A method and device for detecting a forgery attack on an industrial field voice interaction link

Abstract

The application discloses a kind of industrial field voice interaction link's forgery attack detection method and device, it is related to industrial high-end intelligent safety equipment technical field, and its technical solution key points include the following steps: obtaining original audio signal to be detected;The original audio signal is input into pre-trained cross-language self-supervised speech model, and the frame-level hidden layer representation of each Transform layer of the cross-language self-supervised speech model is extracted;The improved residual coding module is used to extract the multi-scale features of the fusion features by the back-end classification network, and the local geometric structure and global context dependence are cooperatively modeled by the double-branch graph attention module to the multi-scale features, and the audio authenticity confidence score is output;According to the audio authenticity confidence score, it is determined that the original audio is real voice or fake voice, and the effect is that various voice forgery attacks can be stably identified in low signal-to-noise ratio industrial field environment.

Description

Technical Field

[0001] This invention relates to the field of high-end intelligent security equipment technology in industry, and more specifically, to a method and device for detecting spoofing attacks on voice interaction links in industrial settings. Background Technology

[0002] Most existing industrial voice detection devices on the market use single-layer voice feature extraction chips, which limit the hardware computing power and signal processing unit performance. These devices are unable to adapt to complex industrial environments such as strong mechanical noise in workshops and multipath interference from underground wireless communication. In factory automated production lines and underground explosion-proof communication scenarios, wireless transmission multipath interference and high-power mechanical noise severely interfere with the signal characterization extraction of low-end detection devices with built-in single-layer voice processing chips. The insufficient computing power of low-end hardware makes it easy to miss subtle acoustic fingerprints of forged voices, and the device cannot distinguish between noise-distorted real human voices and maliciously forged voices. Traditional fixed-weight feature fusion methods cannot dynamically adjust the contribution of each layer of features to frame-by-frame industrial interference. Features at levels severely contaminated by noise continuously interfere with detection operations, reducing detection accuracy in extreme industrial scenarios. Furthermore, traditional detection equipment only carries simple, shallow feature inference units. The hardware cannot simultaneously and in parallel model local acoustic details and long-term global features of the voice, making it difficult for the device to simultaneously resist the dual security threats of industrial wireless channel interference and voice forgery, resulting in poor equipment detection stability. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method and apparatus for detecting forgery attacks on industrial field voice interaction links.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for detecting spoofing attacks on industrial field voice interaction links includes the following steps:

[0006] Acquire the raw audio signal to be detected;

[0007] The original audio signal is input into a pre-trained cross-lingual self-supervised speech model, and the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model are extracted.

[0008] Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representation of each Transformer layer is used as the key vector and value vector. The fusion weight of each layer is adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representation of each Transformer layer is weighted and fused to obtain the fused feature.

[0009] The fused features are input into the backend classification network. The backend classification network extracts multi-scale features from the fused features through an improved residual coding module, and then uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, outputting an audio authenticity confidence score.

[0010] The original audio is determined to be either real or fake speech based on the audio authenticity confidence score.

[0011] Preferably, the method further includes performing a quantitative evaluation of the discrimination potential of the frame-level hidden layer representations of each Transformer layer, calculating the maximum deviation of the cumulative distribution function of the true and false sample feature distributions of each layer, and screening hidden layer representations with discrimination potential based on a critical threshold of a preset significance level.

[0012] The hidden layer representations with discriminative potential are used as the frame-level hidden layer representations of each layer in the input frame-level attention fusion module.

[0013] Preferably, the critical threshold The calculation formula is:

[0014] ;

[0015] in, and These represent the sample sizes of the real and fake samples, respectively. significance level The corresponding critical value coefficient.

