Privacy number anti-fraud system and method based on machine learning large model and ASR

Abstract

The invention relates to the technical field of communication service and cloud communication, and particularly discloses a privacy number anti-fraud system based on a machine learning large model and ASR, which comprises a data statistics preprocessing module, a machine learning module, an automatic voice recognition module, an automatic inspection module, a call content recognition and judgment module, a service alarm module and a service management and control module. The invention further discloses a privacy number anti-fraud method based on the machine learning large model and the ASR, and the privacy number anti-fraud method adopts the anti-fraud system. According to the privacy number anti-fraud system and method based on the machine learning large model and the ASR, fraud-related call behavior characteristics can be actively mined, high-risk fraud-related numbers can be fully automatically intercepted, and fraud behaviors of frauders by using a privacy number platform can be effectively suppressed.

Description

Technical Field

[0001] The present invention relates to the field of communication services and cloud communication technology, and in particular to a privacy number anti-fraud system and method based on a large machine learning model and ASR. Background Art

[0002] As platform customers pay more and more attention to user privacy information, the business volume of the privacy number operation platform has exceeded 10 million per day. Faced with 10 million data, it poses a huge challenge to business risk control, abnormal call discovery and call link tracking. Fraudsters take advantage of the opportunity to defraud users on the privacy number platform. The call words are hidden, making it difficult for business operators to identify. Existing technical means lack effective grouping of abnormal call records, and abnormal call link tracking is in a vacant state.

[0003] The existing anti-fraud methods mainly involve issuing fraud tickets after the fact, manually tracing back suspicious calls, and then blacklisting the called numbers and freezing the customer accounts in the corresponding calls. The existing methods are inefficient and lagging, and cannot effectively intercept fraudulent calls in a timely manner, leaving huge security risks and loopholes. Summary of the invention

[0004] The present invention aims to solve at least one of the above-mentioned technical problems, and provides a privacy number anti-fraud system and method based on a large machine learning model and ASR, which can actively explore the characteristics of fraudulent call behavior, automatically intercept high-risk fraudulent numbers, and effectively curb the behavior of fraudsters using privacy number platforms to commit fraud.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is: a privacy number anti-fraud system based on a machine learning large model and ASR, comprising:

[0006] The data statistics preprocessing module is responsible for refining the database-based call record data and completing the entire data extraction, conversion and loading process to achieve data preprocessing;

[0007] Machine learning module: responsible for clustering the pre-processed data, predicting and classifying the call groups according to key indicator data, and dividing the calls into normal call groups and abnormal call groups;

[0008] The automatic speech recognition module automatically recognizes the abnormal call groups predicted by the machine learning module and outputs text results;

[0009] The automatic inspection module has the function of creating scheduled tasks, and can pull the call recordings of the call group classified as fraudulent in the machine learning module into the automatic voice recognition module, and then schedule the recognized text content to the keyword recognition and large model judgment tasks, so that the keyword recognition and large model judgment models can execute tasks regularly and learn the text content;

[0010] The call content recognition and judgment module has the functions of keyword recognition and large model judgment. The keyword recognition function can identify the text content of abnormal call group conversations and judge whether the conversations contain high-risk fraud-related content. The large model judgment function can make semantic understanding of the text results output by the automatic voice recognition module and judge whether they meet the characteristics of fraud-related calls.

[0011] The business alarm module uses the process engine to aggregate machine learning data and content recognition results, evaluate risk levels and classifications, and then push the alarm content to the business control module;

[0012] The business control module responds with different control strategies based on the alarm level and alarm classification of the business alarm module, and also performs supplementary maintenance on the customer portrait system.

[0013] Preferably, the key indicator data include call duration, connection rate, and short call ratio, and calls with a duration of less than 10s, a connection rate of less than 40%, and a short call ratio of more than 50% are classified as abnormal calls.

[0014] Preferably, the risk levels are divided into level one risk, level two risk, and level three risk. Level one risk is prohibited and illegal activities, level two risk is suspected fraud, level three risk is high-risk business, and level four risk is malicious harassment.

[0015] Preferably, the management and control strategy includes: when the operation receives alarms at various levels, the manual review priority of the business is arranged according to the risk level, that is, the level 1 risk and level 2 risk businesses are reviewed first. If the review results are correct, the level 1 risk and level 2 risk customers will have their business directly frozen and taken offline. For level 3 and level 4 risks, if the review results are correct, the business will be taken offline as appropriate after communication with the sales responsible.

