Audio histogram shape watermarking method capable of resisting synchronous attack

Abstract

The invention relates to the technical field of digital watermarking, in particular to an anti-synchronous attack audio histogram shape watermarking method, which comprises the following steps of: dividing a pre-selected audio sample according to the number of pre-allocated bins to obtain a first time domain histogram; carrying out segmented DWT (Discrete Wavelet Transform) on the first time domain histogram, and carrying out inverse DWT to obtain an embedded watermark audio; obtaining a second time domain histogram corresponding to the watermark-embedded audio, and carrying out the segmented DWT conversion on the second time domain histogram to obtain a second time domain histogram of a low-frequency domain; and performing information extraction on the second time domain histogram of the low-frequency domain to obtain an embedded watermark. According to the method, various stretching attacks can be resisted, the robustness is higher after synchronous attacks such as MP3 compression, resampling and stretching or conventional signal operation is carried out, and watermarks can be effectively extracted.

Description

Technical Field

[0001] The present invention relates to the technical field of digital watermarking, in particular to an audio histogram shape watermarking method resistant to synchronization attacks. Background Art

[0002] In recent years, novel digital watermarking algorithms have been continuously proposed, gradually filling the gaps in the digital watermarking field. However, synchronization attacks remain a difficult problem to solve in traditional audio watermarking. Many scholars have studied audio characteristics and discovered that the shape and mean of its histogram in the time domain are invariant to synchronization attacks. By leveraging this property, a technique has been proposed to embed watermarks by modifying the histogram shape. However, existing algorithms that use histogram shape modification for watermark embedding lack detailed consideration of the histogram shape design, resulting in a large amount of wasted embedding space, ultimately limiting the improvement of embedding capacity.

[0003] Robustness to synchronization attacks and MP3 compression is an important part of the performance of audio robust watermarking algorithms. Currently, most audio watermarking algorithms can only target one of these attacks.

[0004] An earlier robust audio histogram watermarking algorithm, proposed by Xiang et al., was based on time-domain statistical features to resist synchronization attacks. This algorithm primarily exploits the theoretically invariant relationship between the number of audio samples of different amplitudes before and after linear scaling. The watermark is embedded by modifying the shape ratios between histograms using three statistical histogram sample bins (hereinafter referred to as "bins"). Subsequently, Xiang et al. proposed extending this time-domain invariance to the frequency domain using a piecewise DWT approach, addressing the previous algorithm's poor robustness to Gaussian and compression attacks. Hu et al. proposed a high-capacity robust watermarking algorithm based on the DCT transform domain, embedding the watermark in the low-frequency band of the DCT of the audio frame. Liang et al. improved on the audio histogram by introducing a high-order difference statistical model. The histogram can be treated as a robust feature and shifted to embed the watermark sequence. Using the inverse operation of the histogram shift, the original audio file can be losslessly restored. Research on histogram-based robust watermarking is currently ongoing. Based on the above existing histogram correlation methods, they cannot achieve a good balance between high capacity and high robustness.

[0005] In the current research progress, traditional audio robust watermarking technology still has strong development potential and application scenarios. At the same time, with the advancement of communication technology, the requirements for information capacity and security are gradually increasing. Summary of the Invention

[0006] The purpose of the present invention is to propose a watermark technology based on statistical histograms that has strong robustness against common attacks and synchronous attacks and has a high embedding capacity. By reasonably modifying the shape of the low-frequency domain statistical histogram of the audio carrier, the embedding capacity is greatly improved while ensuring the audio auditory quality and robustness, and solves the problem that the traditional histogram watermark method is difficult to embed when the number of samples is insufficient or the amount of watermark data is too large. Therefore, the idea embodied in the present invention has a great inspirational effect on watermark technology in scenarios where both capacity and robustness are required. The present invention plays a significant role in the integrity authentication of digital media, and this role comes from its ability to effectively resist various attacks, such as MP3 compression, time domain stretching, filtering, Gaussian noise, and volume changes.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] The audio histogram shape watermark method resistant to synchronization attacks includes:

[0009] Dividing the pre-selected audio samples according to the number of pre-allocated bins to obtain a first time domain histogram;

[0010] Performing a piecewise DWT transform and an inverse DWT transform on the first time domain histogram to obtain an embedded watermark audio;

[0011] Obtaining a second time domain histogram corresponding to the watermarked audio, and performing the piecewise DWT transform on the second time domain histogram to obtain a second time domain histogram in the low frequency domain;

[0012] Information is extracted from the second time domain histogram in the low frequency domain to obtain an embedded watermark.