[0016] Preferably, the frame-level attention fusion module employs a multi-head attention mechanism for each time frame. The calculation process for single-head attention includes:

[0017] The extracted first Constructing query vectors from hidden layer representations at the frame level ,by Layer-by-frame hidden layer representation construction of key vectors Sum value vector ;

[0018] Calculate the first by scaling the dot product attention. Output of each attention head: ;

[0019] in, ; The dimension of the key vector. Indicates the first The attention head corresponds to the first The value vector of the t-th time frame output by the Transformer layer;

[0020] The fused features of this time frame are obtained by concatenating the outputs of all attention heads and performing a linear transformation: , , , For learnable projection matrix;

[0021] The fusion features of each time frame The fused features are obtained by aggregating along the time dimension.

[0022] Preferably, the improved residual coding module is built on the Res2Net architecture. By inputting the fused features into the residual coding module, the fused features are grouped by channel and subjected to hierarchical convolution to obtain hierarchical convolutional features.

[0023] An attention feature fusion module is embedded in the residual module. The resulting hierarchical convolutional features are dynamically weighted through the attention mechanism. The contribution of each group of features is adaptively adjusted according to the input feature distribution. Non-key features are suppressed while highlighting key features related to forgery, resulting in dynamically weighted features for each group.

[0024] By a top-down approach, the dynamically weighted features output from each layer of Res2Net are aggregated to obtain multi-scale features that contain global information across multiple time series and frequency scales.

[0025] Preferably, the bi-branch graph attention module takes multi-scale features as input, and the bi-branch graph attention module includes a global branch and a local branch:

[0026] The global branch calculates global attention weights by performing dot product operations between nodes and the softmax activation function on multi-scale features, thereby achieving attention interaction among all nodes in the global scope and obtaining the global branch output.

[0027] The local branch constructs an adjacency matrix by calculating pairwise cosine similarity for multi-scale features, and applies a mask to the matrix to highlight the weight priority of adjacent nodes. The pairwise cosine similarity calculation yields the local branch output.

[0028] By combining the global branch output and the local branch output, and collaboratively modeling the global and local dependencies between nodes, the audio authenticity confidence score is obtained.

[0029] Preferably, the adjacency matrix of the local branches is constructed using the Top-K algorithm, specifically as follows: ;in, For nodes The set of similarities with all other nodes. The preset number of adjacent nodes;

[0030] The constructed adjacency matrix Used to calculate the output of the local branch.

[0031] Preferably, the formula for calculating cosine similarity is: ,in and The first The and the first The feature vector of each node.

[0032] Preferably, it further includes:

[0033] The input speech segment is processed into a fixed duration to obtain the processed training audio.

[0034] The RawBoost algorithm is used to enhance the processed training audio. White noise filtered by finite-length unit impulse response is added to simulate real-world interference to obtain the enhanced training audio.

[0035] The enhanced training audio input detection process is forward-propagated to obtain the detection loss.

[0036] In the first N training rounds, all weights of the pre-trained cross-lingual self-supervised speech model are frozen based on the detection loss, and only the parameters of the backend classification network are updated;

[0037] Starting from the N+1th training epoch, the weights of the pre-trained cross-lingual self-supervised speech model are adjusted with a first learning rate based on the detection loss, while the parameters of the backend classification network are updated with a second learning rate, where the first learning rate is less than the second learning rate.

[0038] A device for detecting spoofing attacks on industrial field voice interaction links includes:

[0039] Acquisition module: Acquires the raw audio signal to be detected;

[0040] Extraction module: Input the original audio signal into the pre-trained cross-lingual self-supervised speech model and extract the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model;

[0041] Fusion Module: Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representations of each Transformer layer are used as the key vector and value vector. The fusion weights of each layer are adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representations of each Transformer layer are weighted and fused to obtain the fused features.

[0042] Output module: The fused features are input into the backend classification network. The backend classification network extracts multi-scale features from the fused features through an improved residual coding module, and uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, and outputs the audio authenticity confidence score.

