Method and apparatus for identifying sensitive information in audio data

Abstract

Embodiments of the present specification provide a sensitive information identification method and device in audio data, the method comprising: obtaining first audio data to be identified; comprising a plurality of data points arranged in time sequence, the data points having corresponding numerical values; inputting the plurality of data points in time sequence to a pre-trained identification model, and outputting, by the identification model, a predicted value of a next data point after a preset time window based on the numerical values of the data points in the preset time window; determining a difference between the predicted value of any data point in the plurality of data points and the numerical value of the data point; dividing the data points with a difference less than a first threshold value into sensitive information, and dividing the data points with a difference greater than or equal to the first threshold value into non-sensitive information, to obtain initial identification categories corresponding to the plurality of data points respectively; and optimizing the initial identification categories by using a statistical test method to obtain optimized identification categories corresponding to the plurality of data points respectively. A higher identification accuracy can be achieved with lower computing power consumption.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of computers, and more particularly to methods and apparatus for identifying sensitive information in audio data. Background Technology

[0002] During the digital transformation of enterprises, organizations, or institutions, a large amount of data is generated, some of which is sensitive information. Sensitive information, also known as sensitive data or private data, refers to data that, if leaked, could cause serious harm to society or individuals. If sensitive information is leaked, it may result in the disclosure of confidential information or financial losses for enterprises, organizations, or institutions, and may lead to illegal or unethical activities. Therefore, enterprises, organizations, or institutions want to identify which data assets are sensitive in order to protect and trace this sensitive data.

[0003] Data assets include audio data, which is typically unstructured. Current technologies for identifying sensitive information in audio data often suffer from low accuracy or excessive computational costs. Summary of the Invention

[0004] This specification describes one or more embodiments of a method and apparatus for identifying sensitive information in audio data, which can achieve a high recognition accuracy with low computing power consumption.

[0005] Firstly, a method for identifying sensitive information in audio data is provided, the method comprising:

[0006] Acquire the first audio data to be identified; the first audio data includes multiple data points arranged in time sequence, and each data point has its corresponding value;

[0007] The multiple data points are input into a pre-trained recognition model in the order of the time sequence. The recognition model outputs the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window. The recognition model is a statistical learning model.

[0008] Determine the difference between the predicted value of any data point among the plurality of data points and the actual value of that data point;

[0009] Data points with a difference less than a first threshold are initially classified as sensitive information, and data points with a difference greater than or equal to the first threshold are initially classified as non-sensitive information, thereby obtaining the initial identification categories corresponding to the multiple data points; the identification category is used to indicate whether the corresponding data point belongs to sensitive information.

[0010] The initial identification category is optimized using statistical testing methods to obtain the optimized identification category corresponding to each of the multiple data points.

[0011] In one possible implementation, the recognition model is trained in the following manner:

[0012] Acquire second audio data; the second audio data includes multiple sensitive data points belonging to sensitive information arranged in chronological order, and each sensitive data point has its corresponding value;

[0013] Multiple sensitive data points are input into the recognition model to be trained in the order of the time sequence. The recognition model outputs the predicted value of the next sensitive data point after the preset time window based on the value corresponding to the sensitive data point within the preset time window.

[0014] Determine the difference between the predicted value of any one of the plurality of sensitive data points and the value of that sensitive data point, and determine the total prediction loss based on the difference corresponding to each sensitive data point;

[0015] With minimizing the total prediction loss as the training objective, the model parameters of the recognition model are adjusted to obtain the trained recognition model.

[0016] Furthermore, the second audio data is composed of several audio segments spliced ​​together, and each data point in the audio segment is sensitive information.

[0017] In one possible implementation, the statistical testing method includes an ordinal probability ratio test; the optimization of the initial identification category using the statistical testing method includes:

[0018] Obtain the first preset number of data points from the plurality of data points;

[0019] The optimized recognition category of the first preset number of data points is determined by using the sequential probability ratio test method;

[0020] The initial identification category of the first preset number of data points is updated to the corresponding optimized identification category.