[0013] Optionally, acquiring the first time domain histogram includes:

[0014] Preprocess the initial audio and calculate the absolute mean of the audio;

[0015] Selecting an embedding range, i.e., the audio sample, based on the absolute mean of the audio and a preset constant value λ;

[0016] The number of bins is allocated according to the number of watermark bits to be embedded; 3 bins are used to embed 2 bits;

[0017] The samples are divided according to the number of bins to obtain the time domain histogram.

[0018] Optionally, obtaining the watermark-embedded audio includes:

[0019] Performing a piecewise DWT transform on the first time domain histogram, recording the breakpoints between each discontinuous bin, and obtaining a first time domain histogram in the low-frequency domain;

[0020] Performing information embedding on the first time domain histogram in the low frequency domain, modifying the shape of the bin, and obtaining a modified first time domain histogram;

[0021] The modified first time domain histogram is combined with the recorded breakpoints to perform an inverse secondary DWT transformation to obtain the embedded watermark audio.

[0022] Optionally, performing a piecewise DWT transform on the first time domain histogram includes:

[0023] The audio samples [-λA, λA] are considered to have been assigned to bins of equal width; a two-level DWT transform is performed on the samples in the first bin, that is, the range [-λA, -λA +M), where M represents the width of each bin; a two-level DWT transform is performed on the samples in the second bin, that is, the range [-λA + M, -λA +2M), and each subsequent bin interval is increased by M on the previous bin range, and so on to obtain the interval of each bin; a DWT transform is performed on the piecewise function within each M range separately to obtain the first time domain histogram of the low frequency domain.

[0024] Optionally, methods for modifying the shape of a bin include:

[0025] The bin intervals are grouped into groups of 3, and each group is used to embed 2 bits of information. Different embedding strategies are used for the four different 2-bit information units in the watermark sequence: 00, 01, 10, and 11, where:

[0026] When the current embedded information is 00, modify the bin shape to low-medium-high;

[0027] When the current embedding information is 01, modify the shape of the middle bin to be the lowest;

[0028] When the current embedding information is 10, modify the shape of the middle bin to be the highest;

[0029] When the current embedding information is 11, modify the shape to high-medium-low;

[0030] Low, medium, and high represent the size relationship of three consecutive bin intervals in terms of sample quantity.

[0031] Optionally, obtaining a second time domain histogram corresponding to the watermarked audio includes:

[0032] Preprocessing the watermarked audio and calculating the absolute mean of the watermarked audio;

[0033] Selecting an extraction range based on the absolute mean of the embedded watermark audio and a preset constant value λ;

[0034] A binning operation is performed on the samples within the extraction range to obtain the second time domain histogram.

[0035] Optionally, extracting information from the second time domain histogram in the low frequency domain includes:

[0036] Extract watermark information with 3 bins as a group, and assume that the number of samples in each bin is a, b, and c respectively:

[0037] if , then the current bin group watermark information is 11;

[0038] if , then the current bin group watermark information is 00;

[0039] if , then the current bin group watermark information is "10";

[0040] if , then the current bin group watermark information is "01".

[0041] Optionally, the absolute mean of the audio is:

[0042] ;

[0043] Among them, A is the absolute mean of the audio, N is the length of the original audio, is the i-th sample in each bin.

[0044] The beneficial effects of the present invention are:

[0045] This invention, based on histogram shape, significantly improves the robustness of the algorithm by transferring the histogram's time-domain features to the frequency domain using a segmented DWT. Compared to previous algorithms, this invention leverages the diversity of histogram shapes to achieve higher-capacity embedding with the same number of bins, thereby resolving the difficulty of embedding watermarks in certain voice files due to the small number of samples. This has significant implications for research on robust watermarking using histogram modification.

[0046] The present invention utilizes the robustness of the low-frequency domain coefficients after DWT transformation against attacks such as compression and Gaussian noise. Through the segmented DWT method, the time domain stretching invariance is extended to the frequency domain while retaining the histogram shape, which not only improves resistance to conventional attacks but also retains robustness against stretching attacks.

[0047] The present invention can effectively extract watermark information under different signal processing conditions, such as Gaussian noise and MP3 compressed sound attacks, which meets the requirements of daily digital forensics and digital authentication. At the same time, the algorithm can be applied to different audios and achieves good results on different audios. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 Schematic diagram of a flow chart of a robust watermark embedding algorithm for resisting synchronization attacks using audio histogram shape according to an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of a watermark extraction process according to an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of the histogram shape that needs to be modified for different embedded information according to an embodiment of the present invention;

[0052] Figure 4 Schematic diagram of time domain histogram segmented DWT according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] This embodiment proposes an audio histogram shape watermarking method that is resistant to synchronization attacks, including two parts: robust watermark embedding and robust watermark extraction;

[0056] Robust watermark embedding part: divide the pre-selected audio samples according to the number of pre-assigned bins and obtain the first time domain histogram;

[0057] Performing a piecewise DWT transform and an inverse DWT transform on the first time domain histogram to obtain an embedded watermark audio;

[0058] Robust watermark extraction part: obtaining a second time domain histogram corresponding to the watermarked audio, and performing the piecewise DWT transform on the second time domain histogram to obtain a second time domain histogram in the low frequency domain;

[0059] Information is extracted from the second time domain histogram in the low frequency domain to obtain an embedded watermark.

