Anti-fraud model training and anti-fraud identification processing method and device and readable storage medium

Abstract

The invention provides an anti-fraud model training method and device, an anti-fraud recognition processing method and device and a readable storage medium, and the anti-fraud model training method comprises the steps: obtaining operator multi-dimensional data and user communication text data; performing feature extraction and fusion processing based on the operator multi-dimensional data and the user communication text data to obtain multi-modal behavior features; based on the multi-modal behavior characteristics, model training is carried out through a contrast learning mechanism, an anti-fraud model is obtained, and the anti-fraud model is used for carrying out anti-fraud recognition processing. The cross-channel fraud recognition capability of the model is improved, and the detection of complex fraud behaviors is enhanced. In addition, richer information is provided in the aspect of capturing text details and behavior characteristics, and the recognition accuracy of the model in different scenes is improved. In addition, the distinguishing ability of the model to fraud and non-fraud behaviors is remarkably improved. The training mode strengthens the perception ability of the model to the abnormal behavior mode.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to an anti-fraud model training and anti-fraud identification processing method, device and readable storage medium. Background Art

[0002] In recent years, with the popularization of the Internet and the development of information technology, online fraud incidents have emerged one after another, causing huge economic losses to the general public. Among them, telecommunications fraud is a relatively common fraud method. Criminals illegally defraud the legitimate interests of the people through SMS fraud, telephone fraud, phishing and other methods.

[0003] At present, traditional anti-fraud methods mainly rely on manual identification and prevention, which are inefficient and difficult to cope with increasingly complex online fraud methods.

[0004] Therefore, how to improve the efficiency and accuracy of anti-fraud processing has become a problem that needs to be solved. Summary of the invention

[0005] The technical problem to be solved by the present application is to provide an anti-fraud model training and anti-fraud identification processing method, device and readable storage medium to solve the problems existing in the prior art in view of the above-mentioned deficiencies in the prior art.

[0006] In a first aspect, the present application provides an anti-fraud model training method, the method comprising:

[0007] S1. Obtaining multi-dimensional data of operators and user communication text data;

[0008] S2. Extracting and fusing the operator's multidimensional data and the user's communication text data to obtain multimodal behavior features;

[0009] S3. Based on the multimodal behavior characteristics, model training is performed through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

[0010] In some embodiments, in S1:

[0011] The operator multi-dimensional data includes: user information, SMS records, call records, base station locations, and network access logs;

[0012] The user communication text data includes: SMS data, call data, and traffic data.

[0013] In some embodiments, S2 includes:

[0014] S21, extracting features from user communication text data through the BERT model and the ERNIE model to obtain text features;

[0015] S22, extracting and encoding the operator's multi-dimensional data through a Transformer encoder to obtain a multi-dimensional representation of user behavior;

[0016] S23. Through a cross-layer connection mechanism, the text features and the multi-dimensional representation of the user behavior are fused to obtain multimodal behavior features.

[0017] In some embodiments, S21 includes:

[0018] The BERT model is used to encode the input user communication text data with a bidirectional Transformer structure, and the text context information is processed through a multi-head attention mechanism and a feedforward neural network to generate a deep semantic representation at the vocabulary level.

[0019] The ERNIE model is used to introduce entity information from an external knowledge base into user communication text data to obtain the potential entity relationship recognition results in fraudulent texts.

[0020] The text features include the deep semantic representation and potential entity relationship recognition results.

[0021] In some embodiments, in S3, the contrastive learning mechanism includes:

[0022] Maximize the similarity of positive sample pairs, and minimize the similarity of negative sample pairs;

[0023] Among them, the positive sample pair is the fraudulent text and the corresponding behavior data, and the negative sample pair is the non-fraudulent text and the corresponding behavior data.

[0024] In a second aspect, the present application provides an anti-fraud identification processing method, the method comprising:

[0025] Obtaining data of an object to be identified;

[0026] Inputting the object data into a pre-trained anti-fraud model to obtain an anti-fraud recognition result of the object data;

[0027] Wherein, the anti-fraud model is trained by the above-mentioned anti-fraud model training method.

