CN107564536A - AMR pitch delay steganalysis methods based on difference Markov transition probability features in pitch delay subframe group group - Google Patents

Abstract

The invention discloses an AMR pitch delay steganographic analysis method based on the Markov transfer probability feature of the difference value within a pitch delay subframe group. Aiming at two steganographic algorithms for modifying pitch delay in speech coding, this method proposes the AMR steganalysis feature based on the Markov transition probability of the intra-group difference of pitch delay subframes, uses support vector machine for classification and prediction, and realizes AMR pitch-oriented Delayed steganalysis method. The advantage of the present invention is that, for the two existing pitch-delay-oriented steganography methods, when the relative embedding rate is 50%, the detection rate can reach more than 95%.

Description

AMR basis based on the Markov transition probability feature of intragroup difference of pitch delay subframe group Tone Delay Steganalysis Method

technical field

The invention relates to the technical field of digital media processing, in particular to the technical field of multimedia information content security for judging whether AMR voice is steganographically steganographically.

technical background

In recent years, with the continuous maturity of mobile communication means, communication manufacturers have comprehensively promoted communication technology means for mobile Internet, the demand for voice communication services is increasing, and the requirements for the coding quality of voice communication and network bandwidth occupancy have been comprehensively improved. . In order to meet the needs of mobile voice communication, in October 1999, 3GPP (3rdGeneration Partnership Project) designated AMR (Adaptive Multi-Rate, Adaptive Multi-Rate) voice coding as the voice compression coding standard for mobile Internet voice communication, making AMR coding in Widely used in GSM, TDMA, UMTS and VoLTE. At present, all kinds of mobile terminal equipment, such as Samsung, Apple, Huawei and other mobile phone manufacturers set AMR as the voice format of mobile phones. At the same time, various communication apps installed on smart terminals, such as WeChat, QQ and Skype, etc., also use AMR encoding in the voice chat and voice message functions. With the popularization of AMR encoding standard, AMR voice provides a new space for steganography.

Steganography algorithms for AMR speech gradually appear. These algorithms are mainly integrated in the encoding process of AMR speech. By adjusting and modifying relevant parameters during the encoding process, the parameters contain secret information to complete the embedding. Based on the coding characteristics of AMR speech, the existing steganographic algorithms mainly involve the linear prediction stage, adaptive codebook search stage and fixed codebook modulation stage in speech compression coding. Among them, for the steganographic algorithm in the AMR speech adaptive codebook search stage, fine-tuning the pitch delay in the encoding process to hide the secret information is difficult to detect, and there are few corresponding steganographic analysis algorithms.

In the adaptive codebook search stage of AMR speech coding, the pitch delay is the prediction result of the speech pitch period, and there is short-term stability between pitch delays in voiced segments. However, the existing pitch-delay-oriented steganography method achieves embedding of secret information by controlling the search range of pitch delay, which will lead to changes in the short-term stability of pitch delay in voiced segments. Based on the above considerations, using the Markov transition probability based on the intra-group difference of the pitch delay subframe as a feature, the effective detection of the AMR pitch delay steganography method can be realized.

The steganalysis detection method of the present invention is based on AMR speech, and it is necessary to introduce the encoding and decoding principles of AMR speech before presenting the content of the invention.

AMR voice compression coding is a hybrid voice compression coding technology based on ACELP. Its characteristic is that the appropriate voice compression coding rate can be selected according to the specific situation of the audio to be compressed and the communication environment. Figure 1 is a schematic diagram of the principle of AMR coding. The basic principle is to select the appropriate codebook from the adaptive codebook and the fixed codebook according to the criterion of minimizing the weighted mean square error between the synthesized speech and the original speech when encoding the original speech frame by frame. The codebook vector is used to replace the residual signal, and the address and gain of the code vector, as well as the parameters of each filter are quantized and encoded, and then transmitted to the receiving end. The sampling frequency of the input voice of the AMR encoder is 8kHz, and the voice signal is a 16-bit quantized and uncompressed linear PCM code, which is coded in a frame of 20ms. One frame contains 160 sampling points, and each frame is divided into 4 parts with a unit of 5ms. subframes. According to the functions realized by it, it can be divided into three parts, including the linear prediction stage, the adaptive codebook search stage and the fixed codebook modulation stage. The following is an explanation of the technical terms involved:

1. Compressed audio: Refers to audio that has undergone lossy compression, such as MP3, WMA, and AMR are all lossy compressed audio.

