Method and system for automatically identifying fraud-related numbers based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and particularly discloses an artificial intelligence-based automatic identification method and system for fraud-related numbers, which breaks through the limitation that a traditional method depends on single-dimensional data through a data acquisition module, comprehensively describes a number behavior mode through multi-source data fusion, reduces erroneous judgment and missed judgment caused by one-sided data, and improves the identification accuracy of fraud-related numbers. A data foundation is laid for accurate identification; the feature library construction module introduces a knowledge graph technology, constructs the number, the associated object and the behavior event into a graph, and combines data cleaning and feature extraction and screening to form a dynamically updated fraud-related feature library; the method not only captures the surface behavior characteristics of the number, such as the call duration and the short message content, but also can mine the hidden fraud relation network to update the novel fraud characteristics in time, thereby solving the problem that the traditional method lags in the recognition of the complex fraud mode, and improving the timeliness and depth of the feature library.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method and system for automatically identifying fraudulent numbers. Background Art

[0002] With the rapid development of telecommunications network technology, telecommunications network fraud methods are constantly being updated. Fraudulent numbers, as an important carrier of fraud, are becoming increasingly difficult to identify.

[0003] Traditional methods of identifying fraudulent numbers, such as those based on blacklist and whitelist matching and simple rule-based judgment, have problems such as recognition lag, high rates of misjudgment and missed judgment, and difficulty in dealing with new and complex fraud patterns.

[0004] In the existing technology, some identification methods rely on single-dimensional data, such as call frequency and SMS content keywords, which cannot fully characterize the behavioral characteristics of fraudulent numbers.

[0005] In terms of model construction, it lacks the ability to deeply learn complex behavioral patterns, and has poor adaptability in the face of ever-evolving fraud methods. There are also gaps in the post-identification linkage disposal and cross-domain collaboration links, making it difficult to form a complete anti-fraud closed loop.

[0006] Therefore, there is an urgent need for an automatic identification method and system for fraudulent numbers that integrates multi-dimensional data, has intelligent learning and dynamic optimization capabilities, and can achieve full-process anti-fraud collaboration. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for automatically identifying fraudulent numbers based on artificial intelligence to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solutions: an artificial intelligence-based method and system for automatically identifying fraudulent numbers, the system comprising a data acquisition module, a feature library construction module, an intelligent recognition module, a decision output module, a feedback optimization module, an early warning push module, a linkage disposal module, and a cross-domain collaboration module;

[0009] The data collection module is used to collect basic attribute information of numbers, communication behavior data, and network behavior data from multiple source channels such as telecommunications networks, user behavior logs, and authorized social platforms;

[0010] The feature library construction module is used to clean, extract and filter the collected data, and introduce knowledge graph technology to build a dynamically updated fraud-related feature library;

[0011] The intelligent recognition module is used to build a deep learning model, use the data in the fraud feature library to train the model, and input the features of the number to be identified into the trained model to calculate the probability of fraud and perform preliminary identification;

[0012] The decision output module is used to perform a secondary verification on the preliminary recognition results of the intelligent recognition module in combination with the artificial expert rule base, and output the final fraud-related recognition decision;

[0013] The feedback optimization module is used to collect verification data of recognition results using blockchain technology to optimize the deep learning model and fraud feature library;

[0014] The warning push module is used to push graded fraud warnings through multiple channels based on the recognition results and number usage scenarios;

[0015] The linkage processing module is used to establish linkage with the anti-fraud processing system, synchronize information on fraudulent numbers, trigger the processing process and receive feedback on the processing results;

[0016] The cross-domain collaboration module is used to share fraud-related information and conduct joint prevention and control with anti-fraud platforms in the financial, e-commerce, and social industries through secure data sharing protocols and privacy computing technology.

[0017] Preferably, the data collection module includes a telecommunications data interface unit, a log collection unit and a social data collaboration unit;

[0018] The telecommunications data interface unit is used to connect to the telecommunications operator system to obtain basic number attributes and communication data;

[0019] The log collection unit is used to collect network behavior logs of user terminals;

[0020] The social data collaboration unit performs collaborative calculations of encrypted feature vectors with multiple social platforms based on a federated learning framework.