[0028] In a third aspect, the present application provides an anti-fraud model training device, the device comprising:

[0029] A first data acquisition module, which is configured to acquire operator multi-dimensional data and user communication text data;

[0030] A data processing module, which is configured to perform feature extraction and fusion processing based on the operator's multi-dimensional data and the user's communication text data to obtain multimodal behavior features;

[0031] The model training module is configured to perform model training based on the multimodal behavior characteristics through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

[0032] In a fourth aspect, the present application provides an anti-fraud identification processing device, the device comprising:

[0033] A second data acquisition module, configured to acquire object data to be identified;

[0034] The anti-fraud identification module is configured to input the object data into a pre-trained anti-fraud model to obtain an anti-fraud identification result of the object data.

[0035] In a fifth aspect, the present application provides an anti-fraud device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the above-mentioned anti-fraud model training method or anti-fraud identification processing method.

[0036] In a sixth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the anti-fraud model training method or the anti-fraud identification processing method is implemented.

[0037] The anti-fraud model training and anti-fraud identification processing method, device and readable storage medium provided by the present application, wherein the anti-fraud model training method includes: obtaining operator multidimensional data and user communication text data; performing feature extraction and fusion processing based on the operator multidimensional data and user communication text data to obtain multimodal behavior features; based on the multimodal behavior features, performing model training through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing. The present application combines operator multidimensional data (such as SMS records, call records, network access logs, and base station locations, etc.) with text data to achieve a comprehensive judgment of fraudulent behavior through multimodal fusion technology. This method improves the model's ability to identify cross-channel fraud, not only limited to text, but also captures user behavior patterns, and enhances the detection of complex fraudulent behavior. In addition, a cross-layer connection mechanism is introduced to fuse the text representations generated by BERT and ERNIE at different levels with the multimodal features of operator data. This cross-layer information fusion method provides richer information in capturing text details and behavioral characteristics, and improves the recognition accuracy of the model in different scenarios. In addition, a contrastive learning mechanism was used during model training, which significantly improved the model's ability to distinguish between fraudulent and non-fraudulent behaviors by maximizing the similarity of positive sample pairs (such as fraudulent text and related operator behavior data) and minimizing the similarity of negative sample pairs. This training method strengthens the model's ability to perceive abnormal behavior patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0039] Figure 1 A schematic diagram of an anti-fraud model training method provided in an embodiment of the present application;

[0040] Figure 2 A schematic diagram of an anti-fraud identification processing method provided in an embodiment of the present application;

[0041] Figure 3 A specific example diagram of anti-fraud identification processing provided in an embodiment of the present application;

[0042] Figure 4 A schematic diagram of an anti-fraud model training device provided in an embodiment of the present application;

[0043] Figure 5 A schematic diagram of an anti-fraud identification processing device provided in an embodiment of the present application;

[0044] Figure 6 A schematic diagram of an anti-fraud device provided in an embodiment of the present application.

[0045] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the technical solution of the present application, the implementation mode of the present application will be further described in detail below with reference to the accompanying drawings.

[0047] It should be understood that the specific embodiments and drawings described herein are only used to explain the present application, rather than to limit the present application.

[0048] It can be understood that, in the absence of conflict, the various embodiments in the present application and the various features in the embodiments can be combined with each other.

[0049] It can be understood that, for the convenience of description, the drawings of the present application only show the parts related to the present application, while the parts unrelated to the present application are not shown in the drawings.

[0050] It can be understood that each unit and module involved in the embodiments of the present application may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0051] It can be understood that the terms "first", "second", etc. in the embodiments of the present application are used to distinguish different objects, or to distinguish different processing of the same object, rather than to describe a specific order of objects.

[0052] It is understandable that, in the absence of conflict, the functions and steps marked in the flowcharts and block diagrams of the present application may occur in an order different from that marked in the drawings.

[0053] It is understood that the flowcharts and block diagrams of the present application illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present application. Among them, each box in the flowchart or block diagram may represent a unit, module, program segment, code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or by a combination of hardware and computer instructions.

[0054] It can be understood that the units and modules involved in the embodiments of the present application can be implemented by software or hardware, for example, the units and modules can be located in a processor.

