A Blind EMD-based Audio Watermarking using Quantization

Transcription

1 768 A Blind EMD-based Audio Watermaring using Quantization Chinmay Maiti 1, Bibhas Chandra Dhara 2 Department of Computer Science & Engineering, CEMK, W.B., India, chinmay@cem.ac.in 1 Department of Information Technology, Jadavpur University, W.B., India, bibhas@it.jusl.ac.in 2 ABSTRACT In this paper, a blind audio watermaring algorithm in time domain using empirical mode decomposition (EMD) and quantization is presented. The host audio signal is partitioned into non-overlapping segments. Then, each segment is decomposed into finite number of intrinsic mode functions (IMFs) using EMD. In the proposed algorithm, the first IMF has been considered for embedding the watermar. The watermar bit is embedded into the audio segment by quantizing norm of the extremes of the first IMF using quantization process. The size of the watermar is equal to the number of segments of the host audio signal. The watermar is extracted from watermared audio signal blindly in the extraction process. To enhance the security of the watermar, Arnold transformation technique is used to scramble the watermar before embedding it. The experimental evaluation shows that proposed method achieves high imperceptibility and high capacity. Moreover, the present algorithm is also robust under different audio attacs such as mp3 compression, requantization, amplitude reverse, Additive White Gaussian noise, cropping, etc. Keywords: Audio watermaring, Empirical mode decomposition, Security, Audio attacs, Robustness 1. INTRODUCTION The rapid growths of the internet and multimedia technology have allowed transmission and distribution of the digital multimedia data such as image, video, audio easily and efficiently to distant places. However, this convenience allows unauthorized copying and distribution of the digital multimedia data. At the same time, the problem of copyright protection is increasing day by day. Copyright protection of multimedia data can be possible by two different ways: Cryptography and Watermaring. In cryptography, data encrypted to prevent the unauthorized access but once the data are decrypted then everyone can access it. In watermaring, watermar information is embedded into the digital multimedia data (also called host data signal) and it will be in the host data forever. Whenever, there are disputes regarding ownership of the data, the authorized owner can extract the watermar to establish his/her ownership. Therefore, the watermaring can be effective solution for copyright protection of the digital multimedia data. An effective audio watermaring algorithm must satisfy the following properties: (i) Imperceptibility: after embedding, the quality of the host audio signal should not be affected. According to IFPI [1], SNR of the watermared audio signal should be greater than 20 db. (ii) Capacity: the amount of information that can be embedded into the audio signal per unit of time. The capacity should be more than 20 bit per second (bps). (iii) Robustness: ability to extract the watermar from a watermared audio signal after signal processing attacs.(iv)security: the watermar can be extracted only by authorized users. All watermaring algorithms mae suitable trade-off among these conflict properties. Many algorithms in audio watermaring have been presented to address these challenges. All audio watermaring algorithms can be classified into two categories: time domain [2], [3], [4], [5] and transform domain [6], [7], [8]. In the survey, it is found that transform domain methods are more robust than time domain methods. Embedding watermar into audio signals is more difficult than the image or video signals [9], due to higher sensitivity of Human Auditory System (HAS) compared to that of Human Visual System (HVS). Many audio watermaring algorithms have been implemented in wavelet domain [1], [7] as the wavelet coefficients contain the multiple spectrums of multiple band frequencies [10]. A limitation of wavelet based approach is that the basis functions are fixed and not necessarily match all real signals. For this reason, many researchers use Empirical Mode Decomposition (EMD) [11] to develop new audio watermaring techniques. EMD is a method which can brea down a signal into a number of IMFs and a residual. Then, signal can be easily reconstructed by adding the IMFs and the residual. The main advantage of EMD is that, there is no priori choice of basis functions and it is fully data-driven method. Thus, EMD is adaptive and highly efficient for perceptual transparency (inaudibility) and robustness. In the literature, there are few audio watermaring algorithms have been found based on EMD [12], [13],

