Fraud identification method and device based on audio stream, storage medium and related equipment

Abstract

According to the fraud recognition method and device based on the audio streams, the storage medium and the related equipment provided by the invention, the multiple text segments which are uploaded by the client and are formed by converting the multiple segments of audio streams generated in the voice chat process of the target user are obtained; according to the method, one or more text fragments in all the text fragments can be pre-recognized, when it is recognized that one of the text fragments has a potential fraud behavior, the text fragment serves as a potential fraud fragment, and context information of the potential fraud fragment is extracted from all the text fragments; and then performing multi-dimensional fraud analysis on the potential fraud fragments and the context information according to a preset fraud analysis strategy, performing comprehensive judgment on the fraud analysis result of each dimension, and determining whether the target user has a fraud behavior or not according to the comprehensive judgment result. According to the process, the fraud analysis result of each dimension is comprehensively judged, so that the fraud recognition accuracy is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of text processing technology, and in particular to a fraud identification method, device, storage medium and related equipment based on audio stream. Background Art

[0002] In recent years, the situation of online fraud has become more serious, various fraud methods have emerged in an endless stream, and fraud technology has also developed rapidly. In order to protect the legitimate rights and interests of users, it is very important to accurately identify online fraud.

[0003] At present, when identifying various fraud methods, most of them are based on artificial intelligence. For example, the existing technology uses machine learning and deep learning algorithms to automatically extract and learn fraud features from massive data, thereby improving recognition efficiency and accuracy, or uses natural language processing technology to conduct in-depth analysis of text and voice data.

[0004] However, most of the existing technologies are identification methods used in multiple scenarios. When it comes to a specific scenario, such as a voice stream chat scenario, they are unable to accurately and efficiently identify fraudulent behavior in that scenario. Summary of the invention

[0005] The purpose of this application is to solve at least one of the above-mentioned technical defects, especially the technical defect that the prior art cannot accurately and efficiently identify fraudulent behavior in voice stream chat scenarios.

[0006] The present application provides a fraud identification method based on audio stream, the method comprising:

[0007] Acquire multiple text segments uploaded by the client, which are generated by converting multiple audio streams generated by the target user during the voice chat process;

[0008] Pre-identify one or more text segments from all text segments, and when one of the text segments is identified as potentially fraudulent, identify the text segment as a potentially fraudulent segment, and extract context information of the potentially fraudulent segment from all text segments;

[0009] A multi-dimensional fraud analysis is performed on the potential fraud fragments and the context information according to a preset fraud analysis strategy, and after a comprehensive judgment is made on the fraud analysis results of each dimension, it is determined whether the target user has engaged in fraudulent behavior based on the comprehensive judgment result.

[0010] Optionally, the pre-identifying one or more text segments among all the text segments includes:

[0011] Put each text segment into the first cache queue in sequence according to their respective time order;

[0012] Sequentially extracting multiple text segments from the first cache queue and putting them into the second cache queue until the second cache queue is full, wherein the number of cells in the second cache queue is an odd number greater than 0;

[0013] Extracting a text segment at a middle position of the second cache queue for pre-identification, and determining whether the text segment contains potential fraudulent behavior;

[0014] If not, each text segment in the second cache queue is moved one grid to the left, and the process of extracting the text segment in the middle of the second cache queue for pre-identification and subsequent steps is returned until all text segments are identified, or one of the text segments contains potential fraud.

[0015] Optionally, extracting the text segment at the middle position of the second cache queue for pre-identification and determining whether the text segment contains potential fraudulent behavior includes:

[0016] Extracting the text segment at the middle position of the second cache queue, and inputting the extracted text segment into a pre-configured single text segment fraud model to obtain a pre-recognition score output by the single text segment fraud model;

[0017] comparing the pre-identification score to a preset pre-identification fraud score threshold;

[0018] If the pre-identification score is less than the pre-identification fraud score threshold, determining that the extracted text segment does not contain potential fraudulent behavior;

[0019] If the pre-identification score is not less than the pre-identification fraud score threshold, it is determined that the extracted text segment contains potential fraudulent behavior.

[0020] Optionally, extracting context information of the potential fraudulent segment from all text segments includes:

[0021] The other text segments in the second cache queue except the potential fraud segment are used as context information of the potential fraud segment.

[0022] Optionally, performing a multi-dimensional fraud analysis on the potential fraud segment and the context information according to a preset fraud analysis strategy, and obtaining fraud analysis results of each dimension, includes:

[0023] Performing fraud key point analysis on the potential fraud fragment and the context information to obtain a fraud key point analysis result;

[0024] Performing behavior pattern analysis on the potential fraud fragment and the context information to obtain a behavior pattern analysis result;

[0025] Perform fraud word density analysis on the potential fraud segment and the context information to obtain a fraud word density analysis result.

[0026] Optionally, the performing fraud key point analysis on the potential fraud segment and the context information to obtain a fraud key point analysis result includes:

[0027] Identify multiple pre-set scam points;

[0028] The potential fraudulent segment and the context information are analyzed to determine whether they contain one or more fraudulent points and the first occurrence frequency of the included fraudulent points to obtain a fraudulent point analysis result.

[0029] Optionally, performing behavior pattern analysis on the potential fraudulent fragment and the context information to obtain a behavior pattern analysis result includes:

[0030] Identify multiple preset behavior patterns;

[0031] The potential fraud fragment and the context information are analyzed to determine whether they contain one or more behavior patterns and a second occurrence frequency corresponding to the contained behavior patterns, to obtain a behavior pattern analysis result.

[0032] Optionally, performing fraud word density analysis on the potential fraud segment and the context information to obtain a fraud word density analysis result includes:

[0033] At least one fraudulent word is determined according to the potential fraudulent segment and the context information, and the fraudulent word density of the potential fraudulent segment and the context information is analyzed according to the at least one fraudulent word to obtain a fraudulent word density analysis result.