[0021] Preferably, the feature library construction module includes a data preprocessing unit, a feature engineering unit and a knowledge graph construction unit;

[0022] The data preprocessing unit is used to perform data cleaning and format standardization operations;

[0023] The feature engineering unit uses natural language processing and statistical analysis techniques to extract and filter features;

[0024] The knowledge graph construction unit constructs numbers, associated objects and behavioral events into a knowledge graph to mine the characteristics of fraud-related relationship networks.

[0025] Preferably, the intelligent recognition module adopts an improved recurrent neural network and introduces a generative adversarial network for adversarial training;

[0026] The intelligent identification module includes a fraud probability threshold adjustment unit;

[0027] The fraud probability threshold adjustment unit is used to dynamically adjust the fraud probability determination threshold according to historical recognition data.

[0028] Preferably, the decision output module includes a rule extraction unit and a rule updating unit;

[0029] The rule extraction unit uses natural language processing technology to extract rules from fraud case reports and public security notice texts;

[0030] The rule updating unit dynamically updates the rule base according to the result of the rule extracting unit.

[0031] Preferably, the feedback optimization module adopts a consortium chain architecture and jointly maintains a distributed ledger of identification result verification data with the telecom operator and regulatory agency nodes;

[0032] The feedback optimization module includes a model retraining scheduling unit;

[0033] The model retraining scheduling unit automatically triggers online or offline retraining of the deep learning model according to the data update situation.

[0034] Preferably, the early warning push module includes a risk level unit and a multi-channel push unit;

[0035] The risk level unit assesses the risk level based on the probability of fraud and behavioral characteristics;

[0036] The multi-channel push unit provides differentiated warnings for different risk levels through telecommunication network signaling control channels, smart terminal operating system interfaces, and third-party application message push interfaces.

[0037] Preferably, the linkage handling module includes a handling decision configuration unit and an effect evaluation unit;

[0038] The handling decision configuration unit configures number blocking and transaction interception handling strategies according to the type and risk level of the fraudulent number;

[0039] The effect evaluation unit evaluates the anti-fraud treatment system's treatment effect in real time from multiple dimensions, including the ban success rate and the fund interception rate, and feeds the evaluation results back to the feedback optimization module;

[0040] Preferably, the cross-domain collaboration module adopts a federated transfer learning framework to achieve the transfer of fraud-related feature knowledge between anti-fraud platforms in different industries;

[0041] The cross-domain collaboration module includes a collaboration effect evaluation unit;

[0042] The collaborative effect evaluation unit is used to quantitatively evaluate the joint prevention and control coverage, response speed, and collaborative efficiency of fraud risk prevention and control platforms in various industries, and optimize data sharing and collaborative strategies based on the evaluation results;

[0043] An artificial intelligence-based method for automatically identifying fraudulent numbers, the steps are as follows:

[0044] S1: The data collection module collects number-related data from multiple sources;

[0045] S2: The feature library construction module processes the collected data and constructs a fraud feature library;

[0046] S3: The intelligent recognition module trains a deep learning model and performs preliminary fraud identification on the number to be identified;

[0047] S4: The decision output module verifies the preliminary recognition results in combination with the artificial expert rule base and outputs the final decision;

[0048] S5: Feedback optimization module optimizes the model and feature library based on recognition verification data;

[0049] S6: The warning push module pushes graded warnings based on the recognition results and scenarios;

[0050] S7: The linkage processing module links with the anti-fraud processing system and processes feedback;

[0051] S8: The cross-domain collaboration module enables the sharing and joint prevention and control of fraud-related information in multiple industries.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention, through the data acquisition module, breaks through the limitation of traditional methods that rely on single-dimensional data. Through multi-source data fusion, it comprehensively depicts the behavior patterns of numbers, reduces misjudgments and missed judgments caused by one-sided data, and lays a data foundation for accurate identification; and the feature library construction module introduces knowledge graph technology to construct numbers, related objects, and behavioral events into a graph, and combines data cleaning, feature extraction and screening to form a dynamically updated fraud-related feature library; it not only captures the surface behavioral characteristics of numbers, such as call duration and text message content, but also mines hidden fraud-related relationship networks to timely update new fraud features, solving the problem of traditional methods lagging behind in identifying complex fraud patterns, and improving the timeliness and depth of the feature library.

