A method for assisting in the assessment of fraud risks in telecommunications services based on local AI agents and general large models

Abstract

This invention discloses an auxiliary assessment method for fraud-related risks in telecommunications services based on a local AI agent and a general-purpose large model, relating to the fields of artificial intelligence and telecommunications security risk control. In a local intranet environment, this invention builds an AI agent based on an existing general-purpose large model, constructing a localized private knowledge base for fraud-related risks. After receiving the telecommunications service fraud-related risk assessment requirements input by the operator, the AI ​​agent automatically generates a list of required assessment materials in conjunction with the knowledge base, performs compliance verification and fraud risk identification on the locally uploaded business materials, and outputs an AI-assisted assessment conclusion annotated with compliance basis and risk issues. Finally, a human reviews the risk assessment. The materials and related content used in the assessment are used and cleared immediately, valid only in the current assessment and not used for the AI ​​agent's autonomous learning. General content can only be updated by the administrator in the local private knowledge base. The entire process involves local closed-loop data processing, without external transmission, thus preventing data leakage. This invention achieves integrated support capabilities for material guidance, material verification, and risk assessment without requiring modifications to a general-purpose model. It compensates for the uneven skill levels of assessment personnel through standardized processes, and solves the problems of low efficiency, inconsistent standards, and insecure cloud-based assessment data. It is applicable to various telecommunications business fraud risk compliance assessment scenarios.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence technology, telecommunications network security and risk compliance assessment technology, specifically to an auxiliary assessment method for fraud risk assessment in telecommunications services based on a local AI agent and a general large model. It is applicable to fraud risk compliance assessment in various scenarios such as number account opening, communication services, voice outbound calls, value-added service activation, and business changes in the telecommunications industry, and can effectively solve the problem of non-standard assessment caused by the uneven level of assessment personnel. Background Technology

[0002] Telecommunications services encompass multiple core scenarios, including number registration, communication services, outbound voice calls, activation of value-added services, and service changes. These scenarios present numerous risks of telecommunications fraud and high-risk fraudulent activities. Currently, the industry's assessment of fraud risks in telecommunications services relies entirely on manual review against anti-fraud compliance clauses, with varying levels of expertise among assessors. This lack of standardized and intelligent auxiliary assessment solutions makes it difficult to guarantee the quality of assessments.

[0003] Existing technologies have the following significant drawbacks, making it difficult to meet the compliance and efficiency requirements of fraud risk control in the telecommunications industry: S1. The fraud compliance assessment clauses are complex and highly specialized, and the varying levels of assessment personnel make it difficult to quickly compile complete supporting materials for specific assessment requirements during manual review. This not only results in low efficiency but also easily leads to omissions and incomplete summaries. The results from assessment personnel of different levels vary significantly, seriously affecting the progress and quality of the assessment; S2. During manual material verification, the varying levels of assessment personnel lead to strong subjectivity. Different assessment personnel have different understandings of the assessment standards, resulting in significant deviations in the assessment results. The probability of misjudgments and omissions remains high, making it difficult to ensure the consistency and accuracy of the assessment; S3. If a public cloud-based large-scale model is used for fraud risk analysis, sensitive materials related to telecommunications business (such as user information, business documents, communication records, etc.) need to be uploaded to the cloud, posing security risks of sensitive data leakage, data breaches, and compliance breaches, which does not comply with the telecommunications industry's data security management standards; S4. Existing fraud risk assessment solutions are polarized: either fully automated, comprehensive risk screening (prone to AI decision-making black boxes, difficult-to-trace misjudgments, and requiring extensive model development), or purely manual system standardization (inefficient and inconsistent due to varying assessment personnel skill levels). There is no standardized technical solution based on a general-purpose model, locally built with an intelligent agent and private knowledge base, simultaneously achieving "material requirement generation, material compliance verification, and specific risk-assisted assessment." This fails to address the quality issues arising from differences in assessment personnel skill levels. Furthermore, while existing general-purpose models possess mature semantic understanding, rule matching, and content analysis capabilities, they lack a dedicated process design for localized security loops and specific rule-based targeted assessments to address the varying skill levels of assessment personnel. This makes it difficult to compensate for differences in assessment personnel skill levels through standardized auxiliary methods, thus hindering the stability of assessment quality. Given the shortcomings of existing technologies, there is an urgent need for a fraud risk assessment method that balances efficiency, compliance, and security, without requiring complex model development. This method would address the shortcomings of varying assessment personnel skill levels through standardized auxiliary processes, resolve the pain points of current manual assessments and the security risks of cloud-based assessments, and ensure the consistency of assessment quality. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an auxiliary assessment method for fraud risk evaluation in telecommunications services based on a local AI agent and a general large model, specifically achieving the following objectives: s1. No need to train algorithms or modify the structure of the general-purpose model itself; just build an AI agent on the local / internal network based on the existing general-purpose model, and pair it with a localized private knowledge base of fraud risks, reducing implementation costs and difficulty of implementation; s2. To automatically generate the required assessment materials, verify the compliance of the provided materials, and output standardized auxiliary assessment conclusions for specific fraud risk assessment requirements (adaptable to one or multiple requirements) based on specific fraud risk assessment requirements. By standardizing the process, it can make up for the shortcomings of uneven assessment personnel, improve assessment efficiency and standardization, and ensure consistent assessment quality. s3. All calculations and data interactions are completed locally on the intranet in a closed loop, eliminating the risk of sensitive data leakage and complying with the data security compliance requirements of the telecommunications industry; s4. Clearly define the boundaries between AI-assisted analysis and human final review. AI should only perform auxiliary analysis, and the final risk assessment should be confirmed by humans. Technical solution