2 769 [14]. In [12], a blind audio watermaring method based on EMD and psychoacoustic model is proposed. The watermar bits are embedded by modifying the final residual of the audio signal. This method is not robust to many attacs. A. N. K. Zaman et al [13] have presented an audio watermaring algorithm using EMD and Hilbert Transformation (HT), where watermar is embedded into the IMF containing highest energy. This method is not robust to attacs as highest energy IMF can be a high frequency mode. In [14], a blind audio watermaring scheme is presented based on EMD. A binary watermar is embedded along with the synchronization codes into the extremes of the last IMF. This method is robust under attacs and preserving the quality of the watermared audio signal. The capacity of this method is which is very low. In this paper, a new blind watermaring algorithm based on EMD and quantization is proposed. In the embedding process, the extremes of the first IMF of the audio signal have used as number of extremes decreases with the order of mode [15]. Here, maximum number of extremes have used for embedding the watermar. The watermar bit is embedded by modifying the norm of the extremes using user defined quantization parameter. The rest of the paper is organized as follows. Section 2 gives an overview of the empirical mode decomposition process. The Arnold technique is described in Section 3.The proposed algorithm is presented in Section 4. Section 5 illustrates the experimental results. The performance of the proposed algorithm is analyzed in Section 6. And finally the concluding remars are given in section EMD PROCESS A detailed mathematically formulated description of the empirical mode decomposition (EMD) has been found in [11]. By using the EMD process, any complex multicomponent signal is decomposed into a finite number of intrinsic mode functions (IMFs), each of which contains only one oscillatory mode at any time instances. An intrinsic mode function is a function that satisfies two conditions: 1) in the whole data set, the number of extremes and the number of zero crossing must either equal or differ at most by one; and 2) at any point, the envelop defined by the local maxima and the envelop defined by the local minima is zero. An algorithm, nown as the EMD decomposes a signal into finitely many IMFs. The EMD which extracts all IMFs from a signal s 0 (t), can be described as follows [11] Step 1: The first IMF component g 1 (t) from the signal s 0 (t). a) Determine all the local extremes of the signal s 0 (t). b) Construct upper envelope for all the local maxima and lower envelope for all the local minima using cubic splines. c) Compute the mean of upper and lower envelope values and denoted the mean as m 1 (t). d) Evaluate the difference between the signal s 0 (t) and m 1 (t) as below h 1 (t)= s 0 (t) - m 1 (t) e) If h 1 (t) is not an IMF then replace the signal s 0 (t) by h 1 (t) and repeat the above strategy for h 1 (t) [from steps 1(a) to 1(d) ] f) The shifting process stops until the resulting difference between the mean of two envelope values in Step 1(c) and the signal s 0 (t) in Step 1(a) is an IMF. The first IMF of the original signal s 0 (t) is denoted as g 1 (t). Step 2: Let s 1 (t) = s 0 (t) - g 1 (t). If g 1 (t) becomes a monotonic function, then stop the decomposition process, Otherwise, replace s 0 (t) by s 1 (t) and repeat Step 1 to find next IMF from s 1 (t). From a given signal s 0 (t), obtain set of signals g (t) and s (t), such that g (t) is obtained from s -1 (t), and s (t) = s -1 (t) - g (t) for 1 r, and s r (t) does not have any IMFs. The s r (t) is called the residual signal. The superposition of all IMFs and the residual reconstruct the original signal s 0 (t). The empirical mode decomposition of the original signal s 0 (t) is define as follows: s t t r 0 = g + sr (1) = 1 The attractive characteristics of IMFs [14] are: (i) IMFs are nearly orthogonal to each other and all have nearly zero mean. (ii) The number of extremes is reduced when going from lower order IMF to the higher order IMF. (iii) The IMFs are completely described by their extremes and it could be recovered from them. (iv) The higher order IMFs are the low frequency components and vice versa. (v) The EMD process is totally adaptive as it is fully data-driven. These important characteristics motivated us to select the EMD process and embed watermar into the extremes of the first IMF. 3. ARNOLD TRANSFORM It is very common to scramble an image for maing it into a meaningless image in the domain of information

