various online security applications including media copyright protection, broadcast monitoring, and covert communication [1]. The media copyright protection or video/audio antipiracy over Internet usually has to meet three contradictory requirements [2 4]: robustness, capacity, and fidelity. Correspondingly the digital watermark must provide enough information embedding capacity without downgrading the perceptual quality of host media and perform robustly against common signal processing and malicious attacks. Different types of schemes have been developed to effectively embed watermark information into audio signals. The efforts in [5 7] endeavor to modulate the watermark information bits to the phase components of the audio signal, utilizing the fact that human ears are insensitive to slight phase changes. Orthogonal frequency-division multiplexing concept is introduced in [7] to allow for controllable embedding payload and perceptual distortion. Patchwork modulation embeds watermark bits by modifying the patches, which are defined by a group of coefficients with relatively stable statistical characteristics. Authors in [4] use DCT coefficients for patchwork watermark embedding, achieving high robustness against desynchronization attacks. Quantization index modulation methods proposed in [8,9] can achieve high embedding capacity as they are less affected by the host signal properties. They usually adopt dither modulation that only require simple quantization calculations. Certain features of the audio signal can also be used for watermark embedding, including empirical mode decomposition [10,11], singular value decomposition [12], and log coordinate mapping features [13]. Spread spectrum (SS) modulation [14 16] is an advanced watermarking technique with the advantages of easy realization, strong robustness, and selfsynchronization. By modulating one watermark bit with a relatively long pseudo-noise (PN) sequence, the watermark signal power is spread across the whole spectrum band of host audio signal. So it would be very difficult to remove the watermark signal without damaging the host audio integrity. However, traditional SS watermarking methods severely face the host interference problem [17 19]. As in the case of audio watermarking, the audio signal is very sensitive to additive manipulations. So the Copyright 2016 John Wiley & Sons, Ltd. 4691

2 Host cancelation-based spread spectrum watermarking R. Li et al. watermark embedding power has to be suppressed at very low level to guarantee the audio fidelity. As a result, the host audio signal behaves as a strong noise interference as regard to the watermark signal, which severely downgrades the watermark decoding performance at the correlation extractor. Many approaches have been made to solve the host interference problem and improve the SS audio watermarking performance. One major category contains the informed embedding strategies. As the watermark embedder has the foreknowledge of the host signal, effective watermark embedding can be achieved by removing host interference during watermark modulation. In [18], an embedding strategy called the improved SS (ISS) modulation is proposed. The benefit of known host information is exploited at the watermark embedder to suppress host signal projection on the watermark signal. By subtracting the host signal projection on the PN sequence at the watermark embedder, the host interference can be largely removed. Authors in [17] apply similar informed embedding method. They make the improvement that host interference is compensated not only at the embedding location but also at the neighboring samples. Thus, the watermarked audio is more robust against translation-type attacks. These precompensating schemes can largely remove host signal interference. However, the optimization of error probability is performed under the restriction of limited average distortion, the compensating will introduce localized distortions that may downgrade the perceptual quality. Spread spectrum audio watermarking performance can also be improved by other methods, including approaches at the decoding side [20,21] and the embedding side [22]. In [20], correlation is performed on a selective portion rather than the whole block of watermarked signal, achieving a higher watermark-to-cover signal power ratio at the extractor. Authors in [21] improve the SS watermark decoding performance by introducing adaptive normalization coefficients for correlation calculating and adaptive thresholds for making compare decisions. In [22], the watermark embedding locations are no longer fixed. Maximum energy regions in the frequency spectrum of the cover signal are chosen to embed watermark information. Authors in [14] utilize a vector space projection approach and present both the watermark signal and the host carrier signal as vectors. These methods partially reduce the negative influence of host audio signal. However, they do not achieve total host interference elimination and their performance is rather limited. This paper aims to provide a high-fidelity and highperformance audio watermarking scheme. We utilize the fact that audio signals fit into the autoregressive (AR) model and contain large parts of redundancy power. Instead of compensating the whole power of host audio interference, we show that the embedder can largely reduce the compensating distortion by removing the data redundancy in the host signal. Also, by introducing a twostep correlation method, robust watermark extraction is achieved. The rest of this paper is organized as follows. Section 2 introduces the basics of SS-based audio watermarking. Section 3 presents the redundancy-removing-based ISS (RRISS) strategy, while Section 4 conducts performance analysis correspondingly. In Section 5, we propose the watermark embedding and extraction system based on RRISS strategy, and in Section 6, we illustrate the experiment results and robustness tests using real audio signals. Finally, Section 7 concludes the paper. 