[0034] Optionally, analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain a fraudulent word density analysis result includes:

[0035] Determining the total number of words in the potential fraudulent segment and the context information;

[0036] The fraudulent word density of the potential fraudulent segment and the context information is calculated according to the at least one fraudulent word and the total number of words to obtain a fraudulent word density analysis result.

[0037] Optionally, analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain a fraudulent word density analysis result includes:

[0038] Determine the word spacing between the potential fraudulent segment and each fraudulent word in the context information;

[0039] The fraud word density of the potential fraud segment and the context information is calculated according to the word spacing between each fraud word to obtain a fraud word density analysis result.

[0040] Optionally, after comprehensively judging the fraud analysis results of each dimension, determining whether the target user has engaged in fraudulent behavior according to the comprehensive judgment result includes:

[0041] Determine the number of features that meet the characteristics of fraudulent behavior based on the fraud analysis results of each dimension;

[0042] Comparing the number of features with a preset feature number threshold;

[0043] If the number of features is less than the threshold of the number of features, it is determined that the target user does not have any fraudulent behavior;

[0044] If the number of features is not less than the feature number threshold, it is determined that the target user has committed fraud.

[0045] Optionally, after comprehensively judging the fraud analysis results of each dimension, determining whether the target user has engaged in fraudulent behavior according to the comprehensive judgment result includes:

[0046] Determine the characteristic values ​​corresponding to different fraud behavior characteristics based on the fraud analysis results of each dimension, and determine the overall fraud probability based on the characteristic values ​​corresponding to each fraud behavior characteristic;

[0047] Comparing the overall fraud probability with a preset overall fraud probability threshold;

[0048] If the overall fraud probability is less than the overall fraud probability threshold, it is determined that the target user does not have any fraudulent behavior;

[0049] If the overall fraud probability is not less than the overall fraud probability threshold, it is determined that the target user has committed fraud.

[0050] The present application also provides a fraud identification device based on audio stream, comprising:

[0051] A text segment acquisition module is used to acquire multiple text segments uploaded by the client, which are generated by converting multiple audio streams generated by the target user during the voice chat process;

[0052] A potential fraud identification module is used to pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, the text segment is regarded as a potential fraud segment, and context information of the potential fraud segment is extracted from all text segments;

[0053] The fraud behavior identification module is used to perform multi-dimensional fraud analysis on the potential fraud fragments and the context information according to a preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment results.

[0054] The present application also provides a computer-readable storage medium, in which computer-readable instructions are stored. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the fraud identification method based on audio stream as described in any of the above embodiments.

[0055] The present application also provides a computer device, comprising: one or more processors, and a memory;

[0056] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the fraud identification method based on audio stream as described in any one of the above embodiments are performed.

[0057] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0058] The audio stream-based fraud identification method, device, storage medium and related equipment provided by the present application can first obtain multiple text fragments uploaded by the client, which are formed by converting multiple audio streams generated by the target user during the voice chat process. Since the client has converted the audio streams in the voice chat process into text fragments, the server only needs to identify fraud based on the text fragments, and there is no need to receive a large number of audio streams uploaded by the client, thereby saving a lot of costs; then, the present application can pre-identify one or more text fragments among all text fragments, and when one of the text fragments is identified as having potential fraudulent behavior, the text fragment is used as a potential fraud fragment, and the context information of the potential fraud fragment is extracted from all text fragments, so that high-accuracy fraud data can be effectively recalled while reducing false calls, thereby quickly identifying fraudulent behavior; finally, the present application can perform multi-dimensional fraud analysis on the potential fraud fragments and contextual information according to the preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment result. This process conducts multi-dimensional fraud identification based on the characteristics of the voice stream during real-time audio chat, and makes a comprehensive judgment on the fraud analysis results of each dimension, thereby effectively improving the accuracy of fraud identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application 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, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0060] Figure 1 A flowchart of a fraud identification method based on audio stream provided in an embodiment of the present application;

[0061] Figure 2 A schematic diagram of a process in which a client converts an audio stream into a text segment provided in an embodiment of the present application;

[0062] Figure 3 A schematic diagram of a process for pre-recognizing one or more text segments provided in an embodiment of the present application;

[0063] Figure 4 A schematic diagram of a text segment recognition process in a text buffer pool 2 provided in an embodiment of the present application;

[0064] Figure 5 A schematic diagram of the process of calculating the distance between fraudulent words provided in the embodiment of the present application;

[0065] Figure 6A schematic diagram of a comprehensive determination result of a fraud analysis result provided in an embodiment of the present application;

[0066] Figure 7 Another schematic diagram of a comprehensive determination result of a fraud analysis result provided in an embodiment of the present application;

[0067] Figure 8 A schematic diagram of the structure of a fraud identification device based on audio stream provided in an embodiment of the present application;

[0068] Fig. 9 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0070] In one embodiment, Figure 1 As shown, Figure 1 A flowchart of a fraud identification method based on an audio stream provided in an embodiment of the present application; the present application provides a fraud identification method based on an audio stream, the method may include:

[0071] S110: Acquire multiple text segments uploaded by the client, which are formed by converting multiple audio streams generated by the target user during the voice chat.

[0072] In this step, when providing voice services, the client can enable the fraud identification function according to relevant regulations or with the user's consent, so as to better provide services to users and protect their legitimate rights and interests.

[0073] Specifically, in the present application, during a voice chat with a target user, the client can convert multiple audio streams generated during the voice chat into text segments, and then upload each text segment to the server so that the server can identify whether the target user has committed fraud based on each text segment.

[0074] The target user can be a user who has enabled the fraud identification function by default or by the user, or a user with a low credit rating identified by the platform. The specific setting can be made according to the actual situation and is not limited here. After the client obtains the multiple audio streams generated by the target user during the voice chat, the voice conversion model can be used to convert the multiple audio streams into corresponding text segments.

