Audio Watermarking Based on Multiple Echoes Hiding for FM Radio

Transcription

1 INTERSPEECH 2014 Audio Watermarking Based on Multiple Echoes Hiding for FM Radio Xuejun Zhang, Xiang Xie Beijing Institute of Technology Abstract An audio watermarking system based on multiple echoes hiding is designed for frequency modulation (FM) broadcasting. The embedded watermarking information of 16 bits can be recovered from the receiver by recording the broadcast for any 1 second. On the premise of guaranteeing the high imperceptibility and robustness, the multiple echoes scheme we proposed has a higher capacity. The proposed system is tested on a semi-physical platform with the real channel transmission and the recovery rate reaches higher than 98%. Some attacks, such as white noise, filtering, re-sampling, adding an echo and re-quantization, are used to test the robustness. The subjective test result shows that the system has a good imperceptibility. Index Terms: echo hiding, audio watermarking, FM radio 1. Introduction In recent years, there have been increasing researches interesting in delivering information by audio for its convenience and low cost. Audio watermarking is a hopeful technique to solve this problem. From the technical aspect, audio watermarking aims to embed proprietary data into the audio object. When necessary, the owners can extract these data. For example, the bank card number can be embedded into the audio, and we can accomplish electronic payment by just playing the watermarked audio. The available audio watermarking schemes can be classified into the following four categories: phase coding [1], spread spectrum modulation [2], patchwork [3], and echo hiding [4-6]. Among them, echo hiding algorithm has natural advantages in imperceptibility. In addition, the advantages are also reflected in low complexity, no additive noise, blind detection, low requirements for synchronization act which makes it very attractive to researchers. An effective and practical audio watermarking method should possess three important traits: imperceptibility, robustness, and information capacity. The sound in a real life we heard is a sound with multiple echoes for sound in a real room including multiple reflections from the walls and other objects in the room [7]. That is multiple echoes can provide a more natural sound quality than only a single echo. So the echo-hiding watermarking methods in [4] [6] which utilized a single echo has a relatively large damage on audio quality. In order to guarantee better imperceptibility, the multiple echoes system is proposed by Xulai Cao [8]. A challenging task for watermarking system based on multiple echoes scheme is the trade-off between robustness and capacity. Especially in the application of delivering information, the capacity should be paid more attention. In this paper, the theoretical upper limit of the echoes number in multiple echoes system is pointed out by our analysis. Based on this, we designed an echo hiding system with multiple echoes which can be used in FM radio. The multiple echoes scheme we proposed embeds a symbol which includes 16 bits binary number into 1 segment. Compared with the multiple echoes scheme in [8] which embedded 1 bit binary number into 1 segment, our algorithm achieves higher capacity. Our system is tested on a semi-physical platform with the real channel transmission and shows a better performance. The paper is organized as follows. Section 2 briefly introduces the echo hiding schemes. An echo hiding system is proposed in Section 3. Section 4 gives the experiment results. Conclusions are given in section Echo hiding scheme The basic idea of the echo hiding scheme is to add an echo which is inaudible by human ear to the original audio. It takes advantage of the temporal masking effect of human ear. The traditional echo kernel is expressed as h() =() ( ) (1) where h() is the echo kernel, δ(n) represents the unit-impulse function, and α denotes the amplitude of the echo. The delay d of the echo is decided by the values of the embedded bit. Let x(n) represent the host signal, the watermarked signal can be expressed as () =() h() =() ( ) (2) The specific process of the encoding is described as follows Step 1: Divide the host signal x(n) into several consecutive segments which contain the same number of samples. Step 2: Embed 1 bit into each segment by the different delay: represents bit 0, and represents bit 1. Step 3: Combine the segments with hiding message into a