[0054] 2. In this invention, the intelligent recognition module uses an improved recurrent neural network and introduces a generative adversarial network for adversarial training. Combined with a dynamic threshold adjustment mechanism, this module, through deep learning technology, can automatically learn fraudulent behavior patterns. The simulated samples generated by adversarial training effectively enhance the robustness of the model, enabling it to effectively respond to evolving fraud methods. In addition, the dynamic threshold adjustment mechanism optimizes the judgment criteria based on historical recognition data, further improving the model's adaptability. Compared with traditional rule-matching methods, this module achieves significantly improved recognition accuracy.

[0055] 3. In this invention, the decision output module combines a manual expert rule base to perform a secondary verification of the model's initial recognition results. The feedback optimization module uses blockchain technology to collect verification data and dynamically optimize model parameters and feature libraries. This dual mechanism of "model recognition + rule verification" reduces the misjudgment rate and avoids potential deviations that may occur in a single model. The blockchain ensures the authenticity and traceability of feedback data. Combined with online and offline retraining, it achieves continuous iteration of recognition capabilities, solving the problem that traditional methods cannot dynamically optimize, and forming a closed loop of "recognition-verification-optimization."

[0056] 4. This invention uses an early warning push module to issue warnings based on risk levels, and a linkage disposal module to work with the anti-fraud disposal system to trigger the blocking and interception processes. The cross-domain collaboration module uses federated transfer learning to share fraud-related information with the financial and e-commerce industries, forming a complete closed loop from identification to disposal. High-risk numbers issue real-time mandatory warnings, while medium and low-risk numbers receive gentle reminders, improving warning efficiency. Cross-industry data sharing and joint prevention and control break down information silos, achieve cross-domain interception of fraud risks, build a multi-dimensional anti-fraud defense line, and significantly improve overall anti-fraud effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 This is a system architecture diagram of the present invention;

[0059] Figure 2 This is a diagram of the data acquisition module architecture of the present invention;

[0060] Figure 3 This is a diagram showing the architecture of the feature library building module of the present invention;

[0061] Figure 4 This is a diagram of the decision output module architecture of the present invention;

[0062] Figure 5 This is the architecture diagram of the early warning push module of the present invention;

[0063] Figure 6 This is the framework diagram of the linkage processing module of the present invention. DETAILED DESCRIPTION

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0065] See also Figures 1 to 6 , the present invention provides a technical solution:

[0066] For example 1, please refer to Figure 1 The present invention provides an artificial intelligence-based method and system for automatically identifying fraudulent numbers. The system includes a data acquisition module, a feature library construction module, an intelligent recognition module, a decision output module, a feedback optimization module, an early warning push module, a linkage disposal module, and a cross-domain collaboration module.

[0067] The data collection module is used to collect basic attribute information, communication behavior data, and network behavior data of numbers from multiple sources such as telecommunications networks, user behavior logs, and authorized social platforms;

[0068] The feature library construction module is used to clean, extract and filter the collected data, and introduce knowledge graph technology to build a dynamically updated fraud-related feature library;

[0069] The intelligent recognition module is used to build a deep learning model, train the model using data from the fraud feature library, and input the features of the number to be identified into the trained model to calculate the probability of fraud and perform preliminary identification.

[0070] The decision output module is used to perform a secondary verification of the preliminary recognition results of the intelligent recognition module in combination with the artificial expert rule base, and output the final fraud-related recognition decision;

[0071] The feedback optimization module is used to collect verification data of recognition results using blockchain technology to optimize deep learning models and fraud feature libraries;

[0072] The warning push module is used to push graded fraud warnings through multiple channels based on identification results and number usage scenarios;

[0073] The linkage processing module is used to establish linkage with the anti-fraud processing system, synchronize information on fraudulent numbers, trigger the processing process and receive feedback on the processing results;

[0074] The cross-domain collaboration module is used to share fraud-related information and conduct joint prevention and control with anti-fraud platforms in the financial, e-commerce, and social industries through secure data sharing protocols and privacy computing technologies.

