Insurance false recording evidence detection method and device and storage medium

Abstract

The invention provides an insurance false recording evidence detection method and device and a storage medium. Relates to the technical field of false recording detection. The method comprises the following steps: inputting a to-be-detected recording evidence into a false recording detection model to analyze whether the to-be-detected recording evidence is forged or not; the false recording detection model comprises a Sinc convolution layer, a mapping layer, a residual error convolution encoder, an aggregation module, a time-frequency graph attention network and a classification full-connection network. A sound spectrogram generated by the Sinc convolution layer is mapped into a residual convolution encoder, and the output of each level of residual convolution encoding module of the residual convolution encoder is aggregated through an aggregation module to obtain an aggregation feature; the time-frequency graph attention network extracts time domain feature nodes and frequency domain feature nodes from the time domain and the frequency domain of the stacked aggregation features through graph attention and a graph pooling layer; further obtaining a fusion node; and the classification full-connection network outputs a real or false prediction result based on average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of false recording detection, and in particular to a method, device and storage medium for detecting false recording evidence in insurance. Background Art

[0002] With the rapid development of audio generative neural network technology, the quality of audio synthesis has been significantly improved, and it has become increasingly difficult to distinguish true and false audio. This has brought huge risks to audio recognition technology. Insurance claims require insurance customers to provide sufficient evidence to the insurance company. After the insurance company verifies the authenticity of the submitted evidence, it will further proceed with the claim. Among the evidence provided by insurance customers or insurance agents, audio or video evidence is an important component of restoring the scene. It is possible to generate false audio evidence by imitating the voice characteristics of witnesses through audio generative neural networks. In the absence of mutual corroboration of multiple audio evidence, high-quality false audio evidence is difficult to distinguish through human hearing. For insurance companies, a large number of accident cases involve massive amounts of audio evidence, and the false detection of audio evidence requires a lot of energy from identification personnel. Therefore, a method for detecting false audio evidence in insurance is needed to accurately detect forgery. Summary of the invention

[0003] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides a method, device and storage medium for detecting false insurance recording evidence.

[0004] In a first aspect, the present invention provides a method for detecting false insurance recording evidence, comprising:

[0005] The recording evidence to be detected is sampled through a sliding window, and the sound data in the sliding window is input into the constructed false recording detection model to analyze whether it is forged;

[0006] The fake recording detection model includes: a Sinc convolution layer that processes sound data through a set of finite impulse response filters with partially trainable parameters to obtain a sound spectrogram, and all sound spectrograms of the recording evidence are mapped to a residual convolution encoder by a mapping layer, and the residual convolution encoder includes a plurality of residual convolution encoding modules connected in series, and the outputs of the residual convolution encoding modules of each level of the residual convolution encoder are aggregated through an aggregation module to obtain an aggregate feature of the recording evidence;

[0007] The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted; the average, maximum value and fusion node of the time domain feature nodes and the frequency domain feature nodes are obtained;

[0008] The average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes are input into the classification fully connected network, and the classification fully connected network outputs a true or false prediction result.

[0009] Furthermore, the formula for processing sound data x[n] of each finite impulse response filter is:

[0010]

[0011] in, The hth finite impulse response filter in the Sinc convolutional layer processes the input sound data to obtain the basic features of the sound data;

[0012] w[n] is the Hamming window, a and b are constants, and b is the length of the finite impulse response filter;

[0013] is the hth finite impulse response filter, f 1 h and are trainable parameters of the h-th finite impulse response filter, representing the low cutoff frequency and the high cutoff frequency of the h-th finite impulse response filter respectively;

[0014] The output of all finite impulse response filters The stacking forms a two-dimensional sound spectrum.

[0015] Furthermore, the residual convolutional coding block includes: a series of convolutional layers, batch normalization layers, Selu activation function layers, convolutional layers, and a jump chain spanning the series of convolutional layers, batch normalization layers, Selu activation function layers, and convolutional layers.

