Pitch period modulation information hiding detection method based on Bayesian network

Abstract

The invention relates to the technical field of information hiding, and discloses a pitch period modulation information hiding detection method based on a Bayesian network. From the perspective of different network structures, the invention achieves the best effect by using a joint network combining three networks, namely, intra-frame, inter-frame and self-network. From the perspective of detection performance at different embedding rates, the detection accuracy of the method of the invention at a lower embedding rate is better than some existing steganalysis methods. From the perspective of the size of PDBNS, the smaller the number of Bayesian networks, the shorter the system detection time. When the number of Bayesian networks is small, the method still has a higher detection accuracy, and can realize real-time detection of the pitch period modulation information hiding method. From the perspective of detection performance at different time lengths, the method of the invention has a 100% detection rate of audio from 1 second to 10 seconds at an embedding rate of 1.0, which is better than some existing algorithms.

Description

Technical Field

[0001] The invention relates to the technical field of information hiding, and in particular to a pitch period modulation information hiding detection method based on a Bayesian network. Background Art

[0002] In recent years, with the widespread popularity of mobile devices, network compressed voice based on VoIP (Voice Over Internet Protocol, IP voice transmission) has become one of the most commonly used network communication methods. At the same time, voice security issues occur from time to time, such as forgery and deception, privacy leakage, covert communication, etc. Voice covert communication is a technology that embeds secret information into voice calls to evade regulators. Illegal elements can use voice to transmit secret information to carry out illegal and criminal activities, endangering national security and social stability. Therefore, in order to prevent the illegal abuse of covert communication in network compressed voice, it is particularly important to perform steganalysis on covert communication in voice.

[0003] Although the existing pitch period modulation steganalysis method can achieve a high detection accuracy, there is still room for improvement, especially in the case of low embedding rate and short duration, the steganalysis performance has greater room for improvement. This application document takes G.723.1 6.3kbit / s speech coding as an example, analyzes the phenomenon that the conditional probability of the pitch period codeword changes greatly before and after steganalysis, and based on this characteristic, proposes a new pitch period steganalysis method based on Bayesian network. Summary of the invention

[0004] The present invention provides the following technical solution: a pitch period modulation information hiding detection method based on a Bayesian network, comprising the following main steps;

[0005] Step S1: constructing a pitch delay spatiotemporal model and quantizing the index;

[0006] Step S2: constructing and training a pitch delay Bayesian network set;

[0007] Step S3: Steganalysis detection process.

[0008] Preferably, the pitch delay space-time model construction and index quantization in step S1 include pitch delay space-time model construction and pitch delay space-time model index quantization.

[0009] Preferably, the pitch delay space-time model is constructed to describe the local periodicity of the speech signal, which is mainly manifested in the correlation of the code element values ​​within and between frames, that is, there is a correlation between different code elements in the same frame and there is a correlation between code elements between adjacent frames. The pitch delay space-time model consists of vertices and directed edges, the vertices represent different code elements, and the directed edges describe the correlation relationship between two code elements, and its definition is shown in formula (1-1);

[0010]

[0011] In formula (1-1), D is the entire network, V is the vertex set, each vertex represents a pitch prediction codeword of the audio sequence, that is, vi,k represents the kth pitch subframe codeword in the i-th frame of the speech sequence, and E is the edge set, each edge represents the correlation between two pitch prediction codewords of the audio sequence, that is,<vx,p, vy,q> It means that there is a directed edge from vertex vx,p to vertex vy,q, referred to as edge. When x=y, the edge set E represents the correlation between code elements within the frame. There are 12 edges in total, named as pitch delay intra-frame network, referred to as intra-frame network; when y=x+1, the edge set represents the correlation between code elements between frames. There are 16 edges in total, named as pitch delay inter-frame network, referred to as inter-frame network.

[0012] Preferably, the pitch delay spatiotemporal model index quantization analyzes the code element correlation change before and after the pitch period steganography to find the intra-frame and inter-frame steganography sensitive features, thereby performing efficient information hiding detection. This application document defines this intra-frame and inter-frame steganography sensitive feature as the pitch delay spatiotemporal model transfer index, referred to as the transfer index;

[0013] The pitch delay intra-frame and inter-frame network can describe the transfer relationship of each code element in the compressed speech code stream segment, but the transfer strength relationship of each edge is different. This section quantifies the transfer index to explore the feature strength relationship between different edges in the network.

[0014] For each edge, according to the different value ranges of vx,p and vy,q, an n×m transfer index matrix M will be generated, as shown in formula (1-2);

[0015]

[0016] Where n is the maximum value that vx,p can take, m is the maximum value that the code element corresponding to vertex vy,q can take, Pi,j represents the probability that vy,q=j when vx,p=i, and the calculation method is shown in formula (1-3);

[0017]

[0018] The transfer index matrix R of network compressed speech will change before and after steganography, and the corresponding Pi,j will also change accordingly. The carrier transfer index matrix Rc and the steganographic transfer index matrix Rs are calculated using the unsteganographic samples and the steganographic samples respectively. The corresponding transfer probabilities are recorded as and The residual of the steganographic transfer index matrix is ​​obtained. Due to the selection of different vertices, the size of the transfer index matrix R will change, making it difficult to directly compare the residuals of different edges, which will affect the correct characterization of the strength of the transfer index. Therefore, a regularization process is required. The method for regularizing the transfer probability residual is shown in formula (1-4).