[0075] For example 2, please refer to Figures 2 to 3 As shown, the data collection module includes a telecommunications data interface unit, a log collection unit, and a social data collaboration unit;

[0076] The telecommunications data interface unit is used to connect with the telecommunications operator system to obtain basic number attributes and communication data;

[0077] The log collection unit is used to collect network behavior logs of user terminals; the log collection unit is used to capture number-related behavior data from server logs, including but not limited to access records and transaction records;

[0078] The social data collaboration unit is based on the federated learning framework and collaborates with multiple social platforms to calculate encrypted feature vectors. The social data collaboration unit is used to cooperate with social platforms to obtain the social behavior data of the number on the social platform, such as friend relationships and published content.

[0079] The feature library construction module includes data preprocessing unit, feature engineering unit and knowledge graph construction unit;

[0080] The data preprocessing unit is used to perform data cleaning and format standardization operations;

[0081] Data preprocessing unit formula:

[0082] P pre =D raw →D clean , D clean =σ;

[0083] Among them, P pre For the pre-processing process, D raw is the original data, D clean For cleaned data;

[0084] This formula can effectively clean the raw data, remove noise and irrelevant information, and convert it into high-quality data with a unified format, providing a solid foundation for subsequent feature engineering and knowledge graph construction.

[0085] The formula not only ensures the accuracy and efficiency of data processing, but also protects user privacy through encryption, in line with the core principles of federated learning. Through the application of this formula, the system can more accurately extract features related to fraudulent behavior and improve the accuracy and efficiency of identification.

[0086] The feature engineering unit uses natural language processing and statistical analysis techniques to extract and screen features; the feature engineering unit is responsible for extracting valuable features from the preprocessed data. These features can reflect the social behavior patterns and potential risks of the number. By using machine learning algorithms and statistical methods, the feature engineering unit can screen out the most representative features and provide key input for subsequent model training.

[0087] The knowledge graph construction unit constructs numbers, related objects and behavioral events into a knowledge graph to mine the characteristics of fraud-related relationship networks.

[0088] The knowledge graph construction unit integrates information from different data sources to form a structured knowledge network, which can clearly display the association relationship between numbers, the attributes of associated objects, and the timeline and impact range of behavioral events; in the knowledge graph, each number, associated object and behavioral event is regarded as a node, and the relationship between them is regarded as an edge. Through the complex network structure and attribute information, the knowledge graph construction unit can reveal potential fraudulent behavior patterns and association rules. This step provides rich contextual information and association features for subsequent fraudulent number identification, further enhancing the accuracy and reliability of identification.

[0089] The data collection module breaks through the limitations of traditional methods that rely on single-dimensional data, and through the fusion of multi-source data, a more comprehensive and accurate description of number characteristics is achieved.

[0090] Moreover, the data eraser module can integrate data resources from multiple channels such as communication operators, social media, and e-commerce platforms, and use advanced data cleaning and integration technologies to eliminate redundant information and retain key features.

[0091] Cross-domain data fusion not only enhances the generalization ability of the model, but also significantly improves the accuracy of identifying fraudulent numbers.

[0092] At the same time, the data collection module continuously monitors the updates of the data source, allowing the model to learn the latest fraud methods and characteristics in real time and maintain the cutting-edge nature of its recognition capabilities.

[0093] The feature library construction module introduces knowledge graph technology, which can construct numbers, related objects, and behavior time into a graph, and then combine data cleaning, feature proposal and screening to form a dynamically updated fraud-related feature library, solving the problem of traditional methods lagging behind in identifying complex fraud patterns, and improving the timeliness and depth of the feature library.

[0094] For example three, please refer to Figures 4 to 6 As shown in the figure, the intelligent recognition module adopts an improved recurrent neural network and introduces a generative adversarial network for adversarial training;

[0095] The intelligent identification module includes a fraud probability threshold adjustment unit;

[0096] Threshold dynamic adjustment formula:

[0097]

[0098] Among them, θ thresh is the fraud probability determination threshold, is the initial fraud probability threshold, Δθ is the threshold adjustment step, δ(·) is the adjustment function, P history is the historical identification data, γ is the regularization coefficient;

[0099] This formula can dynamically adjust the fraud probability threshold to adapt to the ever-changing fraud methods and patterns. In application, an initial fraud probability threshold is first set as the starting point for identification. As historical identification data accumulates, the threshold is gradually optimized using the adjustment function and threshold adjustment step size. The introduction of the regularization coefficient ensures the stability and accuracy of the adjustment process, avoiding misjudgments or missed judgments caused by data fluctuations.