[0060] Further, obtaining the first time domain histogram includes:

[0061] Preprocess the initial audio and calculate the absolute mean of the audio;

[0062] Based on the absolute mean of the audio and λ, an embedding range, i.e., the audio samples, is selected. λ is a fixed constant value, usually between [2, 4]. The role of λ is to manually control the audio range to be processed based on [-λA, λA]. The larger λ is, the larger the range we choose. However, it should not be too large, as a large range may cause the range to exceed the audio itself, while a small range may result in insufficient number of samples within the range. Therefore, based on experimental results, a relatively good range is [2.4, 4]. The main purpose of selecting this range is to ensure that the number of samples is sufficient to complete the embedding while removing the lower and higher amplitudes in the audio signal. These extreme parts have relatively few samples and are easy to affect the embedding quality.

[0063] The number of bins is allocated according to the number of watermark bits to be embedded; 3 bins are used to embed 2 bits;

[0064] The samples are divided according to the number of bins to obtain the time domain histogram.

[0065] Furthermore, obtaining the watermarked audio includes:

[0066] Performing a piecewise DWT transform on the first time domain histogram, recording the breakpoints between each discontinuous bin, and obtaining a first time domain histogram in the low-frequency domain;

[0067] Modifying the shape of the bins of the first time domain histogram in the low frequency domain to obtain a modified first time domain histogram;

[0068] The modified first time domain histogram is combined with the recorded breakpoints to perform an inverse secondary DWT transformation to obtain the embedded watermark audio.

[0069] Furthermore, performing a piecewise DWT transform on the first time domain histogram includes:

[0070] The audio samples [-λA, λA] are considered to have been assigned to bins of equal width; a two-level DWT transform is performed on the samples in the first bin, i.e., the range [-λA, -λA +2λA / a), and a two-level DWT transform is performed on the samples in the second bin, i.e., the range [-λA +2λA / a, -λA +4λA / a). This is followed by analogy, and a DWT transform is performed separately on each piecewise function within the range of M to obtain the first time domain histogram of the low-frequency domain; wherein M represents the width of each bin, and a is the number of bins.

[0071] Furthermore, methods for modifying the shape of a bin include:

[0072] To embed the watermark information, the bin intervals are grouped into groups of 3, with each group used to embed 2 bits of information. Different embedding strategies are used for the four different 2-bit information units that may exist in the watermark sequence: 00, 01, 10, and 11.

[0073] When the current embedded information is "00", the bin shape is modified to low-medium-high;

[0074] When the current embedded information is "01", the modified shape is b lowest; medium-low-high, high-low-medium, all are possible;

[0075] When the current embedded information is "10", the modified shape is b highest; medium-high-low, low-high-medium, all are acceptable;

[0076] When the current embedded information is "11", the shape is modified to high-medium-low;

[0077] Among them, "low, medium, and high" represent the size relationship of the three consecutive bin intervals in terms of sample quantity. This adjustment strategy aims to achieve robust embedding of watermark information by regulating the sample quantity and shape while maintaining audio quality.

[0078] The robust watermarking algorithm proposed in this embodiment based on the audio histogram mainly utilizes the two features of histogram shape and average value that remain unchanged after linear scaling modification in the time domain, and expands the time domain features to the frequency domain through the segmented DWT method. Compared with the existing audio robust watermarking algorithm, it has stronger robustness against synchronization attacks while also ensuring strong robustness against conventional attacks. At the same time, the distribution between the rationally designed histogram shapes further improves the capacity of embedded watermark information. This embodiment solves the problem between the robustness and capacity of traditional audio. The ideas adopted in this embodiment have a great inspiration for the implementation of digital watermarking algorithms with insufficient capacity in reality. In the context of the development of information technology in the future, the integrity authentication and copyright issues of digital products are bound to become the focus of research. The histogram shape embedding method proposed in this embodiment greatly improves the embedding capacity while ensuring the unchanged audio quality and robustness, and is a new development direction of digital watermarking application technology.