continuous audio signal. We use the power cepstrum of the watermarked signal to extract the watermark. The power cepstrum of the signal is defined as () = () =() h() (3) Then take (2) into (3), we can get [] = () () ( ) ( ) (4) where () is the power cepstrum of the host signal. From (4), we can see the power cepstrum of the watermarked signal appear a peak at the delay offset d. Then, the embedded information can be extracted by checking the amplitude at or. 3. Proposed echo hiding system The diagram of the whole system is demonstrated in Fig.1. The system mainly consists of three parts: the watermark embedded part, FM transmission part, and the watermark extraction part. Copyright 2014 ISCA September 2014, Singapore

2 Original watermark Extracted watermark Figure 1: the block diagram of the system The theoretical upper limit of the echoes number Differing from multiple echoes scheme in [8], the multiple echoes algorithm we proposed embeds a symbol which includes 16 bits binary number instead of 1 bit binary number into 1 segment. The multiple echoes kernel is described as h() =() ( ) (5) where J (J 2) denotes the number of echoes represents the delay of the each echo. The different delays represent different binary number. It is vital to guarantee there is a peak in each delay sample in the power cepstrum of the watermarked signal so that we can extract the watermark exactly. With the increasing of the echoes number, the interference between each echo becomes stronger. So the maximum number of echoes is needed to be anatomized so that we can extract the watermark accurately. Take (5) to (3), we can get () () () { [( ) ( )] } ( ( )( ( )) [ ( )( )]} [ ( ) ( )] (6) We can see that there are two kinds of peaks in the power cepstrum. First is the peaks in the delay offset. with the amplitude 2, which are needed to extract the hiding information, we call them delay peaks. The second is the peaks in (1 < ) with the amplitude, called interference peaks. The number of the interference peaks is. Watermark encoder Extraction Original audio Embedded algorithm Record FM emitter channel FM receiver In the decoder, the watermark is extracted by seeking the location of the peaks positions. That is, we can gain high robustness only by insuring that the amplitude of delay peaks is always larger than the amplitude of interference peaks. Otherwise, the location of the peaks positions we seek is interference peaks position (1 < ) rather than the delay peak position. Assume a limiting case that all the interference peaks appear in the same position, the amplitude will reach the maximum. If the maximized amplitude of the interference peaks is equal to the amplitude of the delay peak, we can t differentiate which are the delay peaks in the power cepstrum. The value of in this situation is the upper limit of the echoes number. That is ( ) Applying (3) to (8) J = = (7) From (8), we can see that the theoretical upper limit of the echoes number is related to the amplitude of the echo. J =4 when is chosen as 0.4. Gruhl et al. have proved that all signals showed acceptable accuracy with a decay rate between 0.3 and 0.4, and even those with exceptional hearing have difficulty resolving the echoes [10] Embedding algorithm We know from the analysis above that the echo kernel can be described as follows h() =() ( ) ( ) (9) Unlike the traditional algorithm, a symbol consists of 16bits binary number is embedded into 1 segment integrally. Set a symbol={a 0, a 1, a 2, a 3, a 4, a 5, a 6, a 7, a 8, a 9, a 10, a 11, a 12, a 13, a 14, a 15 } and 1 segment is (). The symbol is first changed to a 4 4 matrix, then transforms four binary in each column into the corresponding hexadecimal, we can obtain four hexadecimal number [b1, b2, b3, b4], where [0,15], l [1,4]. The next task is to embed each hexadecimal number into each echo. It is formulated as follows. Step 1: Calculate the delay of the first echo =M (10) where is the minimum value of. ( F 1) is the minimum interval of, we call it resolution. Table.1 shows the relationship between and delay. Table 1. Relationship between and. b 1 d 1 0 M 1 MF 15 M15F Step 2: Work out the rest of the delays = (11) = (12) = (13) From (10)-(13), we can know that is always bigger than, and the minimum interval between them is ( 1), we call it bracer. This algorithm ensures that the delays of four-way echoes signals is increasing, that is d 4 >d 3 >d 2 >d 1. Thus, the peaks of the power cepstrum in the delays are easily distinguished, which is shown in Fig.2. Then the watermark information is extracted by the value of corresponding. Step 3: The watermarked signal is the temporal convolution of () and h() () = () h() (14) Fig.3 illustrates how watermark is embedded into a host signal. (8) 1387