[0016] Furthermore, the aggregation module includes: a global average pooling layer, the global average pooling layer is connected to two fully connected layers, a ReLU activation function is set between the two fully connected layers, the latter fully connected layer is followed by a Sigmoid activation function, and the two fully connected layers cooperate with the ReLU activation function to convert the global average pooling layer based on the output E of the residual convolutional coding module at each level l ,l∈{1,2,...L}, the vector generated by GAP is compressed and expanded, and then processed by Sigmoid activation function to generate weight vector W. The elements W in weight vector W l The output E of the residual convolutional coding module at the corresponding level l After weighting, the weighted result E' of the output of the residual convolutional coding module is obtained l , the weighted result E' l Accumulate and get the aggregate feature E A .

[0017] Furthermore, graph attention generates an attention map based on frequency domain and time domain features and initialized attention weights. Graph attention expands the frequency domain or time domain features, and the expanded result is matrix multiplied with its transpose and input into the linear layer for processing, and then activated by the Tanh activation function. The activation result is weighted with the initialized attention weight, and then divided by the heat, and processed by the softmax layer to obtain the attention map; graph attention projects the attention map and frequency domain or time domain features through the linear layer respectively, and adds the projection results of the frequency domain or time domain features of the weighted attention map and the projection results of the frequency domain or time domain features of the unweighted attention map, and performs batch normalization and SELU activation function processing on the combined results; the graph pooling layer calculates the weight of each frequency domain and time domain feature node through the linear layer, and converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores.

[0018] Furthermore, the topk frequency-domain feature nodes and the topk time-domain feature nodes are mapped to the same dimension through two linear layers and feature concatenation is performed. The concatenation results are subjected to homogeneous and heterogeneous attention through two sets of parallel front-end heterogeneous graph attention, two graph pooling layers, and back-end heterogeneous graph attention:

[0019] The previous heterogeneous graph attention expands the dimension of the splicing result, and the expanded splicing result is matrix multiplied with its transpose and input into the linear layer for processing, and then activated by the Tanh activation function. The activation result is weighted with two initialized isomorphic attention weights and one initialized heterogeneous attention weight, and then divided by the heat. After processing through the softmax layer, an attention map containing two isomorphic attention maps and one heterogeneous attention map is obtained. The two isomorphic attention maps correspond to the association between time domain feature nodes and the association between frequency domain nodes, respectively, and the heterogeneous attention map corresponds to the association between time domain feature nodes and frequency domain feature nodes; the previous heterogeneous graph attention projects the attention map and the splicing result through the linear layer respectively, and adds the projection result of the splicing result of the weighted attention map and the projection result of the splicing result of the unweighted attention map, and performs batch normalization and SELU activation function processing on the combined result; and the next two graph pooling layers following the previous heterogeneous graph attention respectively perform SELU activation function on the SELU activation function The processed time domain feature nodes and frequency domain feature nodes are screened; the front-level heterogeneous graph attention transmits the two sets of results generated based on the two isomorphic attention graphs to the graph pooling layer for processing. The graph pooling layer calculates the weight of each frequency domain and time domain feature node through the linear layer, and converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores; the front-level heterogeneous graph attention transmits the results generated based on the heterogeneous attention graph to the back-level heterogeneous graph attention, and the two graph pooling layers transmit the selected frequency domain and time domain feature nodes to the back-level heterogeneous graph attention respectively; the back-level heterogeneous graph attention also generates frequency domain and time domain feature nodes based on the two isomorphic attention graphs, and generates a fusion node based on a heterogeneous attention graph; the frequency domain feature nodes, time domain feature nodes, and fusion nodes of the two parallel branches are fused respectively, and the maximum and average values ​​of the frequency domain feature nodes and time domain feature nodes after fusion are calculated respectively.

[0020] Furthermore, the training process of the fake recording detection model includes:

[0021] Collect real voice data from samplers of different ages, genders, accents, and timbres, and imitate the false voice data generated by the corresponding samplers through a voice cloning model; perform data enhancement on the collected voice data by adding noise; and configure labels indicating the true and false attributes of the real voice data and the false voice data to obtain training data;

[0022] The fake recording detection model is trained using the training data, and the cross entropy loss between the predicted label and the true label is used as the loss function. The parameters of the Sinc convolution layer, the residual convolution encoder, the time-frequency graph attention network and the classification fully connected network are updated through the gradient descent algorithm; the gradient descent algorithm adopts adam optimization, and the learning rate adopts cosine annealing.