2. Cover: Carrier audio, that is, audio without embedding secret information.

3. Stego: Confidential audio, that is, audio with secret information embedded.

4. Subframe group: Each frame of AMR voice includes 4 subframes, the first subframe and the second subframe form a subframe group, and the third subframe and the fourth subframe form a subframe group.

5. Intra-group relationship: Indicates the relationship between the pitch delay of the previous subframe and the pitch delay of the next subframe in the subframe group.

Contents of the invention

The present invention aims at the problem that current AMR-oriented steganalysis methods are relatively lacking, and realizes steganalysis oriented to AMR pitch delay.

The technical scheme of the present invention adopts the Markov transition probability based on the intra-group difference of the pitch delay subframe as the short-term stability evaluation standard to extract features, and uses the SVM classifier to perform binary classification on the Cover and Stego samples of AMR speech. 1, a kind of AMR pitch delay steganographic analysis method based on pitch delay subframe group difference Markov transition probability feature, it is characterized in that, comprising:

Step 1, based on the Markov transition probability of the difference in the pitch delay sub-frame group as the short-term stability evaluation standard to extract features, specifically including:

Step 1.1, build the difference value in the pitch delay subframe group group, in the AMR voice compression coding, the voice of 20ms is a frame, and a frame of voice is divided into 4 subframes by 5ms, and each subframe has a pitch delay; In a speech frame, the pitch delays of the _second subframe T2 and the _fourth subframe T4 are respectively based on the _first subframe T1 and the _third subframe T3, and are obtained according to correlation within a certain interval; as As shown in Figure 2, in the speech frame, the _first subframe T1 and the _second subframe T2 are used as a subframe group, and the _third subframe T3 and the _fourth subframe T4 are used as another subframe group;

For the convenience of description, the pitch delay sequence of AMR speech is represented as P=(p ₁₁ ,p ₁₂ ,...,p _t1 , _pt2 ,...,p _N1 ,p _N2 ) in subframe group, and N is The total number of subframe groups in AMR speech, i is the index value of the subframe group arranged in time, p _t1 and p _t2 are the first subframe and the second subframe in the subframe group t respectively; due to speech voiced The segment pitch delay has short-term stability, and the calculation of the intra-subframe group difference is used to measure the stability between two subframes in the subframe group; the calculation formula of the intra-subframe group difference D _intra is shown in Equation 1 , t∈[1,N]:

D _intra (t) = p _t1 -p _t2 Formula 1

Step 1.2, constructing the Markov transition probability feature of the intra-group difference of pitch delay subframes: the function of the Markov transition probability matrix is to measure the probability of a variable transitioning between different states, so it can be used to represent the continuous pitch delay subframe The change of the difference value within the frame group group; calculate the Markov transition probability of the difference value within the sub-frame group of the pitch delay as the feature of steganalysis, describe the difference between the difference value of the pitch delay sub-frame group between the Cover speech and the Stego speech, and realize the the distinction between the two;

The Markov transition probability feature of the pitch delay subframe group intra-group difference is defined as M1 _intra , and is calculated according to formula 2: when the value of M1 _intra (i, j) is the intra-group difference D _intra (t) is i, D _intra ( t+1) is the Markovian transition probability of j, t is the index on the time sequence of intra-group differences, S ₀ is the total number of calculated sub-frame group intra-group differences

According to AMR coding, in a frame of speech, the pitch delays of the second subframe T ₂ and the fourth subframe T ₄ are based on the first subframe T ₁ and the third subframe T ₃ respectively, according to the correlation within a certain interval Therefore, for different code rates, the interval of the intra-group difference is fixed, the range of the intra-group difference D _intra , and the corresponding M1 _intra feature dimension, by extracting the intra-group difference of the AMR voice pitch delay subframe group Value Markov transition probability as a steganographic classification feature for pitch delay modification;

Step 2. Use the SVM classifier to perform binary classification on the Cover and Stego samples of the AMR speech and perform steganalysis detection.