[0005] A method for assisting in the assessment of fraud risks in telecommunications services based on a local AI agent and a general large model includes the following steps, each completed in a closed loop locally / within the intranet, without accessing the public network or transmitting any sensitive data to the outside: S1. In a local or intranet isolated environment, an AI scheduling agent is built based on an existing general large model. The AI ​​agent is used to schedule the semantic understanding and rule matching capabilities of the general large model to achieve full-process collaboration in material list generation, material verification, and auxiliary conclusion output. The AI ​​agent does not learn the materials and related content input for this assessment. A private knowledge base for fraud risks in telecommunications services is constructed and stored locally. The knowledge base includes fraud risk assessment standards, compliance requirements, material verification specifications, and risk judgment basis for various scenarios of telecommunications services. It can only be dynamically updated by the administrator and is not updated by the AI ​​agent learning the input content of this assessment. All models, agents, and knowledge bases are deployed locally on the intranet, without accessing the public cloud or establishing data transmission links with external networks. S2. Receive compliance requirements for assessing telecommunications business fraud risks input by operators through their local terminals. These requirements are specific fraud risk control clauses (single or multiple), such as: "Determine if the current account opening service involves multiple cards under one ID for fraud purposes," or "Determine if the current outbound voice call service involves fraudulent tactics involving impersonation of public security, procuratorate, or court officials." This does not include full or batch risk screening requirements. A standardized auxiliary process is used to avoid assessment biases caused by varying levels of assessment personnel skill. S3. The local AI agent utilizes the semantic parsing and rule matching capabilities of a general-purpose model, and searches the local private knowledge base. For the specific assessment requirements input (single or multiple), it accurately analyzes and outputs all necessary materials, supplementary supporting materials, and corresponding compliance requirements (such as material format, filling standards, and verification points) required to complete the risk assessment. This provides clear guidance to assessment personnel, avoiding omissions or non-compliance issues due to varying assessment personnel skill levels.

[0006] Operators upload the corresponding business evaluation materials to their local terminals according to the material list output in step 3. All materials are only valid for this evaluation process and are subject to the principle of "use and clear immediately." They are only temporarily stored, retrieved, and processed locally / on the intranet and are not transmitted to the external public network, third-party cloud, or external terminals to avoid the risk of sensitive data leakage during the transmission process. After the materials are uploaded, they are only used as temporary data for this evaluation and are automatically cleared immediately after the evaluation is completed, leaving no redundant material data. Among them, general content such as common fraud judgment rules and compliance requirements are not updated by the AI ​​agent learning the content of the input materials. Only the local private knowledge base can be updated by the administrator with exclusive permissions.

[0007] The AI ​​agent leverages the content analysis and rule comparison capabilities of a general-purpose large model to verify uploaded materials in three aspects, against specific evaluation requirements and compliance rules within a local private knowledge base: first, completeness verification, confirming whether the materials are complete and without omissions; second, compliance comparison, confirming whether the materials meet compliance requirements and are filled out correctly; and third, risk and suspicion identification, combining fraud-related characteristics in the knowledge base to locate potential fraud risks and non-compliant content in the materials, forming a standardized verification detail to reduce the impact of varying assessment personnel levels on the verification results.