[0023] Furthermore, the noise is added by any one or a combination of the following:

[0024] Linear and nonlinear convolution noise are added to simulate distortion using the following formula:

[0025]

[0026] Among them, x 1 [n] is the sound signal after adding linear and nonlinear convolution noise, is the gain of the i-th order FIR filter, which contains N 1 FIR filter of order, each FIR filter contains N 2 Notch filters, each with different center frequency, bandwidth and notch coefficient. is the notch coefficient of the jth notch filter. The jth notch filter is input to the i-th order FIR filter at x i [n] Nearby x i [nj] according to the notch coefficient Generate notch;

[0027] Impulse noise related to the sound data is added by the following formula:

[0028]

[0029] Among them, x 2 [n] is the sound signal after the impulse noise related to the sound signal, g 2 is the pulse gain, D[n] is the random pulse coefficient generated by logarithmic distribution at the specified sampling point, and P is the set of sampling points uniformly distributed in the sound data x[n];

[0030] Additional noise unrelated to the sound data is added using the following formula:

[0031]

[0032] Among them, x 3 [n] is the sound signal with additional noise that is irrelevant to the sound data, g 3 is the gain of the FIR filter. The FIR filter contains N 3 There are notch filters, each with different center frequency, bandwidth, and notch coefficient. k is the notch coefficient of the kth notch filter. The kth notch filter is at wn[nk] near the white noise wn[n] of the input FIR filter according to the notch coefficient β k Generates a notch.

[0033] In a second aspect, the present invention provides an insurance false recording evidence detection device, comprising: at least one processing unit, the processing unit and a storage unit are interconnected via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insurance false recording evidence detection method is implemented.

[0034] In a third aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for detecting false insurance recording evidence is implemented.

[0035] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0036] The sound spectrogram generated by the Sinc convolution layer in the fake recording detection model of the present application is mapped into the residual convolution encoder, and the outputs of the residual convolution coding modules at each level of the residual convolution encoder are aggregated through the aggregation module to obtain aggregated features; the Sinc convolution layer is designed to effectively capture the basic features of the sound data to form a sound spectrogram, thereby improving the model's ability to process and analyze sound data; the residual convolution encoder models the forgery clues in each sound spectrogram from different levels through stacked residual convolution coding modules, the residual convolution coding module at the front stage extracts low-order forgery clues, and the residual convolution coding module at the back stage extracts high-order forgery clues, which helps to extract forgery clue features of each order from the sound spectrogram.

[0037] The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted. The time-frequency graph attention network associates the forgery clues in the time-frequency domain through the isomorphic and heterogeneous graph attention to better extract the forgery clues and ensure the accuracy of the analysis.

[0038] This application uses an AI model to detect recorded evidence of insurance claim cases, provide technical support for the verification of insurance recorded evidence, and the automated analysis process is conducive to improving the efficiency of verification and identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A flowchart of a method for detecting false insurance recording evidence provided by an embodiment of the present invention;

[0042] Figure 2 An architectural diagram of a false recording detection model provided by an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of a residual convolutional coding module provided by an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of an aggregation module provided in an embodiment of the present invention;

[0045] Figure 5 A schematic diagram of graph attention in a time-frequency graph attention network provided by an embodiment of the present invention;

[0046] Figure 6 An architectural diagram of the heterogeneous graph attention part in the time-frequency graph attention network provided by an embodiment of the present invention;

[0047] Figure 7 A schematic diagram of heterogeneous graph attention in a time-frequency graph attention network provided by an embodiment of the present invention;

[0048] Figure 8 A schematic diagram of a device for detecting false insurance recording evidence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0051] Example 1

[0052] like Figure 1 and Figure 2 As shown, the technology of the present invention implements a method for detecting false insurance recording evidence, and the present invention includes the following steps:

[0053] This application obtains the audio recording evidence involved in each insurance claim case;

[0054] The recording evidence to be detected is sampled through a sliding window, and the sound data x[n] in the sliding window is input into the constructed false recording detection model to analyze whether it is forged.