In the above-mentioned AMR pitch delay steganalysis method based on the pitch delay subframe group difference Markov transition probability feature, the step 2 specifically includes:

Step 2.1, classifier training, specifically includes:

Step 2.1.1, input WAV samples, generate cover samples and corresponding stego samples respectively, and extract pitch delay subframe group difference Markov transition probability classification features according to the method in 1;

Step 2.1.2, after the process of 2.1.1, obtain the training set samples and the same number of encrypted samples of two different embedding algorithms, and then randomly select different numbers of stego samples and cover samples and use the SVM classifier to train the steganalysis model ;

Step 2.2, steganalysis detection, specifically includes:

The process of performing steganalysis detection using the above steganalysis model includes the following steps:

Step 2.2.1, extracting the steganalysis feature set of the sample to be tested;

In step 2.2.2, the features are input into the constructed steganalysis model to obtain the steganographic judgment result of the sample.

Description of drawings

Figure 1 is a flow chart of AMR encoding and decoding.

Fig. 2 is a schematic diagram of subframe group division and intra-group difference.

Fig. 3a is the pitch delay distribution of the carrier audio and secret audio and the difference distribution within the pitch delay sub-frame group (the pitch delay distribution of Cover speech and Stego speech).

Fig. 3b shows the pitch delay distribution of the carrier audio and the secret audio and the difference distribution within the pitch delay sub-frame group (the distribution of the difference within the pitch delay sub-frame group of Cover speech and Stego speech).

Fig. 4a is the Markov transition probability feature (Cover speech) of the pitch delay sub-frame intra-group difference between the carrier audio and the secret audio.

Fig. 4b is the Markov transition probability feature (Stego voice) of the pitch delay sub-frame group differences between the carrier audio and the secret audio.

Fig. 5 is the steganalysis training and detection framework of the present invention.

detailed description

1. Feature extraction method

1.1 Feature extraction

1.1.1 Constructing the intra-group difference value of the pitch delay subframe group

In AMR speech compression coding, 20ms of speech is one frame, and one frame of speech is divided into 4 subframes in units of 5ms, and each subframe has a pitch delay. In one speech frame, the pitch delays of the second subframe T ₂ and the fourth subframe T ₄ are respectively based on the first subframe T ₁ and the third subframe T ₃ , and are obtained according to correlation within a certain interval. As shown in FIG. 2 , in the speech frame, the _first subframe T1 and the second subframe T2 are _regarded as a subframe group, and the _third subframe T3 and the _fourth subframe T4 are regarded as another subframe group.

For the convenience of description, the pitch delay sequence of AMR speech is represented as P=(p ₁₁ ,p ₁₂ ,...,p _t1 , _pt2 ,...,p _N1 ,p _N2 ) in subframe group, and N is The total number of subframe groups in AMR speech, i is the index value of the subframe group arranged in time, p _t1 and p _t2 are the first subframe and the second subframe in the subframe group t, respectively. Since the pitch delay of voiced voiced segment has short-term stability, calculating the intra-subframe group difference is used to measure the stability between two subframes within the subframe group. The formula for calculating the intra-subframe difference D _intra is shown in Equation 1, t∈[1,N]:

D _intra (t) = p _t1 -p _t2 (Formula 1)

1.1.2 Construct the Markov transition probability feature of the difference value within the pitch delay subframe group

The role of the Markov transition probability matrix is to measure the probability of the variable transitioning between different states, so it can be used to represent the change of the difference within the group of consecutive pitch delay subframes. Calculate the Markov transition probability of the intra-group difference of the pitch delay sub-frame group as the feature of steganalysis, describe the difference between the pitch delay sub-frame group difference between the Cover speech and the Stego speech, and realize the distinction between the two.

The Markov transition probability feature of the pitch delay subframe group intra-group difference is defined as M1 _intra , and is calculated according to formula 2: when the value of M1 _intra (i, j) is the intra-group difference D _intra (t) is i, D _intra ( t+1) is the Markovian transition probability of j, t is the index on the time sequence of intra-group differences, S ₀ is the total number of calculated sub-frame group intra-group differences

According to AMR coding, in a frame of speech, the pitch delays of the second subframe T ₂ and the fourth subframe T ₄ are based on the first subframe T ₁ and the third subframe T ₃ respectively, according to the correlation within a certain interval sexual acquisition. Therefore, for different code rates, the interval of the intra-group difference is fixed, and the range of the intra-group difference D _intra and the corresponding M1 _intra feature dimension are shown in Table 1.