[0008] Based on the verification results in step 5, the AI ​​agent generates structured AI-assisted assessment results corresponding to specific assessment requirements (single or multiple), clearly determining whether the submitted materials "meet," "partially meet," or "do not meet" the fraud risk assessment requirements; it also marks the specific risk issues, corresponding compliance basis, and actionable rectification suggestions; all auxiliary assessment results are clearly marked "only AI-assisted reference conclusions, not replacing human judgment on the final fraud nature of the assessment," clearly defining the auxiliary role of AI, providing standardized references for assessors, compensating for the shortcomings of varying assessor levels, and ensuring the standardization of assessment conclusions.

[0009] Risk control auditors (assessors) review and revise the AI-generated auxiliary assessment conclusions and verification details, and combine them with their own professional judgment to make a final fraud risk assessment. The standardized auxiliary content output by AI can effectively reduce the probability of misjudgment and omission due to the uneven skill levels of assessors, and improve the quality of the final review. Temporary cached data generated during the assessment process (such as uploaded business materials and intermediate verification data) strictly follows the principle of "use and clear immediately," and is only valid for this assessment process. It is automatically cleared immediately after the assessment is completed, and only the assessment number, specific assessment requirements (one or more), and the final review conclusion are archived to complete the entire process security loop and reduce the security risks caused by data retention. It should be noted that the materials and content used in this assessment are not used for the AI ​​agent's autonomous learning. General fraud assessment norms and standards can only be updated by the administrator in the local private knowledge base. Beneficial effects

[0010] Compared with the prior art, the present invention has the following significant advantages: s1. Low implementation cost and easy implementation: It reuses existing commercial general-purpose large models, without the need to develop or optimize large model algorithms, and does not involve the development of complex models. It only requires building intelligent agents and private knowledge bases locally, adapting to the existing intranet environment of the telecommunications industry, and can be implemented quickly. s2. Addressing the pain points of manual assessment: A single workflow achieves three core functions (generating a material list based on specific assessment requirements, verifying the compliance of existing materials, and outputting specific risk assessment conclusions). It can adapt to single or multiple assessment requirements. Through standardized auxiliary processes, it compensates for the shortcomings of varying assessment personnel skill levels, solves the problems of slow manual material sorting, inconsistent verification standards, large workload, and easy omissions and misjudgments, improves assessment efficiency and standardization, and ensures consistent assessment quality. s3. High data security compliance: All reasoning, data reading, knowledge base calling, and content analysis are completed locally on the intranet in a closed loop, without uploading to the external cloud or interacting with the public network, blocking the path of sensitive data transmission, eliminating the risk of leakage of business materials and sensitive user information from the architectural level, and complying with the "Data Security Law", "Personal Information Protection Law" and the data security and anti-fraud compliance requirements of the telecommunications industry; s4. Clear boundary between human and machine: Strictly define the boundary between AI assistance and human final review. AI only completes the auxiliary work of material sorting, compliance analysis and risk warning, and does not replace human to make the final fraud determination. It belongs to the scope of technical solution. s5. Strong scalability and practicality: The local private knowledge base supports dynamic addition and updating of fraud assessment rules and risk standards, which can quickly adapt to constantly evolving new fraud risk scenarios (such as new fraudulent scripts and new fraud business models), adapt to various telecommunications business fraud risk assessment scenarios, flexibly handle one or multiple specific assessment requirements, and effectively adapt to the current situation of varying assessment personnel levels, ensuring that assessment personnel of different levels can output standardized and accurate assessment results; it should be noted that the materials and related content used in the assessment process are cleared as soon as they are used, are only valid in this assessment, and are not used for the autonomous learning of the AI ​​agent, and the update of general content can only be completed by the administrator. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0012] In this embodiment, the general-purpose large model used is an existing publicly available commercial general-purpose large model (such as Wenxin Yiyan, Tongyi Qianwen, Qwen, etc.). No algorithm training, structural modification or functional optimization is performed on the model itself. All innovations lie in the complete set of technical processes for localized intelligent agent scheduling, knowledge base matching and human-machine collaborative evaluation.