[0055] In the specific implementation process, the false recording detection model includes:

[0056] Sinc convolution layer, the Sinc convolution layer processes the input sound data through a set of finite impulse response filters with some trainable parameters to obtain a sound spectrogram.

[0057] The formula for processing sound data x[n] of each finite impulse response filter is:

[0058]

[0059] in, The h-th finite impulse response filter in the Sinc convolution layer processes the input sound data to obtain basic features of the sound data; the basic features are the h-th row of the sound puzzle.

[0060] w[n] is the Hamming window, a and b are constants, and b is the length of the finite impulse response filter;

[0061] is the ith finite impulse response filter, f 1 i and are trainable parameters of the i-th finite impulse response filter, representing the low cutoff frequency and the high cutoff frequency of the i-th finite impulse response filter respectively.

[0062] The output of all finite impulse response filters The two-dimensional sound spectrogram is formed by stacking in the row direction. The Sinc convolution layer is designed to effectively capture the basic features of the sound data to form the sound spectrogram FM, thereby improving the model's ability to process and analyze sound data.

[0063] The sound spectrogram FM output by the Sinc convolution layer is mapped by the mapping layer and then input into the residual convolution encoder. The residual convolution encoder includes several residual convolution coding modules connected in series. The residual convolution encoder models the forgery clues in each sound spectrogram FM from different levels through stacked residual convolution coding modules. The residual convolution coding module at the front stage extracts low-order forgery clues, and the residual convolution coding module at the back stage extracts high-order forgery clues, which helps to extract forgery clue features of each order from the sound spectrogram.

[0064] During the specific implementation process, see Figure 3 As shown, the residual convolutional coding block includes: a series of convolutional layers, batch normalization layers, Selu activation function layers, convolutional layers, and a jump chain spanning the series of convolutional layers, batch normalization layers, and Selu activation function layers. The jump chain adds and fuses the output of the residual convolutional coding block of the previous stage with the output of the residual convolutional coding block of the next stage, and the previous stage information is always transmitted to avoid gradient disappearance. It is assumed that the residual convolutional encoder includes L layers of residual convolutional coding modules, and the output of the lth layer of residual convolutional coding modules is represented as E l .

[0065] The output E of each level residual convolutional coding module of the residual convolutional encoder l ,l∈{1,2,...L} is aggregated through the aggregation module to obtain the aggregated features of the recording evidence.

[0066] During the specific implementation process, see Figure 4 As shown, the aggregation module includes: a global average pooling layer, the global average pooling layer is connected to two fully connected layers, a ReLU activation function is set between the two fully connected layers, the latter fully connected layer is followed by a Sigmoid activation function, and the two fully connected layers cooperate with the ReLU activation function to convert the global average pooling layer based on the output E of the residual convolutional coding module at each level l ,l∈{1,2,...L}, the vector generated by GAP is compressed and expanded, and then processed by Sigmoid activation function to generate weight vector W. The elements W in weight vector W l The output E of the residual convolutional coding module at the corresponding level l After weighting, the weighted result E' of the output of the residual convolutional coding module is obtained l , the weighted result E' l Accumulate and get the aggregate feature E A .

[0067] The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted; the average, maximum value and fusion node of the time domain feature nodes and the frequency domain feature nodes are obtained.

[0068] In the specific implementation process, the time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively. In the specific implementation process, refer to Figure 5 As shown in the figure, graph attention generates an attention map based on frequency domain and time domain features and initialized attention weights. Specifically, graph attention expands the frequency domain or time domain feature nodes, and the expanded result is matrix multiplied with its transpose and input into the linear layer for processing, and then activated by the Tanh activation function. The activation result is weighted with the initialized attention weight, and then divided by the heat, and the attention map is obtained after processing through the softmax layer. Graph attention projects the attention map and frequency domain or time domain features through the linear layer respectively, and adds the projection results of the frequency domain or time domain features of the weighted attention map and the projection results of the frequency domain or time domain features of the unweighted attention map. The combined results are batch normalized and processed by the SELU activation function. The graph pooling layer calculates the weight of each frequency domain or time domain feature node through the linear layer, converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores.