[0016] The present application also discloses a privacy number anti-fraud method based on a machine learning large model and ASR, using the above anti-fraud system, including the following steps:

[0017] (1) The data statistics preprocessing module extracts, converts, and loads basic call record information, and outputs key call feature indicator data to the machine learning module;

[0018] (2) performing cluster analysis on the call characteristic index data output by the data preprocessing module through the machine learning module, and dividing the call group into a normal call group and an abnormal call group;

[0019] (3) recognizing the abnormal call group call recording file output by machine learning into text content through the automatic speech recognition module;

[0020] (4) Perform keyword recognition on the text content through the keyword recognition function of the call content recognition and judgment module. If the text content hits the keyword, the hit keyword is extracted. If the text content does not hit the keyword, the classification is manually reviewed to see if it is correct. If the classification is incorrect, the model is optimized. If the classification is correct, new keywords are marked. And the large model judgment function of the call content recognition and judgment module is used to judge whether the text content conforms to abnormal calls. If the text content conforms to abnormal calls, the judgment result is output. If the text content does not conform to abnormal calls, the judgment is manually reviewed to see if it is correct. If the judgment is wrong, the model is optimized, and then the text learning results and machine learning results are aggregated.

[0021] (5) The business alarm module determines the alarm level of the aggregated results after machine learning and content recognition, and outputs the alarm level and category. The business control module makes corresponding policy controls based on the alarm level and category, and performs supplementary maintenance on the customer portrait system.

[0022] Preferably, the model tuning method is to manually review inaccurate call cases to determine whether the specific call scenarios of the call cases are consistent with the actual reported scenarios. If they are consistent, the keywords in the scenario are manually screened and the new keywords are maintained in the corresponding scenario keyword library. If they are inconsistent, the keywords are maintained in the reported and actual inconsistent keyword library.

[0023] The beneficial effect is: compared with the prior art, the privacy number anti-fraud system and method based on machine learning big model and ASR of the present invention is based on the machine learning of indicator data and the support of big model for automatic speech recognition technology, which takes a two-pronged approach, has high efficiency in screening abnormal call groups, and is more accurate in identifying and judging abnormal call content. By preprocessing tens of millions of call records, the machine learning module classifies the call groups according to the call characteristics of the call groups, and quickly screens out call groups suspected of fraud. Keyword recognition and big model judgment perform secondary verification and judgment on the call content of these screened out suspected fraud call groups. In this way, fraud call numbers can be identified more accurately, and the control response speed of abnormal call groups can be accelerated, thereby curbing the occurrence of fraud cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings, wherein:

[0025] Figure 1 This is a schematic diagram of the privacy number anti-fraud system architecture based on a machine learning large model and ASR of the present invention;

[0026] Figure 2Output flow chart for discovering and identifying fraudulent call groups;

[0027] Figure 3 This is a timing diagram for managing fraudulent calls. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] It should be noted that when a component is referred to as being "fixed to" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "connected to" another component, it may be directly connected to the other component or there may be a central component at the same time. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component at the same time. When a component is referred to as being "set in the middle", it does not only mean being set in the middle, as long as it is not set at both ends within the range defined by the middle. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0031] like Figures 1 to 3 As shown, the present application discloses a privacy number anti-fraud system based on a machine learning large model and ASR, including:

[0032] The data statistics preprocessing module is responsible for refining the database-based call record data and completing the entire data extraction, conversion and loading process to achieve data preprocessing;

[0033] Machine learning module: responsible for cluster analysis of pre-processed data, predicting and classifying call groups according to key indicator data, and dividing calls into normal call groups and abnormal call groups. Key indicator data include call duration, connection rate, short call ratio, etc. The abnormal call group is predicted based on these key indicators. The abnormal call group includes those with call duration less than 10s, connection rate less than 40%, and short call ratio more than 50%.

[0034] The automatic speech recognition module automatically recognizes the abnormal call groups predicted by the machine learning module and outputs text results;

[0035] The automatic inspection module has the function of creating scheduled tasks, and can pull the call recordings of the call group classified as fraudulent in the machine learning module into the automatic voice recognition module, and then schedule the recognized text content to the keyword recognition and large model judgment tasks, so that the keyword recognition and large model judgment models can execute tasks regularly and learn the text content;

[0036] The call content recognition and judgment module has the functions of keyword recognition and large model judgment. The keyword recognition function can identify the text content of abnormal call group conversations and judge whether the conversations contain high-risk fraud-related content. The large model judgment function can make semantic understanding of the text results output by the automatic voice recognition module and judge whether they meet the characteristics of fraud-related calls.