2. SPREAD SPECTRUM WATERMARKING AND IMPROVED SPREAD SPECTRUM STRATEGY 2.1. Spread spectrum watermarking A basic SS-based watermarking scheme involves the watermark information b, a key-generated PN sequence u and the host signal x. Here, u and x are vectors of length N where larger vector length means higher SS gain. The watermark information b is in bipolar form, taking its value from {+1, 1}. Then the watermarked signal can be represented as s = x + bu. After passing through additive noise channel, the watermark detector receives y = x + bu + n, which contains three different components: the host audio signal, the watermark signal, and the channel noise. For the watermark embedder, the watermark component u should be limited in power for fidelity requirements, whereas for the watermark extractor, the power of u is crucial for successful extraction and the host signal x and channel noise n should be suppressed as much as possible. Correlation is typically used at the receiver to extract the watermark information from the watermarked and noisecontaminated audio signal. In blind extraction scenario, the original audio signal x is unavailable. The sufficient statistic r is calculated as [18] r = hy, ui hx+bu+n, ui = = b + x + n (1) hu, ui kuk where x, hx, ui/kuk and n, hn, ui/kuk denote the interferences form host signal and channel noise, respectively. The embedded information bit b can be estimated simply by Ob = sign(r). Extraction error probability for SS watermarking can be further derived [18], assuming host signal x and channel noise n have independent white Gaussian distribution and with variances of x 2 and 2 n respectively: p = 1 2 erfc s! Nu 2 2(n x ) Note that the power level of the watermark signal is much lower than that of the host signal because of the fidelity requirement. From (2), the host signal interference will (2) 4692 Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.

3 R. Li et al. Host cancelation-based spread spectrum watermarking result in high decoding error rate or low embedding capacity as x 2 and 2 n are much higher than 2 u. So the simple SS watermarking does not achieve very good performance for practical applications Improved spread spectrum strategy The ISS strategy is introduced in [18] to resolve the host interference problem for SS-based watermarking schemes. ISS strategy applies an informed embedding method to remove host interference at the watermark embedder. The linear form of the informed embedder is commonly adopted that simply subtracts the host projection on the PN sequence from the watermarked signal: s = x +( b x)u (3) The parameter is introduced here to compensate host interference x. When equals to 1, the host interference is completely removed. The parameter limits the average embedding distortion. The ISS strategy optimizes and for lowest decoding error probability under the constraint of limited average distortion. The average distortion of the ISS embedding should be the same as traditional SS. Theoretically, the pre-compensate modulation of ISS watermarking should bring several orders of magnitude of performance gain as regard to error probability. However, the ISS modulation has several drawbacks for real audio watermarking applications: Firstly, the ISS modulation focuses on embedding and does not take extraction optimization into consideration. Common processing attacks like filtering would damage the pre-compensating structure. Secondly, the ISS modulation assumes independent white Gaussian property for both the audio signal and the watermark signal, which is not the case for real audio signals. Thirdly, the ISS strategy uses the limitedaverage-distortion principle. However, the localized distortion introduced by the ISS compensation may be very large that would result in perceivable difference although the average distortion power is the same as the traditional SS watermarking. For SS-based audio watermark embedding, the SS sequence is usually perceptually shaped to guarantee the perceptual quality of watermarked audio. The psychoacoustic model is widely employed in many watermarking schemes for watermark shaping [20,21]. The psychoacoustic model provides a clear boundary for the watermark embedding capacity. This fact is shown in Figure 1 in both time and frequency domain. We can see that the embedding distortion allowed by the psycho-acoustic model is entirely decided by the localized spectrum property of host audio signal. Here, we focus on the effect ofpre-compensation on watermark embedding. The ISS modulationunder perceptual shaping is rewritten as s = x +(Q b Qx)u s (4) Here, u s denotes the shaped watermark signal and Qx, hx, ui/hu s, ui is the host projection, Q is chosen under ISS s average distortion rule. To analyze the localized embedding effect of pre-compensation of ISS modulation, we define the relative distortion as the ratio between the distortion results with and without ISS modulation: D ISS, k( Q b Qx)us k kbu s k = hx, ui ˇ Q b hu s, ui ˇ (5) Figure 1. Demonstration of watermark shaping with the psycho-acoustic model: (a) original audio signal sampled at 44.1 khz, (b) perceptually shaped watermark signal, and (c) host audio power spectrum (thinner line) and masking threshold (bold line) calculated by the psycho-acoustic model. Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd. 4693