[0069] like Figure 6 As shown in the figure, the topk frequency domain feature nodes and the topk time domain feature nodes are mapped to the same dimension through two linear layers respectively and feature concatenation is performed. The concatenation result contains both frequency domain feature node information and time domain feature node information. The concatenation results are subjected to homogeneous and heterogeneous attention through two sets of parallel front-stage heterogeneous graph attention, two graph pooling layers and back-stage heterogeneous graph attention.

[0070] Specifically, Figure 7As shown in the figure, the previous heterogeneous graph attention expands the concatenation result, and the concatenation result after the expansion is matrix multiplied with its transpose and then input into the linear layer for processing, and then activated by the Tanh activation function. The activation result is weighted with two initialized homogeneous attention weights and one initialized heterogeneous attention weight, and then divided by the heat. After processing through the softmax layer, an attention map containing two homogeneous attention maps and one heterogeneous attention map is obtained. The two homogeneous attention maps correspond to the association between time domain feature nodes and the association between frequency domain nodes, respectively, and the heterogeneous attention map corresponds to the association between time domain feature nodes and frequency domain feature nodes. The previous heterogeneous graph attention projects the attention map and the concatenation result through the linear layer respectively, adds the projection result of the concatenation result of the weighted attention map and the projection result of the concatenation result of the unweighted attention map, and performs batch normalization and SELU activation function processing on the combined result. The two graph pooling layers following the previous heterogeneous graph attention respectively filter the time domain feature nodes and frequency domain feature nodes processed by the SELU activation function. The previous heterogeneous graph attention transmits the two sets of results generated based on the two homogeneous attention graphs to the graph pooling layer for processing. The graph pooling layer calculates the weight of each frequency domain and time domain feature node through the linear layer, converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores. The previous heterogeneous graph attention transmits the results generated based on the heterogeneous attention graph to the next heterogeneous graph attention. The two graph pooling layers transmit the selected frequency domain and time domain feature nodes to the next heterogeneous graph attention. The next heterogeneous graph attention also generates frequency domain and time domain feature nodes based on the two homogeneous attention graphs, and generates a fusion node based on a heterogeneous attention graph; the frequency domain feature nodes, time domain feature nodes, and fusion nodes of the two parallel branches are fused separately, and the maximum and average values ​​of the frequency domain feature nodes and time domain feature nodes after fusion are calculated respectively.

[0071] The time-frequency graph attention network associates forgery clues in the time-frequency domain through homogeneous and heterogeneous graph attention to better extract forgery clues and ensure the accuracy of analysis.

[0072] The average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes are input into the classification fully connected network, and the classification fully connected network outputs a true or false prediction result.

[0073] In order to enable the fake recording detection model to detect true and false recordings, it is necessary to train the fake recording detection model. The process includes:

[0074] Prepare training data: collect real voice data from samplers of different ages, genders, accents, and timbres, and use a voice cloning model to imitate the fake voice data generated by the corresponding samplers. An exemplary voice cloning model is the so-vits model. Perform data enhancement on the collected voice data by adding noise. In the specific implementation process, the method of adding noise adopts any one or a combination of the following:

[0075] Linear and nonlinear convolution noise are added to simulate distortion using the following formula:

[0076]

[0077] Among them, x 1 [n] is the sound signal after adding linear and nonlinear convolution noise, is the gain of the i-th order FIR filter, which contains N 1 FIR filter of order, each FIR filter contains N 2 Notch filters, each with different center frequency, bandwidth and notch coefficient. is the notch coefficient of the jth notch filter. The jth notch filter is input to the i-th order FIR filter at x i [n] Nearby x i [nj] according to the notch coefficient Generate notch;

[0078] Impulse noise related to the sound data is added by the following formula:

[0079]

[0080] Among them, x 2 [n] is the sound signal after the impulse noise related to the sound signal, g 2 is the pulse gain, D[n] is the random pulse coefficient generated by logarithmic distribution at the specified sampling point, and P is the set of sampling points uniformly distributed in the sound data x[n];

[0081] Additional noise unrelated to the sound data is added using the following formula:

[0082]

[0083] Among them, x 3 [n] is the sound signal with additional noise that is irrelevant to the sound data, g 3 is the gain of the FIR filter. The FIR filter contains N 3 There are notch filters, each with different center frequency, bandwidth, and notch coefficient. kis the notch coefficient of the kth notch filter. The kth notch filter is at wn[nk] near the white noise wn[n] of the input FIR filter according to the notch coefficient β k Generates a notch.

