Speech steganography positioning method and device based on dynamic mask bidirectional state encoder

Abstract

This invention discloses a speech steganalysis localization method and apparatus based on a dynamic mask bidirectional state encoder, belonging to the field of speech processing technology. The method includes: data preprocessing and feature fusion steps, localization network construction steps, network training steps, and steganalysis localization steps. Specifically, it captures the long-distance bidirectional temporal dependency of the pitch delay sequence through a dynamic mask bidirectional state encoder and dynamically generates a frame-level confidence mask to suppress unreliable frame interference. It utilizes fuzzy multi-head self-attention to simulate steganalysis noise uncertainty and combines a von Neumann boundary low-rank adaptive linear layer to alleviate data imbalance, thereby achieving sub-frame-level steganalysis probability prediction. This invention achieves accurate sub-frame-level steganalysis embedding location, with a finer localization granularity than existing frame-level methods. It maintains low false alarm rates and low false alarm rates in Chinese and English datasets of different steganalysis durations under sudden steganalysis scenarios, making it suitable for tracing the source of hidden information in VoIP voice forensics.

Description

Technical Field

[0001] This invention relates to the field of speech processing technology, and in particular to a speech steganalysis localization method and apparatus based on a dynamic mask bidirectional state encoder. Background Technology

[0002] With the popularization of Voice over Internet Protocol (VoIP), steganography based on speech coding parameters has become a research hotspot in the field of information hiding due to its strong real-time performance and good concealment, but it has also brought serious security challenges. In mainstream speech coding standards (such as AMR-NB and G.729), the pitch delay parameter of the Adaptive Codebook (ACB) is often used as an embedding carrier of secret information because of its quasi-periodicity and dynamic update mechanism. Taking the ACB domain steganography method proposed by Huang et al. (Huang Y, Liu C, Tang S, et al. Steganography integration into a low-bit rate speech codec[J]. IEEE transactions on informationforensics and security, 2012, 7(6): 1865-1875.) as an example, this method embeds 1 bit of secret information by modifying the least significant bit (LSB) of the integer pitch delay of each subframe, while the fractional pitch delay remains unchanged. Since the human ear is not sensitive to minute changes in pitch delay within ±1, this steganography method can achieve highly covert communication while ensuring speech quality, making it one of the common means of illegal covert communication.

[0003] However, most existing ACB domain steganalysis detection technologies only aim at "steganography presence detection," meaning they can only determine whether a piece of speech contains secret information. This leads to a situation where, in practical evidence-gathering scenarios such as tracing the location of embedded secret information and extracting evidence, existing technologies cannot answer the core question of "where is the secret information embedded?" resulting in a lack of location capabilities and insufficient evidentiary value.

[0004] Currently, some methods for ACB domain steganalysis detection are based on statistical features of integer pitch delay (LiS B, Jia YZ, Fu JY, et al. Detection of pitch modulation information hiding based on codebook correlation network [J]. Chinese Journal of Computers, 2014, 37 (10): 2107-2117.), second-order difference features (Ren Y, Yang J, Wang J, et al. AMR steganalysis based on second-order difference of pitch delay [J]. IEEE Transactions on InformationForensics and Security, 2017, 12 (6): 1345-1357.), and fusion features of fractional pitch delay (TianH, Wu J, Yan Y, et al. Detection of pitch modulation information hiding based on codebook correlation network [J]. Chinese Journal of Computers, 2022, 45 (6): (1308-1325.) and subframe-level integer pitch delay and frame-level fractional pitch delay are spatiotemporal feature fusion (Tian H, Qiu Y, MazurczykW, et al. STFF-SM: steganalysis model based on spatial and temporal featurefusion for speech streams [J]. IEEE / ACM Transactions on Audio, Speech, and Language Processing, 2023, 31: 277-289.). However, different steg durations and speech content are also one of the problems causing the detection model to fail in localization. That is, when a steg detection model trained only on long-term embeddings or specific languages ​​is directly used for short-term burst embeddings or cross-language speech, the model performance will drop sharply. This is because the success of a typical steg detection model depends on the steg trace having a sufficiently long accumulation time window, but in actual covert communication, attackers often use burst patterns to evade detection, which leads to insufficient temporal context. At the same time, since steg samples are far fewer than normal samples, data imbalance further exacerbates the overfitting of the localization model. Summary of the Invention

[0005] The purpose of this invention is to provide a speech steganalysis localization method based on a dynamic mask bidirectional state encoder. This method utilizes a bidirectional selective state space network and a fuzzy attention mechanism to constrain the training of the steganalysis localization network. It captures long-distance temporal dependencies through bidirectional state space scanning and injects learnable Gaussian noise into the attention score to simulate the uncertainty of steganalysis noise, thus avoiding poor localization results due to weak traces and blurred boundaries in single frames. Based on this, it maximizes the amplification of steganalysis anomalies accumulated across frames and then combines dynamic frame-level confidence masks to suppress interference from unreliable frames, constructing a sub-frame-level steganalysis localization network to achieve accurate localization of ACB domain steganalysis embedding bits.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] On the one hand, a speech steganalysis localization method based on a dynamic mask bidirectional state encoder includes the following steps:

[0008] For each subframe in the speech coding stream, extract its integer pitch delay and fractional pitch delay. Concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector. Stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence. Generate corresponding subframe-level labels based on the steganography mask to obtain labeled parameter sequence samples.

[0009] A localization network is constructed, comprising a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence. The gated feedforward enhancement module contains two cascaded gated feedforward network units to enhance local nonlinear features. The dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks to capture long-range bidirectional temporal dependencies in the pitch delay sequence and dynamically generate frame-level masks. The fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise. The output localization layer comprises, in sequence, a von Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, outputting the steganalytic probability for each subframe.

[0010] Labeled parameter sequence samples are input into the localization network in batches for training to obtain a trained localization network;

[0011] The speech samples to be localized are acquired, and after feature concatenation and normalization, the extracted features are input into the trained localization network. The steganalysis probability of each sample subframe is output, and subframes with probabilities greater than a preset threshold are identified as steganalysis embedding bits, generating subframe-level localization mask results.