4 Host cancelation-based spread spectrum watermarking R. Li et al. many source coding standards for speech, audio, image, and video data compression. This provides an intuitive way to suppress host interference: apply data redundancy removing at the watermark embedder and extractor. The audio signal data redundancy can be expressed using the AR model. Let {x t M, x t M+1 x t 1, x t }beany M + 1 adjacent samples from the audio signal x. Then the M-order AR process modeling the audio signal can be expressed using the following equation: x t = w k x t k + v t (6) Figure 2. Distribution of relative distortion introduced by precompensation. Using audio signal sampled at 44.1 khz. Watermark inserted with the watermark-to-signal ratio at 23 db and at different modulation speed with pseudo-noise (PN) length 512, 1024, and And scale factor is set to be 0.50, 0.65, and 0.8 respectively to achieve limited average distortion. Equation 5 gives a good measurement of localized distortion. If the value of D ISS is less than 1, then the pre-compensation does not introduce more localized distortion. The distribution of relative distortion D ISS is experimented on real audio signals, and the result is illustrated in Figure 2. It is easy to see that ISS modulation leads to excessive localized distortions, which will dissatisfy the perceptual boundary of the psycho-acoustic model. And the localized distortions become larger as the insertion speed increases. This will limit the ISS watermarking performance under high-capacity applications where ISS modulation was originally intended for. Note that the ISS modulation works under the assumption that host signals fit into Gaussian model. In such assumption, the data redundancy of the host audio signals is neglected. Actually, real audio signals usually have very large dynamic range and contain large parts of redundant power. As stated in the following section, exploiting such redundancy at the embedder and extractor can help to develop a high transparency and high-performance watermarking scheme. Here, the M coefficients w ={w k } are used to model the relationship between neighboring samples, as w k denotes the correlation of two samples that are distanced by k samples. v t is the estimation residual that can be treated as data sampled from a white noise signal v and is usually much lower in power than the host signal x. From (6), the host signal can be treated as two parts: the deterministic (DC) component and the stochastic (SC) component. DC component usually contains the major power of host signal, it is made up with data that can be totally predetermined by the last M samples. SC component v contains data that is totally independent with adjacent samples; it appears to be more randomly distributed and has a near-gaussian property. So the data redundancy of the host signal is contained mainly in the DC component. Linear prediction filtering (LPF) is typically used to remove data redundancy in media signals. LPF estimates the prediction coefficients {w k } by minimizing the power of residual signal: v = L(x). The power ratio between x and v is defined as the redundancy suppression gain: G = x 2/2 v. LPF can be efficiently realized using the Levinson Durbin algorithm, which calculates the prediction coefficients recursively. Real audio signals usually do not strictly fit into a fixed-order AR model, smaller prediction residual can be 3. REDUNDANCY-REMOVING- BASED IMPROVED SPREAD SPECTRUM WATERMARKING 3.1. Data redundancy removing As shown earlier, the ISS modulation pre-compensates host signal projection on the PN sequence and solves the host interference problem and improves watermark decoding performance. However, this will introduce localized distortion that will downgrade the perceptual quality of the host audio. On the other hand, media signals usually contain large parts of redundant power. This fact is utilized by Figure 3. The suppression gain using different prediction order Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.

5 R. Li et al. Host cancelation-based spread spectrum watermarking achieved using a higher order of LPF. However, suppression gain tends to become flat when the order becomes higher. As demonstrated in Figure 3, most redundancy power in audio signals can be removed by 10th-order prediction or less Redundancy-removing-based improved spread spectrum strategy The basic idea of RRISS strategy is to combine host interference cancelation with data redundancy removing at the watermark embedder and extractor. We first show the basic form of RRISS strategy for the watermark embedder. s = x + Qbu s Qb = b + RR(x, u, u s ) Here, RR() denotes the compensation procedure of host interference. At the watermark extractor, the watermarked audio is preprocessed with linear prediction filtering to remove redundant host power. Denote the LPF filtering as L(), and the received watermark is y = x + Qbu s + n, then the residual signal is achieved as (7) y z = L(y) =QbL(u s )+L(x)+L(n) (8) The residual signal is then further passed through correlation detector to decode the watermark information: r = hqb(u s )+(x)+l(n), L(u)i = bhl(u s ), L(u)i + hl(n), L(u)i + hl(x), L(u)i + RRhL(u s ), L(u)i We can see from (9) the correlation result contains four parts: the watermark information, channel noise, host interference, and host compensation. So the host interference can be totally removed by setting the combination of the last two parts into 0: (9) 4. PERFORMANCE OF REDUNDANCY-REMOVING-BASED IMPROVED SPREAD SPECTRUM 4.1. Embedding distortion The RRISS strategy embed watermark information with minimized interference by