[0021] Furthermore, the sequential probability ratio test method includes:

[0022] Based on the assumption that the noise follows a Gaussian distribution, the first probability of the first preset number of data points being noise is determined;

[0023] Using the predicted values ​​output by the recognition model, a second probability is determined that the first preset number of data points are noise;

[0024] Take the logarithm of the ratio of the first probability to the second probability to obtain the logarithmic result;

[0025] The optimized identification category corresponding to the first preset number of data points is determined by comparing the logarithmic result with the second threshold.

[0026] Furthermore, the second threshold is pre-selected from a number of data points that are 0 and greater than 0.

[0027] In one possible implementation, the statistical test method includes a cumulative sum method; the optimization of the initial identification category using the statistical test method includes:

[0028] By using a cumulative sum method, the statistical value of the multiple data points is accumulated and calculated to determine the time point of abrupt change in the statistical value;

[0029] Based on the initial identification category of the first data point before the time point, the optimized identification category of the second data point after the time point is determined; the optimized identification category of the second data point is different from the initial identification category of the first data point.

[0030] In one possible implementation, acquiring the first audio data to be identified includes:

[0031] The original audio is input into the Wiener filter according to the time sequence, and the first audio data is output through the Wiener filter.

[0032] In one possible implementation, the method further includes:

[0033] Based on the optimized recognition categories corresponding to the multiple data points, sensitive information audio segments are extracted from the first audio data.

[0034] Secondly, a device for identifying sensitive information in audio data is provided, the device comprising:

[0035] The acquisition unit is used to acquire first audio data to be identified; the first audio data includes multiple data points arranged in time sequence, and each data point has its corresponding value;

[0036] The prediction unit is used to input multiple data points acquired by the acquisition unit into a pre-trained recognition model in the order of the time sequence. The recognition model outputs the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window. The recognition model is a statistical learning model.

[0037] A determining unit is used to determine the difference between the predicted value of any data point among the plurality of data points obtained by the prediction unit and the value of that data point;

[0038] The identification unit is used to initially classify data points whose difference is less than a first threshold as sensitive information, and data points whose difference is greater than or equal to the first threshold as non-sensitive information, thereby obtaining the initial identification categories corresponding to the multiple data points respectively; the identification category is used to indicate whether the corresponding data point belongs to sensitive information.

[0039] An optimization unit is used to optimize the initial identification category obtained by the identification unit using statistical testing methods, so as to obtain the optimized identification category corresponding to the multiple data points respectively.

[0040] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0041] Fourthly, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method of the first aspect.

[0042] The method and apparatus provided in the embodiments of this specification first acquire first audio data to be identified; the first audio data includes multiple data points arranged in chronological order, each data point having a corresponding numerical value; then the multiple data points are input into a pre-trained recognition model in chronological order, and the recognition model outputs the predicted value of the next data point after the preset time window based on the numerical value of the data point within the preset time window; the recognition model is a statistical learning model; next, the difference between the predicted value of any data point and the numerical value of the data point is determined; then, data points with a difference less than a first threshold are initially classified as sensitive information, and data points with a difference greater than or equal to the first threshold are initially classified as non-sensitive information, thereby obtaining the initial recognition categories corresponding to the multiple data points; the recognition categories are used to indicate whether the corresponding data point belongs to sensitive information; finally, the initial recognition categories are optimized using statistical testing methods to obtain the optimized recognition categories corresponding to the multiple data points. As can be seen from the above, the embodiments of this specification do not require converting audio into text before identifying sensitive information. Instead, they directly perform time series calculations on the audio and use statistical learning models and statistical testing methods to identify sensitive information in the audio, thereby achieving a high recognition accuracy with low computational power consumption. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this specification;

[0045] Figure 2 A flowchart of a method for identifying sensitive information in audio data according to one embodiment is shown;

[0046] Figure 3 A schematic block diagram of a device for identifying sensitive information in audio data according to one embodiment is shown. Detailed Implementation