[0043] Judgment module: Determines whether the original audio is real or fake speech based on the audio authenticity confidence score.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] Compared to chip testing equipment, this invention boasts multiple hardware and system advantages. It is equipped with high-end hardware such as a multi-layer parallel NPU inference chip, a custom attention ASIC, and a multi-scale FPGA accelerator card, enabling parallel extraction of multi-layer acoustic hardware representations of speech. The system incorporates a hierarchical feature filtering coprocessor, which adaptively reduces the computational power ratio of feature layers heavily polluted by industrial noise, amplifying subtle acoustic fingerprints of forged speech. A dedicated dual-branch graph attention hardware unit synchronously and in parallel processes global and local speech correlation features, with hardware computing power capable of withstanding strong noise in workshops and multipath interference from underground wireless systems. The entire high-end equipment is an explosion-proof industrial system, allowing direct deployment in harsh industrial environments with low signal-to-noise ratios. The hardware's detection accuracy and environmental robustness are superior to traditional low-end speech detection equipment, enabling stable interception of all types of speech forgery attacks and ensuring the security of industrial speech interaction links. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the modules in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the hidden layer result in an embodiment of the present invention;

[0049] Figure 4 This is a hardware connection diagram according to an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the high-end device with a built-in hierarchical inference unit (NPU) in an embodiment of the present invention. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0054] Reference Figures 1-5 As shown.

[0055] The embodiments further illustrate the forgery attack detection method and device for industrial field voice interaction links proposed in this invention.

[0056] A method for detecting spoofing attacks on industrial field voice interaction links includes the following steps:

[0057] Acquire the raw audio signal to be detected;

[0058] The original audio signal is input into a pre-trained cross-lingual self-supervised speech model, and the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model are extracted.

[0059] Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representation of each Transformer layer is used as the key vector and value vector. The fusion weight of each layer is adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representation of each Transformer layer is weighted and fused to obtain the fused feature.

[0060] The fused features are input into the back-end classification network. The back-end classification network extracts multi-scale features from the fused features through an improved residual coding module. It then uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, and outputs the audio authenticity confidence score.

[0061] The original audio is determined to be either real or fake speech based on the audio authenticity confidence score.

[0062] It also includes a quantitative evaluation of the discrimination potential of the frame-level hidden layer representations of each Transformer layer, calculation of the maximum deviation of the cumulative distribution function of the feature distribution of true and false samples in each layer, and screening of hidden layer representations with discrimination potential based on the critical threshold of the preset significance level.

[0063] The hidden layer representations with discriminative potential are used as the frame-level hidden layer representations of each layer in the input frame-level attention fusion module.

[0064] After completing the extraction of frame-level hidden layer representations for each Transformer layer, a quantitative evaluation process for discrimination potential is initiated. The evaluation object is the individual frame-level hidden layer representation corresponding to each layer. The calculated metric is the maximum deviation between the cumulative distribution functions of real and fake speech samples at the corresponding layer. This value quantifies the potential of the corresponding layer's features to distinguish between real and fake audio. After obtaining the maximum deviation for each layer, a pre-set critical threshold is retrieved. The maximum deviation calculated for each layer's features is compared with this critical threshold one by one. Hidden layer representations whose maximum deviation meets the threshold condition and possesses the potential to distinguish between real and fake speech are selected, while redundant layer representations with insufficient discrimination potential are eliminated. After the entire selection process is completed, all original Transformer layer frame-level hidden layer representations are no longer used. Instead, only the selected hidden layer representations with discrimination potential are used as input material for the frame-level layer attention fusion module.

[0065] Critical threshold The calculation formula is:

[0066] ;

[0067] in, and These represent the sample sizes of the real and fake samples, respectively. significance level The corresponding critical value coefficient.

[0068] This critical threshold is used to compare the maximum deviation of the true and false sample feature distributions calculated by the frame-level hidden layer representations of each Transformer layer, thereby screening hidden layer representations with discriminative potential.