[0079] Specifically, in this embodiment, the audio histogram shape watermark method against synchronization attack, the watermark embedding process is as follows: Figure 1 The specific steps are as follows:

[0080] 1. Initial audio Perform preprocessing and calculate the absolute mean of the audio signal, which is defined as the normalized sum of the values ​​of the initial audio signal over a duration. Since the absolute mean of the audio signal is always close to zero before and after the attack, we can apply the absolute mean of the audio signal to one of the two invariant features proposed to better estimate the statistical properties of the audio signal. Figure 3 A bin is a group of embedded information. If the number of samples in a group is too small, the embedding quality will be seriously affected. Therefore, the embedding range [-λA, λA] is selected based on the original audio mean A and λ. The range of λ is generally [2.4, 3]. In this embodiment, 2.4 is selected. A larger λ represents a larger selection range. A larger number of samples means stronger robustness, but the signal-to-noise ratio (SNR) will also be lower. Therefore, it is necessary to select an appropriate value to balance SNR and robustness. Specifically:

[0081] 101. Assume binary watermark sequence , original audio Expressed as , where The length of the embedded information, N represents the length of the original audio, Refers to the specific i-th watermark in the sequence W. The absolute average value A is calculated as the sum of the absolute values ​​of the audio samples over their duration:

[0082] ;

[0083] 2. Allocate the number of bins based on the number of watermark bits to be embedded. This embodiment uses 3 bins to embed 2 bits. Under normal circumstances, the number of bins required = (number of watermark bits / 2) 3. Divide the samples according to the number of bins to obtain the time domain histogram H; specifically:

[0084] 201. The embedded watermark is Lw bits, and the algorithm requires (Lw / 2) 3 bins are used to complete the embedding. According to the mean A calculated in step 1, the embedding range [-λA, λA] is selected and divided into (Lw / 2) 3+3 bin parts, each bin width , assign the samples within the range to the corresponding bin according to their amplitude, and obtain the time domain histogram , L=(Lw / 2) 3+3; among them, It refers to the specific i-th bin in the bin set H.

[0085] 202. The extra three bins are due to the small number of samples in the bins near the middle of the histogram, so they are excluded and not used. Furthermore, within the same audio range, a larger bin width results in a larger number of samples, which leads to a more stable and robust histogram shape, but also reduces the SNR. Therefore, a trade-off must be made between robustness, capacity, and SNR. Experiments show that in this embodiment, when the watermark is embedded with 80 bits of information, the SNR is generally between 42dB and 45dB.

[0086] 3. Time domain histogram Perform statistics, and perform secondary DWT transformation on samples within the same bin range. Record the breakpoints between each discontinuous bin to ensure that the subsequent inverse transformation can proceed normally, and obtain the histogram of the low-frequency domain. ; Specifically:

[0087] 301. In step 2, the division of the sample area has been completed. We need to perform a piecewise DWT transform on the histogram. The purpose of doing this is, on the one hand, because the frequency domain has a stronger resistance to Gaussian and compression attacks, and on the other hand, to preserve the histogram shape characteristics and extend the invariance of the time linear scaling of the histogram shape to the DWT domain.

[0088] 302. Assume that the target range [-λA, λA] samples have been assigned to bins of equal width. Perform a two-level DWT transformation on the samples in the first bin, i.e., the range [-λA, -λA +M). Then perform a two-level DWT transformation on the samples in the second bin, i.e., the range [-λA + / M, -λA+2M). The same process is repeated. Each bin interval increases by M on the previous bin range. For example, the third bin is [-λA +2M, -λA +3M), and the fourth bin is [-λA +3M, -λA +4M). Figure 4 As shown in the example, Belongs to [-1.5M, -0.5M), , , belongs to [-0.5M , 0.5M), Belongs to [0.5M, 1.5M), and it is necessary to perform a DWT transformation on each piecewise function within the range of M, and obtain , L=(Lw / 2) 3+3. At the same time, because samples are usually randomly distributed in audio, the samples in the range are not necessarily continuous in the time domain. It is necessary to record the breakpoint coordinates between discontinuous samples during the transformation for use in the subsequent inverse transformation; represents the sample at the beginning of the function, Indicates samples whose amplitude falls at the -0.5M boundary. 、 、 、 Indicates samples whose amplitude falls at the 0.5M boundary. represents samples with amplitudes falling at 0, Represents the frequency domain function of the specific i-th bin in the frequency domain function set obtained after DWT transformation of L bins in the M range.

[0089] 4. According to Figure 3 As shown, every 2 bits of watermark are embedded to modify the shape of 3 bins to meet the expected target, and the histogram after modification is obtained. ; Specifically:

[0090] 401. After completing the DWT transformation, the frequency domain histogram shape is obtained. The shape of each group is modified according to the corresponding watermark information with 3 bins as a group. The specific modified shape is as follows: Figure 3Assume that the number of samples in each bin is a, b, and c, and the thresholds in the algorithm are T=[2, 3], T1=2.5, and T2=3. The change rules are as follows:

[0091] Assuming the current embedded information w(i) is "00", modify the shape to low-medium-high:

[0092] ;

[0093] represents the i-th sample in the first bin before modification, represents the i-th sample in the second bin before modification, is the i-th sample in each bin, The corresponding modified value, M, is the bin width in the time domain histogram. However, due to the secondary DWT transformation, according to the characteristics of the signal from the time domain to the frequency domain, the bin width in the frequency domain will become twice the original, that is, 2M. In the formula, a total of I1 + I2 samples in the bin are changed, and I1 and I2 samples are transferred to bin2 and bin3 respectively. The samples are transferred from a to c to ensure If not satisfied, adjust , transfer I1 samples in a to b, and get the adjusted and ;ensure If not satisfied, adjust , transfer I2 samples in a to c, and get the adjusted and ;in, Indicates the threshold set in advance, T1=2.5.