Table 1 Range of intra-group difference D _intra , feature dimension of M1 _intra

By extracting the Markov transition probability of the intra-group difference of pitch delay subframes in AMR speech, it is used as the steganographic classification feature for pitch delay modification.

1.2 Analysis of characteristic principle

Since the pitch period of the voiced speech segment has short-term stability, the pitch delay is the prediction result of the pitch period in the speech coding process, so the pitch delay of the voiced segment also has short-term stability, and the pitch delay between adjacent pitch delays should be relatively stable , the difference between the two pitch delays in the subframe group should be small. The existing steganography method for pitch delay realizes steganography by controlling the search range of pitch delay in the encoding process. The steganography process leads to changes in the pitch delay value obtained by searching, the short-term stability between adjacent pitch delays in the voiced segment will be destroyed, and the difference between the two pitch delays in the subframe group will become larger.

Fig. 3 shows the distribution of the pitch delay sequence and intra-subframe difference D _intra of a voiced segment of Cover speech and Stego speech. Among them, the Cover voice is an AMR voice file generated by a public AMR encoder with a code rate of 12.2kbps, and the Stego voice is an AMR voice sample generated after embedding with the maximum capacity using the steganography algorithm in literature [1]. Figure 3(a) shows that the pitch delay between subframes is smoother for Cover speech than for Stego speech. Figure 3(b) shows the distribution of the intra-subframe group difference D _intra in Cover speech and Stego speech. It can be seen that the two have obvious differences. In Cover speech, the difference is small and the distribution is concentrated, and the continuous difference changes little. There is a phenomenon that the continuous difference is 0, but in Stego voice, the value of the difference distribution is large, the continuous difference changes greatly, and the case of 0 difference is very small. The comparison results show that the steganographic process does destroy the short-term stability between adjacent pitch delays in voiced segments.

The function of the Markov transition probability matrix is to measure the probability of the variable transitioning between different states, so it can be used to represent the change between continuous differences and calculate the Markov transition probability of the difference within the pitch delay subframe group As a feature, describing the difference between the Cover voice and the Stego voice can realize the distinction between the Cover voice and the Stego voice.

For the Cover voice and the Stego voice, the Markov transition probability distribution of the intra-group differences of the pitch delay sub-frames are calculated respectively, as shown in Fig. 4 . Among them, the Cover voice is an AMR voice file generated by a public AMR encoder with a code rate of 12.2kbps, and the Stego voice is an AMR voice sample generated after embedding with the maximum capacity using the steganography algorithm in literature [1]. It can be seen from the results that in the Cover speech, the feature distribution has an obvious peak in the central part, while in the Stego speech the feature distribution is relatively flat, so the Markov transition probability of the pitch delay subframe group difference in the Cover speech is the same as that of the Stego speech. There is a clear difference in speech.

From the comparison results in Figure 3 and Figure 4, it can be seen that it is effective to use the Markov transition probability feature of the pitch delay subframe group intra-group difference as a classification feature to distinguish cover audio from stego audio.

2. Steganalysis detection

2.1 Classifier Training

Step 2.1.1, input WAV samples, generate cover samples and corresponding stego samples respectively, and extract the Markov transition probability classification feature of pitch delay sub-frame group difference value according to the method in 1.

Step 2.1.2, after the process of 2.1.1, obtain the training set samples and the same number of encrypted samples of two different embedding algorithms, and then randomly select different numbers of stego samples and cover samples and use the SVM classifier to train the steganalysis model .

2.2 Steganalysis detection

The process of performing steganalysis detection using the above steganalysis model includes the following steps:

Step 2.2.1, extracting the steganalysis feature set of the sample to be tested.

In step 2.2.2, the features are input into the constructed steganalysis model to obtain the steganographic judgment result of the sample.

2.3 Experimental results of steganalysis

In order to verify the effectiveness of the algorithm of the present invention, the present invention aims at the steganalysis model trained by different steganography methods, and the experimental results are shown in Table 1 and Table 2. Among them, TPR represents the probability that the secret audio (Stego) is detected as Stego, and TNR represents the detection rate that the carrier audio (Cover) is detected as Cover.

Experimental result shows, under 12.2kbps, 10.2kbps, 7.95kbps, 7.40kbps, 6.70kbps and 5.90kbps encoding mode, classification model of the present invention has to existing two kinds of steganography methods that are modified for scale factor band code book. Good detection ability, when the relative embedding rate is 50%, the detection rate can reach more than 95%.

