RU2390054C2 - Method of identifying audio file containing digital watermark - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: invention relates to methods of identifying audio files (AF) containing digital watermarks (DW). The format of the audio file is determined before multilevel discrete wavelet transformation. If the format is compressed, the compressed audio file is converted to an analogue signal. After that voltage of the input analogue signal is converted to digital pulses in a digital reading device, where the said pulses are converted to a code which corresponds to the voltage value of the input signal. Based on multilevel discrete wavelet transformation, a characteristic vector is formed, the Mahalanobis distance to centre of gravity of the first (audio file contains a digital watermark) and second (audio file does not contain digital watermark) classes is calculated and the class to which the analysed audio file belongs is determined from the minimum calculated distance.

EFFECT: possibility of identifying compressed and uncompressed format audio files containing digital watermarks when there is no prior information on presence of a watermark in the given audio file and information on the rule for embedding the digital watermark.

2 dwg, 2 tbl

Description

The invention relates to the field of steganography, and in particular to methods for identifying sound files (PF) containing a digital watermark (CEH), and can be used to search for PF of various formats containing additional digital information in the form of a CEH in the absence of a priori information about its law embedding and presence in the SF.

A known method for identifying PFs containing CEHs (See Özer N., Avcibas I., Memon N., Sankur B., "Detection of audio covert channels using statistical footprints of hidden messages", Digital Signal Processing 16, 2006, p.389 -401), which includes the step of embedding additional information in the form of a CEH into the sound signal, the step of filtering the sound signal in the wavelet transform domain, the step of calculating the sound signal quality metrics, the step of generating the feature vector, the classifier training step, and the classification step.

The disadvantages of this method are the low probability of correct identification and the inability to analyze ZF compressed formats.

The closest in technical essence to the claimed invention (prototype) is a method for identifying a digital image containing a CEH (see RF Patent No. 2304306, class IPC G06K 9/00, published in 2006), which consists in the fact that the document embed additional information consisting of two types of CEHs, perform multilevel discrete wavelet transform, calculate high-order statistical characteristics from the distribution of wavelet coefficients at different levels of wavelet transform, then form their own character -terrorist vector, wherein in the "training" form an array of characteristic vectors from the training set of images, trained classifier, and in the "Analysis", classification is carried out images.

The disadvantage of this method is the lack of identification of ZF containing CEH.

The technical result to which the invention is directed is the development of a method for identifying sound files containing a digital watermark, providing the ability to analyze ZF both compressed and uncompressed formats, while the probability of correct identification should not be lower than that of the prototype.

The technical result is achieved by the fact that in the known method of identifying a digital image containing a CEH, which consists in the fact that additional information consisting of two types of CEH is embedded in a document, a multi-level discrete wavelet transform is performed, high-order statistical characteristics are calculated from the distribution of wavelet coefficients at different levels of the wavelet transform, then they form their own characteristic vector, while in the "training" mode they form an array of their own hara the characteristic vectors of images from the training set, the classifier is trained, and in the “analysis” mode, the images are classified, in addition, after embedding additional information consisting of two types of CEH in the SF, the format of the SF is determined, if the format is compressed, then commutation to the decoder is performed, when this, the decoder performs the conversion of the compressed audio file into an analog signal, and then in the digital reading device, the voltage of the input analog signal is converted to digital pulses, which are converted into a code corresponding to the value of the input signal voltage, if the format is uncompressed, then they are switched to a discrete multilevel wavelet transform block.

Due to the new set of essential features in the method, it is possible to identify sound files of both compressed (MP3, WMA) and uncompressed (WAV, PCM) formats containing CEH, in the absence of a priori information about the presence of CEH in the ZF and the law of its embedding.

The analysis of the prior art made it possible to establish that analogues that are characterized by a combination of features identical to all the features of the claimed technical solution are absent, which indicates the compliance of the claimed method with the condition of patentability “novelty”.