[0012] Preferably, the preset number of subframes is 4; the integer pitch delay and fractional pitch delay of every 4 consecutive subframes are concatenated to form an 8-dimensional feature vector.

[0013] Preferably, the feature fusion layer maps the input features to a high-dimensional space, specifically including:

[0014] The 8-dimensional feature input is upscaled to a 128-dimensional high-dimensional space through linear mapping.

[0015] Preferably, in the gated feedforward enhancement module, each gated feedforward network unit includes three parallel paths; the first path is an identity mapping, which transmits the output features of the feature fusion layer to the output end; the second path is the backbone nonlinear transformation branch, where the output features of the feature fusion layer are first expanded in dimension by the first linear layer, and then nonlinear transformation features are introduced through the activation function GELU; the third path is a gated branch, where the output features of the feature fusion layer are expanded by the gated linear layer and then compressed to the (0,1) interval by the activation function Sigmoid to generate a gated signal; the nonlinear transformation features output by the second path are multiplied element-wise with the gated signal output by the third path, and the multiplication result is then mapped back to the same dimension space as the input by the second linear layer; the features retained by the identity mapping of the first path are added to the mapping result after the second linear layer to form a residual connection.

[0016] Preferably, each dynamic mask bidirectional state block includes a forward state space unit, a backward state space unit, a dynamic frame-level confidence prediction header, a diagonal jump connection, a residual connection, and layer normalization;

[0017] The forward state space unit performs a forward scan of the output of the gated feedforward enhancement module to obtain the forward output sequence. ;

[0018] The backward state space unit performs a reverse scan of the output of the gated feedforward enhancement module to obtain the backward output sequence. ;

[0019] The dynamic frame-level confidence prediction head is a two-layer fully connected network that takes the current frame features as input and outputs a one-dimensional confidence score. After calculating the confidence score for all frames, a frame-level mask is obtained. ;

[0020] The final output of each dynamic mask bidirectional state block is as follows:

[0021] ;

[0022] in, Representation layer normalization; This indicates the output of the gated feedforward enhancement module; shift indicates a right shift operation by one bit. Indicates learnable diagonal parameters; This indicates that the sequence is flipped along the time axis; This indicates a residual connection.

[0023] Preferably, the fuzzy multi-head self-attention module includes several attention heads, each of which adds Gaussian noise with learnable parameters to the scaled dot product attention score; then, the outputs of each attention head are concatenated along the feature dimension and linearly transformed by the output projection layer to generate the final attention representation.

[0024] Preferably, the Neumann boundary low-rank adaptive linear layer NB-LoRA constructs a semi-orthogonal matrix through Cayley transformation and applies norm constraints to the singular values. Its weight matrix is ​​constructed as follows:

[0025] ;

[0026] in, and These are the two projection matrices in the low-rank decomposition; It is a diagonal matrix; This is the projection matrix on the input side; This is the projection matrix on the output side.

[0027] Preferably, during the training of the localization network, a fusion loss function is used to constrain the network, wherein the fusion loss function is a weighted sum of label smoothing binary cross-entropy loss and Focal loss.

[0028] On the other hand, a speech steganalysis localization device based on a dynamic mask bidirectional state encoder includes:

[0029] The data preprocessing and feature fusion module is used to extract the integer pitch delay and fractional pitch delay of each subframe in the speech coding stream, concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector, stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence, generate corresponding subframe-level labels according to the steganography mask, and obtain labeled parameter sequence samples.

[0030] A localization network construction module is used to construct a localization network, which includes a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence. The feature fusion layer maps input features to a high-dimensional space. The gated feedforward enhancement module contains two cascaded gated feedforward network units to enhance local nonlinear features. The dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks to capture long-range bidirectional temporal dependencies in the pitch delay sequence and dynamically generate frame-level masks. The fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise. The output localization layer includes a Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, outputting the steganalytic probability of each subframe.

[0031] The localization network training module is used to input labeled parameter sequence samples into the localization network in batches for training, and obtain the trained localization network.

[0032] The steganalysis localization module is used to acquire the speech samples to be localized, perform feature concatenation and normalization, input the extracted features into the trained localization network, output the steganalysis probability of each sample's subframe, determine the subframes with probabilities greater than a preset threshold as steganalysis embedding bits, and generate subframe-level localization mask results.

[0033] On the other hand, an electronic device includes:

[0034] One or more processors;

[0035] Storage device for storing one or more programs.

[0036] When the one or more programs are executed by the one or more processors, the one or more processors implement the speech steganalysis localization method based on a dynamic mask bidirectional state encoder.

[0037] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention presents a speech steganalysis localization method based on a dynamic mask bidirectional state encoder, which can achieve high-precision steganalysis localization at the sub-frame level. Specifically, it employs a bidirectional state encoder to capture long-distance bidirectional temporal dependencies in the pitch delay sequence, solving the problem of weak and difficult-to-detect steganalysis traces in a single frame. Furthermore, it introduces a dynamic frame-level mask to adaptively suppress interference from unreliable frames such as unvoiced sounds and noise, effectively improving robustness to short-term burst steganalysis. A fuzzy multi-head self-attention module is used to simulate the uncertainty of steganalysis noise, effectively handling the boundary ambiguity between steganalysis features and natural speech fluctuations. A von Neumann boundary low-rank adaptive linear layer applies low-rank and norm constraints to the fully connected layer, suppressing overfitting caused by data imbalance and improving the generalization ability of the localization network model. The localization granularity is finer than existing frame-level detection methods, achieving excellent localization performance on both Chinese and English datasets. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the speech steganalysis localization method based on a dynamic mask bidirectional state encoder provided in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the localization network model of the method in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of a gated feedforward network unit provided in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the structure of the dynamic mask bidirectional state encoder provided in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of the structure of the fuzzy multi-head self-attention module provided in an embodiment of the present invention;

[0044] Figure 6 This is a schematic diagram of the output positioning layer provided in an embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the positioning results in an embodiment of the present invention;

[0046] Figure 8 A structural block diagram of a speech steganalysis localization device based on a dynamic mask bidirectional state encoder provided in an embodiment of the present invention;

[0047] Figure 9 This is a schematic diagram of the structure of a computer device suitable for implementing the electronic devices of the present application. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] In the description of this invention, it should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0050] like Figure 1 As shown in the figure, this embodiment discloses a speech steganalysis localization method based on a dynamic mask bidirectional state encoder, which includes the following steps.