compensating the SC component of host audio signal. Same as (5), we define the relative distortion as the ratio of distortion with and without RRISS modulation: D RRISS, k(b RR)us k kbu s k = ˇ hl(x), L(u)i ˇb hl(u s ), L(u)i ˇ (12) Compared with the relative distortion of ISS modulation in (5), we can see that the major difference is that RRISS calculates the host projection after the LPF. As analyzed in the earlier section, LPF filtering removes most of host audio redundancy power. On the other hand, as the watermark signal is basically a randomly generated PN sequence and contains little data redundancy, so the LPF filtering does not damage the watermark signal power. The relative distortion is simulated using real audio signal. Applying the same parameters as in Figure 2, the distribution of the localized distortion under RRISS modulation is demonstrated in Figure 4. Comparing Figure 4 to Figure 2, we find that the new RRISS modulation is much lower in localized distortions. This is essential in real audio applications to maintain the high quality of the host audio. As regarding the average distortion, we do not apply the limited average distortion rule. However, because the localized distortion is kept very low, the average distortion is just slightly enlarged. hl(x), L(u)i + RRhL(u s ), L(u)i = 0 (10) Finally, we achieve the RRISS modulation strategy: s = x + Qbu s Qb = b hl(x), L(u)i hl(u s ), L(u)i (11) Note here RRISS does not introduce the average distortion rule used by ISS. Technically this could be performed, but would lost its meaning as the distortion allowed by the psycho-acoustic model is locally determined (Figure 1). Ignoring the average distortion rule also brings extra benefit in computational simplicity as the new strategy does not have to calculate the statistical parameters of the watermark and host signal except localized computation. Figure 4. Distribution of relative distortion introduced by Redundancy-removing-based improved spread spectrum strategy (RRISS) modulation. Using audio signal sampled at 44.1 khz. Watermark embedded with the watermark-to-signal ratio at 23 db and at different modulation speed with pseudonoise (PN) length 512, 1024, and 2048, respectively. Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd. 4695

6 Host cancelation-based spread spectrum watermarking R. Li et al. So the perceptual quality of the watermark embedding is guaranteed Host suppression analysis In our watermarking scheme, linear prediction filtering is applied both at the embedder and extractor. By performing host interference cancelation and data redundancy removing, host suppression is achieved for better performance. Linear prediction filtering is performed by solving the Wiener Hopf equation: Rw f = r (13) where R is the autocorrelation matrix of the filtered signal and r is the autocorrelation vector. 2 3 r 0 r 1 r M 1 r 1 r 0 r M 2 R = r M+1 r M+2 r 0 r = r 1, r 2 r M T (14) For a signal that fits into the AR model (6), the result of Wiener Hopf equation w f is the prediction coefficients {w k, k =1,2M}. After the audio signal is watermarked and passed through the communication channel, the property of the signal is altered. Assume the watermark signal and channel noise have independent Gaussian property {u 2, 2 n }. The autocorrelation matrix of original signal x and the received signal y satisfies: R y = R x + diag( 2 u + 2 n ) r y = r x (15) So after LPF filtering, the prediction residual is derived as ( M ) e yt = wy t k + v t + Qbu s t X Qb w k u s t k (19) + n t w k n t k From (19), we can see DC component of the host signal that contains the major power is successfully removed. The residual signal contains the watermark signal, the channel noise and host interference form the SC component. Note that the watermark signal in (19) is in filtered form, which contains not only the watermark power u t but also the power from the last M watermark samples. To achieve better performance, matched filtering is preferred that means the PN sequence for the receiver should also be filtered using the same LPF filter. We can derive the noise property at the extractor from (19): ( M ) e yt = v t RR u s t X w k u s t k + w k y t k ( M ) (20) + n t w k n t k + b u s t X w k u s t k Here, the first two parts do not contribute to the correlation detector because of the compensation. The noise come from two parts: channel noise and prediction coefficient mismatch. The watermark-to-noise ratio (WNR) for detection is calculated as where diag means a diagonal matrix. Then the Wiener Hopf equation can be rewritten as o R x w x = nr x + diag(u n ) w y = r x (16) So the prediction coefficients mismatch can be derived as n var u s t P o M w k u s t k WNR = var nn t P M w k n t k + P o M w k y t k u 2 s n x {P M w 2 k }/{1 + P M w 2 k } (21) w = w x w y = R 1 x diag(2 n )w y (17) So under light attack channel, the prediction coefficients remains nearly the same. Substitute y = x + Qbu s + n into (6) we obtain: ( M ) y t = w k y t k + v t + Qbu s t X Qb w k u s t k (18) + n t w k n t k Here, we can see that very little power of host signal exists at the extractor. The only part of host signal still exists because of the prediction deviation after attack channel, which is relatively small. From the earlier analysis, we can see that under noisefree channel, the RRISS-based watermarking does not suffer from host interference and achieves no error. Under noise attack channel, the host interference for RRISSbased watermarking is largely suppressed. Moreover, RRISS achieves this performance under the advantage of limited localized distortion Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.