[0069] The frame-level attention fusion module employs a multi-head attention mechanism for each time frame. The calculation process for single-head attention includes:

[0070] The extracted first Constructing query vectors from hidden layer representations at the frame level ,by Layer-by-frame hidden layer representation construction of key vectors Sum value vector ;

[0071] Calculate the first by scaling the dot product attention. Output of each attention head: ;

[0072] in, ; The dimension of the key vector. Indicates the first The attention head corresponds to the first The value vector of the t-th time frame output by the Transformer layer;

[0073] The fused features of this time frame are obtained by concatenating the outputs of all attention heads and performing a linear transformation: , , , For learnable projection matrix;

[0074] The fusion features of each time frame The fused features are obtained by aggregating along the time dimension.

[0075] Extracting the Lth frame-level hidden layer representation to generate the current time frame Corresponding query vector Each Layer-by-frame hidden layer representation, corresponding to the generated key vector Sum value vector Then, a scaled dot product attention operation is performed to solve for the first... The output of the attention head will be obtained by multiplying the weight coefficients of each layer with the matched value vector and then summing them up. Each attention point is focused on a time frame. Output After completing the single-head output calculation for all attention heads, output all attention heads. to Perform a stitching operation, and the stitched result is compared with the learnable output projection matrix. Multiplication completes a linear transformation, thus yielding the first... Single-frame fusion features corresponding to each time frame Collect the single-frame fusion features corresponding to each of all time frames. The fusion of all single-frame features is aggregated along the time dimension to obtain the overall fusion features, which are then fed into the backend classification network for further processing.

[0076] The improved residual coding module is built on the Res2Net architecture. By inputting fused features into the residual coding module, the fused features are grouped by channel and subjected to hierarchical convolution to obtain hierarchical convolutional features.

[0077] An attention feature fusion module is embedded in the residual module. The resulting hierarchical convolutional features are dynamically weighted through the attention mechanism. The contribution of each group of features is adaptively adjusted according to the input feature distribution. Non-key features are suppressed while highlighting key features related to forgery, resulting in dynamically weighted features for each group.

[0078] By a top-down approach, the dynamically weighted features output from each layer of Res2Net are aggregated to obtain multi-scale features that contain global information across multiple time series and frequency scales.

[0079] The fused features are input into the residual encoding module. First, channel grouping is performed on the input fused features. Then, hierarchical convolution operations are applied to the grouped features, generating hierarchical convolutional features. The residual module within the module embeds an attention feature fusion module. This module uses an attention mechanism to dynamically weight the hierarchical convolutional features. The weighting process adaptively adjusts the contribution of each group of features based on the distribution of the current input features. The process amplifies key features related to audio forgery and weakens irrelevant non-key features. After processing, the dynamically weighted features of each group are output. After obtaining the dynamically weighted features output from each layer of the Res2Net architecture, the dynamically weighted features of each layer are aggregated from top to bottom. After aggregation, multi-scale features containing multi-temporal information and multi-frequency scale global information are generated and output to the dual-branch graph attention module for computation.

[0080] The bi-branch graph attention module takes multi-scale features as input and includes a global branch and a local branch:

[0081] The global branch calculates global attention weights by performing dot product operations between nodes and the softmax activation function on multi-scale features, thereby achieving attention interaction among all nodes in the global scope and obtaining the global branch output.

[0082] The local branch constructs an adjacency matrix by calculating pairwise cosine similarity for multi-scale features, and applies a mask to the matrix to highlight the weight priority of adjacent nodes. The local branch output is obtained by calculating pairwise cosine similarity.

[0083] By integrating global branch outputs and local branch outputs, and collaboratively modeling the global and local dependencies between nodes, an audio authenticity confidence score is obtained.