[0019]

[0020] Preferably, the pitch lag Bayesian network set construction and training in step S2 includes pitch lag Bayesian network set construction and pitch lag Bayesian network set training.

[0021] Preferably, the pitch delay Bayesian network set construction comprises the following steps:

[0022] Step S(1): Divide the training sample category C into steganographic samples (denoted as 1) and non-steganographic samples (denoted as 0);

[0023] Step S(2): construct different Bayesian networks according to the histogram of each pitch period codeword in the frame, connect each pitch period codeword in the frame as a sub-node of the sample category and add them to the network respectively. The Bayesian network of this part has only two nodes and one edge.

[0024] Step S(3): construct a different Bayesian network based on each edge of the pitch delay spatiotemporal model; connect vx,p as a child node of the sample category and add them to the corresponding network respectively, then connect vy,q as a child node of the sample category C and vx,p, connect C and vy,q and vx,p and vy,q respectively and add them to the corresponding network respectively. The Bayesian network of this part has three nodes and two edges.

[0025] Preferably, the pitch delay Bayesian network set training is used to train the pitch delay Bayesian network set by training samples before steganalysis classification, the sample category C is recorded as X0, the four pitch period code elements (ACL0, ACL1, ACL2, ACL3) of the current subframe are recorded as X1, X2, X3, X4, the four pitch period code elements of the next subframe are recorded as X5, X6, X7, X8, the values ​​of these random variables are recorded as x0, x1, x2, x3, x4, x5, x6, x7, x8, respectively, then x0∈{0,1}, x1, x3, x5, x7∈{0,1,…,127}, x2, x4, x6, x8∈{0,1,2,3}, the obtained conditional probability is shown in formula (2-1);

[0026]

[0027] In the above formula, Xi represents a random variable, pa(Xi) represents the parent node of the random variable Xi, P(Xi |pa(Xi)) represents the probability of the random variable Xi occurring under the probability of the parent node pa(Xi) occurring, and the PDBNS training process is the process of learning conditional probability. After learning, each node in each Bayesian network can obtain a CPT, which records the probability of a certain codeword being affected by the category and other codeword values. The size of the CPT is determined by the value range of the node and the value range of all the parent nodes of the node.

[0028] Preferably, the steganalysis detection process in step S3 includes a feature extraction process and a steganalysis classification process.

[0029] Preferably, the feature extraction process includes using a speech segment containing n frames, inputting its code elements into PDBNS in code stream order, and according to the code element features of the audio segment, each Bayesian network of PDBNS will output two feature indexes CI, including a predicted positive sample feature index CI0 and a negative sample feature index CI1, which are used to characterize the similarity between the input speech segment and the corresponding category of the network. The higher the predicted feature index, the greater the possibility of the corresponding category. The CI calculation method defined in the present invention is shown in formula (3-1);

[0030]

[0031] The characteristic index CI in formula (3-1) represents the sum of conditional probabilities corresponding to the values ​​of vx,p, and vy,q when the speech takes a fixed value at the root node C;

[0032] For a network compressed speech, we first need to extract its pitch delay value, and then input the pitch delay into PDBNS. Each Bayesian network will calculate two characteristic indexes CI, CI0 and CI1. Arrange the output multiple characteristic indexes in order to obtain the characteristic vector. The formula for generating the characteristic index vector CIV is shown in formula (3-2);

[0033]

[0034] In formula (3-2), and Represents the two characteristic indices output by the ith trained Bayesian network of PDBNS. The network order does not change during the entire steganalysis classification process, that is, the characteristic indices of the generated characteristic index vectors at corresponding positions are the results of the same network output, so that classification can be performed correctly according to the features.

[0035] Preferably, the steganalysis classification process requires the preparation of unsteganographic compressed speech, steganalysis compressed speech samples and compressed speech to be tested for training for steganalysis classification. First, the PDBNS is trained using the existing classified training set to obtain network parameters; then the features of the training set are extracted, that is, a vector composed of feature indices of the training set samples on PDBNS is output for training the SVM classifier; then the features of the unknown category steganalysis compressed speech are extracted, and the feature vector of the sample to be tested on PDBNS is output; finally, the extracted training set feature vector and the sample to be tested feature vector are sent to the SVM classifier for classification, and the classification result is output to complete the steganalysis of the sample to be tested.