[0075] Indicatively, Figure 2 As shown, Figure 2 A schematic diagram of a process in which a client converts an audio stream into a text segment provided in an embodiment of the present application; Figure 2 In the model, when the audio stream enters the model, the model will first segment the audio stream according to the duration T, such as every 10 seconds, then T=10, and then convert each audio stream into text according to the ASR algorithm. Figure 2 After the audio clip on the left is input, the ASR algorithm converts the audio into text on the right: Sogou Voice. After each audio stream is converted into a text segment, a corresponding timestamp can be added after the text segment, so that the order of each text segment can be determined based on the timestamp.

[0076] S120: Pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, treat the text segment as a potential fraudulent segment, and extract context information of the potential fraudulent segment from all text segments.

[0077] In this step, after obtaining multiple text fragments formed by converting multiple audio streams generated by the target user during the voice chat uploaded by the client through S110, the present application can pre-identify one or more text fragments among all text fragments, and when one of the text fragments is identified as having potential fraud, the text fragment is used as a potential fraud fragment, and context information corresponding to the potential fraud fragment is extracted from all text fragments, and then whether there is fraud is further analyzed based on the potential fraud fragment and the context information.

[0078] Specifically, after obtaining multiple text fragments, the present application can pre-identify one or more text fragments among all the text fragments. For example, the present application can extract text fragments that may contain fraudulent behavior from all the text fragments according to the characteristics of the voice stream, and pre-identify them. The pre-identification process can use relevant models for identification, or other methods for identification. The specific settings can be made according to the actual situation and are not limited here. When it is determined after preliminary identification that there is fraudulent behavior, the text fragment can be used as a potential fraudulent fragment, and subsequent verification operations can be continued after obtaining its context information; if it is initially identified that there is no fraudulent behavior, other text fragments can continue to be extracted from all the text fragments for pre-identification until all the text fragments are identified, or potential fraudulent fragments are identified. In this way, fraud data with high accuracy can be effectively recalled while reducing false calls, and fraudulent behavior can be quickly identified.

[0079] S130: Perform multi-dimensional fraud analysis on potential fraud segments and context information according to a preset fraud analysis strategy, and after making a comprehensive judgment on the fraud analysis results of each dimension, determine whether the target user has committed fraudulent behavior based on the comprehensive judgment result.

[0080] In this step, after the potential fraud fragments are identified through S120 and the context information corresponding to the potential fraud fragments are extracted from all text fragments, the present application can perform multi-dimensional fraud analysis on the potential fraud fragments and context information according to the preset fraud analysis strategy, and then obtain the fraud analysis results of each dimension. Then, the present application can also make a comprehensive judgment on the fraud analysis results of each dimension, and determine whether the target user has committed fraudulent behavior based on the comprehensive judgment results.

[0081] It is understandable that when the present application performs pre-identification through S120 and identifies that one of the text segments contains potential fraudulent behavior, in order to further verify whether the identification result is accurate, the present application can obtain the context information of the potential fraudulent segment, and combine its context information to perform multi-dimensional fraud analysis, where the dimensions can be set according to the characteristics of the voice stream. For example, the present application can analyze the key points of fraud in the potential fraud segment and context information, such as age, location, character relationships, etc., and can also analyze the behavior patterns in the potential fraud segment and context information, such as repeated action recognition, repeated confirmation recognition, etc., and can also analyze the density of fraudulent words in the potential fraud segment and context information, or analyze other features in the potential fraud segment and context information. The specific settings can be made according to the actual situation and are not limited here.

[0082] After the present application conducts a multi-dimensional fraud analysis on potential fraud fragments and contextual information according to a preset fraud analysis strategy, in order to further improve the accuracy of fraud identification, the present application can make a comprehensive judgment on the fraud analysis results of each dimension. For example, the present application can score the credibility of the fraud analysis results of each dimension, and then determine whether there is fraud based on the comprehensive score; the final fraud judgment can also be made through algorithms such as the gpdt decision tree, or other methods can be used to determine whether there is fraud. The specific choice can be made based on the actual situation and is not limited here.

[0083] In the above embodiment, first, multiple text segments formed by converting multiple audio streams generated by the target user during the voice chat process uploaded by the client can be obtained. Since the client has converted the audio streams during the voice chat process into text segments, the server only needs to identify fraud based on the text segments, without receiving a large number of audio streams uploaded by the client, thereby saving a lot of costs; then, the present application can pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, the text segment is used as a potential fraud segment, and the context information of the potential fraud segment is extracted from all text segments, so that fraud data with high accuracy can be effectively recalled while reducing false calls, and fraudulent behavior can be quickly identified; finally, the present application can perform multi-dimensional fraud analysis on the potential fraud segments and context information according to the preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment result. This process performs multi-dimensional fraud identification according to the characteristics of the voice stream in the real-time audio chat process, and comprehensively judges the fraud analysis results of each dimension, thereby effectively improving the accuracy of fraud identification.

[0084] In one embodiment, pre-identifying one or more text segments among all text segments in S120 may include:

[0085] S121: Put each text segment into the first cache queue in sequence according to their respective time sequence.

[0086] S122: extracting multiple text segments from the first cache queue in sequence and putting them into the second cache queue until the second cache queue is full, wherein the number of cells in the second cache queue is an odd number greater than 0.

[0087] S123: extracting the text segment at the middle position of the second cache queue for pre-identification, and determining whether the text segment contains potential fraudulent behavior.

[0088] S124: If not, each text segment in the second cache queue is moved one grid to the left, and the process of extracting the text segment in the middle of the second cache queue for pre-identification and subsequent steps is returned until all text segments are identified, or one of the text segments contains potential fraud.

[0089] In this embodiment, when pre-identifying one or more text fragments among all text fragments, the present application can first put each text fragment into a first buffer queue in chronological order, and then extract multiple text fragments from the first buffer queue in turn and put them into the second buffer queue until the second buffer queue is full. Before that, the present application stipulates that the number of grids in the second cache queue is an odd number greater than 0, so that it is possible to identify whether there is potential fraud through the middle position of the second cache queue, and when there is potential fraud, the context information is obtained through the second cache queue, and other key points of fraud are analyzed on the context information.