7 R. Li et al. Host cancelation-based spread spectrum watermarking 5. OUR PROPOSED AUDIO WATERMARKING SCHEME 5.1. Structure of the watermarking scheme The proposed watermarking scheme contains the highquality embedder and the blind extractor. Figure 5 illustrates the block diagram of the watermark embedding system, where the watermark information is inaudibly inserted into the host audio signal. The watermark embedder combines the perceptual analysis and shaping methods with the proposed RRISS modulation. The perceptual model from the MPEG-1 standard [23] is adopted here for perceptual quality control, making sure the embedded watermark signal is imperceptible. The Levinson Durbin algorithmbased RRISS strategy helps to suppress the embedding distortion. The watermark extractor introduces a two-step correlative detection method. It is the combination of the traditional simple correlation detector and the linear prediction filtering-based preprocessing. The extraction system block diagram is shown in Figure 6. The linear prediction filtering removes data redundancy of the received watermarked audio signal, canceling the negative effect of the host audio interference. The same PN sequence as the embedder must be used to successfully extract the embedded watermark information Watermark embedding procedure The perceptual shaping process is performed in the subband domain using the perceptual analysis results from the fast Fourier transform (FFT) domain. The subband transform maps time-domain signal into 32 equal-width Figure 5. The watermark embedding system. Figure 6. The watermark extraction system. Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd. 4697

8 Host cancelation-based spread spectrum watermarking R. Li et al. subbands. It has lower frequency resolution than the FFT; however, it can effectively reduce the computation load. The detailed implementation of the watermark embedding procedure is described as follows: (1) Perform perceptual analysis on the host audio signal. The power spectrum of one audio frame x = {x i 1 i N f } is calculated using FFT. Tonal and nontonal maskers {M kt, M kn } are extracted from the power spectrum samples {X k 1 k N f }. The spectral masking threshold {TF k 1 k N f }is generated by combing the masking effect of each individual masker, and it is further mapped to the subband domain {TB n 1 n 32}. {M kt, M kn }=EX({X k }) {TF k }= X MT (M i ) + X MN (M i ) (22) i2{kt} i2{kn} where {kt} and {kn} are the positions of the tonal and nontonal maskers, respectively. The maskers extraction function EX() and the tonal and nontonal masking effect functions MT(), MN() are all realized using the psycho-acoustic model [23]. The perceptual analysis and shaping process is demonstrated on the left part of Figure 7. An example of the perceptual analysis results is demonstrated in Figure 8. The thick line shows the calculated spectral masking threshold. Signals under the masking threshold are inaudible to the human ear. (2) Perceptual shaping of the watermark signal. Subband analysis SA() [23] is applied on a segment u = {u i 1 i N g } from the PN sequence. In the subband domain, the signal power in each subband {UB m,n } is limited using the subband masking threshold. Then the subband synthesis method SS() [23] is applied to transform the shaped subband samples {US m,n } back into time-domain signal u s ={u s i 1 i N g}. {UB m,n }=SA({u i }) US m,n = sign UB m,n *TBn {u i }=SA {UN m,n } 1 i N g,1 m N b,1 k 32 (23) where N b = N g /32 is the sample number in each subband. (3) Modulate the shaped PN sequence with watermark information b and add it linearly to the host audio signal. Thus, the modulated audio signal s = x + bu s is generated. Then analyze the watermarked signal using the Levinson Durbin algorithm, which generates the prediction coefficients w ={w j 1 j N w } for the linear prediction filter L(), which can be treated as a simple FIR filter for each calculation: Figure 7. Perceptual analysis and shaping process. Ou =L(u) Ou i = X j w j *u(i j) (24) (4) Adjust the embedded signal using using the RRISS strategy. The RRISS strategy calculates the host projection on the watermark using the method in (11). Then the watermark embedding is finished by adding the modulated audio signal with the compensate signal cu s. hl(x), L(u)i c = hl(u s ), L(u)i (25) s i = x i +(b + c)u s i 5.3. Watermark extraction procedure Figure 8. Masking effect analysis of host audio signal. For watermark extraction, the PN sequence u is generated using the same method as the embedder, which also requires the same key: KeyExtr = KeyEmb. The two Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.