[0084] The real sound data and the fake sound data are provided with labels indicating their true and false attributes.

[0085] The fake recording detection model is trained using the training data, and the cross entropy loss between the predicted label and the true label is used as the loss function. The parameters of the Sinc convolution layer, the residual convolution encoder, the time-frequency graph attention network and the classification fully connected network are updated through the gradient descent algorithm; the gradient descent algorithm adopts adam optimization, and the learning rate adopts cosine annealing.

[0086] Example 2

[0087] See also Figure 8 As shown, an embodiment of the present invention provides an insurance false recording evidence detection device, including: at least one processing unit, the processing unit and the storage unit are interconnected through a bus unit, and the storage unit is a computer-readable storage medium, which can be used to store software programs, computer executable programs and modules, such as the software programs, computer executable programs and modules corresponding to an insurance false recording evidence detection method in an embodiment of the present invention. The processing unit implements the above-mentioned insurance false recording evidence detection method by running the software programs, computer executable programs and modules stored in the storage unit, including:

[0088] The recording evidence to be detected is sampled through a sliding window, and the sound data x[n] in the sliding window is input into the constructed false recording detection model to analyze whether it is forged;

[0089] The fake recording detection model includes: a Sinc convolution layer that processes sound data through a set of finite impulse response filters with partially trainable parameters to obtain a sound spectrogram, and all sound spectrograms of the recording evidence are mapped to a residual convolution encoder by a mapping layer, and the residual convolution encoder includes a plurality of residual convolution encoding modules connected in series, and the outputs of the residual convolution encoding modules of each level of the residual convolution encoder are aggregated through an aggregation module to obtain an aggregate feature of the recording evidence;

[0090] The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted; the average, maximum value and fusion node of the time domain feature nodes and the frequency domain feature nodes are obtained;

[0091] The average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes are input into the classification fully connected network, and the classification fully connected network outputs a true or false prediction result.

[0092] Of course, the computer program stored in the storage unit of the insurance false recording evidence detection device provided by the embodiment of the present invention is not limited to the method operation described above, and can also execute the relevant operations in the insurance false recording evidence detection method provided by any embodiment of the present invention.

[0093] Example 3

[0094] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the method for detecting false insurance recording evidence is implemented, including:

[0095] The recording evidence to be detected is sampled through a sliding window, and the sound data in the sliding window is input into the constructed false recording detection model to analyze whether it is forged;

[0096] The fake recording detection model includes: a Sinc convolution layer that processes sound data through a set of finite impulse response filters with partially trainable parameters to obtain a sound spectrogram, and all sound spectrograms of the recording evidence are mapped to a residual convolution encoder by a mapping layer, and the residual convolution encoder includes a plurality of residual convolution encoding modules connected in series, and the outputs of the residual convolution encoding modules of each level of the residual convolution encoder are aggregated through an aggregation module to obtain an aggregate feature of the recording evidence;

[0097] The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted; the average, maximum value and fusion node of the time domain feature nodes and the frequency domain feature nodes are obtained;

[0098] The average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes are input into the classification fully connected network, and the classification fully connected network outputs a true or false prediction result.

[0099] A computer-readable storage medium provided by an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in a method for detecting false insurance recording evidence provided by any embodiment of the present invention.