[0055] This application provides an anti-fraud model training method and an anti-fraud identification processing method. This application is based on the text anti-fraud model training process of BERT and ERNIE, combining the operator's multi-dimensional data (including SMS content, call records, network access logs, user location, etc.) with text data, and uses multimodal fusion technology to achieve more accurate fraud behavior identification. By applying the trained anti-fraud model to anti-fraud identification processing, the efficiency and accuracy of anti-fraud processing can be effectively improved.

[0056] The main ideas of this application are:

[0057] First, the anti-fraud model uses BERT to generate deep semantic representation of text, and introduces entity and phrase information from the external knowledge base through ERNIE to enhance the ability to recognize hidden patterns and entity relationships in fraudulent texts.

[0058] Subsequently, the multidimensional data provided by the operator is processed through feature extraction and cleaning to form behavioral features corresponding to the text data, which are encoded through the Transformer structure. The text features and the multimodal features of the operator data are fused through a cross-layer connection mechanism, and the text information output by each layer is combined with the behavioral features to generate a comprehensive user behavior representation.

[0059] Finally, through the contrastive learning mechanism, the model further optimizes the ability to distinguish between positive sample pairs (such as fraudulent text and related operator behavior data) and negative sample pairs (non-fraudulent text and normal behavior data). Through the joint training process, the model's loss function integrates text loss and operator data loss, improving the model's performance in identifying complex fraud scenarios, especially for cross-channel fraud (such as SMS fraud, phone fraud, phishing, etc.) with higher accuracy and robustness.

[0060] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0061] The present application provides an anti-fraud model training method and an anti-fraud identification and processing method. The working process of the method can be implemented by electronic devices, such as computers, handheld smart terminals, etc. For the convenience of explanation, the embodiments of the present application are described with the method execution subject being a computer.

[0062] Figure 1 A schematic diagram of the anti-fraud model training method provided in the embodiment of the present application, such as Figure 1 As shown, the present application provides an anti-fraud model training method, the method includes S1-S4, which are as follows:

[0063] S1. Obtaining multi-dimensional data of operators and user communication text data;

[0064] Among them, operator multi-dimensional data refers to all data authorized by the user side, and user communication text data refers to some public data on the Internet and some operator-specific data.

[0065] In some embodiments, in S1: the operator multi-dimensional data includes: user information, SMS records, call records, base station locations, and network access logs;

[0066] Among them, user information includes the user's personal information, such as name, gender, age, ID number, contact information, user level, place of origin, etc. This information helps operators better understand and serve users.

[0067] Among them, the SMS log records the content, time, and other information of the SMS messages sent and received by the user.

[0068] Among them, the call records include information such as the duration of the user's call, the time of the call, the other party's mobile phone number, and the type of call (calling or called).

[0069] The base station location refers to the location information of the base station to which the user is connected when using the mobile network service. This information can be used to locate the approximate location of the user.

[0070] Among them, network access logs record the data generated by users when using mobile networks, such as Internet access time, websites visited, applications used, data traffic, and other information.

[0071] In some embodiments, in S1: the user communication text data includes: SMS data, call data, and traffic data.

[0072] Among them, SMS data mainly refers to the content of SMS messages sent and received by users. This data is in text format and can be used to analyze users' communication behaviors and preferences.

[0073] Among them, call data is the text information in the call record, such as call time, call duration, call object, call content, etc.

[0074] Among them, traffic data refers to the data traffic information generated by users when using mobile networks, including Internet access time, websites and applications visited, traffic consumed, etc.

[0075] During the model training process, by introducing labels based on operator data (such as call patterns, SMS keywords, etc.), the model's ability to identify cross-channel fraud (SMS, phone calls, and online fraud) is improved.

[0076] S2. Extracting and fusing the operator's multidimensional data and the user's communication text data to obtain multimodal behavior features;

[0077] In some embodiments, S2 includes:

[0078] S21, extracting features from user communication text data through the BERT model and the ERNIE model to obtain text features;

[0079] S22, extracting and encoding the operator's multi-dimensional data through a Transformer encoder to obtain a multi-dimensional representation of user behavior;

[0080] S23. Through a cross-layer connection mechanism, the text features and the multi-dimensional representation of the user behavior are fused to obtain multimodal behavior features.