[0051] S101, extract the integer pitch delay and fractional pitch delay of each subframe in the speech coding stream, concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector, stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence, generate the corresponding subframe-level label according to the steganography mask, and obtain the labeled parameter sequence sample.

[0052] This embodiment will use AMR-NB speech encoding as an example for explanation, including obtaining the encoding parameters of each frame in the AMR-NB encoded speech stream, and using a sliding window to obtain the encoding parameter sequence with a length of 50 frames, i.e., 1.0 second of speech duration; the window length is based on the AMR-NB frame structure, i.e., each frame contains 4 subframes in 20ms, and 50 frames cover 200 subframes, which is sufficient to capture the short-term stability of pitch delay and cross-frame accumulation anomalies, and meets the temporal accumulation requirements of steganographic traces and the computational efficiency of real-time evidence collection; then the integer pitch delay and fractional pitch delay of every four consecutive subframes are concatenated into an 8-dimensional feature vector and Z-score normalization is performed.

[0053] This step specifically includes: setting up burst steganography scenarios, fusing integer and fractional pitch delays, and Z-score normalization.

[0054] S1011, Burst Steganography Scenario Setup. To simulate the non-uniform and bursty characteristics of information transmission in actual covert communication when constructing the training and test sets, a burst steganography scenario is introduced. The burst steganography scenario is controlled by two core parameters: "steganography interval (Gap)" and "steganography duration (T)" to control the macroscopic distribution of the mask. Gap represents the idle time length between two consecutive embedding bursts, and T represents the steganography duration within each burst. In this embodiment, the steganography interval is set to 60s, the steganography duration is 0.1s~1.0s, the step size is 0.1s, and the steganography duration is random, meaning the steganography duration is not fixed and can be any length from 0.1s to 1.0s. Starting from the first subframe, a subframe segment of length Gap is skipped, during which no steganography is performed; then, a subframe segment of length T is continuously labeled, at which point the embedding rate within this time length is 100%, indicating the entry into a steganography burst; the "skip-label" process is repeated until all subframes are covered.

[0055] Binary mask generation. A binary positioning mask is set to indicate which subframes participated in the embedding. "1" indicates that the subframe was used to carry secret information, and "0" indicates that the subframe did not embed secret information. Let the speech sample be... ,have , , The number of speech samples. For each speech sample, a corresponding subframe-level mask sequence is generated based on the burst steganography parameters. The mask length is equal to the total number of subframes contained in that speech sample.

[0056] Sliding window processing. To segment long-duration speech streams into analysis units of uniform scale and increase the number of samples available for model training, a sliding window processing method is applied to the generated speech coding parameter sequence and the corresponding mask sequence. The window length is set to 1.0 second, corresponding to 200 subframes (50 frames), and the window step size is set to 20 milliseconds, corresponding to 4 subframes (one complete speech frame). During processing, fixed-length window segments are sequentially extracted starting from the beginning of the mask sequence. For each window, it is checked whether the mask within the window contains at least one subframe marked with "1", i.e., whether the window contains steganalytic samples. Only windows that meet this condition are considered valid and are retained for subsequent model training and testing.

[0057] For a valid window, the subframe-level mask of that window needs to be aligned with the encoding parameters stored frame by frame. Since the smallest unit of ACB field steganography is the subframe, the mask uses the subframe as the smallest annotation unit. However, speech encoding parameters are usually organized and stored in the parameter file frame by frame (each frame contains 4 subframes) during actual encoding and decoding. Therefore, the conversion relationship between subframe index and frame index (the first subframe index) needs to be considered. Frame corresponding to subframe index to The process involves determining the 50 speech frames covered by the window and extracting the corresponding 50-frame parameter sequence from the parameter file. The parameter file is text obtained by decoding and parsing the AMR-NB encoded speech stream. Each frame's parameters include Linear Prediction (LSP) coefficients, Adaptive Codebook (ACB) parameters, and Fixed Codebook (FCB) parameters. The ACB parameters include integer pitch delay and fractional pitch delay, as well as adaptive codebook gain. The core purpose of extracting the frame parameter rows is to spatiotemporally align the frame-by-frame encoded parameters with the subframe-by-subframe labeled steganalytes, constructing "feature-label" paired samples for subsequent localization network training. Ultimately, each effective window contains 50 frames of ACB parameters and corresponding 50×4 subframe-level labels.

[0058] S1012, Integer and fractional pitch delay fusion. Obtain the ACB parameter sequence of each effective window after sliding window processing, fuse the integer and fractional pitch delays of every four consecutive subframes to form a fixed-length input feature sequence, and perform feature normalization processing.

[0059] (1) Feature vector construction. For each effective window after sliding window processing, the integer pitch delay and fractional pitch delay of every four consecutive subframes (i.e., one speech frame) are concatenated into an 8-dimensional feature vector:

[0060] ;

[0061] Here, superscripts 0-3 represent the four subframe numbers within that frame. Since the fractional pitch delay has a small range, typically -3 to 3, and has a deterministic encoding mapping relationship with the integer part, using both as input features allows the model to learn the anomaly in the joint integer-fractional distribution caused by steganography. Therefore, each effective window contains 50 frame-level vectors. Stacked in chronological order, forming a feature matrix. The corresponding subframe-level tag. The mask fragment corresponding to the sliding window is directly provided as a supervision signal to guide the model in learning subframe-level steganographic embedding bits.