[0047] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0048] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this specification. This implementation scenario involves the identification of sensitive information in audio data. Sensitive information, also known as sensitive data or private data, refers to data that, if leaked, may cause serious harm to society or individuals. It is understood that audio data can correspond to an audio file; for example, a file ending in .mp3 is an audio file. An audio file may or may not contain sensitive information. Audio data includes multiple data points arranged chronologically. In this embodiment, it is necessary to identify whether each data point constitutes sensitive information. (Refer to...) Figure 1 This diagram illustrates a time series of audio data, comprising multiple data points arranged chronologically. Each data point corresponds to a specific moment on the horizontal axis and a numerical value on the vertical axis. The values ​​corresponding to each data point constitute a time series. A time series is a sequence of data arranged according to the chronological order of the values ​​of the same statistical indicator.

[0049] In the embodiments of this specification, by performing time-series calculations on the audio data, the identification category corresponding to each data point can be obtained. This identification category is used to indicate whether the corresponding data point belongs to sensitive information. For example, if the data point between t1 and t2 is identified as not belonging to sensitive information, the data point between t3 and t4 belongs to sensitive information, and the data point between t5 and t6 does not belong to sensitive information, the data point that does not belong to sensitive information is marked with an identification category of 0, and the data point that belongs to sensitive information is marked with an identification category of 1.

[0050] Figure 2 This diagram illustrates a method for identifying sensitive information in audio data according to one embodiment. The method can be based on... Figure 1 The implementation scenario is shown. For example... Figure 2 As shown, the method for identifying sensitive information in audio data in this embodiment includes the following steps: Step 21, acquiring first audio data to be identified; the first audio data includes multiple data points arranged in chronological order, each data point having its corresponding value; Step 22, inputting the multiple data points into a pre-trained recognition model in chronological order, the recognition model outputting the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window; the recognition model is a statistical learning model; Step 23, determining the difference between the predicted value of any data point and the value of that data point; Step 24, initially classifying data points with a difference less than a first threshold as sensitive information, and initially classifying data points with a difference greater than or equal to the first threshold as non-sensitive information, thereby obtaining the initial recognition categories corresponding to the multiple data points respectively; the recognition category is used to indicate whether the corresponding data point belongs to sensitive information; Step 25, optimizing the initial recognition categories using statistical testing methods to obtain the optimized recognition categories corresponding to the multiple data points respectively. The specific execution methods of the above steps are described below.

[0051] First, in step 21, the first audio data to be identified is obtained; the first audio data includes multiple data points arranged in chronological order, and each data point has its corresponding value. It can be understood that the values ​​corresponding to each data point can form a time series.

[0052] In one example, obtaining the first audio data to be identified includes:

[0053] The original audio is input into the Wiener filter according to the time sequence, and the first audio data is output through the Wiener filter.

[0054] In this example, a Wiener filter is used to reduce ambient noise in the original audio. The Wiener filter separates ambient noise from the microphone matrix signal as much as possible from the original audio. The first audio data is used as preprocessed time-series data for the recognition model's prediction, which helps ensure the accuracy of subsequent predictions.

[0055] Then, in step 22, the multiple data points are input into a pre-trained recognition model according to the time sequence. The recognition model then outputs the predicted value of the next data point after the preset time window based on the values ​​corresponding to the data points within that preset time window. The recognition model is a statistical learning model. It is understood that the prediction of the recognition model is related to the time sequence of each data point; this prediction is a time-series-based prediction.

[0056] In the embodiments described in this specification, the recognition model is a statistical learning model. A statistical learning model extracts statistical and mathematical features from a set of data points to achieve functions such as regression and prediction that cannot be achieved through hard-coded rules alone.

[0057] For example, if the length of the preset time window is 100, the recognition model can obtain the predicted value of the data point at time 101 based on the data points from time 0 to 100.

[0058] In one example, the recognition model is trained in the following manner:

[0059] Acquire second audio data; the second audio data includes multiple sensitive data points belonging to sensitive information arranged in chronological order, and each sensitive data point has its corresponding value;

[0060] Multiple sensitive data points are input into the recognition model to be trained in the order of the time sequence. The recognition model outputs the predicted value of the next sensitive data point after the preset time window based on the value corresponding to the sensitive data point within the preset time window.