[0084] After inputting multi-scale features, the global branch first performs calculations, performing dot product operations on all feature nodes within the multi-scale features, and then introducing a softmax activation function for normalization. This calculates the global attention weights, which are then used to perform global attention interaction calculations on all nodes. The global branch output is generated after this calculation. The local branch calculates the cosine similarity between each pair of nodes within the same multi-scale feature set, constructs an adjacency matrix based on the similarity values, and applies a mask to the adjacency matrix. This masking process increases the priority of the weights corresponding to adjacent nodes. After the masking operation, feature calculations are performed based on the pairwise cosine similarity to generate the local branch output. After obtaining the global and local branch outputs respectively, a fusion operation is performed on the outputs of the two branches. The fused features can collaboratively model the global and local geometric dependencies between all feature nodes. After the fusion operation, an audio authenticity confidence score is output to determine the authenticity of the audio.

[0085] The adjacency matrix of the local branches is constructed using the Top-K algorithm, specifically as follows: ;in, For nodes The set of similarities with all other nodes. The preset number of adjacent nodes;

[0086] The constructed adjacency matrix Used to calculate the output of local branches.

[0087] After the adjacency matrix is ​​constructed, it is substituted into the local branch operation flow and used as the basic data for the operation to complete the calculation of the local branch output.

[0088] The formula for calculating cosine similarity is: ,in and The first The and the first The feature vectors of each node. The resulting similarity values ​​will be used in the construction of the adjacency matrix.

[0089] Also includes:

[0090] The input speech segment is processed into a fixed duration to obtain the processed training audio.

[0091] The RawBoost algorithm is used to enhance the processed training audio. White noise filtered by finite-length unit impulse response is added to simulate real-world interference to obtain the enhanced training audio.

[0092] The enhanced training audio input detection process is forward-propagated to obtain the detection loss.

[0093] In the first N training rounds, all weights of the pre-trained cross-lingual self-supervised speech model are frozen based on the detection loss, and only the parameters of the backend classification network are updated;

[0094] Starting from the N+1th training epoch, the weights of the pre-trained cross-lingual self-supervised speech model are adjusted with a first learning rate based on the detection loss, while the parameters of the backend classification network are updated with a second learning rate, where the first learning rate is less than the second learning rate.

[0095] The original input speech segment undergoes duration processing to adjust it to a fixed duration, generating training audio with basic processing. The RawBoost algorithm is then used to perform data augmentation on this training audio, adding white noise filtered by a finite-length impulse response (FIR) to simulate various environmental interferences present in real-world applications. The augmented training audio is then output. This augmented training audio is fed into the detection process for forward propagation, calculating the corresponding detection loss value, which serves as the basis for subsequent model parameter updates. The model training process is divided into two phases with differentiated parameter updates. The first phase, the first N training epochs, freezes all pre-trained cross-lingual self-supervised speech model weights based on the detection loss, and only updates the parameters within the backend classification network. The second phase begins from the N+1th training epoch. This phase still uses the detection loss as the optimization basis, setting two different learning rates to update the network parameters. The first learning rate is used to fine-tune the weights of the pre-trained cross-lingual self-supervised speech model, while the second learning rate is used to simultaneously update the parameters of the backend classification network.

[0096] A device for detecting spoofing attacks on industrial field voice interaction links includes:

[0097] Acquisition module: Acquires the raw audio signal to be detected;

[0098] Extraction module: Input the original audio signal into the pre-trained cross-lingual self-supervised speech model and extract the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model;

[0099] Fusion Module: Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representations of each Transformer layer are used as the key vector and value vector. The fusion weights of each layer are adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representations of each Transformer layer are weighted and fused to obtain the fused features.

[0100] Output module: The fused features are input into the backend classification network. The backend classification network extracts multi-scale features from the fused features through an improved residual coding module. It then uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, and outputs the audio authenticity confidence score.

[0101] Judgment module: Determines whether the original audio is real or fake speech based on the audio authenticity confidence score.