[0100] In this way, the system can continuously learn and improve its recognition capabilities, providing users with more accurate and efficient fraudulent number identification services.

[0101] The fraud probability threshold adjustment unit is used to dynamically adjust the fraud probability determination threshold based on historical recognition data.

[0102] The decision output module includes a rule extraction unit and a rule updating unit;

[0103] The rule extraction unit uses natural language processing technology to extract rules from fraud case reports and public security notices;

[0104] The rules include but are not limited to the key features of fraudulent behavior, common phrases, and number characteristics, providing an important basis for subsequent identification of fraudulent numbers. The rule extraction unit can automatically identify and summarize the general patterns and characteristics of fraudulent behavior through deep learning and analysis of large amounts of text data, thereby greatly improving recognition efficiency and accuracy. At the same time, the unit can also continuously update and optimize the rule base based on new fraud case reports and public security notifications, ensuring that the recognition system can keep up with changes in fraud methods and always maintain a high degree of sensitivity and accuracy.

[0105] The rule updating unit dynamically updates the rule base according to the result of the rule extracting unit.

[0106] The rule update unit not only focuses on new features and new tactics of fraudulent behavior, but also monitors and analyzes changes in number characteristics to ensure the comprehensiveness and pertinence of the rule base. Through continuous rule updates, the fraudulent number identification system can continuously improve its identification capabilities and provide users with a more reliable and secure communication environment.

[0107] The feedback optimization module adopts a consortium chain architecture and jointly maintains a distributed ledger of identification result verification data with telecom operators and regulatory agency nodes;

[0108] The feedback optimization module ensures the security and credibility of identification results and verification data through the consensus mechanism of the alliance chain. After identifying a fraudulent number, the feedback optimization module will upload the identification results and related verification data to the alliance chain. Telecom operators and regulatory agency nodes can review and verify these data to further confirm the accuracy of the identification results. The feedback optimization module also uses the smart contract function of the alliance chain to realize the automated processing and feedback of identification results, improving processing efficiency and accuracy. Through the distributed ledger technology of the alliance chain, the feedback optimization module can ensure the immutability and traceability of identification results and verification data, providing strong data support for subsequent fraudulent number governance.

[0109] The feedback optimization module includes a model retraining scheduling unit;

[0110] The model retraining scheduling unit automatically triggers online or offline retraining of the deep learning model based on data updates.

[0111] The model retraining scheduling unit can intelligently determine when the model needs to be updated to adapt to the ever-changing fraud methods. When new fraudulent language or behavioral characteristics are identified and accumulated to a certain level, the model retraining scheduling unit will automatically start the retraining process. Online retraining allows the model to gradually learn and absorb new data without interrupting service, maintaining a keen insight into the latest fraud methods; offline retraining is suitable for scenarios with large-scale data updates or model structure adjustments, ensuring that the model can more accurately identify fraudulent numbers after comprehensive optimization. Through automated retraining scheduling, the fraudulent number identification system can continue to evolve and effectively respond to various new fraud threats.

[0112] The early warning push module includes a risk level unit and a multi-channel push unit;

[0113] The risk level unit assesses the risk level based on the probability of fraud and behavioral characteristics;

[0114] Risk level calculation formula:

[0115] R level =μ·P fraud +(1-μ)·R behavior ;

[0116] Among them, R level is the risk level, μ is the weight coefficient, P fraud is the probability of fraud, R behavior Score behavioral characteristics; this formula can accurately calculate the risk level of fraudulent numbers, providing a scientific basis for subsequent early warning push. The calculated risk level will directly affect the working strategy of the multi-channel push unit, ensuring that high-risk numbers receive timely and effective intervention, while low-risk numbers may adopt a milder reminder method, which not only improves the accuracy of the early warning, but also avoids unnecessary interruptions; in addition, the formula also has good scalability and flexibility, and can adjust the weight coefficient according to actual needs to adapt to the ever-changing fraudulent behavior patterns.