[0037] The business alarm module uses the process engine to aggregate machine learning data and content recognition results to evaluate risk levels and classifications, and then pushes the alarm content to the business control module. The risk levels are divided into level 1 risk, level 2 risk, and level 3 risk. Level 1 risk is prohibited and illegal, level 2 risk is suspected fraud, level 3 risk is high-risk business, and level 4 risk is malicious harassment.

[0038] The business control module responds with different control strategies based on the alarm level and alarm classification of the business alarm module, and also performs supplementary maintenance on the customer portrait system. The following is the alarm level classification:

[0039] Risk level classification Detailed classification

[0040] Level 1 prohibited pornographic services, religious propaganda, illegal drugs, reactionary speeches, etc.

[0041] Second level suspected of fraud: financial fraud, fraudulent rhetoric, cashback, payment links, etc.

[0042] Level 3 high-risk businesses: financial collection, legal consulting, medical beauty, POS, game marketing, tobacco / alcohol / tea, etc. Level 4 malicious harassment activity promotion, blind calling, repeated harassment, vulgar insults, etc.

[0043] Management and control strategy: When the operation receives alarms at various levels, the manual review priority of the business is arranged according to the risk level, that is, the first-level risk and second-level risk businesses are reviewed first. If the review results are correct, the first-level and second-level risk customers will have their business directly frozen and taken offline. For level 3 and level 4 risks, if the review results are correct, the sales responsible will be communicated with and the business will be taken offline as appropriate.

[0044] In addition, the present application also discloses a privacy number anti-fraud method based on a machine learning large model and ASR, using the above anti-fraud system, including the following steps:

[0045] (1) The data statistics preprocessing module extracts, converts, and loads basic call record information, and outputs key call feature indicator data to the machine learning module;

[0046] (2) performing cluster analysis on the call characteristic index data output by the data preprocessing module through the machine learning module, and dividing the call group into a normal call group and an abnormal call group;

[0047] (3) recognizing the abnormal call group call recording file output by machine learning into text content through the automatic speech recognition module;

[0048] (4) The keyword recognition function of the call content recognition and judgment module is used to recognize keywords in the text content. If the text content hits the keywords, the hit keywords are extracted. If the text content does not hit the keywords, the classification is manually reviewed to see if it is correct. If the classification is incorrect, the model is optimized. If the classification is correct, new keywords are marked. And the large model judgment function of the call content recognition and judgment module is used to judge whether the text content meets the requirements of abnormal calls. If the text content meets the requirements of abnormal calls, the judgment result is output. If the text content does not meet the requirements of abnormal calls, the judgment is manually reviewed to see if it is correct. If the judgment is wrong, the model is optimized, and then the text learning results and the machine learning results are aggregated. Among them, the optimization method is to manually review inaccurate call cases to determine whether the specific call scenarios of the call cases are consistent with the actual reported scenarios. If they are consistent, the keywords in the scenarios are manually screened and the new keywords are maintained in the corresponding scenario keyword library. If they are inconsistent, the keywords are maintained in the keyword library of reported and actual inconsistency.

[0049] (5) The business alarm module determines the alarm level of the aggregated results after machine learning and content recognition, and outputs the alarm level and category. The business control module makes corresponding policy controls based on the alarm level and category, and performs supplementary maintenance on the customer portrait system.

[0050] The privacy number anti-fraud system and method based on machine learning big model and ASR of the present invention is based on the machine learning of indicator data and the support of big model for automatic speech recognition technology. It takes a two-pronged approach, has high efficiency in screening abnormal call groups, and is more accurate in identifying and judging abnormal call content. By preprocessing tens of millions of call records, the machine learning module classifies the call groups according to the call characteristics, and quickly screens out call groups suspected of being involved in fraud. Keyword recognition and big model judgment perform secondary verification and judgment on the call content of these screened out suspected fraud call groups. In this way, fraud call numbers can be identified more accurately, and the control response speed of abnormal call groups can be accelerated, thereby curbing the occurrence of fraud cases.

[0051] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Any modification or equivalent substitution that does not depart from the spirit and scope of the present invention shall be included in the scope of the technical solution of the present invention.