[0062] (2) Z-score normalization. Calculate the mean and standard deviation for all samples in the training set, and then perform Z-score normalization on the input sequence:

[0063] ;

[0064] in, The feature matrix generated in the feature vector construction step, and These represent the mean and standard deviation of each dimension of the training set. To prevent division by zero by small constants. This operation eliminates the influence of amplitude differences between different speech samples, making model training more stable.

[0065] S102, Construct a localization network, which includes a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence.

[0066] like Figure 2 As shown, the feature fusion layer maps input features to a high-dimensional space; the gated feedforward enhancement module contains two cascaded gated feedforward network units to enhance local nonlinear features; the dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks to capture long-range bidirectional temporal dependencies in the pitch delay sequence and dynamically generate frame-level masks; the fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise; the output localization layer sequentially includes a von Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, outputting the steganalytic probability of each subframe.

[0067] Specifically, the integer pitch delay variation caused by ACB domain steganography is extremely small, only ±1. This weak perturbation highly overlaps with normal pitch fluctuations in the original 8-dimensional parameter space, making it difficult for the model to perceive effectively. The feature fusion layer upscales the 8-dimensional input to a 128-dimensional high-dimensional space through linear mapping, dispersing the subtle differences caused by steganography into a high-dimensional manifold, providing more discriminative feature representations for subsequent modules.

[0068] In the feature fusion layer, the normalized ACB input feature sequence is used. As input, it is mapped from the original 8-dimensional feature space to a high-dimensional fusion space through a linear mapping, and the output dimension is... .

[0069] The specific implementation is as follows.

[0070] First, define the learnable weight matrix. and bias vector Weight matrix and bias vector The initial values ​​are initialized using Kaiming uniform initialization (adapted for the GELU activation function) and automatically updated and optimized during training via gradient descent. Then, the input features are... Perform a linear transformation:

[0071] ;

[0072] in, This layer maps the original 8-dimensional ACB parameters to a high-dimensional space, providing richer representation capabilities for subsequent deep feature extraction.

[0073] Furthermore, steganalytic traces manifest as localized, minute numerical fluctuations in the time domain, which are difficult to distinguish effectively from normal coding fluctuations using conventional linear transformations. Moreover, the requirements for nonlinear transformations vary significantly across different speech segments. The gated feedforward enhancement module introduces a gating mechanism, dynamically controlling the intensity of the nonlinear transformation through a learnable gating signal. This allows the model to enhance nonlinearity in voiced segments with stable pitches to capture subtle anomalies, while suppressing nonlinearity in randomly fluctuating unvoiced segments to avoid misjudgments.

[0074] The gated feedforward enhancement module comprises two cascaded gated feedforward network units, such as... Figure 3 As shown, each unit implements feature enhancement using a three-way parallel data flow structure. The input features of the module are the output features of the feature fusion layer. This tensor contains 50 time steps, each corresponding to a 128-dimensional feature vector. Input features The input features are simultaneously fed into three parallel computation paths. The first path is an identity mapping, directly applying the input features... The signal is transmitted to the output as a reference signal for the residual connection. The second path is the main nonlinear transform branch, with input characteristics... First, a first linear layer expands the dimension from 128 to 512 to provide a wider representation space. Then, a non-linear transformation feature is introduced through the GELU activation function, which preserves non-zero gradients in the negative region, helping to maintain feature richness. The third path is a gated branch, which takes the input features... After being extended to 512 dimensions by another independent gated linear layer, it is compressed to the (0,1) interval by the activation function Sigmoid to generate a gated signal. Each element of this signal independently controls the nonlinear transformation intensity of the corresponding position feature.

[0075] Subsequently, the nonlinear transformation feature output from the second path is multiplied element-wise with the gated signal output from the third path. This allows each feature dimension to independently determine its contribution to the final output. When the value is close to 1, the feature fully participates in the discrimination; when the value is close to 0, the feature is suppressed, thus achieving adaptive response to different speech segments. The multiplication result is then passed through a second linear layer to map the dimension from 512 back to 128, restoring it to the same dimensional space as the input.

[0076] Finally, the original input features preserved by the first path identity mapping are... The result is added to the mapping result after the second linear layer to form a residual connection. This structure ensures the original input features are preserved. Information can be directly passed to subsequent network layers, avoiding the gradient vanishing problem caused by deep stacking, and enabling the gated feedforward enhancement module to stably update parameters during training. Two gated feedforward enhancement modules are stacked in series, with the output of the first module serving as the input of the second identical module, further enhancing the model's ability to fit complex nonlinear steganalytic patterns. In actual operation, when faced with voiced speech segments with stable pitch, the gate signal tends to enhance the nonlinear transformation to capture subtle steganalytic traces; when faced with randomly fluctuating unvoiced speech segments, the gate signal tends to suppress the nonlinear transformation to avoid misjudging natural jitter as steganalytic anomalies.

[0077] Features are obtained through alternating processing by two gated feedforward network units. This allows the network to effectively suppress random fluctuations in unvoiced segments while perceiving minute steganalytic distortions, enabling the subsequent dynamic mask bidirectional state encoder module to achieve a higher steganalytic trace signal-to-noise ratio at the feature level.

[0078] Furthermore, steganographic traces of pitch delays require accumulation over a time window of tens of frames to become detectable, but unidirectional temporal models, constrained by gradient limitations, struggle to capture long-range dependencies. Simultaneously, random fluctuations in unreliable frames such as unvoiced or silent frames severely contaminate temporal features. The dynamic mask bidirectional state encoder achieves bidirectional long-range dependency capture through forward and backward state space units, allowing weak anomalies in a single frame to be accumulated and amplified within the complete temporal context. It also incorporates a dynamic frame-level confidence prediction head to generate frame-level masks in real-time to suppress interference from unreliable frames.

[0079] In the dynamic mask bidirectional state encoder, the output of the gated feedforward enhancement module will be... As input, such as Figure 4 As shown, long-distance bidirectional temporal dependencies in the pitch delay sequence are captured by two stacked dynamic mask bidirectional state blocks, and frame-level masks are dynamically generated to suppress interference from unreliable frames. The specific implementation is as follows.