[0090] For example, if Figure 3 As shown, Figure 3 A schematic diagram of a process for pre-recognizing one or more text segments provided in an embodiment of the present application; Figure 3 In this application, the text segments decoded by ASR can be placed into the text buffer pool 1 in chronological order. This method can solve the problem of data discard caused by the inconsistency between the ASR decoding speed and the single text fraud identification speed. If the ASR decoding speed is too fast, the decoded audio segment text can be pushed into the text buffer pool 1 from right to left, and the text segments in the text buffer pool 1 can be automatically cleaned up when they are stored overtime. For example, this application can set the timeout threshold to T_th = 10 minutes. If the timestamp of the text segment in the text buffer pool 1 is greater than the current time, and the exceeded time is greater than T_th, the text segment in the text buffer pool 1 will be cleared.

[0091] Furthermore, when the present application performs fraud pre-identification on a single audio clip text in text buffer pool 1, text clips in text buffer pool 1 can be taken from left to right and put into text buffer pool 2 in sequence, and the positions of the text clips taken out in text buffer pool 1 are cleared. Then, the present application can perform single text clip fraud pre-identification on the middle position of text buffer pool 2. For example, assuming that the position of text clip i is 0, when the length of text buffer pool 2 is 5, the i+2 position is taken for single text clip fraud pre-identification. If the length is 7, the i+3 position is taken for single text clip fraud pre-identification.

[0092] Furthermore, after the present application extracts the text segment at the middle position of the text buffer pool 2 for pre-identification, if it is determined that the text segment does not have potential fraud, the text segment in the text buffer pool 2 is moved one grid to the left, and the text segment at the i+2 position is continuously identified until all the text segments in the text buffer pool 2 are identified, or one of the text segments has potential fraud. Furthermore, when all the text segments in the text buffer pool 2 are identified, if there are still text segments in the text buffer pool 1, the text segments can continue to be extracted from the text buffer pool 1 to the text buffer pool 2, and the text segment at the middle position of the text buffer pool 2 can be pre-identified. If there is no text segment at the middle position of the text buffer pool 2, the identification is terminated. If a text segment with potential fraud is identified during this process, the identification can also be terminated, and the potential fraud can continue to be verified.

[0093] In one embodiment, extracting the text segment at the middle position of the second cache queue for pre-identification in S123 and determining whether the text segment has potential fraudulent behavior may include:

[0094] S1231: Extract the text segment at the middle position of the second cache queue, and input the extracted text segment into a pre-configured single text segment fraud model to obtain a pre-recognition score output by the single text segment fraud model.

[0095] S1232: Compare the pre-identification score with a preset pre-identification fraud score threshold.

[0096] S1233: If the pre-identification score is less than the pre-identification fraud score threshold, it is determined that the extracted text segment does not contain potential fraudulent behavior.

[0097] S1234: If the pre-identification score is not less than the pre-identification fraud score threshold, it is determined that the extracted text segment contains potential fraudulent behavior.

[0098] In this embodiment, when pre-identifying the text segment at the middle position of the second cache queue, the present application can first extract the text segment at the middle position of the second cache queue, and then input the extracted text segment into a pre-configured single text segment fraud model to identify the pre-identification score of the text segment through the single text segment fraud model. Then, the present application can compare the pre-identification score with the preset pre-identification fraud score threshold. If the pre-identification score is less than the pre-identification fraud score threshold, it is determined that the extracted text segment does not have potential fraud behavior; if the pre-identification score is not less than the pre-identification fraud score threshold, it is determined that the extracted text segment has potential fraud behavior.

[0099] In a specific implementation, the present application can input the extracted text fragment into a pre-configured single text fragment fraud model. If the pre-recognition score output by the single text fragment fraud model is lower than the pre-recognition fraud score threshold Th_score_1, for example, the present application can set 0<=Th_score_1<=1, such as Th_score_1=0.5. The specific value can be set according to the actual situation and is not limited here. When it is determined that the pre-recognition score is lower than the pre-recognition fraud score threshold, the text fragment in the second cache queue is moved one grid to the left, and then the text fragment in the middle position is continued to be recognized. If the pre-recognition score output by the single text fragment fraud model is greater than or equal to the pre-recognition fraud score threshold Th_score_1, the text fragment is determined to be a potential fraud fragment. At this time, the context information corresponding to the text fragment can be obtained, and fraud behavior identification can be performed based on the potential fraud fragment and the context information.

[0100] In one embodiment, extracting the context information of the potential fraudulent segment from all text segments in S120 may include:

[0101] S125: Using other text segments in the second cache queue except the potential fraud segment as context information of the potential fraud segment.

[0102] In this embodiment, after the present application pre-identifies the text segment in the middle position of the second cache queue and determines that there is a potential fraud segment, the other text segments in the second cache queue except the potential fraud segment can be directly used as the context information of the potential fraud segment. This can not only quickly obtain the context information for fraudulent behavior identification, but also effectively recall fraud data with high accuracy while reducing false calls, thereby quickly identifying fraudulent behavior.

[0103] In one embodiment, in S130, a multi-dimensional fraud analysis is performed on the potential fraud segment and the context information according to a preset fraud analysis strategy, and fraud analysis results of each dimension are obtained, which may include:

[0104] S131: Analyze the potential fraudulent clips and the context information to obtain a fraudulent key point analysis result.

[0105] S132: Performing behavior pattern analysis on the potential fraudulent fragment and the context information to obtain a behavior pattern analysis result.

[0106] S133: Perform fraud word density analysis on the potential fraud segment and the context information to obtain a fraud word density analysis result.

[0107] In this embodiment, when performing multi-dimensional fraud analysis on potential fraud segments and context information, the potential fraud segments and context information can be respectively subjected to fraud key point analysis, behavior pattern analysis and fraud word density analysis, so that a comprehensive judgment can be made by combining the fraud analysis results of the three.