[0061] Determine the difference between the predicted value of any one of the plurality of sensitive data points and the value of that sensitive data point, and determine the total prediction loss based on the difference corresponding to each sensitive data point;

[0062] With minimizing the total prediction loss as the training objective, the model parameters of the recognition model are adjusted to obtain the trained recognition model.

[0063] Furthermore, the second audio data is composed of several audio segments spliced ​​together, and each data point in the audio segment is sensitive information.

[0064] In the embodiments of this specification, the aforementioned audio segments may come from the same or different audio files. For example, audio segment 1 and audio segment 2 come from audio file 1, audio segment 3 comes from audio file 2, and audio segment 4 and audio segment 5 come from audio file 3. Audio segments 1, 2, 3, 4, and 5 can be concatenated to form the second audio data.

[0065] In addition, the following processing steps may be involved in training the recognition model.

[0066] First, prepare a dataset consisting of multiple raw audio recordings, which may contain sensitive data. The sources include at least one of the following: raw audio obtained through open-source channels; manually generated raw audio; automatically generated raw audio, such as raw audio produced by using software to pronounce sensitive text content; and partially authorized raw audio. In this embodiment, the specific content of the sensitive data is not limited; for example, a contract appearing in the audio may be considered sensitive data according to classification and grading rules.

[0067] Then, the dataset is divided into training set, validation set and test set according to the proportions.

[0068] Next, the ambient noise in the original audio is reduced using a Wiener filter to separate the ambient noise and microphone matrix signal from the original audio as much as possible. The filtered audio data is then used as preprocessed time series data for training the subsequent recognition model, which helps to improve the model training effect and ensure the recognition accuracy of the trained recognition model.

[0069] The filtered training, validation, and test sets are then used to train a recognition model. This model is applied to a set of time-series audio data to obtain a recognition model that can hit sensitive information. A machine learning-based regression algorithm is used, applying the mean spared error loss function. For lightweight and efficient operation, the underlying mathematical formulas are directly implemented.

[0070]

[0071] Where k is the time index of the data point within the preset time window, K is the length of the preset time window, n is the time index of the data point to be predicted, and w k To identify the model parameters, x k y(n) represents the data point within the preset time window or the value of the data point, and y(n) represents the data point to be predicted or the predicted value of the data point.

[0072]

[0073] w(i+1)=w(i)+2×μ×e(i)×x(i);

[0074]

[0075] Where w(i) represents a vector, w i L represents an element in the vector, and L and K have the same meaning. μ is a hyperparameter.

[0076] e(i) = d(i) - y(i);

[0077] Where d(i) is the value of data point i, y(i) is the predicted value of data point i, and e(i) is the difference between the two.

[0078] In the embodiments described in this specification, the time series can be iterated using the least mean square error (LMS) optimization criterion to adjust the model parameters of the recognition model. Parameter tuning methods may include adjusting the number of iterations (epochs) based on the audio volume. Smaller audio volumes require more epochs. For example, audio with millions of data points would require approximately 20-30 epochs. A set of K-order optimization parameters W[k] is obtained, and applying the last parameter w will yield better results.

[0079] After the recognition model is trained, the parameter W[k] is used to predict the test data points. Since non-sensitive data lacks a training process, the error generated during the application will be relatively large. By comparing the error e with the preset threshold, some data points that meet the threshold error can be selected. These data points are identified as sensitive data and marked as 1. Other data are non-sensitive data and marked as 0.

[0080] Next, in step 23, the difference between the predicted value and the actual value of any one of the plurality of data points is determined. It is understood that the values ​​of different data points are typically different.

[0081] This embodiment utilizes the difference in prediction accuracy between sensitive and non-sensitive information by the identification model to identify sensitive information. For sensitive information, higher prediction accuracy results in a smaller difference, while for non-sensitive information, lower prediction accuracy results in a larger difference. Next, in step 24, data points with differences less than a first threshold are initially classified as sensitive information, and data points with differences greater than or equal to the first threshold are initially classified as non-sensitive information, thereby obtaining initial identification categories corresponding to the multiple data points. These identification categories indicate whether the corresponding data point belongs to sensitive information. It is understood that the initial identification categories may not be entirely accurate, and further optimization can be performed on the initial identification categories subsequently.