[0036] Beneficial Effects

[0037] Compared with the prior art, the present invention provides a pitch period modulation information hiding detection method based on Bayesian network, which has the following beneficial effects:

[0038] From the perspective of different network structures, the present invention achieves the best effect by using a joint network combining intra-frame, inter-frame and self networks; from the perspective of detection performance at different embedding rates, the detection accuracy of the method of the present invention at a lower embedding rate is better than some existing steganalysis methods; from the perspective of the size of PDBNS, the smaller the number of Bayesian networks, the shorter the system detection time, and when the number of Bayesian networks is small, it still has a high detection accuracy, and can realize real-time detection of the pitch period modulation information hiding method; from the perspective of detection performance at different time lengths, the detection rate of the method of the present invention is 100% for audio from 1 second to 10 seconds at an embedding rate of 1.0, which is better than some existing algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1A schematic diagram of a pitch delay intra-frame and inter-frame network of a pitch period modulation information hiding detection method based on a Bayesian network of the present invention;

[0040] Figure 2 A schematic diagram of the intra-frame and inter-frame transfer index of the pitch period modulation information hiding detection method based on the Bayesian network of the present invention;

[0041] Figure 3 A schematic diagram of sorted network transfer indexes of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0042] Figure 4 A schematic diagram of a concentrated part of a pitch delay Bayesian network of a pitch period modulation information hiding detection method based on a Bayesian network of the present invention;

[0043] Figure 5 A conditional probability diagram of some nodes of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0044] Figure 6 It is a schematic diagram of characteristic indexes of unsteganographic speech and steganographic speech samples at different embedding rates of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0045] Figure 7 It is a schematic diagram of the feature extraction steps of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0046] Figure 8 It is a schematic diagram of the steganalysis classification process of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0047] Fig. 9 It is a schematic diagram of the detection accuracy rate (%) of a single Bayesian network under the embedding rate of the pitch period modulation information hiding detection method based on the Bayesian network of the present invention;

[0048] Fig.10 It is a schematic diagram of the detection accuracy (%) of PDBNS composed of different numbers of Bayesian networks under the embedding rate of the pitch period modulation information hiding detection method based on Bayesian network of the present invention;

[0049] Fig.11 It is a schematic diagram of the steganalysis accuracy (%) at different embedding rates of the pitch period modulation information hiding detection method based on the Bayesian network of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0051] Embodiment 1;

[0052] Please read Figure 1 ; A pitch period modulation information hiding detection method based on a Bayesian network includes pitch delay space-time model construction and exponential quantization. The pitch delay space-time model construction and exponential quantization include pitch delay space-time model construction and pitch delay space-time model exponential quantization. The pitch delay space-time model construction is used to describe the local periodicity of speech signals, which is mainly manifested in the correlation between code element values ​​within and between frames, that is, there is correlation between different code elements in the same frame and there is correlation between code elements between adjacent frames. The pitch delay space-time model consists of vertices and directed edges. Vertices represent different code elements, and directed edges describe the correlation relationship between two code elements. Its definition is shown in formula (1-1).

[0053]

[0054] In formula (1-1), D is the entire network, V is the vertex set, each vertex represents a pitch prediction codeword of the audio sequence, that is, vi,k represents the kth pitch subframe codeword in the i-th frame of the speech sequence. E is the edge set, each edge represents the correlation between two pitch prediction codewords of the audio sequence, that is,<vx,p, vy,q> Indicates that there is a directed edge from vertex vx,p to vertex vy,q, referred to as edge. When x=y, the edge set E represents the correlation between the code elements within the frame, with a total of 12 edges, named pitch delay intraframe network, referred to as intraframe network; when y=x+1, the edge set represents the correlation between the code elements between frames, with a total of 16 edges, named pitch delay interframe network, referred to as interframe network. The intraframe network and the interframe network are combined to obtain the pitch delay intraframe interframe network, referred to as intraframe interframe network.

[0055] Embodiment 2:

[0056] like Figure 2 and 3 As shown. The pitch delay spatiotemporal model index quantization analyzes the code element correlation changes before and after the pitch period steganography, finds the intra-frame and inter-frame steganography sensitive features, and then performs efficient information hiding detection. This application document defines this intra-frame and inter-frame steganography sensitive feature as the pitch delay spatiotemporal model transfer index, referred to as the transfer index;

[0057] The pitch delay intra-frame and inter-frame network can describe the transfer relationship of each code element in the compressed speech code stream segment, but the transfer strength relationship of each edge is different. This section quantifies the transfer index to explore the feature strength relationship between different edges in the network.

[0058] For each edge, depending on the value ranges of vx,p and vy,q, an n×m transfer index matrix M will be generated, as shown in formula (1-2).

[0059]

[0060] The value of n is the maximum value that vx,p can take, the value of m is the maximum value that the code element corresponding to the vertex vy,q can take, Pi,j represents the probability that vy,q=j when vx,p=i, and the calculation method is shown in formula (1-3).

[0061]

[0062] The transfer index matrix R of network compressed speech will change before and after steganography, and the corresponding Pi,j will also change accordingly. The carrier transfer index matrix Rc and the steganographic transfer index matrix Rs are calculated using the unsteganographic samples and the steganographic samples respectively. The corresponding transfer probabilities are recorded as and The residual of the steganalysis transfer index matrix is ​​obtained. Since different vertices are selected, the size of the transfer index matrix R will change, making it difficult to directly compare the residuals of different edges, which will affect the correct characterization of the strength of the transfer index. Therefore, a regularization process is required. The method for regularizing the transfer probability residual is shown in formula (1-4).

[0063]

[0064] In this embodiment, 9336 non-steganographic samples and steganographic samples with an embedding rate of 0.1 are randomly selected to calculate the transfer index of different edges between frames within the frame, such as Figure 2 shown.