[0080] Specifically, each dynamic mask bidirectional state block includes a forward state space unit, a backward state space unit, a dynamic frame-level confidence prediction header, a diagonal jump connection, a residual connection, and layer normalization.

[0081] First, the forward state space unit performs a forward scan of the input sequence, and its internal state update formula is:

[0082] ;

[0083] ;

[0084] in, The state matrix, Given the state matrix of the previous time step, the initial state is... It is a zero matrix, in this embodiment ; is the gating coefficient, where This represents a learnable linear projection layer used to project the input... The values ​​are mapped to scalar gate values, then compressed to the (0,1) interval using the Sigmoid function to control historical states. Regarding the current state The extent of the impact That is, from the input sequence The 128-dimensional feature vector of the t-th frame is obtained frame by frame by the preceding module (S1022); For query vectors (Query). Key vector. Each of the three is a value vector, and all three are derived from the input. Obtained through learnable linear projection, with all projection dimensions being... .

[0085] State matrix By iteratively updating at each time step, historical time-series information is continuously accumulated and passed on, resulting in the final output of each time step. All of them incorporate all contextual information from the starting frame to the current frame; the forward output sequence obtained after the forward scan is completed. (its first) The output of each time step is This will participate in the final output formula. calculate, These are internal state variables used for support. Frame-by-frame production; It is the output of the forward scan and directly participates in the final output calculation of the block.

[0086] Then, the backward state space unit processes the input sequence. Perform a reverse scan:

[0087] ;

[0088] in, This indicates that the sequence is flipped along the time axis, and SSD represents the forward scan operation of the forward state space unit. Here, the backward scan is essentially applying the same forward scan operation to both the original sequence and the flipped sequence, achieving bidirectional temporal information capture. The output sequence obtained by the backward scan, its first... The output of each time step is denoted as .

[0089] The dynamic frame-level confidence prediction head is a two-layer fully connected network, using the current frame feature vector. As input, output a one-dimensional confidence score. The frame-level mask is obtained after calculating all frames. :

[0090] ;

[0091] in, Mapping 128 dimensions to 64 dimensions Mapping 64 dimensions to 1 dimension; GELU (Gaussian Error Linear Unit) is the activation function, preserving non-zero gradients in the negative region; The Sigmoid function compresses the output to... The range. This mask. Used to suppress the output of unreliable frames.

[0092] The final output of each dynamic mask bidirectional state block for:

[0093] ;

[0094] in, (Residual connection) directly passes the original input of the block to the output, ensuring that the original feature information can be directly passed without complex temporal transformations, avoiding gradient vanishing; shift is a right shift operation (padded with zeros at the beginning), used to align information between adjacent frames; input With learnable diagonal parameters Element-wise multiplication provides independent skip connection weights for each feature dimension. This learnable diagonal parameter... As trainable parameters of the model, they are initialized with all 1s to ensure that diagonal skip connections remain an identity mapping during the early stages of training, thus avoiding the introduction of additional bias. During model training... Automatic updates and optimizations are achieved through backpropagation and gradient descent algorithms, enabling the network to adaptively adjust the skip connection strength of each feature dimension according to the steganalysis task, thereby selectively preserving and enhancing the original input features; LN stands for layer normalization, which standardizes the sum of the above four terms.

[0095] Furthermore, the pitch parameter changes caused by steganography completely overlap with the normal pitch period fluctuations in numerical range, and there is no clear distinction between the two. Hard decision-making is prone to false alarms or missed detections. The fuzzy multi-head self-attention module injects learnable Gaussian noise into the self-attention score, making the localization network output a softened probability value to express the boundary uncertainty; by using multi-head attention to simultaneously model local transitions and long-period patterns from multiple subspaces, the steganography probability of each subframe is comprehensively determined.

[0096] In the fuzzy multi-head self-attention module, the encoded features are... As input, learnable Gaussian noise is added to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise, such as... Figure 5 As shown.

[0097] First, the input features The query matrix is ​​calculated using three learnable linear projection matrices. Key matrix Sum matrix In this embodiment, the number of heads in the multi-head self-attention system is 8, and the dimension of each head is [dimension not specified]. ,therefore:

[0098] ;

[0099] in, The learnable weight matrix is ​​projected and then split into 8 parts. The head matrix.

[0100] Then, for each attention head, a scaled dot product attention score is calculated, and learnable Gaussian noise is added to the score:

[0101] ;

[0102] Among them, subscript Indicates the first One attention point ( ), For each dimension of attention head, The fuzzy parameter is a learnable parameter, initially set to 0.1. For the standard Gaussian noise matrix (and (The shapes are the same). The purpose of adding noise is to simulate the uncertainty of steganalytic noise, so that the model outputs a softened probability value.

[0103] The outputs of the eight attention heads are concatenated along the feature dimension to obtain... Then through the output projection layer Perform a linear transformation:

[0104] ;

[0105] in, The discard rate for this module is set to 0.1.

[0106] Furthermore, in the output localization layer, a von Neumann boundary low-rank adaptive linear layer is employed. By imposing low-rank constraints and spectral norm constraints on the weight matrix, the effective number of parameters is limited to a compact low-dimensional subspace. This forces the network to focus on the few feature combinations most critical for steganalysis, thereby improving the model's ability to distinguish stegana frames under conditions of extreme imbalance between positive and negative samples, effectively alleviating the overfitting problem. Figure 6 As shown, the specific implementation is as follows.