3 770 security. The objective is that an original image falls into confusion based on the inverse image transformation. So, the actual content is concealed. The Arnold transformation technique [16], [17] which is a 2-D transformation is used to transform one matrix into another. This can be viewed as a discrete chaotic system [18]. Without loss of generality, let H be an image which is a matrix of size N N. The element (i, j) of the matrix H can be transformed to another position (i, j ) using the transformation rule as below. i = A i modn (2) j j Where A = p s r q Here, A is called Arnold matrix. The element of the matrix A, i.e., p, q, r, s are positive integers and the determinant of A must be equal to unity, i.e., A = 1. Due to its simplicity, it is used for pre-processing of the watermar image in digital watermaring. The main idea is to distribute the pixels randomly within the image. The Arnold matrix is used as secret ey in the proposed watermaring algorithm. It is used to improve the security of the watermar even the watermaring scheme is nown. According to the periodicity of Arnold transformation technique, the actual image can be restored after several iterations. 4. THE PROPOSED ALGORITHM The idea of the proposed watermaring algorithm is to embed a binary watermar image into the host audio signal (X) of L samples. The host audio is partitioned into non-overlapping segments having M samples. The number of segments is equal to the size of the watermar. The size of the binary watermar image (W) is P P. i.e., L = M (P P). The watermar is scrambled using Arnold transformation before embedding it into the host audio. In the embedding process, each audio segment is decomposed into IMFs using EMD process described in section 2. The watermar bit is embedded into each segment by quantizing the norm of the extremer of first IMF using user defined quantization parameter. The value of is adapted experimentally. In the extraction process, same value is applied to extract the watermar. The embedding and extraction process of the proposed algorithm are described as follows. 4.1 Embedding Process The bloc diagram of the proposed algorithm is shown in Fig. 1. The algorithmic steps are given below. Step1: The audio signal X is partitioned into nonoverlapping segment SG, =1, 2,,(P P), each of size M. Step2: The watermar image W is scrambled by Arnold transformation technique to obtain scrambled watermar W s. Step3: The EMD process is applied on each segment SG and computes the norm of the extremes of the first IMF as below. Let e = {e 1, e 2,, e q } be the set extremes of the first IMF. The norm is obtained as follows: R = e = q i= 1 e i 2 (3) Step4: The watermar bits are embedded by quantizing the norm using user defined quantization parameter. Let H = R mod Case 1: If watermar bit = 0, then quantize the norm as follows 3 7 R + H if H R = 12 R + H, otherwise 10 Case 2: If watermar bit = 1, then quantize the norm as follows 3 3 R H if H R = 7 R otherwise + H 10 Step5: Then, modify the extremes of the first IMFs using the formula: R e = e (4) R Step6: Compute the modified segment SG w by inverse EMD process Step7: The watermared audio signal X w is obtained from all watermared segments.

4 771 Step4: Then, apply inverse Arnold transformation technique to obtain actual sequence of the extracted watermar image. Fig.1: Watermar Embedding Process 4.2 Extraction Process The bloc diagram of the extraction process of the proposed algorithm is shown in Fig. 2. The algorithmic steps are given below Step1:The watermared audio signal X w (possibly attaced) is partitioned into non-overlapping segment SG w, =1, 2,,P P, each of size M. Step2: EMD process is applied on each segment and the norm of the extremes of first IMF of the segment are computed as follows R = e Now, it is then quantized as follows H = R mod Step3: Finally, watermar bit is computed as follows 0 if H Watermar = 2 1 otherwise Fig.2: Watermar Extraction Process 5. EXPERIMENTAL RESULTS The proposed method is implemented in MATLAB environment. In this section, experimental results are presented to analyze effectiveness of the proposed watermaring algorithm. The results have been enumerated on two host audio signals namely, classical and jazz. The audio signals are 16-bit mono file in the WAVE format and has 44.1 Hz sampling rate. The original binary watermar image of size and its scrambled version are shown in Fig. 3. The host audio signals and its watermared version are shown in Fig. 4 and Fig. 5 respectively. In the proposed algorithm the audio signal is partitioned into non-overlapping segments with 64 samples. In this experiment quantization parameter is set as The effectiveness of the proposed technique is measured in terms of four properties namely, imperceptibility, robustness, payload and security. Due to the embedding of the watermar, the extremes of the IMF signal get modified and hence, the original host signal cannot be retrieved, i.e., proposed watermar method is a lossy method.