Search results for known solutions in this and related fields of technology in order to identify features that match the distinctive features of the claimed object from the prototype showed that they do not follow explicitly from the prior art. The prior art also did not reveal the fame of the distinctive essential features that determine the same technical result, which was achieved in the claimed method. Therefore, the claimed invention meets the condition of patentability "inventive step".

The claimed method is illustrated by drawings, which show:

figure 1 is a General diagram of a method for identifying a ZF containing a digital watermark;

figure 2 is a graph of the probability of correct identification of SFs containing CEH, as a result of the operation of the proposed method.

The claimed method, the scheme of which is presented in figure 1, is carried out in two stages, called "training" and "analysis". The “learning” stage is implemented due to the fact that first, using the block 1, a training sample is formed, including sound files containing randomly generated CEHs, and then, using the block 2, the format of the SF is determined. If the format is compressed, then through block 3 they perform switching to decoder 4. The decoder converts the compressed audio file into an analog signal. This device is implemented on the basis of a single-chip digital audio decoder AT85C51SND3B3. From the output of the decoder, the analog signal enters a digital measuring device 5, the circuit of which is known and described in the book by B.Ya. Avdeev, E.M. Antonyuk, and E.M. Dushin, "Fundamentals of Metrology and Electrical Measurements: A Textbook for High Schools," ed . E.M. Dushina. - 6th ed., Revised. and add. - L .: Energoatomizdat. Leningra. Department, 1987 .-- 480 pp., ill. (p. 255, fig. 8-34). In this case, the voltage of the input analog signal is converted into digital pulses, which are converted into a code corresponding to the value of the voltage of the input signal. If the SF format is not compressed, then they are switched to a discrete multilevel wavelet transform unit 6. In block 6, the digitized signal is subjected to a multilevel one-dimensional discrete wavelet transform (at least five levels of the wavelet transform are necessary) using biorthogonal low-frequency (LF) and high-frequency (HF) filters (Vorobev V.I., Gribunin V.G. Theory and practice of wavelet transform. - St. Petersburg: Military University of Communications, 1999, p. 42-49). It is recommended to use filters with coefficients according to table 1.

Then, in block 7, high-order statistical characteristics are calculated from the distribution of wavelet coefficients at different levels of the wavelet transform. These statistical characteristics are sample mean, sample variance, asymmetry, excess and central moments of even order from the fourth to the thirtieth (Ventzel E.S., Ovcharov L.A. Probability theory and its engineering applications. Textbook for technical colleges. - 2- e ed., ster. - M .: Higher school, 2000. - 480 pp .: ill., pp. 111-121), which are determined by the formulas (1, 2, 3, 4, 5), respectively.

Table 1 The values of the coefficients of the block biorthogonal wavelet filters 3/9 No. LF Coefficients No. HF coefficients p / p filter p / p filter 0 -0.00067974 0 -0.17678 one 0.0020392 one 0.53033 2 0.0050603 2 -0.53033 3 -0.020619 3 0.17678 four -0.014113 5 0.099135 6 0.0123 7 -0.32019 8 0.00205 9 0.94213 10 0.94213 eleven 0.00205 12 -0.32019 13 0.0123 fourteen 0.099135 fifteen -0.014113 16 -0.020619 17 0.0050603 eighteen 0.0020392 19 -0.00067974

- выборочное среднее;Where

- sample mean;

n is the number of coefficients of a given wavelet transform level.

where D _B is the sample variance;

n is the number of coefficients of a given wavelet transform level.

where A _S is the asymmetry of the theoretical distribution;

µ ₃ is the central moment of the third order;

σ ³ - cube of the sample mean square deviation.

where E _k is the excess of the theoretical distribution;

µ ₄ is the central moment of the fourth order;

σ ⁴ - sample mean square deviation to the fourth degree.

where µ _S is the central moment of the Sth order;

n is the number of coefficients of a given wavelet transform level.

In the block for the formation of the characteristic characteristic vector 8, all calculated values of the statistical characteristics are included in the vector (6).

where N is the number of wavelet transform levels.

Thus, the dimension of such a vector depends on the number of levels of the wavelet transform N and is equal to 18N.