[0081] In some embodiments, S21 includes:

[0082] The BERT model is used to encode the input user communication text data with a bidirectional Transformer structure, and the text context information is processed through a multi-head attention mechanism and a feedforward neural network to generate a deep semantic representation at the vocabulary level.

[0083] The ERNIE model is used to introduce entity information from an external knowledge base into user communication text data to obtain the potential entity relationship recognition results in fraudulent texts.

[0084] The text features include the deep semantic representation and potential entity relationship recognition results.

[0085] Specifically, this application uses BERT as a pre-trained language model, encodes the input text with a bidirectional Transformer structure, processes text context information through a multi-head attention mechanism and a feedforward neural network, and generates a deep semantic representation at the vocabulary level; in addition, the ERNIE model is used to introduce entity information from an external knowledge base into the text representation, combined with a multi-granularity information integration strategy to capture the potential knowledge in the fraudulent text.

[0086] In some embodiments, the BERT model is based on a multi-head attention mechanism and a feedforward neural network architecture, and the multi-head attention calculation formula is:

[0087]

[0088] MultiHead(Q,K,V)=Concat(head1,…,head h )W O

[0089] in, is a linear transformation matrix, Q, K, and V represent query, key, and value, respectively. Multi-head attention can enhance the model's ability to capture different semantic levels.

[0090] In some embodiments, the ERNIE model introduces an external knowledge base and combines entity-level and phrase-level masked language models in the pre-training stage to improve the ability to identify potential entity relationships in fraudulent texts; the ERNIE model is trained through phrase-level masking, and its loss function is:

[0091]

[0092] By introducing entity-level knowledge and multi-granularity information, ERNIE can better identify fraudulent methods hidden in fraudulent texts. For example, in combination with external knowledge bases, ERNIE can automatically detect certain keywords in fraudulent speech and their implicit contextual associations, thereby improving the model's ability to capture fraudulent behavior. Entity relationship modeling: By modeling the relationships between fraud-related entities (such as "bank", "verification code", etc.), ERNIE can have a deeper understanding of fraud patterns. For example, in the model, user behavior is mapped to fraud terms to automatically determine whether certain key semantic combinations may be fraudulent.

[0093] In this application, the anti-fraud model introduces multimodal data from operators, including text messages, call records, network access logs, and user location information, and performs joint modeling. The Transformer encoder is used to extract and encode the operator data to generate a multidimensional representation of user behavior. The formula for multimodal data fusion is as follows:

[0094] h multimodal = Transformer([h text ,h call ,h web ])

[0095] Among them, h text , represents the representation of text data, h call , indicates call records, h web Represents a representation of a network access log.

[0096] In this application, the text features output by each layer of BERT and ERNIE are fused with the multimodal behavioral features generated by the operator data through a cross-layer connection mechanism. For fraud text data, BERT can be used as the basic text encoder, ERNIE can be used for knowledge enhancement, and cross-layer connections can be introduced in the network to capture semantic representations at different levels. This architecture can fuse low-level grammatical information with high-level semantic information, thereby achieving accurate recognition of complex fraud texts.

[0097] Among them, the cross-layer connection formula is:

[0098]

[0099] in, Indicates the behavior characteristics of operator data at layer l.

[0100] S3. Based on the multimodal behavior characteristics, model training is performed through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

[0101] In some embodiments, in S3, the contrastive learning mechanism includes:

[0102] Maximize the similarity of positive sample pairs, and minimize the similarity of negative sample pairs;

[0103] Among them, the positive sample pair is the fraudulent text and the corresponding behavior data, and the negative sample pair is the non-fraudulent text and the corresponding behavior data.

[0104] Specifically, the model maximizes the similarity of positive sample pairs (fraudulent text and its behavior data) and minimizes the similarity of negative sample pairs (non-fraudulent text and its behavior data) through contrastive learning mechanism; the loss function of contrastive learning is:

[0105]

[0106] Among them, sin(h i ,h j ) is the similarity measure between samples i and j.