[0108] Among them, when the present application performs a fraud point analysis on potential fraud clips and context information, it can first determine the fraud point to be analyzed, and then search whether the fraud point exists in the potential fraud clip and context information, and the number of times it appears; when performing a behavior pattern analysis on potential fraud clips and context information, it can also first determine the behavior pattern to be analyzed, and then search whether the behavior pattern exists in the potential fraud clip and context information, and the number of times it appears; when performing a fraud word density analysis on potential fraud clips and context information, it can first obtain one or more fraud words in the potential fraud clip and context information, and then calculate the fraud word density of the potential fraud clip and context information, and then obtain the final fraud analysis result.

[0109] In one embodiment, performing fraud key point analysis on the potential fraud segment and the context information in S131 to obtain a fraud key point analysis result may include:

[0110] S1311: Determine multiple preset fraud points.

[0111] S1312: Analyze whether the potential fraud segment and the context information contain one or more fraud points and the first occurrence frequency corresponding to the included fraud points to obtain a fraud point analysis result.

[0112] In this embodiment, when analyzing the key points of fraud in potential fraud clips and contextual information, multiple preset key points of fraud can be determined first, such as location, time, character relationships, inducement, threats, etc., and then the potential fraud clips and contextual information are analyzed to see whether they contain one or more of the key points of fraud and the first frequency of occurrence corresponding to the contained key points of fraud. In this way, the final key points of fraud analysis results can be determined in combination with the key points of fraud that appear and the corresponding first frequency of occurrence.

[0113] For example, if Figure 4 As shown, Figure 4 A schematic diagram of the text segment recognition process in the text buffer pool 2 provided in the embodiment of the present application; when the text segment of the i+2 segment in the text buffer pool 2 of the present application is: "That's easy to find. You find a quiet place. Okay, find a quiet place. I don't understand. No.", the single text segment pre-recognition result is:

[0114] {'model_label': 'Fraud', 'model_score': '0.95', 'textWords': 'That's easy to find. Find a quiet place. Okay. Find a quiet place. I don't understand.'}, after comparing its pre-recognition score with the pre-recognition fraud score threshold, if it is greater than the pre-recognition fraud score threshold, it can be determined that the text segment at the middle position of the current text buffer pool 2 is a potential fraud segment. At this time, the potential fraud segment and context information can be analyzed for fraud points, and the specific analysis is as follows:

[0115] (1) Age analysis: No age-related data are included;

[0116] (2) Location analysis: including quiet places;

[0117] (3) Character relationship analysis: Character relationships are not included;

[0118] (4) Induction analysis: including delivering skin and guiding operation;

[0119] Instructions: 1. Ask him to write it and I will send it together; 2. Find a quiet place; 3. Go back to my room first and don’t go into the egg;

[0120] (5) Threat analysis: No threat data is included.

[0121] According to the analysis results, the first occurrence frequency corresponding to each fraud point can be counted, and then the final fraud point analysis results can be determined according to the fraud points that appear and the corresponding first occurrence frequencies.

[0122] In one embodiment, the behavior pattern analysis of the potential fraudulent segment and the context information in S132 to obtain the behavior pattern analysis result may include:

[0123] S1321: Determine multiple preset behavior modes.

[0124] S1322: Analyze whether the potential fraud segment and the context information contain one or more behavior patterns and the second occurrence frequency corresponding to the contained behavior patterns to obtain a behavior pattern analysis result.

[0125] In this embodiment, when performing behavioral pattern analysis on potential fraudulent clips and contextual information, the present application can pre-set multiple behavioral patterns, such as repeated action recognition, repeated confirmation recognition, etc., and then analyze whether the potential fraudulent clips and contextual information contain one or more of the behavioral patterns and the second occurrence frequency corresponding to the included behavioral patterns. In this way, the final behavioral pattern analysis results can be determined based on the occurring behavioral patterns and the second occurrence frequency corresponding to the behavioral patterns.

[0126] Illustratively, the present application continues to use the above example to illustrate that after determining that the text segment at the middle position of the current text buffer pool 2 is a potential fraud segment, the present application can perform a behavior pattern analysis on the potential fraud segment and context information, and the specific analysis is as follows:

[0127] (1) Repetitive action recognition: Find a quiet place with inconspicuous repetitive actions.

[0128] (2) Repeated confirmation identification: Data related to repeated confirmation is not included.

[0129] According to the analysis results, the second occurrence frequency corresponding to each behavior pattern can be counted, and then the final behavior pattern analysis results can be determined according to the behavior pattern and the corresponding second occurrence frequency. Of course, in addition to the above-mentioned behavior patterns, this application can also set other behavior patterns based on past experience, which can be set according to actual conditions and are not limited here.

[0130] In one embodiment, performing fraud word density analysis on the potential fraud segment and the context information in S133 to obtain a fraud word density analysis result may include:

[0131] S1331: Determine at least one fraudulent word according to the potential fraudulent segment and the context information, and analyze the fraudulent word density of the potential fraudulent segment and the context information according to the at least one fraudulent word to obtain a fraudulent word density analysis result.

[0132] In this embodiment, when performing a fraudulent word density analysis on potential fraudulent clips and context information, the present application can first determine at least one fraudulent word based on the potential fraudulent clips and context information, and then analyze the fraudulent word density of the potential fraudulent clips and context information based on the at least one fraudulent word, so that the final fraudulent word density analysis result can be obtained.

[0133] Among them, when determining the potential fraudulent segment and at least one fraudulent word in the context information, the present application can construct a fraudulent word list, which can be provided by the business or extracted through NER algorithm recognition. For example, the fraudulent word list constructed by the present application through the potential fraudulent segment and context information in the above example is as follows:

[0134]

[0135] Table 1 List of fraud words

[0136] Next, the present application can calculate the density of fraudulent words in potential fraudulent fragments and contextual information based on the fraudulent word list. The calculation process can be performed by using the frequency statistics of fraudulent words, or by using the spacing between fraudulent words to calculate the density of fraudulent words. Other methods can also be used for calculation. The specific settings can be made according to the actual situation and are not limited here.