[0094] 402. Assuming that the current embedded information w(i) is "01", modify the shape to b lowest:

[0095] ;

[0096] Migrate samples from b to a, c to ensure If not satisfied, adjust , put b in samples are transferred to a, The samples are transferred to c, and the adjusted , and ;in, Indicates the threshold set in advance, T2=3.

[0097] 403. Assume that the current embedded information w(i) is "10", and modify the shape to b highest:

[0098] ;

[0099] Migrate samples from a, c to b to ensure If not satisfied, adjust , put a Transfer samples to b and transfer samples to c The samples are transferred to b, and the adjusted , and .

[0100] 404. Assuming the current embedded information w(i) is "00", modify the shape to high-medium-low:

[0101] ;

[0102] in, represents the i-th sample in the 3rd bin before modification;

[0103] Migrate samples from c to a to ensure If not satisfied, adjust , transfer I1 samples in c to b, and get the adjusted and ;ensure If not satisfied, adjust ,Will I2 samples are transferred to a, and the adjusted and .

[0104] 5. Modify the completed histogram Combine the breakpoints recorded in step 3 to perform inverse secondary DWT transformation to obtain the embedded watermark audio ; Specifically:

[0105] 501. After modifying the shapes of each bin group of the histogram, a frequency domain histogram Hw of the embedded information is obtained. Combined with the breakpoints and high and low frequency coefficients between each discontinuous sample recorded in step 3, the modified frequency domain histogram Hw is subjected to an inverse secondary DWT transform to obtain the watermarked audio Iw.

[0106] Furthermore, obtaining a second time domain histogram corresponding to the watermarked audio includes:

[0107] Preprocessing the watermarked audio and calculating the absolute mean of the watermarked audio;

[0108] Selecting an extraction range based on the absolute mean value and λ of the embedded watermark audio;

[0109] A binning operation is performed on the samples within the extraction range to obtain the second time domain histogram.

[0110] Specifically, in this embodiment, the robust watermarking method for resisting synchronization attacks using the audio histogram shape is used. The watermark extraction process is as follows: Figure 2 The specific steps are as follows:

[0111] 6. Preprocess the initial audio I and calculate the absolute mean of the audio. and λ determine the extraction range [-λA , -λA];

[0112] 7. Perform binning on samples in the range [-λA, -λA] to obtain the time domain statistical histogram ;

[0113] 8. Perform a two-level DWT transformation on the samples in the same bin to obtain a histogram in the low-frequency domain ;

[0114] 9. According to the shape of the low-frequency histogram, each 3 bins is a group. Figure 3 Extract information. Specifically:

[0115] 901. After completing the DWT transformation, the frequency domain histogram shape is obtained, and the watermark information is extracted with 3 bins as a group. Assume that the number of samples in each bin group is a, b, and c respectively.

[0116] 902. If , then the current bin group watermark information w'(i) is "11".

[0117] 903. Judge if , then the current bin group watermark information w'(i) is "00".

[0118] 904. If , then the current bin group watermark information w'(i) is "10".

[0119] 905. Judge if , then the current bin group watermark information w'(i) is "01".

[0120] In this example, we used three different audio clips: 90s pop music track 1, 121s rock music track 2, and 150s noisy conversation track 3, as experimental subjects. These three audio groups have different characteristics, which are shared by various types of audio in daily life. Therefore, using these three audio groups as experimental subjects can make the experimental results generalizable. The audio sampling rate selected is 44.1kHz, which is not much different from one audio to another, so this invention can be applied to various audio types.

[0121] The embodiment of the present invention provides an audio robust watermark algorithm that is resistant to synchronization attacks and conventional signal processing, mainly including: watermark embedding, watermark extraction method and histogram shape modification method.