9 R. Li et al. Host cancelation-based spread spectrum watermarking step watermark extraction method combines traditional correlation decoder with additional preprocessing, which performs linear prediction filtering on the received watermarked signal y and the PN sequence u. Oy =L(y) Ou =L(u) (26) Then the embedded watermark information is extracted using simple correlation and symbol judgment: r = hoy, Oui Ob = sign(r) where Ob is the estimated watermark information Computational overhead analysis (27) The computational complexity of the watermarking scheme is mainly influenced by the perceptual analysis method. The computational load of the RRISS modulation strategy proposed in the watermarking scheme can be neglected. The linear prediction filter used in the scheme is typically less than 10th order. Moreover, the Levinson Durbin algorithm calculates the prediction coefficients effectively using the recursive method. From the earlier analysis, the perceptual shaping method uses the subband transform, which has a lower frequency resolution that the FFT used by the perceptual analysis process. So the perceptual shaping process costs relatively less computation. The FFT in the perceptual analysis method makes up the major computational load. Another way to measure the computational complexity of the proposed watermarking scheme is to compare it with the audio coding algorithms. As explained earlier, the psycho-acoustic model in MPEG-1 standard [23] is adopted in our perceptual analysis method, so the complexity of the watermarking scheme should be relatively the same as the audio compression method. where var() denotes the power level calculation and x, s denote the host audio signal and watermarked audio signal, respectively. Note that the SWR value is not a direct evaluation for perceptual quality. However, it reflects the watermark embedding power level and it is a useful parameter for estimating extraction performance. For watermarking schemes that adopt the average distortion rule, the SWR value is the key parameter for embedding optimization. The ODG value, on the other hand, directly indicates the perceptual quality of audio signals. ODG measurement is a very efficient tool typically used in high-quality audio codec evaluations. ODG calculation requires a reference signal and a test signal slightly different from the reference signal. In our case, the reference signal is the host audio signal and the test signal is the watermarked audio signal. The relationship between ODG values and the perceptual quality is shown in Table I. Perceptual advantages of our proposed RRISS modulation is explained in the previous section. To fully evaluate the embedding performance of the proposed RRISS strategy, we use 50 audio segments for SWR and ODG tests. Figure 9 demonstrates the SWR values under different watermark embedding payloads. As we can see, since the RRISS strategy does not introduce the limited average distortion rule, the SWR values are only a little lower than the corresponding SS watermark and slightly decrease as the payload goes up. The corresponding ODG values are shown in Figure 10. We can see that even under high payload up to bps, Table I. ODG value and corresponding audio quality. ODG value Impairment 0.0 Imperceptible 1.0 Perceptible, but not annoying 2.0 Slightly annoying 3.0 Annoying 4.0 Very annoying ODG, objective difference grade. 6. EXPERIMENTAL RESULTS Experiments are performed on real audio signals of different types. These audio signals are in CD quality with 44.1-kHz sampling rate and 16-bit quantization. The performance of the proposed scheme is evaluated in two ways: perceptual performance evaluation and robustness performance evaluation. There are several ways to evaluate the perceptual performance of audio watermarking. The most commonly used parameters in watermarking evaluation are signalto-watermark ratio (SWR) and objective difference grade (ODG) evaluation [24]. SWR is defined as the power ratio between the host signal and the watermark signal: var(x) SWR, 10log var(s-x) (28) Figure 9. Signal-to-watermark ratio (SWR) values comparison under different payload. Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd. 4699

10 Host cancelation-based spread spectrum watermarking R. Li et al. Figure 10. Objective difference grade (ODG) values comparison under different payload. Figure 12. Objective difference grade (ODG) measurement for each individual audio signal under the payload of and pseudo-noise (PN) length is Table II. Robustness tests for N = 4096, SWR =25dB. Attack type Parameters ODG BER:% No attack none AWGN 20 db Amplitude scaling 200% Low-pass filtering 10 khz Resampling khz Requantization 8 bits MP3 compression 64 kbps AAC compression 64 kbps AWGN, additive white Gaussian noise; BER, Bit Error Rate; ODG, objective difference grade. Figure 11. Signal-to-watermark ratio (SWR) measurement for each individual audio signal under the payload of and pseudo-noise (PN) length is the ODG value is still in the high-quality range defined as imperceptible. SWR and ODG test results for each individual audio are illustrated in Figures 11 and 12. The PN sequence length is set as 1024 with a relatively high payload of bps per channel. We can see that different types of audio signals result in different embedding SWR. Also, the ODG value varies between different audio signal. This is understandable because each audio signal has its own characteristics. Extraction performance is evaluated using decoding error probability measurement. The experiments are performed under several common audio signal processing channel attacks to test the robustness of the proposed watermarking scheme: (1) White noise attack. The watermark signal is passed through additive white Gaussian noise channel with 20-dB signal-to-noise ratio. Table III. Robustness tests for N = 1024, SWR =28dB. Attack type Parameters ODG BER:% No attack none AWGN 20 db Amplitude scaling 200% Low-pass filtering 10 khz Resampling khz Requantization 8bit MP3 compression 64 kbps AAC compression 64 kbps AWGN, additive white Gaussian noise; BER, Bit Error Rate; ODG, objective difference grade. (2) Amplitude scaling. The watermarked signal is magnified by 200% in amplitude and crop the samples exceeding the dynamic range. (3) Low-pass filtering. The watermarked signal is filtered with 10-kHz low-pass filter. (4) Resampling. Downsample the watermarked signal with half the original sampling rate (22.05 khz) and then upsample back the sample rate (44.1 khz) Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.