[0107] First, the Neumann boundary low-rank adaptive linear layer (NB-LoRA) constructs a semi-orthogonal matrix through Cayley transformation and applies norm constraints to the singular values. Its weight matrix is ​​constructed as follows:

[0108] ;

[0109] in, and These are the two projection matrices in the low-rank decomposition. For the input feature dimension, To output feature dimensions, It is a diagonal matrix whose diagonal elements are singular value vectors. . Specifically, The projection matrix on the input side is responsible for... 3D input features compressed to Low-rank space; The projection matrix on the output side is responsible for... Mapping of low-rank space to Dimensional output space; In this embodiment, the dimension is a low-rank space. , This refers to the input feature dimension of the NB-LoRA layer (i.e., the output feature dimension of the S1024 module). This represents the output feature dimension of the NB-LoRA layer. and From the free parameters respectively through Cayley transform and Obtain, guarantee [ [A semi-orthogonal matrix; free parameters] and Kaiming uniform initialization is used to generate random values, which are then automatically updated and optimized during model training using backpropagation and gradient descent algorithms. Furthermore, singular value vectors... Scaling to norm bounds after Softmax transformation Inside.

[0110] In this embodiment, rank Norm boundary =0.1, norm type (Frobenius norm). This constraint ensures that the weight matrix satisfies and ,in Represents the rank of a matrix The low-rank constraint is the maximum number of linearly independent row or column vectors in a matrix, reflecting the effective information dimension contained in the matrix. Restrict the weight matrix to In a low-rank subspace, the network is forced to utilize only... The most important feature combination direction is used for discrimination, and the effective parameter quantity is determined by... Compress to In this embodiment, by Each parameter is compressed to The number of parameters has been reduced to about one-tenth of the original, significantly reducing the risk of overfitting and guiding the model to ignore redundant feature fluctuations unrelated to steganalysis, focusing only on the features most sensitive to steganalysis traces. This feature pattern effectively alleviates overfitting caused by data imbalance.

[0111] The NB-LoRA layer will blur the output of the multi-head self-attention module. As input, it is first processed through the weight matrix. Perform a linear transformation and add a bias term Mapping 128 dimensions to 256 dimensions, that is:

[0112] ;

[0113] in, The aforementioned weight matrix transpose, bias term Zero initialization is used, along with the weight matrix. Together with the learnable parameters, they are automatically updated and optimized during model training through backpropagation and gradient descent algorithms, providing an adjustable offset for each dimension of the output space and enhancing the model's fitting ability.

[0114] After passing through the GELU activation function and a Dropout layer with a dropout rate of 0.2, the input is then fed into the final fully connected layer, mapping the 256-dimensional features to a 4-dimensional output, as follows:

[0115] ;

[0116] Where GELU is the Gaussian error linear unit activation function, and Dropout is a random dropout layer. To output a fully connected layer, The unnormalized prediction value for each subframe.

[0117] Finally, the steganography probability of each subframe is obtained by the Sigmoid function, as follows:

[0118] ;

[0119] in, , The function is the Sigmoid function, which maps the input from the real number field to the interval (0,1), giving it a probabilistic meaning. The larger the probability value, the higher the probability that the subframe is embedded with secret information.

[0120] S103, the labeled parameter sequence samples are input into the localization network in batches for training to obtain the trained localization network.

[0121] Specifically, preprocessed labeled ACB parameter sequence samples are input into the localization network in batches for training. During training, a fusion loss function is used to constrain the network, employing a strategy combining label-smoothed binary cross-entropy loss and Focal loss. Label smoothing prevents the model from overconfident in normal samples, while Focal loss automatically reduces the weight of easily classified samples and focuses on difficult-to-classify steganalytic samples. The synergistic effect of these two methods enables the model to effectively learn steganalytic features even under conditions of extreme imbalance between positive and negative samples. Small Gaussian noise is added to the input data during training to force the model to learn more robust feature representations, further improving its generalization ability.

[0122] Specifically, the combined loss function is a weighted sum of the label smoothing binary cross-entropy loss and the Focal loss, as follows:

[0123] ;

[0124] Among them, label smoothing binary cross-entropy loss The calculation method is as follows:

[0125] ;

[0126] In this embodiment, take In this loss function, Use the sample index to traverse all subframes in all training samples; For the first The actual labels of each subframe (1 indicates that the subframe is steganized, and 0 indicates that it is not steganized); The model predicts the first The probability of steganography in each subframe; As a label smoothing factor, by hard label Softening the model to 0 or 1 prevents it from becoming overconfident in normal samples, thereby improving its generalization ability.

[0127] Focal loss The calculation method is as follows:

[0128] ;

[0129] In this embodiment, take , .in The model represents the first The prediction confidence of the true class of each subframe, when the true label is a steganalysis sample. hour Equal to predicted probability When the true label is a normal sample, that is... hour equal ; As the modulation factor, when When the modulation factor approaches 1, it approaches 0, automatically reducing the loss contribution of easily classified samples. When the modulation factor approaches 1, meaning the sample is difficult to classify, the loss contribution is preserved, thus focusing on the difficult-to-classify stegana samples. The category balance factor is used to adjust the weight ratio of positive and negative samples. To focus on parameters, the decay rate of the modulation factor is controlled. Focal loss focuses on hard-to-classify samples, increasing the model's attention to steganalytic samples and alleviating the imbalance between positive and negative samples.

[0130] S104: Obtain the speech sample to be located, perform feature concatenation and normalization processing, input the extracted features into the trained localization network, output the steganalysis probability of each sample's subframe, determine the subframe with a probability greater than a preset threshold as a steganalysis embedding bit, and generate the subframe-level localization mask result.

[0131] Specifically, the target domain speech samples to be localized are acquired. Following steps S101, sliding window segmentation, integer and fractional pitch delay fusion, and Z-score normalization preprocessing are performed. The acoustic features are then input into the localization network trained in S103. The network outputs the steganalysis probability for each subframe. Subframes with probabilities greater than a preset threshold are identified as steganalysis embeddings, generating a subframe-level localization mask. This step applies the trained model to the actual localization task, discretizing continuous probability values ​​into a binary localization mask through threshold determination. A threshold of 0.5 is the unbiased default choice for probability determination; during training, the model is optimized to make the predicted probabilities as close as possible to the true labels.

[0132] It should be noted that this step does not require setting up sudden steganography scenarios or filtering valid windows when preprocessing the speech samples to be located.