[0082] Finally, in step 25, the initial identification categories are optimized using statistical tests to obtain the optimized identification categories corresponding to each of the multiple data points. It is understood that the aforementioned statistical tests may specifically include one or more methods; some data points may have the same initial identification category as their optimized identification category, while others may have a different initial identification category.

[0083] Statistical test: a statistical hypothesis test based on statistics, used to determine whether differences between samples or between a sample and the population are caused by sampling error or by essential differences.

[0084] In one example, the statistical testing method includes the ordinal probability ratio test; the optimization of the initial identification category using the statistical testing method includes:

[0085] Obtain the first preset number of data points from the plurality of data points;

[0086] The optimized recognition category of the first preset number of data points is determined by using the sequential probability ratio test method;

[0087] The initial identification category of the first preset number of data points is updated to the corresponding optimized identification category.

[0088] In this example, to address the impact of slow model start on recognition accuracy, the initialization process involves setting the initial recognition result to 0 or 1 at the beginning of the audio. For instance, choosing one-thousandth of the entire audio as the initial data point, assuming the entire audio contains 1,000,000 data points, and optimizing the first 1,000 data points can improve the recognition accuracy of sensitive data in the initial part of the audio.

[0089] Furthermore, the sequential probability ratio test method includes:

[0090] Based on the assumption that the noise follows a Gaussian distribution, the first probability of the first preset number of data points being noise is determined;

[0091] Using the predicted values ​​output by the recognition model, a second probability is determined that the first preset number of data points are noise;

[0092] Take the logarithm of the ratio of the first probability to the second probability to obtain the logarithmic result;

[0093] The optimized identification category corresponding to the first preset number of data points is determined by comparing the logarithmic result with the second threshold.

[0094] Furthermore, the second threshold is pre-selected from a number of data points that are 0 and greater than 0.

[0095] For example, the order probability ratio test method is specifically the SPRT method, assuming that the statistical model noise is Gaussian noise. Use formula Perform statistical tests on the data points and the predicted output y-values.

[0096] When the second threshold is set to 0, the initial values ​​are selected based on the sign of the J(n) result: Pi is the aforementioned first probability, and Pj is the aforementioned second probability. If the J(n) result is negative, the initial data point is marked as 1, meaning that the audio starts as sensitive data; otherwise, the initial data point is marked as 0, which represents the majority of cases, meaning that the audio starts as noise, silence, or other conversations or sounds.

[0097] In the embodiments described in this specification, the second threshold can be fine-tuned starting from 0, and the second threshold can be selected optimally. For example, the second threshold can be 0.1, in which case the initial part of the audio can be distinguished as sensitive data based on whether the J(n) result is greater than or less than 0.1.

[0098] In one example, the statistical test method includes a cumulative sum method; the optimization of the initial identification category using the statistical test method includes:

[0099] By using a cumulative sum method, the statistical value of the multiple data points is accumulated and calculated to determine the time point of abrupt change in the statistical value;

[0100] Based on the initial identification category of the first data point before the time point, the optimized identification category of the second data point after the time point is determined; the optimized identification category of the second data point is different from the initial identification category of the first data point.

[0101] This example uses the statistical test method of sequential analysis. Sequential analysis: with an uncertain sample size, a small sample is first taken, and the existing results are analyzed to determine whether to continue.

[0102] For example, the cumulative sum (CUSUM) method is used for smoothing and anomaly detection at sensitive data points. Before a signal sequence undergoes a sudden change, the CUSUM statistic g(k) is a variable that fluctuates randomly around 0. When a signal undergoes a sudden change, g(k+) continuously increases with a positive shift, which is a cumulative process. When a negative shift occurs, g(k-) continues to accumulate. When the accumulation reaches a certain level (e.g., exceeding a set threshold), a signal change can be considered to have occurred, thus detecting the transient signal. The time of the signal change can also be calculated based on the delay time d. The CUSUM change point detection experiment based on a sliding window obtains the data change behavior by calculating the difference between data points, thereby identifying and modifying relevant points. For example, a segment of data points before the sensitive data is identified can already be classified as sensitive data, but the error e of the identification model may show a decreasing trend, but not below the threshold. This CUSUM method can process and modify the point values ​​before and after the sensitive data identified by the identification model, thereby further refining the hit range.