[0065] exist Figure 2 In the figure, ACL0, ACL1, ACL2, and ACL3 represent the code elements of the current subframe, ACL0', ACL1', ACL2', and ACL3' represent the code elements of the next subframe, and arrows are used to represent the transition index from one code element to another. The specific values ​​in the arrows represent the values ​​of the transition indexes, and the thickness of the arrows represents the strength of the transition index. A thick arrow represents a strong transition index, and a thin arrow represents a weak transition index.

[0066] pass Figure 2It can be seen that when the vertex vx,p is 7-bit ACL2 and ACL0, the transfer index is stronger; when the vertex vy,q is 2-bit ACL3 and ACL1, the transfer index is stronger; when the vertex vy,q is 7-bit ACL2 and ACL0, the transfer index is weaker. In addition, the transfer index between frames is slightly higher than the transfer index within a frame.

[0067] The transfer indexes of the intra-frame and inter-frame networks are sorted from large to small to obtain the sorted transfer indexes. The edges with strong transfer indexes are named strong transfer index edges, and the edges with weak transfer indexes are named weak transfer index edges. The schematic diagram of the sorted transfer index is shown in Figure 3 shown.

[0068] from Figure 3 It can be seen that the transfer index of 8 edges is lower than 0.03, the transfer index of 12 edges is higher than 0.075, and the strength of the other 8 edges is between 0.03 and 0.075. This transfer index strength relationship can be used to describe the changes of an edge before and after steganography, so as to quickly find out the steganographic sensitive features. When building a network, a suitable threshold can be selected in advance, and some edges can be filtered out through the threshold, that is, when building a network, strong transfer index edges can be selected as much as possible to achieve more efficient steganalysis through a simpler pitch delay space-time model.

[0069] Embodiment 3

[0070] Please read Figure 4 ; A pitch period modulation information hiding detection method based on a Bayesian network, including the construction and training of a pitch delay Bayesian network set, the construction and training of a pitch delay Bayesian network set including the construction of a pitch delay Bayesian network set and the training of a pitch delay Bayesian network set, and the construction of a pitch delay Bayesian network set includes the following steps;

[0071] Step S(1): Divide the training sample category C into steganographic samples (denoted as 1) and non-steganographic samples (denoted as 0);

[0072] Step S(2): construct different Bayesian networks according to the histogram of each pitch period codeword in the frame, connect each pitch period codeword in the frame as a sub-node of the sample category and add them to the network respectively. The Bayesian network of this part has only two nodes and one edge.

[0073] Step S(3): construct a different Bayesian network based on each edge of the pitch delay spatiotemporal model; connect vx,p as a child node of the sample category and add them to the corresponding network respectively, then connect vy,q as a child node of the sample category C and vx,p, connect C and vy,q and vx,p and vy,q respectively and add them to the corresponding network respectively. The Bayesian network of this part has three nodes and two edges.

[0074] The schematic diagram of the pitch delay Bayesian network set is as follows Figure 4 As shown, the figure shows the own network, intra-frame and inter-frame networks representing the histogram features respectively.

[0075] Embodiment 4:

[0076] Please read Figure 5 ; Pitch delay Bayesian network set training is used to train the pitch delay Bayesian network set through training samples before steganalysis classification. Figure 4 The intra-frame network, inter-frame network, and self-network in the subframe are described separately. Sample category C is denoted as X0, and the four pitch period code elements (ACL0, ACL1, ACL2, ACL3) of the current subframe are denoted as X1, X2, X3, X4, and the four pitch period code elements of the next subframe are denoted as X5, X6, X7, X8. The values ​​of these random variables are denoted as x0, x1, x2, x3, x4, x5, x6, x7, x8, respectively. Then x0∈{0,1}, x1, x3, x5, x7∈{0,1,…,127}, x2, x4, x6, x8∈{0,1,2,3}. The obtained conditional probability is shown in formula (2-1).

[0077]

[0078] In the above formula, Xi represents a random variable, pa(Xi) represents the parent node of the random variable Xi, and P(Xi |pa(Xi)) represents the probability of the random variable Xi occurring under the probability of the parent node pa(Xi) occurring. The PDBNS training process is the process of learning conditional probability. After learning, each node in each Bayesian network can obtain a CPT, which records the probability of a certain codeword being affected by the category and other codeword values. The size of the CPT is determined by the value range of the node and the value range of all the parent nodes of the node. Figure 5 It represents the conditional probability of nodes in the non-steganographic samples and the steganographic samples in the partial network.

[0079] exist Figure 5 In the figure, “Cover” represents the carrier, i.e., the non-steganographic sample, “Stego” represents the steganographic, i.e., the steganographic sample, the horizontal axis is the value of the node, and the vertical axis is the conditional probability. It can be seen that under different conditions of steganography and non-steganography, the conditional probabilities of the four nodes are obviously different. Therefore, the conditional probability of the PDBNS trained by the present invention can reflect the changes before and after steganography.