11 R. Li et al. Host cancelation-based spread spectrum watermarking Table IV. Comparison with different audio watermarking methods (BER:%). Scheme Method Capacity None White noise Re- Low-pass MP3 bps quantization filtering compression Proposed Spread spectrum (20 db) 0.12(8 bits) 0.17 (10 khz) 0.04 (64 kbps) Proposed Spread spectrum (20 db) 0.38(8 bits) 2.04 (10 khz) 0.21 (64 kbps) In [25] Spread spectrum (30 db) 2.52(8 bits) 4.36 (12 khz) 3.02 (128 kbps) In [14] Spread spectrum (20 db) 0.00(8 bits) 0.34 ( khz) In [26] Patchwork (20 db) 0.04(8 bits) 0 (8 khz) 0 (128 kbps) In [27] Echo hiding (36 db) 5.6(8 bits) 3.2 (6 khz) 0.5 (-) In [10] EMD (20 db) 0(8 bits) 6 (-) 1 (32 kbps) In [28] Quantization (20 db) 0(8 bits) 0 ( khz) 2 (32 kbps) EMD, empirical mode decomposition. (5) Requantization. Coarsely requantize the 16-bit watermarked signal with 8-bit resolution. (6) MP3 compression. Compress the watermarked signal using 64 kbps MP3 compression. (7) AAC compression. Compress the watermarked signal using 64-kbps AAC compression. Tables II and III demonstrate the robustness test results for the watermarking scheme. Two sets of embedding parameters are tested, one with lower embedding capacity (10.77 bps, n = 4096) and higher average distortion (SWR = 25 db), another one with higher embedding capacity (43.07 bps, n = 1024) and lower average distortion (SWR = 28 db). Note here, we use the limited average distortion principle at the embedder to achieve fixed embedding distortion so that the extraction results are suitable for comparison. As a result, the embedding ODG values are 0.73 and 0.59, respectively. We can see that, for most channel attacks, the ODG values of the audio signal after the attacks are severely downgraded. Which means the watermarked audio signals are perceptually damaged. Still, most watermark information can be successfully extracted. This means the RRISSbased watermarking scheme is robust gainst commonly used signal processing attacks. Also, a tradeozff can be made between the decoding performance and the embedding capacity by adjusting the frame length N, making practical applications more flexible. For applications like convert communications, the watermark signal can survive very terrible communication channels and maintain the integrity of secure information. And for copyright protection, unauthorized third parties cannot completely remove the watermark without causing damages to the host audio signal. The performance of our watermarking scheme is compared with several other audio watermarking schemes. As shown in Table IV, these schemes are realized using different types of methods, including SS, patchwork modulation, echo hiding, empirical mode decomposition, and quantization modulation. The proposed watermarking scheme achieves higher performance than other SS-based schemes and shows advantages in most of the test results. Note that the watermarking scheme [25] achieves high embedding capacity, which is a general property forquantization-based watermarking schemes. SS-based watermarking schemes, on the other hand, have the advantages of high security and good synchronization property. Moreover, the SS watermarking schemes perform better against malicious attacks. 7. CONCLUSIONS In this paper, we develop a host cancelation-based watermarking scheme for online audio anti-piracy. Our scheme innovatively exploits the data redundancy property of audio signal by removing data redundancy at both the watermarked embedder and the extractor using the Levinson Durbin algorithm. Unlike the pervious SS-based watermarking approaches who pay little attention to the data structure of host signal, our proposed RRISS strategy combines the short-time stationary property of audio signals with the ISS pre-compensation method, so that the host interference is removed at the watermark embedder with minimized localized distortions. Experiments with real audio signals clearly demonstrate the advantages of our proposed scheme and its potential to practical applications. REFERENCES 1. Mazurczyk W, Szczypiorski K, Tian H, Liu Y. Trends in modern information hiding: techniques, applications and detection. Security and Communication Networks 2013; 6(11): Cvejic N, Drajic D, Seppanen T. Audio watermarking: more than meets the ear. In Recent Advances in Multimedia Signal Processing and Communications. Springer, 2009; Ghebleh M, Kanso A, Own HS. A blind chaos-based watermarking technique. Security and Communication Networks 2014; 7(4): Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd. 4701