After that, through block 9, switching to the block for generating an array of characteristic eigenvectors from the training set 10 is performed, where the array of eigenvector characteristic vectors of sound files from the training set is generated. For this, by means of block 11, switching to block 2 is carried out until an array of eigenvectors of all sound files from the training sample is generated (the training sample must contain at least 285 SFs that do not contain CEH and at least 285 SF containing CEH , moreover, all SFs of this training sample should be of the same format and the CEH should be built in according to different laws).

After the formation of an array of eigenvector characteristic vectors of all sound files from the training set, the resulting array, presented in the form of table 2, is switched to the training unit of classifier 12 for linear discrimination of sound files from the training set into two classes: ZF containing CEH and ZF not containing CEH . After that, the “training” stage ends and the second stage begins - “analysis".

At the “analysis” stage, with the SF selected for analysis (it must be of the same format as the SF from the training set), all the procedures described above are performed, only now, by means of block 9, they are switched to classification block 13, to which at the same time from the classifier training block 12 provide the results of discrimination of all eigenvalue characteristic vectors obtained from the SF from the training sample at the “training” stage.

table 2 Feature Vector Number Tag number in vector one 2 ... j ... n one y ₁₁ y ₁₂ ... y _1j ... y _1n 2 y ₂₁ y ₂₂ ... y _2j ... y _2n

i y _i1 y _i2 ... y _ij ... y _in

N y _N1 y _N2 ... y _Nj ... y _nn

In the classification block ZF 13 on the basis of the calculated Mahalanobis distance according to the formula (7) (Kalugina T.F., Kiselev V.Yu. Mathematical statistics: Textbook / Ivan. State energy University - Ivanovo, 2001. - 321 p., P.245) a decision is made on whether the analyzed sound file belongs to one of the classes: either to the class of PF containing CEH, or to the class of PF not containing CEH.

where d _M (x _i , x _j ) is the Mahalanobis distance,

K ^-1 = C is the matrix inverse to the covariance matrix K calculated from the sample x,

c _pq - elements of the matrix C.

To test the proposed method, a classifier based on discriminant analysis was trained on a sample of 1710 sound files of various formats, 855 of which contained CEH. Using a set of 1800 SFs of the WAV, MP3 and WMA format that do not belong to the training set, 600 of which contained a central digital health center built in by an unknown law (200 SF of each format), the claimed method was tested. The results presented in figure 2, showed that the probability of correct identification of the proposed method is not lower than that of the prototype.

Claims (1)

A method for identifying a sound file containing a digital watermark (CEH), which consists in performing a multi-level discrete wavelet transform, calculating high-order statistical characteristics from the distribution of wavelet coefficients at different levels of the wavelet transform, then generating their own characteristic vector, characterized in that in the "analysis" mode before performing a multilevel discrete wavelet transform, the format of the sound file is determined, if the format is compressed, then perform mutation to the decoder, while the decoder converts the compressed audio file into an analog signal, and then in the digital reading device, the voltage of the input analog signal is converted to digital pulses, which are converted into a code corresponding to the value of the input signal voltage, if the format is uncompressed, then they perform switching to a discrete multilevel wavelet transform block, in addition, after generating its own characteristic vector for it, they add the distances of Mahalanobis to the centers of gravity of the first (the sound file contains the CEH) and the second (the sound file does not contain the CEH) classes, and then decide whether the analyzed sound file belongs to one of the classes to minimize the calculated distance, and in the "training" mode form the training a sample that includes sound files containing randomly built CEHs and not containing those, while the sound files of the training sample must be of the same format, if the format is compressed, then they are switched to the decoder, this, the decoder converts the compressed audio file into an analog signal, and then, in the digital reading device, the voltage of the input analog signal is converted to digital pulses, which are converted into a code corresponding to the value of the input signal voltage, if the format is uncompressed, then they are switched to a discrete block multilevel wavelet transform, and then perform multilevel discrete wavelet transform, in addition, after the formation of the sobs array Of characteristic vectors from the training set, the centers of gravity for two classes of sound files and the covariance matrix are calculated.

Images