[0080] Embodiment five;

[0081] Please read Figure 6 and 7; A pitch period modulation information hiding detection method based on a Bayesian network includes a steganalysis detection process, which includes a feature extraction process and a steganalysis classification process. The feature extraction process includes using a speech segment containing n frames, and inputting its code elements into PDBNS in code stream order. According to the code element features of the audio segment, each Bayesian network of PDBNS will output two feature indexes CI, including a predicted positive sample feature index CI0 and a negative sample feature index CI1. The index is used to characterize the similarity between the input speech segment and the corresponding category of the network. The higher the predicted feature index, the greater the possibility of the corresponding category. The CI calculation method defined in the present invention is shown in formula (3-1).

[0082]

[0083] The characteristic index CI in formula (3-1) represents the sum of conditional probabilities corresponding to the values ​​of vx,p, and vy,q when the speech takes a fixed value at the root node C. Figure 6 The CI index of the unsteganographic speech and the steganographic speech samples with different embedding rates under one of the Bayesian networks is shown, with the horizontal axis representing CI0 and the vertical axis representing CI1;

[0084] Depend on Figure 6 It is known that the two CI indices output by a single Bayesian network will change significantly at different embedding rates. For a network compressed speech, its pitch delay value needs to be extracted first, and then the pitch delay is input into PDBNS. Each Bayesian network will calculate two characteristic indices CI, namely CI0 and CI1. Arrange the output multiple characteristic indices in order to obtain a characteristic vector. The formula for generating the characteristic index vector CIV is shown in formula (3-2);

[0085]

[0086] In formula (3-2), and Represents the two feature indices output by the ith trained Bayesian network of PDBNS. The network order does not change during the entire steganalysis classification process, that is, the feature indices of the generated feature index vectors at corresponding positions are the results of the same network output, so that classification can be performed correctly according to the features. The feature extraction steps are as follows: Figure 7 .

[0087] Embodiment six;

[0088] Please read Figure 8; The process of steganalysis classification requires the preparation of unsteganographic compressed speech for training, steganalysis compressed speech samples and compressed speech to be tested for detection. First, use the existing classified training set to train PDBNS to obtain network parameters; then extract the features of the training set, that is, output the vector composed of the feature indexes of the training set samples on PDBNS, which is used to train the SVM classifier; then extract the features of the unknown category of steganalysis compressed speech, and output the feature vector of the sample to be tested on PDBNS; finally, send the extracted training set feature vector and the sample to be tested feature vector to the SVM classifier for classification, and output the classification result to complete the steganalysis of the sample to be tested. The schematic diagram of the steganalysis classification process is shown below. Figure 8 shown.

[0089] Experimental example;

[0090] Please read Fig. 9 , 10 and 11; the experimental data set contains 41 hours of Chinese speech and 72 hours of English speech, each speech lasts 10 seconds, and the G723.1 high-speed 6.3kb / s speech encoder is used. The Chinese and English data sets are divided into training set, validation set and test set in a ratio of 3:1:1. The experiments are conducted using C++ and Python. This experimental example uses the steganalysis methods proposed in references

[22] ,

[23] and

[24] as comparisons, and analyzes the performance from different network structures, different single Bayesian networks, different numbers of Bayesian networks, different embedding rates and different speech durations.

[0091] (1) Detection performance analysis of different network structures

[0092] PDBNS contains multiple networks, and multiple networks can be combined with each other. Table 6-1 lists the accuracy of different combinations of networks under different embedding rates for Chinese and English datasets.

[0093]

[0094] From Table 6-1, we know that as the embedding rate increases, the number of audio modified codewords increases, so the discrimination rate of different network structures continues to increase. From the accuracy of a single network structure, the detection accuracy of the intra-frame network and the inter-frame network is higher than that of the self-network. At the 10% embedding rate of the Chinese data set, the intra-frame network discrimination success rate reached 87.72%, the inter-frame network discrimination success rate was 86.71%, and the self-network discrimination accuracy was 81.35%. From the combined results of the two network structures, the intra-frame + inter-frame network has the highest accuracy, with a discrimination accuracy of 89.86% at a 10% embedding rate of the Chinese data set, and higher than the accuracy of a single intra-frame and inter-frame network structure; the combination of intra-frame + self-network also achieved good results, and the accuracy was higher than that of a single intra-frame and self-network structure; the accuracy of the inter-frame + self-network was only 86.65% at a 10% embedding rate of the Chinese data set, which was lower than the discrimination rate of 86.71% of the inter-frame network at 10%. Judging from the combination results of the three network structures, the total intra-frame + inter-frame + self-network achieved the best effect among all network combinations, and the English dataset has a similar conclusion.

[0095] (2) Detection performance analysis of different single Bayesian networks

[0096] Please read Fig. 9 , each Bayesian network in PDBNS will have different contributions when performing classification. This section studies the detection performance of different Bayesian networks by studying the detection effect of a single Bayesian network on different Chinese and English data sets at an embedding rate of 0.1. The experimental results are shown in Fig. 9 As shown. And by Fig. 9 It can be seen that different single Bayesian network classifiers show different detection accuracy when detecting the same audio. The accuracy of some classifiers is only 50%, and the best single classifier can achieve an accuracy of more than 80%.