[0117] The multi-channel push unit provides differentiated warnings for different risk levels through telecommunications network signaling control channels, smart terminal operating system interfaces, and third-party application message push interfaces.

[0118] The linkage disposal module includes a disposal decision configuration unit and an effect evaluation unit;

[0119] The disposal decision configuration unit configures number blocking and transaction interception disposal strategies based on the type and risk level of the fraudulent number;

[0120] The effectiveness evaluation unit conducts real-time evaluation of the anti-fraud disposal system's effectiveness from multiple dimensions, including the ban success rate and fund interception rate, and feeds the evaluation results back to the feedback optimization module.

[0121] Ban success rate calculation formula:

[0122]

[0123] Among them, S is the success rate, N blocked The number of fraudulent numbers successfully blocked, N total is the total number of fraudulent numbers that need to be banned.

[0124] This formula can be used to calculate the blocking success rate, thereby quantitatively evaluating the effectiveness of the anti-fraud system in blocking numbers. This success rate metric is a key parameter in measuring system performance. A high success rate means the system is more effective in identifying and blocking fraudulent numbers, reducing the risk of fraud. Combined with the other evaluation dimensions of the fund interception rate, it can comprehensively reflect the overall effectiveness of the anti-fraud system, providing a strong basis for subsequent feedback and optimization.

[0125] Funds interception rate formula:

[0126]

[0127] Among them, S is the success rate, Vintercepted is the amount of fraudulent funds successfully intercepted, V total The total amount of funds involved in the fraudulent activities.

[0128] This formula can be used to calculate the efficiency of intercepting funds for fraudulent numbers, thereby evaluating the actual application effect of the identification method. In actual operation, we input the data of fraudulent numbers into the system, and the system automatically identifies and calculates the number of fraudulent numbers that have been successfully banned. At the same time, based on the total amount of funds involved in the fraudulent behavior, the amount of fraudulent funds that have been successfully intercepted is calculated. Finally, the funds interception rate formula can be used to obtain the funds interception rate, so as to continuously optimize the identification method and system, and improve the recognition efficiency of fraudulent numbers and the funds interception effect.

[0129] The cross-domain collaboration module uses a federated transfer learning framework to transfer fraud-related feature knowledge between anti-fraud platforms in different industries;

[0130] The cross-domain collaboration module includes a collaboration effect evaluation unit;

[0131] The collaborative effect evaluation unit is used to quantitatively assess the joint prevention and control coverage, response speed, and collaborative efficiency of anti-fraud platforms in various industries, and optimize data sharing and collaborative strategies based on the evaluation results.

[0132] Effect coverage calculation formula:

[0133]

[0134] Among them, C is the coverage rate of joint prevention and control, N cross-detected is the number of fraud cases detected by cross-domain collaboration, N total-fraud is the total number of fraud cases.

[0135] The above formula can be used to calculate the actual performance of anti-fraud platforms in various industries in joint prevention and control. This quantitative indicator not only provides an objective basis for evaluating the synergy effect of each platform, but also lays the foundation for subsequent optimization of data sharing strategies and the formulation of more effective collaborative plans. At the same time, the system will automatically adjust the breadth and depth of data sharing and the strategy for collaborative handling of fraud risks based on the evaluation results to ensure that the entire anti-fraud system can operate continuously and efficiently.

[0136] Through the joint operation of multiple modules including intelligent identification module, decision output module, feedback optimization module, early warning push module, linkage disposal module and cross-domain collaboration module, it can conduct deep learning independently and deal with different means and evolved fraud methods. Moreover, through the adjustment of dynamic thresholds, it can judge whether it is a fraud based on historical data, further improving the adaptability of the model. Compared with the traditional rule matching method, the recognition accuracy of this system is significantly improved.

[0137] In addition, combined with the artificial expert rule library in the decision output module, the preliminary results of the model can be verified twice to further avoid fraud. Moreover, through the feedback optimization module and the dynamic optimization model, a dual mechanism of recognition + rule sampling is formed to further reduce the occurrence of misjudgment, avoid the deviation of a single model, solve the problem that traditional methods cannot be dynamically optimized, and form a closed loop of "identification-verification-optimization".