[0137] In one embodiment, in S1331, analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain the fraudulent word density analysis result may include:

[0138] S3311: Determine the total number of words in the potential fraudulent segment and the context information.

[0139] S3312: Calculate the fraudulent word density of the potential fraudulent segment and the context information according to the at least one fraudulent word and the total number of words to obtain a fraudulent word density analysis result.

[0140] In this embodiment, when analyzing the fraudulent word density of potential fraudulent segments and context information based on at least one fraudulent word, the present application can first determine the total number of words in the potential fraudulent segments and context information, and then calculate the fraudulent word density of the potential fraudulent segments and context information based on at least one fraudulent word and the total number of words, and then obtain the final fraudulent word density analysis result.

[0141] In a specific implementation, the text of the potential fraudulent snippet and context information of this application is as follows:

[0142] text = 'But you can't give a duplicate name, right? You asked him to write this thing out. I'll send it together. Okay, you need to look for it. It's easy to find. Find a quiet place. Okay, find a quiet place. I don't understand. You wait a moment. Go back to my room first. Don't go in. I'll send these two skin threads to you. I'll send two to each of you, okay?'. This text contains 81 words, of which 13 are scam words, namely:

[0143] 3 free

[0144] 3 quiet

[0145] 4 Find

[0146] 1 egg

[0147] 1 skin

[0148] 1 back

[0149] Therefore, the fraud word density of the potential fraud segment and context information is calculated as follows:

[0150] Fraud word density = number of fraud words / total number of potential fraud fragments and context information

[0151] This calculation method can be used to obtain the fraud word density of potential fraud fragments and context information: density = 13 / 81 = 0.16. Assuming that the fraud word density threshold is: Th_density = 0.05, the current density >Th_density, indicating that it meets the characteristics of high-density fraud keywords.

[0152] In one embodiment, in S1331, analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain the fraudulent word density analysis result may include:

[0153] S3313: Determine the word spacing between the potential fraudulent segment and each fraudulent word in the context information.

[0154] S3314: Calculate the fraudulent word density of the potential fraudulent segment and the context information according to the word spacing between each fraudulent word, and obtain a fraudulent word density analysis result.

[0155] In this embodiment, when analyzing the fraudulent word density of potential fraudulent segments and context information based on at least one fraudulent word, the present application can first determine the word spacing between each fraudulent word in the potential fraudulent segment and context information, and then calculate the fraudulent word density of the potential fraudulent segment and context information based on the word spacing between each fraudulent word, and then obtain the final fraudulent word density analysis result.

[0156] Indicatively, Figure 5 As shown, Figure 5 A schematic diagram of the process of calculating the word spacing between fraudulent words provided in the embodiment of the present application; Figure 5 From the distance between the fraud words in the second row, we can calculate the fraud word density as: (6+4+3+3+1+1) / 6=3, based on which we can get the final fraud word density analysis result.

[0157] In one embodiment, after comprehensively judging the fraud analysis results of each dimension in S130, determining whether the target user has fraudulent behavior according to the comprehensive judgment result may include:

[0158] S310: Determine the number of features that meet the characteristics of fraudulent behavior based on the fraud analysis results of each dimension.

[0159] S311: Compare the number of features with a preset feature number threshold.

[0160] S312: If the number of features is less than the feature number threshold, it is determined that the target user does not have any fraudulent behavior.

[0161] S313: If the number of features is not less than the feature number threshold, it is determined that the target user has engaged in fraudulent behavior.

[0162] In this embodiment, when comprehensively judging the fraud analysis results of various dimensions, a variety of judgment methods can be used to make a comprehensive judgment.

[0163] To elaborate, the present application can determine the number of features that meet the characteristics of fraudulent behavior based on the fraud analysis results of various dimensions, and then compare the number of features with a preset feature number threshold. If the number of features is less than the feature number threshold, it is determined that the target user does not have fraudulent behavior; if the number of features is not less than the feature number threshold, it is determined that the target user has committed fraudulent behavior.

[0164] For example, the present application can obtain fraud analysis results of various dimensions through the above calculation process, and determine the number of features that meet the characteristics of fraud behavior according to the fraud analysis results of various dimensions, such as Figure 6 As shown, Figure 6 A schematic diagram of a comprehensive determination result of a fraud analysis result provided in an embodiment of the present application; Figure 6 It is possible to clearly see whether each fraud behavior feature in the fraud analysis results of each dimension meets the fraud behavior features, and based on the number of features that meet the fraud behavior features and the feature number threshold, it can be determined whether the target user has engaged in fraudulent behavior.

[0165] In one embodiment, after comprehensively judging the fraud analysis results of each dimension in S130, determining whether the target user has fraudulent behavior according to the comprehensive judgment result may include:

[0166] S320: Determine the characteristic values ​​corresponding to different fraud behavior characteristics based on the fraud analysis results of each dimension, and determine the overall fraud probability based on the characteristic values ​​corresponding to each fraud behavior characteristic.

[0167] S321: Compare the overall fraud probability with a preset overall fraud probability threshold.

[0168] S322: If the overall fraud probability is less than the overall fraud probability threshold, it is determined that the target user does not have any fraudulent behavior.

[0169] S323: If the overall fraud probability is not less than the overall fraud probability threshold, it is determined that the target user has engaged in fraudulent behavior.

[0170] In this embodiment, when comprehensively judging the fraud analysis results of various dimensions, an upper limit can be set for the value of each fraud behavior feature, so as to calculate the overall probability of fraud.