Table 2 detects the performance of the steganographic algorithm in the literature [1]

Table 3 detects the performance of the steganographic algorithm in the literature [2]

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (2)

1. a kind of AMR pitch delay steganalysis method based on pitch delay subframe group difference Markov transition probability feature, it is characterized in that, comprising: Step 1, based on the Markov transition probability of the difference in the pitch delay sub-frame group as the short-term stability evaluation standard to extract features, specifically including: Step 1.1, build the difference value in the pitch delay subframe group group, in the AMR voice compression coding, the voice of 20ms is a frame, and a frame of voice is divided into 4 subframes by 5ms, and each subframe has a pitch delay; In a speech frame, the pitch delays of the _second subframe T2 and the _fourth subframe T4 are respectively based on the _first subframe T1 and the _third subframe T3, and are obtained according to correlation within a certain interval; as As shown in Figure 2, in the speech frame, the _first subframe T1 and the _second subframe T2 are used as a subframe group, and the _third subframe T3 and the _fourth subframe T4 are used as another subframe group; For the convenience of description, the pitch delay sequence of AMR speech is represented as P=(p ₁₁ ,p ₁₂ ,...,p _t1 , _pt2 ,...,p _N1 ,p _N2 ) in subframe group, and N is The total number of subframe groups in AMR speech, i is the index value of the subframe group arranged in time, p _t1 and p _t2 are the first subframe and the second subframe in the subframe group t respectively; due to speech voiced The segment pitch delay has short-term stability, and the calculation of the intra-subframe group difference is used to measure the stability between two subframes in the subframe group; the calculation formula of the intra-subframe group difference D _intra is shown in Equation 1 , t∈[1,N]: D _intra (t) = p _t1 -p _t2 Formula 1 Step 1.2, constructing the Markov transition probability feature of the intra-group difference of pitch delay subframes: the function of the Markov transition probability matrix is to measure the probability of a variable transitioning between different states, so it can be used to represent the continuous pitch delay subframe The change of the difference value within the frame group group; calculate the Markov transition probability of the difference value within the sub-frame group of the pitch delay as the feature of steganalysis, describe the difference between the difference value of the pitch delay sub-frame group between the Cover speech and the Stego speech, and realize the the distinction between the two; The Markov transition probability feature of the pitch delay subframe group intra-group difference is defined as M1 _intra , and is calculated according to formula 2: when the value of M1 _intra (i, j) is the intra-group difference D _intra (t) is i, D _intra ( t+1) is the Markovian transition probability of j, t is the index on the time sequence of intra-group differences, S ₀ is the total number of calculated sub-frame group intra-group differences According to AMR coding, in a frame of speech, the pitch delays of the second subframe T ₂ and the fourth subframe T ₄ are based on the first subframe T ₁ and the third subframe T ₃ respectively, according to the correlation within a certain interval Therefore, for different code rates, the interval of the intra-group difference is fixed, the range of the intra-group difference D _intra , and the corresponding M1 _intra feature dimension, by extracting the intra-group difference of the AMR voice pitch delay subframe group Value Markov transition probability as a steganographic classification feature for pitch delay modification; Step 2. Use the SVM classifier to perform binary classification on the Cover and Stego samples of the AMR speech and perform steganalysis detection. 2. a kind of AMR pitch delay steganalysis method based on pitch delay subframe group difference Markov transition probability feature according to claim 1, it is characterized in that, described step 2 specifically comprises: Step 2.1, classifier training, specifically includes: Step 2.1.1, input WAV samples, generate cover samples and corresponding stego samples respectively, and extract pitch delay subframe group difference Markov transition probability classification features according to the method in 1; Step 2.1.2, after the process of 2.1.1, obtain the training set samples and the same number of encrypted samples of two different embedding algorithms, and then randomly select different numbers of stego samples and cover samples and use the SVM classifier to train the steganalysis model ; Step 2.2, steganalysis detection, specifically includes: The process of performing steganalysis detection using the above steganalysis model includes the following steps: Step 2.2.1, extracting the steganalysis feature set of the sample to be tested; In step 2.2.2, the features are input into the constructed steganalysis model to obtain the steganographic judgment result of the sample.