[0138] Furthermore, the early warning push module, linkage disposal module and cross-domain collaboration module form a perfect closed loop of early warning and disposal. Moreover, the cross-domain collaboration module can realize data sharing, break the traditional information island situation, realize cross-regional interception of fraud risks, build a multi-dimensional anti-fraud information network, and improve the comprehensive ban on fraud methods.

[0139] For example 4, please refer to Figures 1 to 6 As shown, an artificial intelligence-based automatic identification method for fraudulent numbers includes the following steps: S1: a data collection module collects number-related data from multiple source channels;

[0140] S2: The feature library construction module processes the collected data and constructs a fraud feature library;

[0141] S3: The intelligent recognition module trains a deep learning model and performs preliminary fraud identification on the number to be identified;

[0142] S4: The decision output module verifies the preliminary recognition results in combination with the artificial expert rule base and outputs the final decision;

[0143] S5: Feedback optimization module optimizes the model and feature library based on recognition verification data;

[0144] S6: The warning push module pushes graded warnings based on the recognition results and scenarios;

[0145] S7: The linkage processing module links with the anti-fraud processing system and processes feedback;

[0146] S8: The cross-domain collaboration module enables the sharing and joint prevention and control of fraud-related information in multiple industries.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based automatic identification system for fraudulent numbers, characterized by: The system includes a data acquisition module, a feature library construction module, an intelligent recognition module, a decision output module, a feedback optimization module, an early warning push module, a linkage disposal module and a cross-domain collaboration module; The data collection module is used to collect basic attribute information of numbers, communication behavior data, and network behavior data from multiple source channels such as telecommunications networks, user behavior logs, and authorized social platforms; The feature library construction module is used to clean, extract and filter the collected data, and introduce knowledge graph technology to build a dynamically updated fraud-related feature library; The intelligent recognition module is used to build a deep learning model, use the data in the fraud feature library to train the model, and input the features of the number to be identified into the trained model to calculate the probability of fraud and perform preliminary identification; The decision output module is used to perform a secondary verification on the preliminary recognition results of the intelligent recognition module in combination with the artificial expert rule base, and output the final fraud-related recognition decision; The feedback optimization module is used to collect verification data of recognition results using blockchain technology to optimize the deep learning model and fraud feature library; The warning push module is used to push graded fraud warnings through multiple channels based on the recognition results and number usage scenarios; The linkage processing module is used to establish linkage with the anti-fraud processing system, synchronize information on fraudulent numbers, trigger the processing process and receive feedback on the processing results; The cross-domain collaboration module is used to share fraud-related information and conduct joint prevention and control with anti-fraud platforms in the financial, e-commerce, and social industries through secure data sharing protocols and privacy computing technologies.

2. The artificial intelligence-based automatic fraud number identification system according to claim 1, characterized in that: The data collection module includes a telecommunications data interface unit, a log collection unit and a social data collaboration unit; The telecommunications data interface unit is used to connect to the telecommunications operator system to obtain basic number attributes and communication data; The log collection unit is used to collect network behavior logs of user terminals; The social data collaboration unit performs collaborative calculations of encrypted feature vectors with multiple social platforms based on a federated learning framework.

3. The artificial intelligence-based automatic fraud number identification system according to claim 1, characterized in that: The feature library construction module includes a data preprocessing unit, a feature engineering unit and a knowledge graph construction unit; The data preprocessing unit is used to perform data cleaning and format standardization operations; The data preprocessing unit formula: P pre =D raw →D clean ,D clean =σ; Among them, P pre For the pre-processing process, D raw is the original data, D clean For cleaned data; The feature engineering unit uses natural language processing and statistical analysis techniques to extract and filter features; The knowledge graph construction unit constructs numbers, associated objects and behavioral events into a knowledge graph to mine the characteristics of fraud-related relationship networks.