[0171] For example, the present application can determine the characteristic values ​​corresponding to different fraud behavior characteristics according to the fraud analysis results of each dimension, and determine the overall fraud probability according to the characteristic values ​​corresponding to each fraud behavior characteristic, as follows:

[0172] (1) The upper limit of the number of features mentioning age is 2. If there are more than 2 mentions of fraudulent age in the potential fraud segment and context information, the upper limit is 2.

[0173] (2) The upper limit of the number of features mentioning the location is 3. If there are more than 3 mentions of the fraud location in the potential fraud segment and context information, then it is set to 3;

[0174] (3) The upper limit of the number of character relationship features is 3. If there are more than 3 fraud character relationships in the potential fraud segment and context information, then it is set to 3;

[0175] (4) The upper limit of the number of induced features is 10. If there are more than 10 fraud inducements in the potential fraud segment and context information, it is set to 10;

[0176] (5) The upper limit of the number of threat features is 2, which is set to 2 if there are more than 2 fraud threats in the potential fraud segment and context information;

[0177] (6) The upper limit of the number of features for repeated action recognition is 3. If there are more than 3 fraudulent repeated action recognitions in the potential fraud clip and context information, it is set to 3;

[0178] (7) The upper limit of the number of features for repeated confirmation is 3. If there are more than 3 fraud repeated confirmations in the potential fraud segment and context information, it is set to 3;

[0179] (8) Due to the sparsity of the fraud word density value, the value needs to be enlarged. For example, the density value is divided by the density threshold to obtain the converted fraud word density value. The current converted word density value is: 0.16 / 0.05=3.2;

[0180] (9) The i+2 segment text fraud score needs to be amplified by 10 times to obtain a converted score of 9.5.

[0181] After the above calculation, we can get Figure 7 The schematic diagram of the comprehensive determination result is shown in FIG. Figure 7 The overall fraud probability can be calculated as follows:

[0182] Score = (0+2+0+6+0+1+0+3.2+9.5) / (2+3+3+10+2+3+3) = 0.83

[0183] So the overall fraud probability is: 0.83

[0184] The overall fraud probability threshold Th_defaud can be obtained through sample statistics, such as Th_defaud = 0.6. After comparing it with the overall fraud probability value, it can be determined whether the target user has fraudulent behavior.

[0185] The following is a description of a fraud identification device based on an audio stream provided in an embodiment of the present application. The fraud identification device based on an audio stream described below and the fraud identification method based on an audio stream described above can refer to each other.

[0186] In one embodiment, Figure 8 As shown, Figure 8 A schematic diagram of a fraud identification device based on an audio stream provided in an embodiment of the present application; the present application also provides a fraud identification device based on an audio stream, which may include a text segment acquisition module 210, a potential fraud identification module 220, and a fraud behavior identification module 230, specifically including the following:

[0187] The text segment acquisition module 210 is used to acquire a plurality of text segments uploaded by the client, which are formed by converting a plurality of audio streams generated by the target user during the voice chat.

[0188] The potential fraud identification module 220 is used to pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, the text segment is identified as a potential fraud segment, and context information of the potential fraud segment is extracted from all text segments.

[0189] The fraud behavior identification module 230 is used to perform multi-dimensional fraud analysis on the potential fraud fragments and the context information according to a preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment results.

[0190] In the above embodiment, first, multiple text segments formed by converting multiple audio streams generated by the target user during the voice chat process uploaded by the client can be obtained. Since the client has converted the audio streams during the voice chat process into text segments, the server only needs to identify fraud based on the text segments, without receiving a large number of audio streams uploaded by the client, thereby saving a lot of costs; then, the present application can pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, the text segment is used as a potential fraud segment, and the context information of the potential fraud segment is extracted from all text segments, so that fraud data with high accuracy can be effectively recalled while reducing false calls, and fraudulent behavior can be quickly identified; finally, the present application can perform multi-dimensional fraud analysis on the potential fraud segments and context information according to the preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment result. This process performs multi-dimensional fraud identification according to the characteristics of the voice stream in the real-time audio chat process, and comprehensively judges the fraud analysis results of each dimension, thereby effectively improving the accuracy of fraud identification.

[0191] In one embodiment, the present application also provides a computer-readable storage medium, in which computer-readable instructions are stored. When the computer-readable instructions are executed by one or more processors, the one or more processors perform the steps of the audio stream-based fraud identification method as described in any of the above embodiments.

[0192] In one embodiment, the present application also provides a computer device, including: one or more processors, and a memory.

[0193] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the fraud identification method based on audio stream as described in any one of the above embodiments are performed.

[0194] Indicatively, Fig. 9 As shown, Fig. 9 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 may be provided as a server. Fig. 9The computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by a memory 301 for storing instructions executable by the processing component 302, such as an application. The application stored in the memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 302 is configured to execute instructions to perform the fraud identification method based on the audio stream of any of the above embodiments.

[0195] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM, or the like.

[0196] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0197] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0198] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0199] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. 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 application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fraud identification method based on audio stream, characterized in that: The method comprises: Acquire multiple text segments uploaded by the client, which are generated by converting multiple audio streams generated by the target user during the voice chat process; Pre-identify one or more text segments from all text segments, and when one of the text segments is identified as potentially fraudulent, identify the text segment as a potentially fraudulent segment, and extract context information of the potentially fraudulent segment from all text segments; A multi-dimensional fraud analysis is performed on the potential fraud fragments and the context information according to a preset fraud analysis strategy, and after a comprehensive judgment is made on the fraud analysis results of each dimension, it is determined whether the target user has engaged in fraudulent behavior based on the comprehensive judgment result.

2. The fraud identification method based on audio stream according to claim 1 is characterized in that: The pre-identification of one or more text segments among all the text segments includes: Put each text segment into the first cache queue in sequence according to their respective time order; Sequentially extracting multiple text segments from the first cache queue and putting them into the second cache queue until the second cache queue is full, wherein the number of cells in the second cache queue is an odd number greater than 0; Extracting a text segment at a middle position of the second cache queue for pre-identification, and determining whether the text segment contains potential fraudulent behavior; If not, each text segment in the second cache queue is moved one grid to the left, and the process of extracting the text segment in the middle of the second cache queue for pre-identification and subsequent steps is returned until all text segments are identified, or one of the text segments contains potential fraud.