[0102] Figure 3 The graph shows the performance test curves for distinguishing between real and fake speech using the built-in NPU hierarchical inference hardware unit of this high-end equipment. The vertical axis represents the Kolmogorov-Smirnov statistic, and the horizontal axis represents the XLS-R Transformer layer number. Three categories of labels are used to distinguish the results in the graph, and the green dashed line represents the significance level. The critical value for a KS score of 0.05 is 0.0092. The red line represents layers whose KS scores exceed the critical value and whose layer discrimination ability is statistically significant, while the blue line represents layers whose KS scores are below the critical value and whose discrimination ability is not significantly different. This graph only shows the red line. Looking at the curve's numerical variation, the KS scores for Transformer layers 0 to 4 remain in the range of 0.4 to 0.55, indicating generally low values. When the layer number increases to 5, the KS score rises significantly, and after layer 6, the curve value jumps sharply to over 0.8. The KS scores for all layers 6 to 23 are far higher than the critical value of 0.0092, indicating that all layers in this range have significant discrimination ability against both fake and real speech. The overall statistical values ​​of layers 6 to 23 remained in the range of 0.8 to 0.9. Among them, the values ​​of layers 6 to 10 slowly increased and reached the peak of the range, while layers 11 to 16 slightly decreased and formed local troughs. The values ​​of layers 17 to 23 slowly rebounded again and remained stable at a high level. The overall trend proves that the shallow Transformer layer of XLS-R has a weak ability to distinguish between real and fake speech, while the deep Transformer layer can stably capture the acoustic feature differences between fake and real speech and has reliable distinguishing power. The hierarchical inference architecture of this high-end device has the hardware computing power foundation for fake speech recognition.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting spoofing attacks on industrial field voice interaction links, characterized in that, Includes the following steps: Acquire the raw audio signal to be detected; The original audio signal is input into a pre-trained cross-lingual self-supervised speech model, and the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model are extracted. Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representation of each Transformer layer is used as the key vector and value vector. The fusion weight of each layer is adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representation of each Transformer layer is weighted and fused to obtain the fused feature. The fused features are input into the backend classification network. The backend classification network extracts multi-scale features from the fused features through an improved residual coding module, and then uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, outputting an audio authenticity confidence score. The original audio is determined to be either real or fake speech based on the audio authenticity confidence score.

2. The method for detecting forgery attacks on industrial field voice interaction links according to claim 1, characterized in that, It also includes a quantitative evaluation of the discrimination potential of the frame-level hidden layer representations of each Transformer layer, calculation of the maximum deviation of the cumulative distribution function of the feature distribution of true and false samples in each layer, and screening of hidden layer representations with discrimination potential based on the critical threshold of the preset significance level. The hidden layer representations with discriminative potential are used as the frame-level hidden layer representations of each layer in the input frame-level attention fusion module.

3. The method for detecting spoofing attacks on industrial field voice interaction links according to claim 2, characterized in that, The critical threshold The calculation formula is: ； in, and These represent the sample sizes of the real and fake samples, respectively. significance level The corresponding critical value coefficient.

4. The method for detecting forgery attacks on industrial field voice interaction links according to claim 1, characterized in that, The frame-level attention fusion module employs a multi-head attention mechanism for each time frame. The calculation process for single-head attention includes: The extracted first Constructing query vectors from hidden layer representations at the layer-frame level ,by Layer-by-frame hidden layer representation construction of key vectors Sum value vector ; Calculate the first by scaling the dot product attention. Output of each attention head: ; in, ; The dimension of the key vector. Indicates the first The attention head corresponds to the first The value vector of the t-th time frame output by the Transformer layer; The fused features of this time frame are obtained by concatenating the outputs of all attention heads and performing a linear transformation: , , , For learnable projection matrix; The fusion features of each time frame The fused features are obtained by aggregating along the time dimension.