3 3.3. Parameter analysis Figure 2: The power cepstrum of 1 second watermarked signal in computer simulation. x(n) * δ(n) αδ(n ) αδ(n ) αδ(n ) αδ(n ) h(n) 1 x(n) α x(n-d 1 ) α x(n-d 2 ) α x(n-d 3 ) α x(n-d 4 ) y(n) d 1 d 2 d 3 d 4 * Figure 3: Schemetic explanation of how watermark is embedded. The scope of interference peak position M M60F3K K 45F3K The scope of delay peak position Figure 4: The position of the two kinds of peaks. t Echo 1 Echo 2 Echo 3 Echo 4 Watermarked signal t Parameter M Form (10) to (13) and the scope of, we know that the maximum of is 60 3, which is also the maximum of all the delays. That is the position of the delay peak is in the scope of [M, M60F3K]. From analysis above, we know that the position of interference peaks are,,,,,. From (10) to (13), we can calculate that =3 ( 2) [K, 15F K] (15) =3 ( 4) [2K,30F2K] (16) =3 ( 6) [3K, 45F 3K] (17) =3 ( 2) [K,15FK] (18) =3 ( 4) [2K, 30F 2K] (19) =3 ( 2) [K,15FK] (20) Form (15) to (20), we know that the position of the interference peak is in the scope of [K,45F3K]. Fig.4 showed that the peaks in the intersection region [M, 45F3K] can t be judged which kind of peak it belongs to. In order to eliminate the cross interference between the delay peak and the interference peak, the parameter M is ascertained as 45F3K1. In this way, there is no intersection region between the two kinds of peaks Parameter F and K The accurate rate ( ) is defined as: = 100% (21) where a symbol contains 16 bits information. Delay time will affect detecting performance and the subjective quality of echoed audio signal. It is difficult to detect the position of echo for too short delay time, and as a result, the is low. The energy of echoes will be less and the subjective quality will be bad for long delay time. According to the methods in [9], delay time belonging to 2ms~6ms is acceptable both for detecting performance and the subjective quality. There are 64 delay positions in our system and the maximum delay time of ensuring the imperceptibility is 6ms. Thus, the maximum temporary resolution is about 0.1ms. In this way, the Min Samples is 10kHz. Here, the host audio we chose is sampled at 44.1 khz sampling rate with 16-bit quantization. From (10) to (13), we know that the minimum delay time of is M which is also the minimum delay time of all echoes and the delay time of is M60F3K which is the maximum delay time. So it needs to meet the two equations M 60F 3K f (22) M f t (23) where is the maximum delay time, t is the minimum delay time and f is the sampling frequency. Applying (22),(23) to (24), the following conclusion is straightforward 17.5F K 44 15F K 29 (24) It is proved that the super parameters F =2,K [1,9]. 4. Experiments In order to estimate the performance of the system, we first test the capacity of the proposed echo hiding scheme. Some attacks are also used to measure the robustness of the algorithm. The subjective test result showed that the embedded 1388

4 algorithm has a good imperceptibility. Then the echo hiding scheme is used on FM broadcast system with a semi-physical platform to estimate the performance of the whole system. Twenty uncompressed audio clips are used as host audio and the parameter F=9. These 20 clips were sampled at 44.1 khz sampling rate with 16-bit quantization and belong to 5 different audio groups as illustrated in Table 2. Table 2. signal used in simulation. Host Signal Genres S 01 -S 04 Chinese Rock S 05 -S 08 Disc Jockey S 09 -S 12 Eastern Classic Songs S 13 -S 16 Foreign Songs S 17 -S 20 Pop Songs 4.1. The performance of echo hiding scheme Transmission rate test The results of the transmission rate comparison between the proposed algorithm and the scheme in [8] are showed in table.3. A symbol contains 16 bits is embedded into one segment in the algorithm we proposed, so it is feasible to control the transmission rate by varying the length of segment. For example, when the length of the segment is changed to 0.5 second, which means we embed a symbol per 0.5 second, thus the transmission rate is 32 bits/s. Obviously, proposed kernel shows a better extraction performance than conventional echo methods especially when we increase the transmission rate. Table 3. Result of transmission rate