[0122] The watermark embedding method provided by the embodiment of the present invention comprises the following steps:

[0123] For original audio Perform pretreatment;

[0124] Binary watermark sequence , original audio Expressed as ,in, is the length of the embedded information, and N represents the length of the audio. The absolute average value A is calculated as the sum of the absolute values ​​of the audio samples over their duration:

[0125] ;

[0126] b) obtaining a time domain histogram by statistics;

[0127] Assuming the embedded watermark is Lw bits, the algorithm requires (Lw / 2) 3 bins are used to complete watermark embedding. Divide the embedding range [-λA, λA] selected in step 1 into (Lw / 2) 3+3 parts, each bin width , the samples within the range are allocated to the corresponding bin according to their amplitude, and the time domain statistical histogram is obtained , L=(Lw / 2) 3+3.

[0128] c) Perform segmented DWT on the time domain histogram;

[0129] After the target range [-λA, λA] samples have been assigned to bins of equal width, each sample in the same i The samples in bin(i) within the range of M are transformed by the secondary DWT. At the same time, the breakpoint coordinates between discontinuous samples are recorded during the transformation for use in the subsequent inverse transformation to obtain the frequency domain statistical histogram , L=(Lw / 2) 3+3.

[0130] d) Modifying the shape of the frequency domain histogram to embed watermark information;

[0131] After completing the DWT transformation, the frequency domain histogram shape is obtained. Each group of shapes is modified according to the corresponding watermark information, with 3 bins as a group. Assuming that the number of samples in each bin is a, b, and c respectively, the modification rules are as follows:

[0132] When the embedded information w(i) is "00", the shape is modified to low-medium-high:

[0133] ;

[0134] Migrate samples in order from a to c to ensure If not satisfied, adjust , transfer I1 samples in a to b, and get the adjusted and ;ensure If not satisfied, adjust , transfer I2 samples in a to c, and get the adjusted and .

[0135] When the embedded information w(i) is "01", the modified shape is b minimum:

[0136] ;

[0137] Migrate samples from b to a, c to ensure If not satisfied, adjust , put b in samples are transferred to a, The samples are transferred to c, and the adjusted , and .

[0138] When the embedded information w(i) is "10", the shape is modified to be the highest b:

[0139] ;

[0140] Migrate samples from a, c to b to ensure If not satisfied, adjust , put a Transfer samples to b and transfer samples to c The samples are transferred to b, and the adjusted , and .

[0141] When the embedded information w(i) is "00", the shape is modified to high-medium-low:

[0142] ;

[0143] Migrate samples from c to a to ensure If not satisfied, adjust , transfer I1 samples in c to b, and get the adjusted and ;ensure If not satisfied, adjust ,Will I2 samples are transferred to a, and the adjusted and .

[0144] e) Extract watermark information:

[0145] After preprocessing the watermarked audio Iw and obtaining the frequency domain histogram Hw with the watermark information, the watermark information is extracted in groups of three bins. Assume that the number of samples in each bin is a, b, and c respectively.

[0146] If , then the current bin group watermark information w'(i) is "11".

[0147] If , then the current bin group watermark information w'(i) is "00".

[0148] If , then the current bin group watermark information w'(i) is "10".

[0149] If , then the current bin group watermark information w'(i) is "01".

[0150] f) Convert the modified frequency domain histogram back to watermarked audio:

[0151] After modifying the shape of each bin group of the histogram, the frequency domain histogram Hw of the embedded information is obtained. Combined with the breakpoints and high and low frequency coefficients between each discontinuous sample recorded in step 3, the modified frequency domain histogram Hw is subjected to an inverse secondary DWT transform to obtain the watermarked audio Iw.

[0152] The experimental results are shown in the following table.

[0153] For audio robust watermarking algorithms, audio with watermark information can be considered to have good robustness if the bit error rate is below 20% after being attacked.

[0154] The experimental results based on audio track 1 are shown in Tables 1 to 5. The results show that the algorithm can resist MP3 compression with a quality factor of 48Kbps, Stretching attack, Attack with amplitude conversion, 30db Gaussian noise and low-pass filter above 4k.

[0155] The experimental results based on audio track 2 are shown in Tables 6 to 10. The results show that the algorithm can resist MP3 compression with a quality factor of 48Kbps, Stretching attack, Attack with amplitude conversion, 30db Gaussian noise and low-pass filter above 4k.

[0156] The experimental results based on audio track 3 are shown in Tables 11 to 15. The results show that the algorithm can resist MP3 compression attacks with a quality factor of 48Kbps, Stretching attack, Attack with amplitude conversion, 30db Gaussian noise and low-pass filter above 4k.