[0107] This step optimizes the model by using contrastive learning mechanism and supervised learning to maximize the semantic difference between fraudulent text and non-fraudulent text and improve the generalization ability of the model.

[0108] This application provides an anti-fraud model training method that combines multi-dimensional data of operators (such as SMS records, call records, network access logs, and base station locations, etc.) with text data to achieve comprehensive judgment of fraudulent behavior through multimodal fusion technology. This method improves the model's ability to identify cross-channel fraud, not only limited to text, but also able to capture user behavior patterns, enhancing the detection of complex fraudulent behavior.

[0109] In addition, a cross-layer connection mechanism is introduced to fuse the text representations generated by BERT and ERNIE at different levels with the multimodal features of operator data. This cross-layer information fusion method provides richer information in capturing text details and behavioral characteristics, and improves the recognition accuracy of the model in different scenarios.

[0110] In addition, a contrastive learning mechanism was used during model training, which significantly improved the model's ability to distinguish between fraudulent and non-fraudulent behaviors by maximizing the similarity of positive sample pairs (such as fraudulent text and related operator behavior data) and minimizing the similarity of negative sample pairs. This training method strengthens the model's ability to perceive abnormal behavior patterns.

[0111] Figure 2 A schematic diagram of the anti-fraud identification processing method provided in the embodiment of the present application, such as Figure 2 As shown, in some embodiments, the present application provides an anti-fraud identification processing method, the method comprising:

[0112] Obtaining data of an object to be identified;

[0113] Inputting the object data into a pre-trained anti-fraud model to obtain an anti-fraud recognition result of the object data;

[0114] The anti-fraud model is trained by the anti-fraud model training method described in the above embodiments.

[0115] Figure 3 A specific example diagram of the anti-fraud identification process provided in the embodiment of the present application, such as Figure 3 As shown, the following is a specific example of anti-fraud identification processing for this application:

[0116] 1. Feature dimension extraction of user-side dimension data in operators; related call features of a single user, and related positioning base station features;

[0117] 2. Feature extraction on the text side: Bert and Ernie perform feature extraction on all collected data;

[0118] 3. Align the dimensions of the features in 1 and 2, and label the information (relevant labeling is based on sensitive information. If the features in 1 and 2 are threshold-fitted with sensitive information at the same time, it can be defined as fraud if it is higher than the predetermined threshold);

[0119] 4. Information below a certain predetermined threshold can be defined as true information; information between the true information threshold and the fraud threshold is defined as neutral. (Neutral information can be hard-coded into the LLM model (determined as true information or fraud information by the LLM model capability))

[0120] It should be understood that, although the various steps in the flowcharts in the above-described embodiments are sequentially displayed according to the indications of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and they can be executed in other orders. Moreover, at least a portion of the steps in the figure may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily to be carried out sequentially, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0121] Figure 4 A schematic diagram of an anti-fraud model training device provided in an embodiment of the present application, such as Figure 4 As shown, the present application provides an anti-fraud model training device, the device comprising:

[0122] A first data acquisition module 11, which is configured to acquire operator multi-dimensional data and user communication text data;

[0123] A data processing module 12, which is configured to perform feature extraction and fusion processing based on the operator multi-dimensional data and user communication text data to obtain multimodal behavior features;

[0124] The model training module 13 is configured to perform model training based on the multimodal behavior characteristics through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

[0125] Regarding the limitation of the anti-fraud model training device, reference may be made to the limitation of the anti-fraud model training method in the above-mentioned embodiments of the present application, which will not be elaborated in this embodiment.

[0126] Figure 5 A schematic diagram of an anti-fraud identification processing device provided in an embodiment of the present application, such as Figure 5 As shown, the present application provides an anti-fraud identification processing device, the device comprising:

[0127] A second data acquisition module 21, which is configured to acquire the object data to be identified;

[0128] The anti-fraud identification module 22 is configured to input the object data into a pre-trained anti-fraud model to obtain an anti-fraud identification result of the object data.

[0129] Regarding the limitation of the anti-fraud identification processing device, reference may be made to the limitation of the anti-fraud identification processing method in the above-mentioned embodiments of the present application, which will not be elaborated in this embodiment.