[0097] (3) Detection performance analysis of different numbers of Bayesian networks

[0098] Please read Fig.10 ; Each Bayesian network in PDBNS will output two feature indices to form a feature vector. The more Bayesian networks there are, the more dimensions of the feature vector will be, and the time required to extract features will also increase. This section uses the quantization index of the pitch delay spatiotemporal network model proposed in this experimental example to selectively construct PDBNS. In an incremental manner, each time the network with the largest corresponding quantization index among the remaining unselected Bayesian networks is selected and added to PDBNS. Finally, each group of PDBNS is tested under the Chinese and English 0.1 embedding rate datasets. The detection accuracy is as follows Fig.10 shown.

[0099] from Fig.10It can be seen that as the number of Bayesian networks in PDBNS increases, the accuracy of the network model gradually increases. The detection accuracy of using only the Bayesian network with the largest quantization index is 76.24%, and the detection accuracy of using the first 12 Bayesian networks with the largest quantization index is 88.38%. Therefore, the time required for system detection can be reduced by selecting an appropriate number of Bayesian networks.

[0100] This experimental example uses VS 2019 to perform time tests on a Windows 10 system with an Intel(R) Xeon(R) Gold 5218R CPU@2.10GHz and 128GB memory. After experiments, using 32 Bayesian networks, including intra-frame, inter-frame and self networks, for classification, the average time required to detect a 10-second speech is 13 milliseconds. Using only one Bayesian network with the largest quantization index for classification, the average time required to detect the same audio is 0.6 milliseconds, and the time taken increases linearly with the increase of PDBNS.

[0101] (4) Detection performance analysis at different embedding rates

[0102] Please read Fig.11 ; The embedding rate is the ratio of the number of embedded bits to the embedding capacity. By reducing the embedding rate, the security of steganography can be improved and the impact of modified codewords on speech quality can be reduced. This experimental example conducts experimental research at 10 different embedding rates (10% to 100%) to study the efficiency of model steganalysis. In the data set, the steganalysis methods of references

[22] ,

[23] and

[24] are used as comparison methods. The experimental results are shown in the figure. Fig.11 shown.

[0103] from Fig.11 As we know, as the embedding rate increases, the number of modified symbols in a single frame of audio increases, and the accuracy of the four steganalysis algorithms increases. Reference

[23] and the experimental example method have better accuracy than reference

[24] and reference

[22] methods in general, and the experimental example method can achieve better results than the reference

[23] method at low embedding rates. On the Chinese dataset, when the embedding rate is 10%, compared with the reference

[24] , reference

[23] and reference

[22] methods, the experimental example method achieves 25.41%, 5.63%, and 0.94% improvements respectively. When the embedding rate is less than 40%, the experimental example method achieves a higher detection rate than the other three methods. When the embedding rate is greater than or equal to 40%, the accuracy of the experimental example method is significantly improved compared with the reference

[24] method, and is equal to the accuracy of the reference

[22] and reference

[23] methods, all of which are close to 100%. Similar conclusions are drawn on the English dataset, and the overall accuracy of the four methods on the English dataset is improved compared with the accuracy of the Chinese method.

[0104] (5) Detection performance analysis of different speech durations

[0105] The longer the time, the more intra-frame and inter-frame relationships can be obtained, thus extracting more features. This section evaluates the PDBNS model proposed in the experimental example from the perspective of speech duration segment length. Experiments were conducted at 10 durations with an embedding rate of 1.0, and the experimental results are shown in Table 6-2.

[0106]

[0107] As can be seen from Table 6-2, the accuracy of the experimental method in each duration under 1.0 embedding rate reaches 100%. The detection rate of the algorithm in reference

[24] in English samples with 1.0 embedding rate and 1 second duration is 87.25%, while the detection rate of the method in reference

[23] is 99.84%.

[0108] Adaptive codebook search obtains pitch delay codewords by predicting the pitch period value. The prediction results of existing pitch period prediction methods have errors and cannot achieve accurate prediction. Based on this feature, the existing information hiding method based on adaptive codebook coefficient modulation fine-tunes the pitch period result in the codebook search stage to achieve the hiding of secret information. Reference

[10] divides the closed-loop pitch period search range of the four subframes of G.723.1 speech coding into two groups, and the secret information bits 0 / 1 correspond to one group of search ranges respectively, thereby achieving information hiding. Reference

[11] adjusts the ±1 / 3 fractional search range of the pitch period of two subframes of G.729a speech coding to embed secret information. Reference

[12] embeds 3-bit secret information in G.723.1 speech coding through a double-layer embedding algorithm to improve embedding efficiency and reduce speech steganography distortion. Reference

[13] replaces the pitch period prediction value less than the error threshold with secret information, thereby achieving covert communication in Speex. Reference

[14] uses fractional pitch delay to embed secret information based on the characteristic that fractional pitch delay remains stable during AMR speech coding, thereby improving the embedding capacity while ensuring speech quality.

[0109] There are also some steganalysis methods for information hiding based on adaptive codebook coefficient modulation. These methods can be divided into two categories: one is the manual feature extraction method based on SVM[18-21], and the other is the automatic feature extraction method based on neural network[22-23]. Early research mainly belongs to the first category. For example, Li Songbin et al.