[0157] Table 1 Bit error rate of audio track 1 when attacked by Gaussian noise (embedded robust watermark is 80 bits)

[0158] Gaussian noise (dB) 30 35 40 45 Bit Error Rate 19% 3% 0% 0%

[0159] Table 2 Bit error rate of audio track 1 when it is attacked by compression factor

[0160] MP3 compression (compression factor) 48 56 64 80 96 128 Bit Error Rate 16% 11.3% 6% 2% 2% 3%

[0161] Table 3 Bit error rate of audio track 1 when it is attacked by amplitude conversion

[0162] Amplitude conversion (%) -40 -30 -20 -10 +10 +20 +30 +40 Bit Error Rate 5% 0% 0% 0% 0% 0% 0% 1.67%

[0163] Table 4 Bit error rate of audio track 1 when it is attacked by the stretch factor

[0164] Stretch (stretch factor %) -30 -20 -15 -10 -5 +5 +10 +15 +20 +30 Bit Error Rate 0% 0% 0% 0% 0% 0% 0% 0% 0% 5%

[0165] Table 5 Bit error rate of audio track 1 when it is attacked by low-pass filtering

[0166] Low-pass filtering 4k 5k 6k 7k 8k Bit Error Rate 7.4% 0% 0% 0% 0%

[0167] Table 6 Bit error rate of audio track 2 when attacked by Gaussian noise (embedded robust watermark is 80 bits)

[0168] Gaussian noise (dB) 30 35 40 45 Bit Error Rate 13.7% 0% 0% 0%

[0169] Table 7 Bit error rate of audio track 2 when it is attacked by compression factor

[0170] MP3 compression (compression factor) 48 56 64 80 96 128 Bit Error Rate 20% 15% 12.7% 8% 5% 0%

[0171] Table 8 Bit error rate of audio track 2 when it is attacked by amplitude conversion

[0172] Amplitude conversion (%) -40 -30 -20 -10 +10 +20 +30 +40 Bit Error Rate 10% 7% 0% 0% 0% 0% 0% 5%

[0173] Table 9 Bit error rate of audio track 2 when it is attacked by the stretch factor

[0174] Stretch (stretch factor %) -30 -20 -15 -10 -5 +5 +10 +15 +20 +30 Bit Error Rate 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%

[0175] Table 10 Bit error rate of audio track 2 when attacked by low-pass filtering

[0176] Low-pass filtering 4k 5k 6k 7k 8k Bit Error Rate 13.7% 4% 2% 2% 2%

[0177] Table 11: Bit error rate of audio track 3 when attacked by Gaussian noise (embedded robust watermark is 80 bits)

[0178] Gaussian noise (dB) 30 35 40 45 Bit Error Rate 15% 8% 0% 0%

[0179] Table 12 Bit error rate of audio track 3 when attacked by compression factor

[0180] MP3 compression (compression factor) 48 56 64 80 96 128 Bit Error Rate 9% 4% 0% 0% 5% 0%

[0181] Table 13 Bit error rate of audio track 3 when it is attacked by amplitude conversion

[0182] Amplitude conversion (%) -40 -30 -20 -10 +10 +20 +30 +40 Bit Error Rate 0% 0% 0% 0% 0% 0% 0% 0%

[0183] Table 14 Bit error rate of audio track 3 when it is attacked by the stretch factor

[0184] Stretch (stretch factor %) -30 -20 -15 -10 -5 +5 +10 +15 +20 +30 Bit Error Rate 0% 0% 0% 0% 0% 0% 0% 0% 0% 0%

[0185] Table 15 Bit error rate of audio track 3 when attacked by low-pass filter

[0186] Low-pass filtering 4k 5k 6k 7k 8k Bit Error Rate 15% 14% 12% 12% 12%

[0187] The robust watermark algorithm of the present invention greatly improves the embedding capacity by reasonably designing the histogram embedding shape, and uses the segmented DWT method to expand the histogram time domain characteristics to the frequency domain to enhance its robustness.

[0188] Robustness against synchronization attacks and MP3 compression is a crucial aspect of the performance of an audio robust watermarking algorithm. Previously, most audio watermarking algorithms could only target one of these attacks. This embodiment achieves robustness against both synchronization and compression attacks, and designs a richer histogram shape, improving embedding capacity while ensuring robustness. Given the same number of histogram samples, the proposed algorithm has a higher embedding capacity than similar algorithms. This feature demonstrates that even for audio carriers with fewer sampling points (e.g., a sampling rate of 8000 Hz), the proposed algorithm still has the ability to embed, thereby enhancing its practical application value.

[0189] Compared with existing methods, this embodiment can resist various stretching attacks and has stronger robustness after being subjected to synchronization attacks such as MP3 compression, resampling and stretching, or conventional signal operations. The watermark can be effectively extracted. Specifically:

[0190] This embodiment, based on histogram shape, significantly improves the robustness of the algorithm by transferring the histogram's time-domain features to the frequency domain using a segmented DWT. Compared to previous algorithms, this invention leverages the diversity of histogram shapes to achieve higher-capacity embedding with the same number of bins, thereby resolving the difficulty of embedding watermarks in certain voice files due to the small number of samples. This has significant implications for research on robust watermarking in the area of ​​histogram modification.