4. The artificial intelligence-based fraudulent number automatic identification system according to claim 1, characterized in that: The intelligent recognition module adopts an improved recurrent neural network and introduces a generative adversarial network for adversarial training; The intelligent identification module includes a fraud probability threshold adjustment unit; The threshold dynamic adjustment formula: Among them, θ thresh is the fraud probability determination threshold, is the initial fraud probability threshold, Δθ is the threshold adjustment step, δ(·) is the adjustment function, P history is the historical identification data, γ is the regularization coefficient; The fraud probability threshold adjustment unit is used to dynamically adjust the fraud probability determination threshold according to historical recognition data.

5. The artificial intelligence-based automatic fraud number identification system according to claim 1, characterized in that: The decision output module includes a rule extraction unit and a rule updating unit; The rule extraction unit uses natural language processing technology to extract rules from fraud case reports and public security notice texts; The rule updating unit dynamically updates the rule base according to the result of the rule extracting unit.

6. The artificial intelligence-based fraudulent number automatic identification system according to claim 1, characterized in that: The feedback optimization module adopts a consortium chain architecture and jointly maintains a distributed ledger of identification result verification data with telecom operators and regulatory agency nodes; The feedback optimization module includes a model retraining scheduling unit; The model retraining scheduling unit automatically triggers online or offline retraining of the deep learning model according to the data update situation.

7. The artificial intelligence-based fraudulent number automatic identification system according to claim 1, characterized in that: The early warning push module includes a risk level unit and a multi-channel push unit; The risk level unit assesses the risk level based on the probability of fraud and behavioral characteristics; The risk level calculation formula is: R level =μ·P fraud +(1-μ)·R behavior ; Among them, R level is the risk level, μ is the weight coefficient, P fraud is the probability of fraud, R behavior score behavioral traits; The multi-channel push unit provides differentiated warnings for different risk levels through telecommunication network signaling control channels, smart terminal operating system interfaces, and third-party application message push interfaces.

8. The artificial intelligence-based automatic fraud number identification system according to claim 1, characterized in that: The linkage disposal module includes a disposal decision configuration unit and an effect evaluation unit; The handling decision configuration unit configures number blocking and transaction interception handling strategies according to the type and risk level of the fraudulent number; The effect evaluation unit evaluates the anti-fraud treatment system's treatment effect in real time from multiple dimensions, including the ban success rate and the fund interception rate, and feeds the evaluation results back to the feedback optimization module; The calculation formula for the ban success rate is: Among them, S is the success rate, N blocked The number of fraudulent numbers successfully blocked, N total is the total number of fraudulent numbers that need to be banned. The fund interception rate formula is: Among them, S is the success rate, V intercepted is the amount of fraudulent funds successfully intercepted, V total The total amount of funds involved in the fraudulent activities.

9. The artificial intelligence-based automatic fraud number identification system according to claim 1, characterized in that: The cross-domain collaboration module uses a federated transfer learning framework to achieve the transfer of fraud-related feature knowledge between anti-fraud platforms in different industries; The cross-domain collaboration module includes a collaboration effect evaluation unit; The collaborative effect evaluation unit is used to quantitatively evaluate the joint prevention and control coverage, response speed, and collaborative efficiency of fraud risk prevention and control platforms in various industries, and optimize data sharing and collaborative strategies based on the evaluation results; The calculation formula of the effect coverage is: Among them, C is the coverage rate of joint prevention and control, N cross-detected is the number of fraud cases detected by cross-domain collaboration, N total-fraud is the total number of fraud cases.

10. An artificial intelligence-based method for automatically identifying fraudulent numbers, according to the artificial intelligence-based system for automatically identifying fraudulent numbers according to any one of claims 1 to 9, characterized in that: The steps of the automatic identification method are as follows: S1: The data collection module collects number-related data from multiple sources; S2: The feature library construction module processes the collected data and constructs a fraud feature library; S3: The intelligent recognition module trains a deep learning model and performs preliminary fraud identification on the number to be identified; S4: The decision output module verifies the preliminary recognition results in combination with the artificial expert rule base and outputs the final decision; S5: Feedback optimization module optimizes the model and feature library based on recognition verification data; S6: The warning push module pushes graded warnings based on the recognition results and scenarios; S7: The linkage processing module links with the anti-fraud processing system and processes feedback; S8: The cross-domain collaboration module enables the sharing and joint prevention and control of fraud-related information in multiple industries.

Images