[0100] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0101] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0102] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0103] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for detecting false insurance recording evidence, characterized in that: include: The recording evidence to be detected is sampled through a sliding window, and the sound data in the sliding window is input into the constructed false recording detection model to analyze whether it is forged; The fake recording detection model includes: a Sinc convolution layer that processes sound data through a set of finite impulse response filters with partially trainable parameters to obtain a sound spectrogram, and all sound spectrograms of the recording evidence are mapped to a residual convolution encoder by a mapping layer, and the residual convolution encoder includes a plurality of residual convolution encoding modules connected in series, and the outputs of the residual convolution encoding modules of each level of the residual convolution encoder are aggregated through an aggregation module to obtain an aggregate feature of the recording evidence; The time-frequency graph attention network takes the maximum along the frequency dimension of the aggregated features, and takes the maximum along the time dimension of the aggregated features to obtain frequency domain features and time domain features respectively. The frequency domain features and time domain features are processed by the graph attention and graph pooling layers respectively, and the topk important frequency domain feature nodes and time domain feature nodes are selected. After the important time domain feature nodes and frequency domain feature nodes are combined, the information or relationship between the frequency domain feature nodes and the time domain feature nodes is extracted and accumulated into the fusion node through the parallel front-stage heterogeneous graph attention, two graph pooling layers and the back-stage heterogeneous graph attention, and the information between the domain feature nodes and the time domain feature nodes is further extracted; the average, maximum value and fusion node of the time domain feature nodes and the frequency domain feature nodes are obtained; The average, maximum and fusion nodes of the time domain feature nodes and the frequency domain feature nodes are input into the classification fully connected network, and the classification fully connected network outputs a true or false prediction result.

2. The method for detecting false insurance recording evidence according to claim 1, characterized in that: The formula for processing sound data x[n] of each finite impulse response filter is: in, The hth finite impulse response filter in the Sinc convolutional layer processes the input sound data to obtain the basic features of the sound data; w[n] is the Hamming window, a and b are constants, and b is the length of the finite impulse response filter; is the hth finite impulse response filter, f1 h and are trainable parameters of the h-th finite impulse response filter, representing the low cutoff frequency and the high cutoff frequency of the h-th finite impulse response filter respectively; The output of all finite impulse response filters The stacking forms a two-dimensional sound spectrum.

3. The method for detecting false insurance recording evidence according to claim 1, characterized in that: The residual convolutional coding block includes: a series of convolutional layers, a batch normalization layer, a Selu activation function layer, a convolutional layer, and a jump chain spanning the series of convolutional layers, the batch normalization layer, the Selu activation function layer, and the convolutional layer.

4. The method for detecting false insurance recording evidence according to claim 1, characterized in that: The aggregation module includes: a global average pooling layer, the global average pooling layer is connected to two fully connected layers, a ReLU activation function is set between the two fully connected layers, the latter fully connected layer is followed by a Sigmoid activation function, and the two fully connected layers cooperate with the ReLU activation function to convert the global average pooling layer based on the output E of the residual convolutional coding module at each level l ,l∈{1,2,...L}, the vector generated by GAP is compressed and expanded, and then processed by Sigmoid activation function to generate weight vector W. The elements W in weight vector W l The output E of the residual convolutional coding module at the corresponding level l After weighting, the weighted result E' of the output of the residual convolutional coding module is obtained l , the weighted result E' l Accumulate and get the aggregate feature E A .

5. The method for detecting false insurance recording evidence according to claim 1, characterized in that: On the one hand, graph attention generates an attention map based on frequency domain and time domain features and the initialized attention weights. Graph attention expands the frequency domain or time domain features, performs matrix multiplication on the expanded result and its transpose, and then inputs it into the linear layer for processing. It is then activated by the Tanh activation function, and the activation result is weighted with the initialized attention weight, and then divided by the heat, and processed by the softmax layer to obtain the attention map. Graph attention projects the attention map and frequency domain or time domain features through linear layers respectively, adds the projection results of the frequency domain or time domain features of the weighted attention map and the projection results of the frequency domain or time domain features of the unweighted attention map together, and performs batch normalization and SELU activation function processing on the combined results; The graph pooling layer calculates the weight of each frequency domain and time domain feature node through the linear layer, converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores.