12 Host cancelation-based spread spectrum watermarking R. Li et al. 4. Xiang Y, Natgunanathan I, Guo S, Zhou W, Nahavandi S. Patchwork-based audio watermarking method robust to de-synchronization attacks. IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP) 2014; 22(9): Arnold M, Chen XM, Baum P, Gries U, Doerr G. A phase-based audio watermarking system robust to acoustic path propagation. IEEE Transactions on Information Forensics and Security 2014; 9(3): Djebbar F, Ayad B, Abed-Meraim K, Hamam H. Unified phase and magnitude speech spectra data hiding algorithm. Security and Communication Networks 2013; 6(8): Garcia-Hernandez JJ, Parra-Michel R, FeregrinoUribe C, Cumplido R. High payload data-hiding in audio signals based on a modified OFDM approach. Expert Systems with Applications 2013; 40(8): Fallahpour M, Megias D. High capacity logarithmic audio watermarking based on the human auditory system. IEEE International Symposium on Multimedia (ISM) 2012: Hu HT, Li WC. A perceptually adaptive and retrievable QIM scheme for efficient blind audio watermarking. International Conference on Information Science and Applications (ICISA), 2012; Khaldi K, Boudraa A. Audio watermarking via EMD. IEEE Transactions on Audio, Speech, and Language Processing 2013; 21(3): Wang XG, Niu PP, Yang HY, Zhang Y, Ma TX. A robust audio watermarking scheme using higher-order statistics in empirical mode decomposition domain. Fundamenta Informaticae 2014; 130(4): Bhat V, Sengupta I, Das A. A new audio watermarking scheme based on singular value decomposition and quantization. Circuits, Systems, and Signal Processing 2011; 30(5): Kang X, Yang R, Huang J. Geometric invariant audio watermarking based on an LCM feature. IEEE Transactions on Multimedia 2011; 13(2): Abubakar AI, Zeki AM, Chiroma H, Muaz SA, Sari EN, Herawan T. Spread spectrum audio watermarking using vector space projections. In Advances in Intelligent Informatics. Springer, 2015; Jia W, Tso FP, Ling Z, Fu X, Xuan D, Yu W. Blind detection of spread spectrum flow watermarks. Security and Communication Networks 2013; 6(3): Zhang Y, Xu Z, Huang B. Channel capacity analysis of the generalized spread spectrum watermarking in audio signals. IEEE Signal Processing Letters 2015; 22(5): Gerek N, Mihcak MK. Generalized improved spread spectrum watermarking robust against translation attacks. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2008; Malvar HS, Florencio DA. Improved spread spectrum: a new modulation technique for robust watermarking. IEEE Transactions on Signal Processing 2003; 51(4): Wei L, Pados DA, Batalama SN, Hu RQ, Medley MJ. Optimal multiuser spread-spectrum data hiding in digital images. Security and Communication Networks 2015; 8(4): Wook Kim H, Choi D, Choi H, Kim T. Selective correlation detector for additive spread spectrum watermarking in transform domain. Signal Processing 2010; 90(8): Li L, Fang X. Adaptive detection for spread spectrum audio watermarking, 2010; Koz A, Delpha C. Adaptive selection of embedding locations for spread spectrum watermarking of compressed audio. Digital Forensics and Watermarking, Springer 2012: Committee I, et al. Coding of moving pictures and associated audio for digital storage media at up to about 1.5 mbit/s-part 3: Audio. ISO/IEC 1993; 11: Lin Y, Abdulla W. Objective quality measures for perceptual evaluation in digital audio watermarking. IET Signal Processing 2011; 5(7): Bhat V, Sengupta I, Das A. An audio watermarking scheme using singular value decomposition and dithermodulation quantization. Multimedia Tools and Applications 2011; 52(2-3): Zhang P, Li Y, Fan Y, Jiang J, Ma X, Hao Q. Robust audio watermarking based on frequencydomain spread spectrum using CAZAC sequence. Sixth International Conference on Advanced Computational Intelligence (ICACI) 2013: Natgunanathan I, Xiang Y, Rong Y, Zhou W, Guo S. Robust patchwork-based embedding and decoding scheme for digital audio watermarking. IEEE Transactions on Audio Robust patchwork-based Speech, and Language Processing 2012; 20(8): Cao X, Zhang L. Researches on echo kernels of audio digital watermarking technology based on echo hiding. IEEE International Conference on Wireless Communications and Signal Processing (WCSP) 2011: Security Comm. Networks 2016; 9: John Wiley & Sons, Ltd.