5 772 (a) (b) Fig.3: Binary Watermar: (a) Original, (b) Scrambled 5.1 Imperceptibility Test For the imperceptibility test of our proposed algorithm, signal-to-noise ratio (SNR) is used. It is defined as follows SNR = 10* log 10 L i1 L i1 x( i) x 2 x( i) w ( i) 2 ( db) Where x(i) and x w (i) are i th sample of the host audio signal and the watermared audio signal respectively. The higher value of SNR refers to the high quality watermared audio signals. That is perceptually no significant difference between original and watermared audio. This experiment results average SNR of the watermared signals is db which is relatively high compare to IFPI standard. 5.2 Robustness Test The following audio attacs are performed to evaluate the robustness of the algorithm. The audio editing and attacing tools used in the experiment are Adobe Audition 1.0 and Gold Wave The different attacs that are performed in this experiment are as follows: i. Additive white Gaussian noise (AWGN): White Gaussian noise is added to the watermared signal. ii. Amplitude reverse: Reverse the signs of the audio sample amplitude. iii. Re-quantization: The 16-bit watermared audio signals are re-quantized down to 8 bits/sample and then bac to 16 bits/ sample. iv. Cropping: Segments of 500 samples are removed from six different positions of the watermared audio signal and subsequently replaced by the segments of the attaced (with additive white Gaussian noise) watermared audio signal. v. Compression: The MPEG-1 layer-3 compression is applied. The watermared audio signal is compressed at the bit rate 128 bps, 64 bps, 48 bps and then decompressed bac to the WAVE format. (5) To measure the robustness of the algorithm under the attacs, bit error rate (BER) and normalized crosscorrelation (NC) are used. Normalized cross-correlation (NC) is used to compute the similarity between the original watermar and the extracted watermar and is defined as follows: W i, jw i, j (6) NC W,W = i i, j j W i, j W i, j Where W and W are the original and extracted watermars respectively, and i, j are positions of the binary watermar image. If NC(W,W ) is close to one, then the similarity between W and W is very high. If NC (W,W ) close to zero, then the similarity between W and W is very low. The bit error rate (BER) is used to compute the watermar detection accuracy after processing of audio attacs. The BER is defined as follows: BER W,W i, j W i, j W i, j (7) = P P The symbol is the XOR operator. Table 1 and Table 2 presented the BER (%), NC values, extracted watermar image from attaced watermared classical and jazz audio respectively. The result shows that the proposed method satisfies acceptable amount of robustness. In Table 3, we have compared the performance of several recent audio watermaring methods against compression, sorted by capacity. The proposed algorithm has the highest capacity. (a) Host classical audio signal

6 773 (b) Host jazz audio signal Fig.4: A plot of the two host audio signals 5.3 Capacity Fig.6: FPE probability versus number of watermar bits Then, the capacity CP (say) is defined as follows. N w CP = bps (8) T s The capacity (also called data payload) of the watermaring algorithm is defined as the number of bits that can be embedded into the audio signal within a unit of time and it is measured in the unit of bps (bits per second). Suppose, the length of host audio signal is T s seconds and the number of watermar bits is N w. Fig.7: FNE probability versus number of watermar bits (a) Watermared classical audio signal For the proposed algorithm, number of watermar bits is 1024 and length of the host audio signal is 1.86 seconds. Therefore, the capacity of the proposed method is bps. This is relatively very high as typical capacity is bps, shown in Table 3.Since capacity of the proposed method is very high compare to other methods, additional information may combined with watermar bits so that original host signal can be recovered after extraction of the watermar, i.e., the proposed method may be further extended to a lossless watermaring method. 