[0013] The specific implementation steps are as follows: S1. Environment setup and knowledge base construction: Deploy an existing commercial general-purpose large model on the intranet server and build an AI agent for scheduling analysis. The AI ​​agent has the scheduling functions of calling the large model, retrieving the knowledge base, receiving materials, and outputting conclusions, and does not learn the materials and related content input during the evaluation process; Import a pre-organized private knowledge base on fraud risks in telecommunications services. The knowledge base contains all fraud judgment rules, material verification standards, and compliance requirements corresponding to scenarios such as number account opening, business processing, voice outbound calls, and value-added service activation, and reserves a dedicated update interface for administrators. It can only be dynamically supplemented and updated by administrators according to new fraud methods, and is not updated by the AI ​​agent learning the evaluation input content; All deployments are completed on the intranet and are not connected to the public network. The standardized architecture provides auxiliary support for evaluators and compensates for differences in the level of evaluators. S2. Obtain Specific Risk Assessment Requirements: Operators (assessors) input specific fraud assessment requirements on their local terminal (internal network access) (e.g., "Determine whether the current account opening business involves multiple cards under one ID for fraud purposes"). This method can be adapted to a single or multiple such specific assessment requirements. Through standardized auxiliary processes, it avoids operational biases caused by varying assessment personnel skill levels. S3. Generate a List of Required Materials for Assessment: The AI ​​agent utilizes the semantic parsing capabilities of a general-purpose model to retrieve relevant rules for "fraud assessment involving multiple cards under one ID" from the local private knowledge base. It outputs a list of materials and compliance standards required for this assessment: Required materials (copy of the account holder's ID card, account opening application form, real-name verification record), supplementary supporting materials (account holder's communication records for the past 3 months); compliance standards (clear and legible copy of ID card, complete and unaltered account opening application form, authentic and valid real-name verification record). This step can generate corresponding material lists for single or multiple specific assessment requirements, providing clear guidance for assessors and avoiding non-standard material preparation due to differences in assessor skill levels. S4. Local Reception of Materials to be Evaluated: According to the material list, the operator uploads the anonymized copy of the account holder's ID card, account opening application form, real-name verification record, and communication records for the past 3 months to the local terminal. All materials strictly adhere to the principle of "use and clear immediately," are only valid during this evaluation process, are only temporarily stored and retrieved on the internal network server, and are not transmitted to the outside. They are temporarily cached after uploading and automatically cleared immediately after the evaluation is completed. The materials and related content uploaded this time are not used for the AI ​​agent's autonomous learning. If there are general fraud judgment rules and standards that need to be updated, only the administrator can update the local private knowledge base.S5. Material Compliance and Fraud Risk Verification: The AI ​​agent calls upon a general model to verify the uploaded materials against the requirements of "one ID card for multiple cards fraud assessment" and knowledge base rules: Completeness verification (confirming that all four types of materials are complete and without omissions); Compliance comparison (confirming that the ID card copy is clear, the application form is filled out completely, and the real-name verification record is valid); Risk point identification (through communication record analysis, it was found that the account holder had applied for 5 numbers from the same operator in the past 3 months, which meets the characteristics of "one ID card for multiple cards" fraud, and the risk point was located). S6. Output Auxiliary Evaluation Conclusion: The AI ​​agent outputs a structured auxiliary evaluation result: "Does not meet the requirements for fraud risk assessment"; Risk issue: The account holder has multiple cards under one ID, which meets the characteristics of fraud risk; Compliance basis: Relevant clauses in the Anti-Telecommunications Network Fraud Law regarding real-name management of account opening and control of multiple cards under one ID; Rectification suggestion: Suspend this account opening business, require the account holder to provide a reasonable explanation, and verify the purpose of the number use; At the same time, it is noted that "This conclusion is only for AI auxiliary reference, and the final fraud determination shall be confirmed by human review"; This step can output corresponding auxiliary conclusions for one or more specific evaluation requirements, providing standardized references for assessors and reducing the conclusion bias caused by the uneven level of assessors. S7. Manual Final Review and Data Loop Archiving: Risk control reviewers (assessors) review the AI-assisted conclusions and verification details, and, based on the account holder's actual situation, confirm that the account opening business involves the risk of fraud involving multiple cards under one certificate, and make a final review conclusion of "prohibiting account opening"; the standardized auxiliary content output by AI effectively compensates for the shortcomings of varying assessment personnel levels, and improves the accuracy and standardization of the final review; after the assessment is completed, the system strictly adheres to the principle of immediate clearing, immediately and automatically clearing the temporarily cached account opening materials and verification intermediate data. Such materials and content are only valid in this assessment and are not used for AI agent learning. Only the assessment number, the specific assessment requirements input in this case, and the final conclusion of "prohibiting account opening" are archived, completing a complete auxiliary assessment process; this process can be adapted to single or multiple specific assessment requirements and the usage needs of assessment personnel of different levels. The updating of general content is the responsibility of the administrator to operate the local private knowledge base.