[0191] This embodiment utilizes the robustness of the low-frequency domain coefficients after DWT transformation against attacks such as compression and Gaussian noise. Through the segmented DWT method, the time domain stretching invariance is extended to the frequency domain while retaining the histogram shape, which not only improves resistance to conventional attacks but also retains robustness against stretching attacks.

[0192] This embodiment can effectively extract watermark information under different signal processing conditions, such as Gaussian noise and MP3 compressed sound attacks, and meets the requirements of daily digital forensics and digital authentication. At the same time, this embodiment can be applied to different audios and achieves good results on different audios.

[0193] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. An audio histogram shape watermarking method resistant to synchronization attacks, characterized in that: include: Divide the pre-selected audio samples according to the number of pre-allocated bins to obtain a first time domain histogram; wherein a bin is a statistical histogram sample interval; Performing a piecewise DWT transform and an inverse DWT transform on the first time domain histogram to obtain an embedded watermark audio; Obtaining a second time domain histogram corresponding to the watermarked audio, and performing the piecewise DWT transform on the second time domain histogram to obtain a second time domain histogram in the low frequency domain; Information is extracted from the second time domain histogram in the low frequency domain to obtain an embedded watermark.

2. The audio histogram shape watermark method for resisting synchronization attacks according to claim 1 is characterized in that: Acquiring the first time domain histogram includes: Preprocess the initial audio and calculate the absolute mean of the audio; Selecting an embedding range, i.e., the audio sample, based on the absolute mean of the audio and a preset constant value λ; The number of bins is allocated according to the number of watermark bits to be embedded; 3 bins are used to embed 2 bits; The samples are divided according to the number of bins to obtain the first time domain histogram.

3. The audio histogram shape watermark method against synchronization attacks according to claim 1, characterized in that: Acquiring the watermarked audio includes: Performing a piecewise DWT transform on the first time domain histogram, recording the breakpoints between each discontinuous bin, and obtaining a first time domain histogram in the low-frequency domain; Performing information embedding on the first time domain histogram in the low frequency domain, modifying the shape of the bin, and obtaining a modified first time domain histogram; The modified first time domain histogram is combined with the recorded breakpoints to perform an inverse secondary DWT transformation to obtain the embedded watermark audio.

4. The audio histogram shape watermarking method against synchronization attacks according to claim 3 is characterized in that: Performing a piecewise DWT transform on the first time domain histogram includes: The audio samples [-λA, λA] are considered to have been assigned to bins of equal width, where A is the absolute mean of the audio and λ is a preset constant value used to control the audio range to be processed; a two-level DWT transform is performed on the samples in the first bin, i.e., the range [-λA, -λA +M), where M represents the width of each bin; a two-level DWT transform is performed on the samples in the second bin, i.e., the range [-λA + / M, -λA +2M), and each subsequent bin interval is increased by M on the previous bin range, and so on to obtain the interval of each bin; a DWT transform is performed on the piecewise function within each M range separately to obtain the first time domain histogram of the low-frequency domain.

5. The audio histogram shape watermark method for resisting synchronization attacks according to claim 3 is characterized in that: Methods for modifying the shape of a bin include: The bin intervals are grouped into groups of 3, and each group is used to embed 2 bits of information. Different embedding strategies are used for the four different 2-bit information units in the watermark sequence: 00, 01, 10, and 11, where: When the current embedded information is 00, modify the bin shape to low-medium-high; When the current embedding information is 01, modify the shape of the middle bin to be the lowest; When the current embedding information is 10, modify the shape of the middle bin to be the highest; When the current embedding information is 11, modify the shape to high-medium-low; Low, medium, and high represent the size relationship of three consecutive bin intervals in terms of sample quantity.

6. The audio histogram shape watermark method for resisting synchronization attacks according to claim 1 is characterized in that: Obtaining a second time domain histogram corresponding to the watermarked audio includes: Preprocessing the watermarked audio and calculating the absolute mean of the watermarked audio; Selecting an extraction range based on the absolute mean of the embedded watermark audio and a preset constant value λ; A binning operation is performed on the samples within the extraction range to obtain the second time domain histogram.

7. The audio histogram shape watermark method for resisting synchronization attacks according to claim 1 is characterized in that: Extracting information from the second time domain histogram in the low frequency domain includes: Extract watermark information with 3 bins as a group, and assume that the number of samples in each bin is a, b, and c respectively: if , then the current bin group watermark information is 11; if , then the current bin group watermark information is 00; if , then the current bin group watermark information is 10; if , then the current bin group watermark information is 01.

8. The audio histogram shape watermark method for resisting synchronization attacks according to claim 2 is characterized in that: The absolute mean value of the audio is: ； Among them, A is the absolute mean of the audio, N is the length of the original audio, is the i-th sample in each bin.

Images