Claims

1. A privacy number anti-fraud system based on a large machine learning model and ASR, characterized in that: include: The data statistics preprocessing module is responsible for refining the database-based call record data and completing the entire data extraction, conversion and loading process to achieve data preprocessing; Machine learning module: responsible for clustering analysis of preprocessed data, predicting and classifying call groups based on key indicator data, and dividing calls into normal call groups and abnormal call groups. The automatic speech recognition module automatically recognizes the abnormal call groups predicted by the machine learning module and outputs text results; The automatic inspection module has the function of creating scheduled tasks, and can pull the call recordings of the call group classified as fraudulent in the machine learning module into the automatic voice recognition module, and then schedule the recognized text content to the keyword recognition and large model judgment tasks, so that the keyword recognition and large model judgment models can execute tasks regularly and learn the text content; The call content recognition and judgment module has the functions of keyword recognition and large model judgment. The keyword recognition function can identify the text content of abnormal call group conversations and judge whether the conversations contain high-risk fraud-related content. The large model judgment function can make semantic understanding of the text results output by the automatic voice recognition module and judge whether they meet the characteristics of fraud-related calls. The business alarm module uses the process engine to aggregate machine learning data and content recognition results, evaluate risk levels and classifications, and then push the alarm content to the business control module; The business control module responds with different control strategies based on the alarm level and alarm classification of the business alarm module, and also performs supplementary maintenance on the customer portrait system.

2. According to claim 1, a privacy number anti-fraud system based on machine learning large model and ASR is characterized in that: The key indicator data include call duration, connection rate, and short call ratio. Calls with a duration of less than 10 seconds, a connection rate of less than 40%, and a short call ratio of more than 50% are classified as abnormal calls.

3. According to claim 1, a privacy number anti-fraud system based on machine learning large model and ASR is characterized in that: The risk levels are divided into level one risk, level two risk, and level three risk. Level one risk is prohibited and illegal activities, level two risk is suspected fraud, level three risk is high-risk business, and level four risk is malicious harassment.

4. According to claim 1, a privacy number anti-fraud system based on machine learning large model and ASR is characterized in that: The management and control strategies include: when the operation receives alarms at various levels, the manual review priority of the business is arranged according to the risk level, that is, the first-level risk and second-level risk businesses are reviewed first. If the review results are correct, the first-level risk and second-level risk customers will have their business directly frozen and taken offline. For level 3 and level 4 risks, if the review results are correct, the sales responsible will be communicated with and the business will be taken offline as appropriate.

5. A privacy number anti-fraud method based on a large machine learning model and ASR, characterized in that: The anti-fraud system according to any one of claims 1 to 4 comprises the following steps: (1) The data statistics preprocessing module extracts, converts, and loads basic call record information, and outputs key call feature indicator data to the machine learning module; (2) performing cluster analysis on the call characteristic index data output by the data preprocessing module through the machine learning module, and dividing the call group into a normal call group and an abnormal call group; (3) recognizing the abnormal call group call recording file output by machine learning into text content through the automatic speech recognition module; (4) Perform keyword recognition on the text content through the keyword recognition function of the call content recognition and judgment module. If the text content hits the keyword, the hit keyword is extracted. If the text content does not hit the keyword, the classification is manually reviewed to see if it is correct. If the classification is incorrect, the model is optimized. If the classification is correct, new keywords are marked. And the large model judgment function of the call content recognition and judgment module is used to judge whether the text content meets the requirements of abnormal calls. If the text content meets the requirements of abnormal calls, the judgment result is output. If the text content does not meet the requirements of abnormal calls, the judgment is manually reviewed to see if it is correct. If the judgment is wrong, the model is optimized, and then the text learning results and machine learning results are aggregated. (5) The business alarm module determines the alarm level of the aggregated results after machine learning and content recognition, and outputs the alarm level and category. The business control module makes corresponding policy controls based on the alarm level and category, and performs supplementary maintenance on the customer portrait system.

6. According to claim 5, a privacy number anti-fraud method based on machine learning large model and ASR is characterized in that: The model tuning method is to manually review inaccurate call cases to determine whether the specific call scenarios of the call cases are consistent with the actual reported scenarios. If they are consistent, the keywords in the scenario are manually screened and the new keywords are maintained in the corresponding scenario keyword library. If they are inconsistent, the keywords are maintained in the reported and actual inconsistent keyword library.

Images