[0133] In this embodiment, the voice sample to be located The acoustic features are obtained by performing sliding window segmentation, integer and fractional pitch delay fusion, and Z-score normalization preprocessing. .Will Input the trained localization network to obtain the prediction probability matrix. ,when At that time, the judgment of the first The first frame If a subframe is a steganographic embedding bit, then it is a normal subframe.

[0134] To evaluate the speech steganalysis localization method based on a dynamic mask bidirectional state encoder, a localization model was trained on a speech steganalysis detection dataset, and its performance was evaluated on a test set. The training and test sets were divided by file prefix numbers. File prefix numbers 1 to 34 were used as the training set (23,988 English speech samples with random steganalysis duration), and file prefix numbers 35 to 42 were used as the test set (5,661 English speech samples with random steganalysis duration). To intuitively evaluate the model's performance, this application adopted mainstream evaluation metrics: accuracy (Acc), F1 score, false negative rate (FNR), and false positive rate (FPR). Accuracy reflects the overall correctness of the model's judgment; F1 score is the harmonic mean of precision and recall, which comprehensively measures the accuracy and completeness of localization; false negative rate reflects the model's sensitivity to steganalysis, and a lower false negative rate means fewer steganalysis subframes are misclassified as normal; false alarm rate measures the model's false alarm level, and a lower false alarm rate means fewer normal subframes are misclassified as steganalysis. Figure 7 Experimental results demonstrate that the proposed speech steganalysis localization method based on a dynamic mask bidirectional state encoder exhibits significant advantages in localization performance. On a 1.0s steganalysis duration dataset (both Chinese and English), the F1 score exceeds 0.98, with a false alarm rate (FAR) below 0.02 and a FAR below 0.02. Even in an extremely short burst scenario of 0.1s, the F1 score remains above 0.88, validating the effectiveness of the dynamic frame-level confidence mask and fuzzy attention mechanism in handling weak steganalysis traces and short-duration burst steganalysis. The model demonstrates highly consistent performance on both Chinese and English datasets, showcasing excellent cross-language generalization capabilities.

[0135] like Figure 8 As shown, this embodiment also discloses a speech steganalysis localization device based on a dynamic mask bidirectional state encoder, comprising:

[0136] The data preprocessing and feature fusion module 801 is used to extract the integer pitch delay and fractional pitch delay of each subframe in the speech coding stream, concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector, stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence, generate corresponding subframe-level labels according to the steganography mask, and obtain labeled parameter sequence samples.

[0137] The localization network construction module 802 is used to construct a localization network, which includes a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence. The feature fusion layer maps input features to a high-dimensional space. The gated feedforward enhancement module contains two cascaded gated feedforward network units to enhance local nonlinear features. The dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks to capture long-range bidirectional temporal dependencies in the pitch delay sequence and dynamically generate frame-level masks. The fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise. The output localization layer includes a von Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, outputting the steganalytic probability of each subframe.

[0138] The localization network training module 803 is used to input labeled parameter sequence samples into the localization network in batches for training, and obtain the trained localization network.

[0139] The steganalysis localization module 804 is used to acquire the speech samples to be localized, perform feature concatenation and normalization processing, input the extracted features into the trained localization network, output the steganalysis probability of each sample's subframe, determine the subframes with probabilities greater than a preset threshold as steganalysis embedding bits, and generate subframe-level localization mask results.

[0140] The specific implementation process of a speech steganalysis localization device based on a dynamic mask bidirectional state encoder is basically the same as that of the above method embodiments, and will not be repeated here. It should be noted that each functional module in the device corresponds to a corresponding step in the method embodiments, and those skilled in the art can understand the specific implementation of the device based on the method flow.

[0141] like Figure 9 As shown, it illustrates a structural schematic diagram of a computer device 900 suitable for implementing electronic devices according to embodiments of the present application. Figure 9 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0142] Figure 9 In this computer device 900, a central processing unit (CPU) 901 and a graphics processing unit (GPU) 902 are included. These units can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 903 or programs loaded from storage section 909 into random access memory (RAM) 904. The RAM 904 also stores various programs and data required for the operation of the device 900. The CPU 901, GPU 902, ROM 903, and RAM 904 are interconnected via bus 905. An input / output (I / O) interface 906 is also connected to bus 905. The following components are connected to the I / O interface 906: an input section 907 including a keyboard, mouse, etc.; an output section 908 including a liquid crystal display (LCD) and speakers, etc.; a storage section 909 including a hard disk, etc.; and a communication section 910 including a network interface card such as a LAN card and a modem, etc. The communication section 910 performs communication processing via a network such as the Internet. A driver 911 may also be connected to the I / O interface 906 as needed. Removable media 912, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on drive 911 as needed so that computer programs read from them can be installed into storage section 909 as needed.

[0143] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including instructions for performing the processes. Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 910, and / or installed from removable medium 912. When the computer program is executed by central processing unit (CPU) 901 and graphics processing unit (GPU) 902, it performs the functions defined in the method of this application.

[0144] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium, a computer-readable medium, or any combination thereof. A computer-readable medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, or any combination thereof. More specific examples of a computer-readable medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution device, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than a computer-readable medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution device, apparatus, or apparatus. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, radio frequency (RF), or any suitable combination thereof.

[0145] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using dedicated hardware-based means to perform the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0147] The modules described in the embodiments of this application can be implemented in software or hardware. These modules can also be located within a processor.

[0148] The above embodiments have illustrated the basic principles and implementation methods of the present invention, aiming to help understand the core concept and key steps of the invention. It should be understood that these embodiments are merely examples and do not limit the scope of application of the present invention. Those skilled in the art, based on their understanding of the concept of the present invention, can make various equivalent improvements to specific steps, parameter configurations, or device structures. These improvements also fall within the protection scope of the present invention, as defined in the appended claims.