[0103] The embodiments in this specification use the principles of time series machine learning combined with statistical hypothesis testing and sequential analysis to achieve good prediction results for sensitive information using a lightweight model.

[0104] In one example, the method further includes:

[0105] Based on the optimized recognition categories corresponding to the multiple data points, sensitive information audio segments are extracted from the first audio data.

[0106] For example, by tagging audio with 0s and 1s, it's possible to determine whether sensitive information is contained and its location, which refers to the time position, down to the second. Based on the sensitive data identification results, sensitive audio content can be directly extracted from audio filtered by a Wiener filter for subsequent analysis, protection, and other data security-related operations.

[0107] The method provided in the embodiments of this specification first acquires first audio data to be identified; the first audio data includes multiple data points arranged in chronological order, each data point having a corresponding numerical value; then the multiple data points are input into a pre-trained recognition model in chronological order, and the recognition model outputs the predicted value of the next data point after the preset time window based on the numerical value corresponding to the data point within the preset time window; the recognition model is a statistical learning model; next, the difference between the predicted value of any data point and the numerical value of that data point is determined; then, data points with a difference less than a first threshold are initially classified as sensitive information, and data points with a difference greater than or equal to the first threshold are initially classified as non-sensitive information, thereby obtaining the initial recognition categories corresponding to the multiple data points; the recognition categories are used to indicate whether the corresponding data point belongs to sensitive information; finally, the initial recognition categories are optimized using statistical testing methods to obtain the optimized recognition categories corresponding to the multiple data points. As can be seen from the above, the embodiments of this specification do not require converting audio into text before identifying sensitive information. Instead, they directly perform time series calculations on the audio and use statistical learning models and statistical testing methods to identify sensitive information in the audio, thereby achieving a high recognition accuracy with low computational power consumption.

[0108] According to another embodiment, a sensitive information identification device in audio data is also provided, which is used to perform the method provided in the embodiments of this specification. Figure 3 A schematic block diagram of a sensitive information identification device in audio data according to one embodiment is shown. Figure 3 As shown, the device 300 includes:

[0109] The acquisition unit 31 is used to acquire the first audio data to be identified; the first audio data includes multiple data points arranged in time sequence, and each data point has its corresponding value;

[0110] Prediction unit 32 is used to input multiple data points acquired by acquisition unit 31 into a pre-trained recognition model in the order of the time sequence, and the recognition model outputs the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window; the recognition model is a statistical learning model.

[0111] The determining unit 33 is used to determine the difference between the predicted value of any data point among the plurality of data points obtained by the predicting unit 32 and the value of that data point;

[0112] The identification unit 34 is used to initially classify data points whose difference is less than a first threshold obtained by the determining unit 33 as sensitive information, and to initially classify data points whose difference is greater than or equal to the first threshold as non-sensitive information, thereby obtaining the initial identification categories corresponding to the plurality of data points respectively; the identification category is used to indicate whether the corresponding data point belongs to sensitive information.

[0113] The optimization unit 35 is used to optimize the initial identification category obtained by the identification unit 34 using statistical testing methods to obtain the optimized identification category corresponding to the multiple data points respectively.

[0114] Optionally, as an embodiment, the recognition model is trained in the following manner:

[0115] Acquire second audio data; the second audio data includes multiple sensitive data points belonging to sensitive information arranged in chronological order, and each sensitive data point has its corresponding value;

[0116] Multiple sensitive data points are input into the recognition model to be trained in the order of the time sequence. The recognition model outputs the predicted value of the next sensitive data point after the preset time window based on the value corresponding to the sensitive data point within the preset time window.