[0130] Figure 6 A schematic diagram of an anti-fraud device provided in an embodiment of the present application, such as Figure 6 As shown, in some embodiments, the present application provides an anti-fraud device, including a memory 32 and a processor 31, the memory stores a computer program, and the processor is configured to run the computer program to execute the anti-fraud model training method and the anti-fraud identification processing method in the above-mentioned embodiments of the present application.

[0131] The memory is connected to the processor, the memory may be a flash memory or a read-only memory or other memory, and the processor may be a central processing unit or a single-chip microcomputer.

[0132] In some embodiments, the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the anti-fraud model training method and the anti-fraud identification processing method in the above-mentioned embodiments of the present application are implemented.

[0133] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable read only memory), flash memory or other memory technology, CD-ROM (Compact Disc Read-Only Memory), digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer.

[0134] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present application, but the present application is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of the present application, and these modifications and improvements are also considered to be within the scope of protection of the present application.

Claims

1. An anti-fraud model training method, characterized in that: The method comprises: S1. Obtaining multi-dimensional data of operators and user communication text data; S2. Extracting and fusing the operator's multidimensional data and the user's communication text data to obtain multimodal behavior features; S3. Based on the multimodal behavior characteristics, model training is performed through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

2. The anti-fraud model training method according to claim 1, characterized in that: In S1: The operator multi-dimensional data includes: user information, SMS records, call records, base station locations, and network access logs; The user communication text data includes: SMS data, call data, and traffic data.

3. The anti-fraud model training method according to claim 1, characterized in that: S2, including: S21, extracting features from user communication text data through the BERT model and the ERNIE model to obtain text features; S22, extracting and encoding the operator's multi-dimensional data through a Transformer encoder to obtain a multi-dimensional representation of user behavior; S23. Through a cross-layer connection mechanism, the text features and the multi-dimensional representation of the user behavior are fused to obtain multimodal behavior features.

4. The anti-fraud model training method according to claim 3, characterized in that: S21, including: The BERT model is used to encode the input user communication text data with a bidirectional Transformer structure, and the text context information is processed through a multi-head attention mechanism and a feedforward neural network to generate a deep semantic representation at the vocabulary level. The ERNIE model is used to introduce entity information from an external knowledge base into user communication text data to obtain the potential entity relationship recognition results in fraudulent texts; The text features include the deep semantic representation and potential entity relationship recognition results.

5. The anti-fraud model training method according to claim 1, characterized in that: In S3, the contrastive learning mechanism includes: Maximize the similarity of positive sample pairs, and minimize the similarity of negative sample pairs; Among them, the positive sample pair is the fraudulent text and the corresponding behavior data, and the negative sample pair is the non-fraudulent text and the corresponding behavior data.

6. An anti-fraud identification processing method, characterized in that: The method comprises: Obtaining data of an object to be identified; Inputting the object data into a pre-trained anti-fraud model to obtain an anti-fraud recognition result of the object data; Wherein, the anti-fraud model is trained by the anti-fraud model training method according to any one of claims 1-5. 7.An anti-fraud model training device, characterized in that: The device comprises: A first data acquisition module, which is configured to acquire operator multi-dimensional data and user communication text data; A data processing module, which is configured to perform feature extraction and fusion processing based on the operator's multi-dimensional data and the user's communication text data to obtain multimodal behavior features; The model training module is configured to perform model training based on the multimodal behavior characteristics through a comparative learning mechanism to obtain an anti-fraud model, and the anti-fraud model is used for anti-fraud identification processing.

8. An anti-fraud identification and processing device, characterized in that: The device comprises: A second data acquisition module, configured to acquire object data to be identified; The anti-fraud identification module is configured to input the object data into a pre-trained anti-fraud model to obtain an anti-fraud identification result of the object data.

9. An anti-fraud device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the anti-fraud model training method according to any one of claims 1 to 5 or the anti-fraud identification processing method according to claim 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the anti-fraud model training method according to any one of claims 1 to 5 or the anti-fraud identification processing method according to claim 6 is implemented.

Images