6. The method for detecting false insurance recording evidence according to claim 5, characterized in that: The topk frequency-domain feature nodes and the topk time-domain feature nodes are respectively mapped to the same dimension through two linear layers and feature concatenation is performed. The concatenation results are subjected to homogeneous and heterogeneous attention through two sets of parallel front-stage heterogeneous graph attention, two graph pooling layers and back-stage heterogeneous graph attention: The previous heterogeneous graph attention expands the dimension of the splicing result, and the expanded splicing result is matrix multiplied with its transpose and then input into the linear layer for processing, and then activated by the Tanh activation function. The activation result is weighted with two initialized homogeneous attention weights and one initialized heterogeneous attention weight, and then divided by the heat. After processing through the softmax layer, an attention map containing two homogeneous attention maps and one heterogeneous attention map is obtained. The two homogeneous attention maps correspond to the association between time domain feature nodes and the association between frequency domain nodes, respectively, and the heterogeneous attention map corresponds to the association between time domain feature nodes and frequency domain feature nodes; the previous heterogeneous graph attention projects the attention map and the splicing result through the linear layer respectively, and adds the projection result of the splicing result of the weighted attention map and the projection result of the splicing result of the unweighted attention map together, and the combined result is batch normalized and processed by the SELU activation function; The previous heterogeneous graph attention transmits the two sets of results generated based on the two homogeneous attention graphs to the graph pooling layer for processing. The graph pooling layer calculates the weight of each frequency domain and time domain feature node through the linear layer, converts the weight of the frequency domain and time domain feature nodes into importance scores through the Sigmoid activation function, and selects the topk frequency domain and time domain feature nodes based on the importance scores; The front-stage heterogeneous graph attention transmits the results generated based on the heterogeneous attention graph to the back-stage heterogeneous graph attention. The two graph pooling layers transmit the selected frequency domain and time domain feature nodes to the back-stage heterogeneous graph attention respectively. The subsequent heterogeneous graph attention also generates frequency domain and time domain feature nodes based on two homogeneous attention graphs, and generates a fusion node based on a heterogeneous attention graph; the frequency domain feature nodes, time domain feature nodes, and fusion nodes of the two parallel branches are fused separately, and the maximum and average values ​​of the frequency domain feature nodes and time domain feature nodes after fusion are calculated respectively.

7. The method for detecting false insurance recording evidence according to claim 1, characterized in that: The training process of the fake recording detection model includes: Collect real voice data from samplers of different ages, genders, accents, and timbres, and imitate the false voice data generated by the corresponding samplers through a voice cloning model; perform data enhancement on the collected voice data by adding noise; and configure labels indicating the true and false attributes of the real voice data and the false voice data to obtain training data; The fake recording detection model is trained using the training data, and the cross entropy loss between the predicted label and the true label is used as the loss function. The parameters of the Sinc convolution layer, the residual convolution encoder, the time-frequency graph attention network and the classification fully connected network are updated through the gradient descent algorithm; the gradient descent algorithm adopts adam optimization, and the learning rate adopts cosine annealing.

8. The method for detecting false insurance recording evidence according to claim 7, characterized in that: Noise can be added by any one or a combination of the following: Linear and nonlinear convolution noise are added to simulate distortion using the following formula: Among them, x1[n] is the sound signal after adding linear and nonlinear convolution noise, is the gain of the i-th order FIR filter, which contains a total of N1-order FIR filters. Each order FIR filter contains N2 notch filters. Each notch filter has different center frequency, bandwidth, and notch coefficient. is the notch coefficient of the jth notch filter. The jth notch filter is input to the i-th order FIR filter at x i [n] Nearby x i [nj] according to the notch coefficient Generate notch; Impulse noise related to the sound data is added by the following formula: Where x2[n] is the sound signal after the impulse noise related to the sound signal, g2 is the pulse gain, D[n] is the random pulse coefficient generated according to the logarithmic distribution at the specified sampling point, and P is the set of sampling points uniformly distributed in the sound data x[n]; Additional noise unrelated to the sound data is added using the following formula: Where x3[n] is the sound signal with additional noise that is unrelated to the sound data, g3 is the gain of the FIR filter, and the FIR filter contains N3 notch filters, each with different center frequency, bandwidth, and notch coefficient. k is the notch coefficient of the kth notch filter. The kth notch filter is at wn[nk] near the white noise wn[n] of the input FIR filter according to the notch coefficient β k Generates a notch.

9. An insurance false recording evidence detection device, characterized in that: include: At least one processing unit, the processing unit and the storage unit are interconnected through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insurance false recording evidence detection method as described in any one of claims 1-8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for detecting false insurance recording evidence as described in any one of claims 1 to 8 is implemented.

Images