[0117] Determine the difference between the predicted value of any one of the plurality of sensitive data points and the value of that sensitive data point, and determine the total prediction loss based on the difference corresponding to each sensitive data point;

[0118] With minimizing the total prediction loss as the training objective, the model parameters of the recognition model are adjusted to obtain the trained recognition model.

[0119] Furthermore, the second audio data is composed of several audio segments spliced ​​together, and each data point in the audio segment is sensitive information.

[0120] Optionally, as an embodiment, the statistical test method includes an ordinal probability ratio test; the optimization unit 35 includes:

[0121] The acquisition unit is used to acquire the first preset number of data points of the plurality of data points;

[0122] The first determining subunit is used to determine the optimized recognition category of the first preset number of data points acquired by the acquiring word unit using the sequential probability ratio test method;

[0123] An update subunit is used to update the initial identification category of the first preset number of data points to the corresponding optimized identification category determined by the first determining subunit.

[0124] Furthermore, the sequential probability ratio test method includes:

[0125] Based on the assumption that the noise follows a Gaussian distribution, the first probability of the first preset number of data points being noise is determined;

[0126] Using the predicted values ​​output by the recognition model, a second probability is determined that the first preset number of data points are noise;

[0127] Take the logarithm of the ratio of the first probability to the second probability to obtain the logarithmic result;

[0128] The optimized identification category corresponding to the first preset number of data points is determined by comparing the logarithmic result with the second threshold.

[0129] Furthermore, the second threshold is pre-selected from a number of data points that are 0 and greater than 0.

[0130] Optionally, as an embodiment, the statistical test method includes a cumulative sum method; the optimization unit 35 includes:

[0131] The second determining subunit is used to perform cumulative calculations on the multiple data points for the statistical quantity using a cumulative sum method to determine the time point of abrupt change in the statistical quantity;

[0132] The third determining subunit is used to determine the optimized identification category of the second data point after the time point based on the initial identification category of the first data point before the time point determined by the second determining subunit; the optimized identification category of the second data point is different from the initial identification category of the first data point.

[0133] Optionally, as an embodiment, the acquisition unit 31 is specifically used to input the original audio into the Wiener filter according to the time sequence, and output the first audio data through the Wiener filter.

[0134] Optionally, as an embodiment, the apparatus further includes:

[0135] The extraction unit is used to extract sensitive information audio segments from the first audio data according to the optimized recognition categories corresponding to the multiple data points.

[0136] The apparatus provided in the embodiments of this specification first acquires first audio data to be identified by acquisition unit 31. The first audio data includes multiple data points arranged in chronological order, each data point having a corresponding numerical value. Then, prediction unit 32 inputs the multiple data points into a pre-trained recognition model in chronological order. The recognition model outputs the predicted value of the next data point after the preset time window based on the numerical value of the data point within the preset time window. The recognition model is a statistical learning model. Next, determination unit 33 determines the difference between the predicted value of any data point and the numerical value of that data point. Then, recognition unit 34 initially classifies data points with differences less than a first threshold as sensitive information and data points with differences greater than or equal to the first threshold as non-sensitive information, thereby obtaining the initial recognition categories corresponding to the multiple data points. The recognition categories are used to indicate whether the corresponding data point belongs to sensitive information. Finally, optimization unit 35 optimizes the initial recognition categories using statistical testing methods to obtain the optimized recognition categories corresponding to the multiple data points. As can be seen from the above, the embodiments of this specification do not require converting audio into text before identifying sensitive information. Instead, they directly perform time series calculations on the audio and use statistical learning models and statistical testing methods to identify sensitive information in the audio, thereby achieving a high recognition accuracy with low computational power consumption.

[0137] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a combination Figure 2 The method described.

[0138] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a combination... Figure 2 The method described.