5.4 Security (b) Watermared jazz audio signal Fig.5: A plot of the two watermared audio signals To improve the security of the watermar, Arnold transformation technique is applied for preprocessing the watermar before actual embedding into the host audio signal. Without Arnold matrix no one can generate the actual sequence of the watermar image. Hence, Arnold matrix is used as the first Secret Key in this wor. Since, this watermar embedding algorithm depends

7 774 completely on the quantization parameter value, so without the quantization value, it is impossible to extract the watermar. Hence, quantization parameter is used as another Secret Key. Therefore, two secret eys have used in this wor which enhance the security of the watermar. 6. PERFORMANCE ANALYSIS Two types of errors may occur while searching Table 1.BER and NC of extracted watermar image from watermared classical audio signal under different attacs Audio attacs No attacs BER (%) NC AWGN Cropping (128 bps) (64 bps) (48 bps) Extracted watermar the watermar sequence: (i) false positive error (FPE) and (ii) false negative error (FNE). These errors are very significant to evaluate the performance of the watermaring system. It is difficult to design an exact probabilistic model of FPE and FNE. Here, a simplified model based on binomial probability distribution, similar to [18], is adopted. 6.1 False Positive Error The FPE is the probability that an un-watermared audio signal is declared as watermared during extraction process. Less FPE probability implies better watermaring algorithm Let N w be the total number of watermar bits, and m be the number of matching bits. The false positive error probability can be evaluated as P FPE = P( P( w, w ) Th no watermar ) (9) Table 2.BER and NC of extracted watermar image from watermared jazz audio signal under different attacs Audio attacs No attacs BER (%) NC AWGN Cropping (128 bps) (64 bps) (48 bps) Extracted watermar Table 3.Comparison of audio watermaring methods, sorted by data capacity Reference Proposed Bhat et al. [9] J. Li et al. [8] L. Wang et al. [12] B. Y. Leiet al.[7] K. Khaldiet al. [14] Method Requantizati on Amplitud e reverse Requantizati on Amplitud e reverse EMD- QIM SVD- QIM LWT- QR-QIM EMDpsychoac oustic LWT- SVD EMD- QIM Capacity ( bps) Robust-ness to (bps)

8 775 Where Th is an application dependent threshold. In most of the practical application, expected value of P FPE is very low value. BER less than 25% can meet this demand and Th is set to 0.75 N w. Therefore, P FPE may be described as N w P Nw FPE = 2 N m0.75n w m w (10) Nw Where is binomial co-efficient. In this scheme, m N w =1024, hence the P FPE is close to 0. Indeed, putting N w =150 in Eq. (10) gives P FPE = , which means the P FPE is zero. Fig. 6 plots the P FPE for N w ϵ (0 100]. It is shown that P FPE approaches to 0 when N w = False Negative Error The FNE is the probability that a watermared audio signal is declared as un-watermared during extraction process. Less FNE probability implies better watermaring algorithm. The false negative error probability can be evaluated as follows: 0.75 Nw 1 m m P = N P 1 P N w FNE w (11) E E m= 0 m Here, P E may be obtained from BER in eq. (7) under different attacs. In this scheme, N w =1024, hence the P FNE is close to 0. Fig. 7 plots P FNE for N w ϵ (00]. It is shown that the P FNE approaches to 0 when N w is larger than CONCLUSION In this paper, a