5. The method for detecting spoofing attacks on industrial field voice interaction links according to claim 4, characterized in that, The improved residual coding module is built on the Res2Net architecture. By inputting the fused features into the residual coding module, the fused features are grouped by channel and subjected to hierarchical convolution to obtain hierarchical convolutional features. An attention feature fusion module is embedded in the residual module. The resulting hierarchical convolutional features are dynamically weighted through the attention mechanism. The contribution of each group of features is adaptively adjusted according to the input feature distribution. Non-key features are suppressed while highlighting key features related to forgery, resulting in dynamically weighted features for each group. By a top-down approach, the dynamically weighted features output from each layer of Res2Net are aggregated to obtain multi-scale features that contain global information across multiple time series and frequency scales.

6. The method for detecting spoofing attacks on industrial field voice interaction links according to claim 5, characterized in that, The bi-branch graph attention module takes multi-scale features as input and includes a global branch and a local branch. The global branch calculates global attention weights by performing dot product operations between nodes and the softmax activation function on multi-scale features, thereby achieving attention interaction among all nodes in the global scope and obtaining the global branch output. The local branch constructs an adjacency matrix by calculating pairwise cosine similarity for multi-scale features, and applies a mask to the matrix to highlight the weight priority of adjacent nodes. The pairwise cosine similarity calculation yields the local branch output. By combining the global branch output and the local branch output, and collaboratively modeling the global and local dependencies between nodes, the audio authenticity confidence score is obtained.

7. The method for detecting spoofing attacks on industrial field voice interaction links according to claim 6, characterized in that, The adjacency matrix of the local branches is constructed using the Top-K algorithm, specifically as follows: ;in, For nodes The set of similarities with all other nodes. The preset number of adjacent nodes; The constructed adjacency matrix Used to calculate the output of the local branch.

8. The method for detecting spoofing attacks on industrial field voice interaction links according to claim 6, characterized in that, The formula for calculating cosine similarity is: ,in and The first The and the first The feature vector of each node.

9. The method for detecting forgery attacks on industrial field voice interaction links according to claim 1, characterized in that, Also includes: The input speech segment is processed into a fixed duration to obtain the processed training audio. The RawBoost algorithm is used to enhance the processed training audio. White noise filtered by finite-length unit impulse response is added to simulate real-world interference to obtain the enhanced training audio. The enhanced training audio input detection process is forward-propagated to obtain the detection loss. In the first N training rounds, all weights of the pre-trained cross-lingual self-supervised speech model are frozen based on the detection loss, and only the parameters of the backend classification network are updated; Starting from the N+1th training epoch, the weights of the pre-trained cross-lingual self-supervised speech model are adjusted with a first learning rate based on the detection loss, while the parameters of the backend classification network are updated with a second learning rate, where the first learning rate is less than the second learning rate.

10. A device for detecting spoofing attacks on industrial field voice interaction links, applied to the method for detecting spoofing attacks on industrial field voice interaction links as described in any one of claims 1-9, characterized in that, include: Acquisition module: Acquires the raw audio signal to be detected; Extraction module: Input the original audio signal into the pre-trained cross-lingual self-supervised speech model and extract the frame-level hidden layer representations of each Transformer layer of the cross-lingual self-supervised speech model; Fusion Module: Through the frame-level attention fusion module, the frame-level hidden layer representation of the last layer is used as the query vector, and the frame-level hidden layer representations of each Transformer layer are used as the key vector and value vector. The fusion weights of each layer are adaptively assigned to each time frame in the hierarchical dimension, and the frame-level hidden layer representations of each Transformer layer are weighted and fused to obtain the fused features. Output module: The fused features are input into the backend classification network. The backend classification network extracts multi-scale features from the fused features through an improved residual coding module, and uses a dual-branch graph attention module to collaboratively model the local geometric structure and global contextual dependencies of the multi-scale features, and outputs the audio authenticity confidence score. Judgment module: Determines whether the original audio is real or fake speech based on the audio authenticity confidence score.

Images