[0139] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0140] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying sensitive information in audio data, the method comprising: Obtain the first audio data to be recognized; The first audio data includes multiple data points arranged in time sequence, and each data point has its corresponding value; The multiple data points are input into a pre-trained recognition model in the order of the time sequence. The recognition model outputs the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window. The recognition model is a statistical learning model. Determine the difference between the predicted value of any data point among the plurality of data points and the actual value of that data point; Data points with a difference less than a first threshold are initially classified as sensitive information, and data points with a difference greater than or equal to the first threshold are initially classified as non-sensitive information, thereby obtaining the initial identification categories corresponding to the multiple data points respectively. The identification category is used to indicate whether the corresponding data point belongs to sensitive information; By using a cumulative sum method, the statistical value of the multiple data points is accumulated and calculated to determine the time point of abrupt change in the statistical value; Based on the initial identification category of the first data point before the time point, determine the optimized identification category of the second data point after the time point; The optimized identification category of the second data point is different from the initial identification category of the first data point, thereby obtaining the optimized identification categories corresponding to the multiple data points respectively.

2. The method as described in claim 1, wherein, The recognition model is trained in the following manner: Acquire second audio data; the second audio data includes multiple sensitive data points belonging to sensitive information arranged in chronological order, and each sensitive data point has its corresponding value; Multiple sensitive data points are input into the recognition model to be trained in the order of the time sequence. The recognition model outputs the predicted value of the next sensitive data point after the preset time window based on the value corresponding to the sensitive data point within the preset time window. Determine the difference between the predicted value of any one of the plurality of sensitive data points and the value of that sensitive data point, and determine the total prediction loss based on the difference corresponding to each sensitive data point; With minimizing the total prediction loss as the training objective, the model parameters of the recognition model are adjusted to obtain the trained recognition model.

3. The method as described in claim 2, wherein, The second audio data is composed of several audio segments spliced ​​together, and each data point in the audio segment is sensitive information.

4. The method of claim 1, wherein, The method further includes: Obtain the first preset number of data points from the plurality of data points; The optimized recognition category of the first preset number of data points is determined by using the sequential probability ratio test method; The initial identification category of the first preset number of data points is updated to the corresponding optimized identification category.

5. The method of claim 4, wherein, The sequential probability ratio test method includes: Based on the assumption that the noise follows a Gaussian distribution, the first probability of the first preset number of data points being noise is determined; Using the predicted values ​​output by the recognition model, a second probability is determined that the first preset number of data points are noise; Take the logarithm of the ratio of the first probability to the second probability to obtain the logarithmic result; The optimized identification category corresponding to the first preset number of data points is determined by comparing the logarithmic result with the second threshold.

6. The method of claim 5, wherein, The second threshold is selected in advance from a number of data points that are 0 and greater than 0.

7. The method of claim 1, wherein, The acquisition of the first audio data to be identified includes: The original audio is input into the Wiener filter according to the time sequence, and the first audio data is output through the Wiener filter.

8. The method of claim 1, wherein, The method further includes: Based on the optimized recognition categories corresponding to the multiple data points, sensitive information audio segments are extracted from the first audio data.

9. A device for identifying sensitive information in audio data, the device comprising: The acquisition unit is used to acquire the first audio data to be recognized. The first audio data includes multiple data points arranged in time sequence, and each data point has its corresponding value; The prediction unit is used to input multiple data points acquired by the acquisition unit into a pre-trained recognition model in the order of the time sequence. The recognition model outputs the predicted value of the next data point after the preset time window based on the value corresponding to the data point within the preset time window. The recognition model is a statistical learning model. A determining unit is used to determine the difference between the predicted value of any data point among the plurality of data points obtained by the prediction unit and the value of that data point; The identification unit is used to initially classify data points whose difference is less than a first threshold as sensitive information, and to initially classify data points whose difference is greater than or equal to the first threshold as non-sensitive information, thereby obtaining the initial identification categories corresponding to the multiple data points respectively. The identification category is used to indicate whether the corresponding data point belongs to sensitive information; An optimization unit is used to perform cumulative calculations on the multiple data points for a statistical measure using a cumulative sum method, determine the time point of abrupt change in the statistical measure, and determine the optimized identification category of the second data point after the time point based on the initial identification category of the first data point before the time point. The optimized identification category of the second data point is different from the initial identification category of the first data point, thereby obtaining the optimized identification categories corresponding to the multiple data points respectively.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-8.

11. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-8.

Images