test. Transmission rate (%) (bits/s) Proposed method in [8] Robustness test In daily life, the audio files often need to carry out several necessary signal processing operations. So it is necessary to test the performance after the signal processing operations which is illustrated in table.4. Table 4. Detail of attack. attack Detailed Description high-pass filtering >5khz Low-pass filtering <5.5khz Resampling 44.1k 22.05k 44.1k Requantization 16bits 8bits 16bits Noise Addition SNR=20db Additional Echo Delay=8ms, α=0.1 Table 5. Result of attack test. attack (%) Proposed method in [8] high-pass filtering Low-pass filtering Resampling Requantization Noise Addition Additional Echo From table.5, we can see the proposed kernel have a better performance to signal processing operations comparing with the conventional echo hiding Subjective test We adopt AB test method to test the quality of watermarked audio. In this test, each of the participants chose the partial audio from the pair of host audio and the watermarked audio. 10 subjects are required to listen 20 audio clip pairs respectively. The result is shown in Table.6. Table 6. Result of subjective test. prefer Watermarked audio number rate The almost half-to-half rate indicates that our watermarking process does not obvious damage on the hearing quality of host audio Performance of the whole system The FM, which is band-limited to below 16kHz, meets the minimum sampling rate. The proposed system is utilized on and the performance showed in Table.7. With the increase of the recording distance, SNR will decrease which inevitable leads to low. As we can see in Table.6, the system has a good performance with the transmission rate of 16 bits/s and the recording distance can reach 4 meters with 90.05%. Table 7. Test result of the whole system. Distance (%) 10cm cm meters meters meters meters meters Conclusion In this paper, we designed an audio watermarking system based on the multiple echoes kernel. Theoretical upper limit of the echoes number and detection gain is derived. Then we describe how we embed watermark into the host audio in detail. Finally, it is used in FM broadcast system to increase the transmission distance and we can recover the embedded information by recording the FM broadcast. On the premise of guaranteeing the high imperceptibility and robustness, the multiple echoes scheme we proposed has a higher capacity. The proposed system is tested on a semi-physical platform with the real channel transmission and experimental results show the system has a good performance. How to differentiate the multiple embedding positions accurately and decrease the computational complexity are the coming problems needed to be solved. 6. Acknowledgements This work is supported by National Nature Science Foundation of China (NSFC): Grant No and No

5 7. References [1] M. Arnold, P. G. Baum, and W. Voeßing, A phase modulation audio watermarking technique, in Proc. Lecture Notes in Computer Science, 2009, vol. 5806, pp [2] A. Valizadeh and Z. J. Wang, Correlation-and-bit-aware spread spectrum embedding for data hiding, IEEE Trans. Inf. Forensics Security, vol. 6, no. 2, pp , Jun [3] H. Kang, K. Yamaguchi, B. Kurkoski, K. Yamaguchi, and K. Kobayashi, Full-index-embedding patchwork algorithm for audio watermarking, IEICE Trans. Inf. Syst., vol. E91-D, no. 11, pp , Nov [4] O. T.-C. Chen and W.-C. Wu, Highly robust, secure, and perceptual quality echo hiding scheme, IEEE Trans. Audio, Speech, Language Process., vol. 16, no. 3, pp , Mar [5] Hyoung Joongn Kim, Yong Hee Choi, A Novel Echo-Hiding Scheme with Backward and Forward Kernels, IEEE transactions on circuits and systems for video technology, 13(8), pp , [6] O. T.-C. Chen and W.-C. Wu, Highly robust, secure, and perceptualquality echo hiding scheme, IEEE Trans. Audio, Speech, Language Process., vol. 16, no. 3, pp , Mar [7] B.-S. Ko, R. Nishimura, and Y. Suzuki, Time-spread echo method for digital audio watermarking, IEEE Trans.Multimedia, vol. 7, no. 2, pp , Apr [8] Xulai Cao, Linghua ZHANG, Researches on Echo Kernels of Audio Digital Watermarking Technology Based on Echo Hiding, International Conference on Wireless Communications and Signal Processing (WCSP 2011), p.5pp., 2011 [9] D. Gruhl and W. Bender, Echo hiding, Proceeding of Information Hiding Workshop, Cambridge, U.K., pp , [10] LI LI, YA-QI SONG, EXPERIMENTAL RESEARCH ON PARAMETER SELECTION OF ECHO HIDING IN VOICE, Eighth International Conference on Machine Learning and Cybernetics (ICMLC), p ,