Claims

1. A speech steganalysis localization method based on a dynamic mask bidirectional state encoder, characterized in that, Includes the following steps: For each subframe in the speech coding stream, extract its integer pitch delay and fractional pitch delay. Concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector. Stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence. Generate corresponding subframe-level labels based on the steganography mask to obtain labeled parameter sequence samples. A localization network is constructed, comprising a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence; the feature fusion layer maps input features to a high-dimensional space; the gated feedforward enhancement module contains two cascaded gated feedforward network units for enhancing local nonlinear features; The dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks, which are used to capture long-distance bidirectional timing dependencies in the pitch delay sequence and dynamically generate frame-level masks. The fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise; the output localization layer sequentially includes a von Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, which outputs the steganalytic probability of each subframe. Labeled parameter sequence samples are input into the localization network in batches for training to obtain a trained localization network; The speech samples to be localized are acquired, and after feature concatenation and normalization, the extracted features are input into the trained localization network. The steganalysis probability of each sample subframe is output, and subframes with probabilities greater than a preset threshold are identified as steganalysis embedding bits, generating subframe-level localization mask results.

2. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, The preset number of subframes is 4; the integer pitch delay and fractional pitch delay of every 4 consecutive subframes are concatenated to form an 8-dimensional feature vector.

3. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 2, characterized in that, The feature fusion layer maps the input features to a high-dimensional space, specifically including: The 8-dimensional feature input is upscaled to a 128-dimensional high-dimensional space through linear mapping.

4. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, In the gated feedforward enhancement module, each gated feedforward network unit includes three parallel paths; the first path is an identity mapping, which transmits the output features of the feature fusion layer to the output end; The second path is the main nonlinear transformation branch. The output features of the feature fusion layer are first expanded in dimension by the first linear layer, and then nonlinear transformation features are introduced through the GELU activation function. The third path is the gated branch. The output features of the feature fusion layer are expanded by the gated linear layer and then compressed to the (0,1) interval by the Sigmoid activation function to generate a gated signal. The nonlinear transformation features output by the second path are multiplied element-wise with the gated signal output by the third path. The multiplication result is then mapped back to the same dimension space as the input by the second linear layer. The features preserved by the identity mapping of the first path are added to the mapping result after the second linear layer to form a residual connection.

5. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, Each dynamic mask bidirectional state block includes a forward state space unit, a backward state space unit, a dynamic frame-level confidence prediction header, a diagonal jump connection, a residual connection, and layer normalization. The forward state space unit performs a forward scan of the output of the gated feedforward enhancement module to obtain the forward output sequence. ; The backward state space unit performs a reverse scan of the output of the gated feedforward enhancement module to obtain the backward output sequence. ; The dynamic frame-level confidence prediction head is a two-layer fully connected network that takes the current frame features as input and outputs a one-dimensional confidence score. After calculating the confidence score for all frames, a frame-level mask is obtained. ; The final output of each dynamic mask bidirectional state block is as follows: ； in, Representation layer normalization; This indicates the output of the gated feedforward enhancement module; shift indicates a right shift operation by one bit. Indicates learnable diagonal parameters; This indicates that the sequence is flipped along the time axis; This indicates a residual connection.

6. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, The fuzzy multi-head self-attention module contains several attention heads. Each attention head adds Gaussian noise with learnable parameters to the scaled dot product attention score. Then, the outputs of each attention head are concatenated along the feature dimension and linearly transformed by the output projection layer to produce the final attention representation.

7. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, The Neumann boundary low-rank adaptive linear layer NB-LoRA constructs a semi-orthogonal matrix through Cayley transformation and applies norm constraints to singular values. Its weight matrix is ​​constructed as follows: ； in, and These are the two projection matrices in the low-rank decomposition; It is a diagonal matrix; This is the projection matrix on the input side; This is the projection matrix on the output side.

8. The speech steganalysis localization method based on a dynamic mask bidirectional state encoder according to claim 1, characterized in that, During the training of the localization network, a fusion loss function is used to constrain the network. The fusion loss function is a weighted sum of the label smoothing binary cross-entropy loss and the Focal loss.

9. A speech steganalysis localization device based on a dynamic mask bidirectional state encoder, characterized in that, include: The data preprocessing and feature fusion module is used to extract the integer pitch delay and fractional pitch delay of each subframe in the speech coding stream, concatenate the integer pitch delay and fractional pitch delay of each consecutive preset subframe to form a feature vector, stack all feature vectors in the sliding window in chronological order and normalize them to obtain the input feature sequence, generate corresponding subframe-level labels according to the steganography mask, and obtain labeled parameter sequence samples. A localization network construction module is used to construct a localization network, which includes a feature fusion layer, a gated feedforward enhancement module, a dynamic mask bidirectional state encoder, a fuzzy multi-head self-attention module, and an output localization layer connected in sequence. The feature fusion layer maps the input features to a high-dimensional space. The gated feedforward enhancement module contains two cascaded gated feedforward network units for enhancing local nonlinear features. The dynamic mask bidirectional state encoder contains several stacked dynamic mask bidirectional state blocks, which are used to capture long-distance bidirectional timing dependencies in the pitch delay sequence and dynamically generate frame-level masks. The fuzzy multi-head self-attention module adds learnable Gaussian noise to the standard scaled dot product attention score to simulate the uncertainty of steganalytic noise; the output localization layer sequentially includes a von Neumann boundary low-rank adaptive linear layer NB-LoRA, a Gaussian error linear unit activation function, a random dropout layer, a fully connected layer, and a Sigmoid function, which outputs the steganalytic probability of each subframe. The localization network training module is used to input labeled parameter sequence samples into the localization network in batches for training, and obtain the trained localization network. The steganalysis localization module is used to acquire the speech samples to be localized, perform feature concatenation and normalization, input the extracted features into the trained localization network, output the steganalysis probability of each sample's subframe, determine the subframes with probabilities greater than a preset threshold as steganalysis embedding bits, and generate subframe-level localization mask results.

10. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the speech steganalysis localization method based on a dynamic mask bidirectional state encoder as described in any one of claims 1 to 8.

Images