new blind audio watermaring algorithm based on empirical mode decomposition and quantization is proposed. The proposed method has very higher capacity compared to recent audio watermaring methods and shows acceptable robustness against different audio attacs. The error rates of the proposed method are very low. The proposed scheme is suitable for application lie copyright protection of digital multimedia data. The proposed method is a lossy one and may be extended to a lossless watermaring method. REFERENCES [1] V. B. K., I. Sengupta, and A. Das, An adaptive audio watermaring based on the singular value decomposition in the wavelet domain. Digital Signal Processing, vol. 20, pp , [2] A. Binny and M. Koilauntla, Hiding secret information Using lsb based audio Steganography, in Proc. of IEEE International Conferences on Soft Computing and Machine Intelligence, 2014, pp [3] P. Hu, D. Peng, Z. Yi, and Y. Xiang, Robust time-spread echo watermaring using characteristics of host signals, Electronics Letters, vol. 52, no. 1, 2016, pp [4] W.-N. Lie and L.-C. Chang, Robust and highquality time-domain audio watermaring on lowfrequency amplitude modification, IEEE Transaction on Multimedia, vol. 8, no. 1,2006, pp [5] S. Subbarayan and S. K. Ramanatha, Effective watermaring of digital audio and image using matlab technique, in Proc. of IEEE International Conferences on Vision, 2009, pp [6] P. K. Dhar and T. Shimamura, A blind lwt-based audio watermaring using fast walshhadamard transform and singular value decomposition, in Proc. of IEEE International Symposium on Circuits and Systems, 2014, pp [7] B. Y. Lei, I. Y. Soon, F. Zhou, Z. Li, and H. Lei, A robust Audio watermaring scheme based on lifting wavelet Transform and singular value decomposition, Signal Processing, vol. 92, 2012, pp [8] J. Li and T. Wu, Robust audio watermaring scheme via qim of correlation coefficients using lwt and qr decomposition, in Proc. of IEEE International Conferences on Informative and Cybernetics for Computational Social Systems, 2015, pp [9] V. B. K., I. Sengupta, and A. Das, An audio watermaring scheme using singular value decomposition and dither-modulation quantization, Multimed Tools Appl, vol. 52, 2011, pp [10] S. P. Mehobroba and T. V. Divya, Secret Audio Watermaring using Empirical Mode Decomposition and Chaotic Map, in Proc. of IEEE International Conferencee on Power, Instrumentation, Control and Computing, 2015, pp [11] N. E. Huang et al., The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis, Proc. R. Soc. Lond. A, vol. 454, 1998, pp

9 776 [12] L. Wang, S. Emmanuel and M. S. Kananalli, EMD and psychoacoustic model based watermaring for audio, IEEE International Conference on Multimedia Expo, 2010, pp [13] A. N. K. Zamam, K. M. I. Khalilullah, M.W. Islam, M. and M. K. I. Molla, A Robust Digital Audio Watermaring Algorithm Using Empirical Mode Decomposition,, in proc. of IEEE CCECE, 2010, pp [14] K. Khaldi and A. O. Boudraa, Audio Watermaring Via EMD, IEEE Transactions on Audio, Speech, and Language Processing, vol. 21, no , pp [15] M. Telespan and B. W. Schuller, Audio Watermaring Based On Empirical Mode Decomposition and Beat Detection, in Proc. of IEEE International Conference on Acoutics, Speech and Signal Processing, 2016, pp [16] J. Li and F. Wu, robust watermaring for text images based on Arnold scrambling and dwt-dft, in Proc. of IEEE International Conferences on Mechatronic Science, Electric Engineering and Computer, 2013, pp [17] X. C. Wen and C. Jing. An anti-statistical analysis LSB Steganography incorporating extended cat mapping. Berlin Heidelberg, 12(5): , [18] M. Fan and H. Wang, Chaos-based discrete fractional Sine transform domain audio watermaring scheme. Comp. Electrical Eng. 35(3), , 2009