[0014] In this embodiment, AI only provides auxiliary analysis and does not learn the materials and related content used in this assessment on its own. The final determination of fraud is completed by humans (assessors). All temporary materials and content during the assessment process are cleared as soon as they are used and are only valid in this assessment. All data is local and closed-loop, without modifying the general model. General fraud judgment content can only be updated by the administrator in the local private knowledge base. Through standardized auxiliary processes, the shortcomings of uneven assessment personnel are effectively made up for.

Claims

1. A method for auxiliary assessment of fraud risk in telecommunications services based on local AI agents and general large models, characterized in that, The process includes the following steps, each completed entirely in a closed loop locally / on the intranet: s1. In a local or intranet-isolated environment, build an AI agent based on a general large model, construct and locally store a private knowledge base for telecommunications business fraud risks; s2. Obtain the input compliance requirements for assessing telecommunications business fraud risks. These requirements are specific fraud risk control clauses, which can be single or multiple, encompassing full-scale and batch risk screening needs, adapting to the current situation of varying assessment personnel skill levels; s3. The AI ​​agent, in conjunction with the private knowledge base, generates a complete list of materials and compliance specifications required to complete the risk assessment based on the assessment requirements, providing clear guidance for assessment personnel and compensating for the shortcomings of varying assessment personnel skill levels; s4. The process begins with: 1) Locally acquiring the business materials to be evaluated. All materials are processed only within the intranet and are not transmitted to public clouds or external terminals. 2) The AI ​​agent utilizes general large-scale model capabilities to verify compliance and identify potential fraud risks in the business materials, comparing them against the evaluation requirements and the private knowledge base. This generates standardized verification details to mitigate the impact of varying assessment skill levels. 3) Outputting AI-assisted evaluation conclusions corresponding to the evaluation requirements. This determines the degree to which the materials meet the requirements and marks risk issues, compliance basis, and rectification suggestions. These conclusions are for reference only and do not constitute a final fraud determination. They provide standardized references for assessors to ensure consistent evaluation quality. 4) A manual final fraud risk assessment confirmation is completed, strictly adhering to the "use and clear" principle. Temporary cached materials and related data from this evaluation process are cleared. These materials and content are only valid for this evaluation and are not used for the AI ​​agent's autonomous learning. The final evaluation results are archived to complete a secure closed loop. General fraud assessment rules and compliance requirements can only be updated by the administrator in the local private knowledge base.

2. The method of claim 1, wherein, The general large model mentioned is an existing commercially available large model. This method does not perform algorithm training, structural modification, or functional optimization on the general large model, but only calls its existing semantic understanding, rule matching, and content analysis capabilities.

3. The method of claim 1, wherein, All model reasoning, data reading, knowledge base retrieval, content analysis, and material transmission are completed locally on the intranet in a closed loop, blocking public network data transmission paths and preventing the leakage of sensitive business materials and user information.

4. The method of claim 1, wherein, The AI ​​agent and general large model only complete auxiliary tasks such as generating material lists, verifying material compliance, identifying risks and doubts, and outputting auxiliary judgment conclusions. The final qualitative judgment of fraud risks is completed by human review (assessment personnel). AI does not replace human judgment. At the same time, it compensates for the shortcomings of varying assessment personnel levels through standardized auxiliary content.

5. The method of claim 1, wherein, The local private knowledge base includes fraud risk assessment standards, compliance requirements, material verification specifications, and risk judgment basis for various scenarios of telecommunications services. It only supports dynamic addition and updating by administrators and does not use AI agents to learn the materials and related content input in this assessment to achieve updates, thus adapting to new fraud risk scenarios.

6. The method of claim 1, wherein, In step (7), after the evaluation is completed, the principle of "use and clear" is strictly followed. Temporarily cached business materials to be evaluated, verification intermediate data and related content used in this evaluation are automatically cleared. Such materials and content are only valid in this evaluation and are not used for the autonomous learning of AI agents. Only the evaluation identifier, evaluation requirements and final review conclusion are retained to reduce the data retention security risk and adapt to the usage scenarios where the evaluation personnel have varying levels of expertise. Updates to general content can only be performed by the administrator on the local private knowledge base.

7. The method according to claim 1, characterized in that, The compliance requirements for assessing fraud risks in telecommunications services in step (2) are specific fraud risk control clauses, which may be single or multiple, but do not include full or batch risk screening requirements. They can avoid assessment bias caused by uneven levels of assessment personnel through standardized auxiliary processes.