3. The fraud identification method based on audio stream according to claim 2 is characterized in that: The extracting of the text segment at the middle position of the second cache queue for pre-identification and determining whether the text segment contains potential fraudulent behavior includes: Extracting the text segment at the middle position of the second cache queue, and inputting the extracted text segment into a pre-configured single text segment fraud model to obtain a pre-recognition score output by the single text segment fraud model; comparing the pre-identification score to a preset pre-identification fraud score threshold; If the pre-identification score is less than the pre-identification fraud score threshold, determining that the extracted text segment does not contain potential fraudulent behavior; If the pre-identification score is not less than the pre-identification fraud score threshold, it is determined that the extracted text segment contains potential fraudulent behavior.

4. The fraud identification method based on audio stream according to claim 2 is characterized in that: The step of extracting context information of the potential fraudulent fragments from all text fragments includes: The other text segments in the second cache queue except the potential fraud segment are used as context information of the potential fraud segment.

5. The fraud identification method based on audio stream according to claim 1 is characterized in that: The method of performing multi-dimensional fraud analysis on the potential fraud fragment and the context information according to the preset fraud analysis strategy and obtaining fraud analysis results of each dimension includes: Performing fraud key point analysis on the potential fraud fragment and the context information to obtain a fraud key point analysis result; Performing behavior pattern analysis on the potential fraud fragment and the context information to obtain a behavior pattern analysis result; Perform fraud word density analysis on the potential fraud segment and the context information to obtain a fraud word density analysis result.

6. The fraud identification method based on audio stream according to claim 5 is characterized in that: The performing fraud key point analysis on the potential fraud segment and the context information to obtain the fraud key point analysis result includes: Identify multiple pre-set scam points; The potential fraudulent segment and the context information are analyzed to determine whether they contain one or more fraudulent points and the first occurrence frequency of the included fraudulent points to obtain a fraudulent point analysis result.

7. The fraud identification method based on audio stream according to claim 5 is characterized in that: The performing behavior pattern analysis on the potential fraudulent fragment and the context information to obtain the behavior pattern analysis result includes: Identify multiple preset behavior patterns; The potential fraud fragment and the context information are analyzed to determine whether they contain one or more behavior patterns and a second occurrence frequency corresponding to the contained behavior patterns, to obtain a behavior pattern analysis result.

8. The fraud identification method based on audio stream according to claim 5 is characterized in that: The performing fraud word density analysis on the potential fraud segment and the context information to obtain a fraud word density analysis result includes: At least one fraudulent word is determined according to the potential fraudulent segment and the context information, and the fraudulent word density of the potential fraudulent segment and the context information is analyzed according to the at least one fraudulent word to obtain a fraudulent word density analysis result.

9. The fraud identification method based on audio stream according to claim 8 is characterized in that: The step of analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain a fraudulent word density analysis result includes: Determining the total number of words in the potential fraudulent segment and the context information; The fraudulent word density of the potential fraudulent segment and the context information is calculated according to the at least one fraudulent word and the total number of words to obtain a fraudulent word density analysis result.

10. The fraud identification method based on audio stream according to claim 8, characterized in that: The step of analyzing the potential fraudulent segment and the fraudulent word density of the context information according to the at least one fraudulent word to obtain a fraudulent word density analysis result includes: Determine the word spacing between the potential fraudulent segment and each fraudulent word in the context information; The fraud word density of the potential fraud segment and the context information is calculated according to the word spacing between each fraud word to obtain a fraud word density analysis result.

11. The fraud identification method based on audio stream according to any one of claims 1 to 10, characterized in that: After comprehensively judging the fraud analysis results of each dimension, determining whether the target user has engaged in fraudulent behavior based on the comprehensive judgment results includes: Determine the number of features that meet the characteristics of fraudulent behavior based on the fraud analysis results of each dimension; Comparing the number of features with a preset feature number threshold; If the number of features is less than the threshold of the number of features, it is determined that the target user does not have any fraudulent behavior; If the number of features is not less than the feature number threshold, it is determined that the target user has committed fraud.

12. The fraud identification method based on audio stream according to any one of claims 1 to 10, characterized in that: After comprehensively judging the fraud analysis results of each dimension, determining whether the target user has engaged in fraudulent behavior based on the comprehensive judgment results includes: Determine the characteristic values ​​corresponding to different fraud behavior characteristics based on the fraud analysis results of each dimension, and determine the overall fraud probability based on the characteristic values ​​corresponding to each fraud behavior characteristic; Comparing the overall fraud probability with a preset overall fraud probability threshold; If the overall fraud probability is less than the overall fraud probability threshold, it is determined that the target user does not have any fraudulent behavior; If the overall fraud probability is not less than the overall fraud probability threshold, it is determined that the target user has committed fraud.

13. A fraud identification device based on audio stream, characterized in that: include: A text segment acquisition module is used to acquire multiple text segments uploaded by the client, which are generated by converting multiple audio streams generated by the target user during the voice chat process; A potential fraud identification module is used to pre-identify one or more text segments among all text segments, and when one of the text segments is identified as having potential fraudulent behavior, the text segment is regarded as a potential fraud segment, and context information of the potential fraud segment is extracted from all text segments; The fraud behavior identification module is used to perform multi-dimensional fraud analysis on the potential fraud fragments and the context information according to a preset fraud analysis strategy, and after comprehensively judging the fraud analysis results of each dimension, determine whether the target user has fraudulent behavior based on the comprehensive judgment results.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the fraud identification method based on audio stream as described in any one of claims 1 to 12.

15. A computer device, characterized in that: include: one or more processors, and memory; The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the fraud identification method based on audio stream as claimed in any one of claims 1 to 12 are performed.

Images