[18] proposed a pitch modulation information hiding detection method based on codebook association network by using the correlation of codebook. This method extracts sample features by constructing a codebook association rule network, performs principal component analysis on the extracted features to reduce the dimension, and sends them to SVM classifier for training and steganalysis. Jia Yizhen et al.

[19] used the characteristics of steganography that the pitch period estimation symbiotic characteristics will change, and constructed feature vectors based on the pitch period symbiotic characteristics to realize the detection of G.729a pitch modulation information hiding. Ren et al.

[20] proposed a pitch modulation information hiding detection method based on MSDPD (Markov transition probability matrix of the Second-orderDifference of Pitch Delay). This method uses MSDPD as the SVM classifier feature and obtains the calibrated MSDPD feature by recompression to further improve the detection accuracy. Based on reference

[20] , reference

[21] combines the first-order Markov transition probability feature and MSDPD in the subframe. Combined with the proposed information hiding detection framework based on pitch modulation. With the development of neural networks and deep learning, pitch period modulation steganalysis methods based on deep learning have emerged one after another. Hu et al. proposed a steganalysis feature fusion network method, which uses long short-term memory to extract pitch period modulation and quantization index modulation features, and uses convolutional neural network to fuse the two features, and realizes information hiding detection through a fully connected network; Li et al. proposed a steganalysis method based on code element embedding, bidirectional long short-term memory and convolutional neural network attention mechanism. This method can detect multiple types of information hiding at the same time, and can also achieve good detection performance for pitch period information hiding.

[0110] In summary, from the perspective of different network structures, the present invention achieves the best effect by using a joint network combining intra-frame, inter-frame and self-networks; from the perspective of detection performance at different embedding rates, the detection accuracy of the experimental example method at a lower embedding rate is better than some existing steganalysis methods; from the size of PDBNS, the fewer the number of Bayesian networks, the shorter the system detection time, and when the number of Bayesian networks is small, it still has a high detection accuracy, and can realize real-time detection of the pitch period modulation information hiding method; from the perspective of detection performance at different time lengths, the detection rate of the experimental example method at an embedding rate of 1.0 for 1 second to 10 seconds of audio is 100%, which is better than some existing algorithms.

[0111] It should be noted that, in this experimental example, relational terms such as first and second, etc., are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0112] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A pitch period modulation information hiding detection method based on Bayesian network, characterized in that: It includes the following main steps; Step S1: constructing a pitch delay spatiotemporal model and quantizing the index; Step S2: constructing and training a pitch delay Bayesian network set; Step S3: Steganalysis detection process; The pitch delay space-time model is constructed to describe the local periodicity of the speech signal, which is mainly manifested in the correlation of the code element values ​​within and between frames, that is, there is correlation between different code elements in the same frame and there is correlation between code elements between adjacent frames. The pitch delay space-time model consists of vertices and directed edges, where the vertices represent different code elements and the directed edges describe the correlation relationship between two code elements, which is defined as: D=<V,E>, V={v i,k },i∈{0,1,2,…},k∈{1,2,3,4} E={<v x,p ,v y,q >|v x,p ,v y,q ∈V,v x,p ≠v y,q },xy∈{0,1} Where D is the entire network, V is the vertex set, each vertex represents a pitch prediction codeword of the audio sequence, that is, vi,k represents the kth pitch subframe codeword in the i-th frame of the speech sequence, and E is the edge set, each edge represents the correlation between two pitch prediction codewords of the audio sequence, that is,<vx,p,vy,q> It means that there is a directed edge from vertex vx,p to vertex vy,q, referred to as edge. When x=y, the edge set E represents the correlation between code elements within the frame. There are 12 edges in total, named as pitch delay intra-frame network, referred to as intra-frame network; when y=x+1, the edge set represents the correlation between code elements between frames. There are 16 edges in total, named as pitch delay inter-frame network, referred to as inter-frame network.

2. The pitch period modulation information hiding detection method based on Bayesian network according to claim 1 is characterized in that: The pitch delay space-time model construction and index quantization in step S1 include pitch delay space-time model construction and pitch delay space-time model index quantization.

3. The pitch period modulation information hiding detection method based on Bayesian network according to claim 2 is characterized in that: The pitch delay spatiotemporal model index quantization analyzes the code element correlation change before and after the pitch period steganography, finds the intra-frame and inter-frame steganography sensitive features, and then performs efficient information hiding detection. This intra-frame and inter-frame steganography sensitive feature is defined as the pitch delay spatiotemporal model transfer index, referred to as the transfer index; The pitch delay intra-frame and inter-frame network can describe the transfer relationship of each code element in the compressed speech code stream segment, but the transfer strength relationship of each edge is different. This section quantifies the transfer index to explore the feature strength relationship between different edges in the network. For each edge, according to the different value ranges of vx,p and vy,q, an n×m transfer index matrix M will be generated, the formula is: The value of n is the maximum value that vx,p can take, the value of m is the maximum value that the code element corresponding to the vertex vy,q can take, and Pi,j represents the probability that vy,q=j when vx,p=i. The calculation formula is: The transfer index matrix R of network compressed speech will change before and after steganography, and the corresponding Pi,j will also change accordingly. The carrier transfer index matrix Rc and the steganographic transfer index matrix Rs are calculated using the unsteganographic samples and the steganographic samples respectively. The corresponding transfer probabilities are recorded as and The residual of the steganalysis transfer index matrix is ​​obtained. Due to the selection of different vertices, the size of the transfer index matrix R will change, making it difficult to directly compare the residuals of different edges, which will affect the correct characterization of the strength of the transfer index. Therefore, a regularization process is required. The formula for regularizing the transfer probability residual is:

4. The pitch period modulation information hiding detection method based on Bayesian network according to claim 1 is characterized in that: The pitch lag Bayesian network set construction and training in step S2 includes pitch lag Bayesian network set construction and pitch lag Bayesian network set training.

5. The pitch period modulation information hiding detection method based on Bayesian network according to claim 4 is characterized in that: The pitch delay Bayesian network set construction comprises the following steps: Step S1: Divide the training sample category C into steganalyzed samples denoted as 1 and non-steganographic samples denoted as 0; Step S2: construct different Bayesian networks according to the histogram of each pitch period codeword in the frame, connect each pitch period codeword in the frame as a sub-node of the sample category and add them to the network respectively. The Bayesian network of this part has only two nodes and one edge; Step S3: Construct different Bayesian networks according to each edge of the pitch delay spatiotemporal model, connect vx,p as the child node of the sample category and add them to the corresponding network respectively, then connect vy,q as the child node of the sample category C and vx,p, connect C and vy,q as well as vx,p and vy,q respectively and add them to the corresponding network respectively. This part of the Bayesian network has three nodes and two edges.

6. The pitch period modulation information hiding detection method based on Bayesian network according to claim 4 is characterized in that: The pitch delay Bayesian network set training is used to train the pitch delay Bayesian network set through training samples before steganalysis classification. The sample category C is recorded as X0. The four pitch period code elements ACL0, ACL1, ACL2, and ACL3 of the current subframe are recorded as X1, X2, X3, and X4. The four pitch period code elements of the next subframe are recorded as X5, X6, X7, and X8. The values ​​of these random variables are recorded as x0, x1, x2, x3, x4, x5, x6, x7, and x8, respectively. Then x0∈{0,1}, x1, x3, x5, x7∈{0,1,…,127}, x2, x4, x6, x8∈{0,1,2,3}, and the formula for the obtained conditional probability is: In the above formula, Xi represents a random variable, pa(Xi) represents the parent node of the random variable Xi, P(Xi pa(Xi)) represents the probability of the random variable Xi occurring under the probability of the parent node pa(Xi) occurring, and the PDBNS training process is the process of learning conditional probability. After learning, each node in each Bayesian network can obtain a CPT, which records the probability of a certain codeword being affected by the category and other codeword values. The size of the CPT is determined by the value range of the node and the value range of all the parent nodes of the node.

7. The pitch period modulation information hiding detection method based on Bayesian network according to claim 1 is characterized in that: The steganalysis detection process in step S3 includes a feature extraction process and a steganalysis classification process.

8. The pitch period modulation information hiding detection method based on Bayesian network according to claim 7 is characterized in that: The feature extraction process includes using a speech segment containing n frames, inputting its codewords into PDBNS in code stream order, and according to the codeword features of the audio segment, each Bayesian network of PDBNS will output two feature indexes CI, including the predicted positive sample feature index CI0 and the negative sample feature index CI1. The index is used to characterize the similarity between the input speech segment and the corresponding category of the network. The higher the predicted feature index, the greater the possibility of the corresponding category. The CI calculation method formula is defined as: The characteristic index CI of the formula represents the sum of conditional probabilities corresponding to the values ​​of vx, p, vy, and q when the speech takes a fixed value at the root node C; For a network compressed speech, we first need to extract its pitch delay value, and then input the pitch delay into PDBNS. Each Bayesian network will calculate two characteristic indexes CI, CI0 and CI1. Arrange the output multiple characteristic indexes in order to obtain the characteristic vector. The formula for generating the characteristic index vector CIV is: In the formula, and Represents the two feature indices output by the ith trained Bayesian network of PDBNS. The network order does not change during the entire steganalysis classification process, that is, the feature indices at the corresponding positions of the generated feature index vector are the results of the same network output, so that classification can be performed correctly according to the features.

9. The pitch period modulation information hiding detection method based on Bayesian network according to claim 7, characterized in that: The steganalysis classification process requires the preparation of unsteganographic compressed speech, steganalysis compressed speech samples and compressed speech to be tested for steganalysis classification. First, the PDBNS is trained using the existing classified training set to obtain network parameters; then the features of the training set are extracted, that is, a vector composed of feature indices of the training set samples on PDBNS is output for training the SVM classifier; then the features of the unknown category steganalysis compressed speech are extracted, and the feature vector of the sample to be tested on PDBNS is output; finally, the extracted training set feature vector and the sample to be tested feature vector are sent to the SVM classifier for classification, and the classification result is output to complete the steganalysis of the sample to be tested.

Images