Customer information examination method, device and equipment and storage medium

Abstract

The application discloses a kind of customer data examination methods, devices, equipment and storage medium, related to data processing field, which comprises: collecting the to-be-examined material of customer for to-be-examined business, and the privacy enhancement OCR verification is carried out to to-be-examined material, to obtain trusted to-be-examined material;Trusted to-be-examined material is input into the comprehensive evaluation model based on federated learning, and the comprehensive characteristic parameter of customer is obtained;According to the comprehensive characteristic parameter of customer, the to-be-examined business of customer is examined and rated, and the examination and rating result is obtained.The scheme provided by the application improves the accuracy and efficiency of customer data examination.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a customer data approval method, apparatus, device, and storage medium. Background Technology

[0002] In industries such as banking, before customers can conduct various transactions, the institutions providing the services need to review the customers' personal information and provide corresponding services based on the review results.

[0003] The current customer data approval process first requires customers to submit various required materials, then verifies the authenticity of these materials and conducts qualification reviews. Additionally, due diligence is performed on the personal data, and risk assessments are conducted to identify potential key risks before an approval result is finally obtained. The entire approval process is complex, time-consuming, and requires significant manual intervention, resulting in low efficiency. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for approving customer information, which improves the accuracy and efficiency of customer information approval.

[0005] According to one aspect of the present invention, a customer data approval method is provided, comprising:

[0006] Collect pending approval materials from clients for business applications, and perform privacy-enhanced OCR verification on these materials to obtain credible pending approval materials;

[0007] The credible materials pending approval are input into a comprehensive evaluation model based on federated learning to obtain the customer's comprehensive characteristic parameters;

[0008] The approval rating is determined based on the customer's comprehensive characteristic parameters.

[0009] In one embodiment, the method further includes:

[0010] Collect multi-source heterogeneous data related to pending approval business;

[0011] The federated learning-based comprehensive evaluation model is dynamically optimized periodically based on multi-source heterogeneous data.

[0012] In one embodiment, pending approval materials from customers for business transactions are collected, and privacy-enhanced OCR verification is performed on these materials to obtain credible pending approval materials, including:

[0013] Collect pending approval materials from customers for pending business and build a knowledge graph. Among them, privacy-enhanced OCR verification is used for pending approval materials uploaded by customers.

[0014] When the required materials for approval are determined to be complete and credible based on the knowledge graph and the approval rules of the pending business, credible materials for approval are obtained.

[0015] In one embodiment, privacy-enhanced OCR verification is performed on the materials uploaded by the customer for approval, including:

[0016] Multimodal verification methods are used to detect counterfeit materials uploaded by customers for approval. These methods include physical feature verification and / or chemical feature verification.

[0017] Logical contradiction identification is performed on the materials to be approved obtained from OCR recognition.

[0018] The verification result is obtained by dynamically scoring the materials to be approved based on the detection results of anti-counterfeiting detection and the identification results of logical contradiction identification.

[0019] Among them, privacy-enhanced OCR verification processes data within a hardware-isolated secure domain.

[0020] In one embodiment, the method further includes:

[0021] A federated learning-based comprehensive evaluation model is trained using a distributed training method. The training data is anonymized and aggregated using a secure protocol.

[0022] In one embodiment, after rating the customer's pending business based on the customer's comprehensive characteristic parameters and obtaining the rating result, the process further includes:

[0023] For approval ratings that are rejected, generate structured reasons for rejection and actionable optimization suggestions.

[0024] According to another aspect of the present invention, a customer data approval device is provided, comprising:

[0025] The data collection module is used to collect the materials pending approval from customers for pending business, and to perform privacy-enhanced OCR verification on the materials pending approval to obtain credible materials pending approval.

[0026] The model evaluation module is used to input credible materials to be approved into a comprehensive evaluation model based on federated learning to obtain the customer's comprehensive characteristic parameters;

[0027] The approval rating module is used to rate the pending business transactions of a customer based on the customer's comprehensive characteristic parameters, and obtain the approval rating result.

[0028] According to another aspect of the present invention, an electronic device is provided, comprising: a processor, and a memory communicatively connected to the processor;

[0029] The memory stores instructions that the computer executes;

[0030] The processor executes computer execution instructions stored in memory to implement the aforementioned customer data approval method.

[0031] According to another aspect of the present invention, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium, wherein at least one processor can read the computer program from the computer-readable storage medium, and the at least one processor implements the above-described customer data approval method when executing the computer program.

[0032] The technical solution of this invention verifies the materials to be approved using privacy-enhanced OCR to obtain credible materials to be approved, and processes the credible materials to be approved using a comprehensive evaluation model based on federated learning to finally obtain the approval rating result. Under the premise of ensuring the security of customers' personal privacy data, it greatly reduces the problems of low approval efficiency and low accuracy of approval results caused by manual participation.

[0033] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a diagram illustrating a traditional customer data approval process.

[0036] Figure 2 A schematic diagram of the property valuation process;

[0037] Figure 3 A flowchart illustrating a customer data approval method provided in an embodiment of the present invention;

[0038] Figure 4 A flowchart illustrating another customer data approval method provided in an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram of a customer data approval device provided in an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0041] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present invention. The acquisition, storage, use, and processing of data in the technical solutions of this application all comply with the relevant provisions of national laws and regulations.

[0042] Current customer data approval processes rely heavily on manual review, resulting in low efficiency. These processes typically occur when customers apply for various services or business applications. The service provider needs to determine whether to offer the service or business based on the submitted personal information, and also to identify the type of service or business to be provided. For example, a bank may be an institution providing such services. In this embodiment, a customer applying for a loan from a bank is used as an example to illustrate the data processing method provided in this embodiment.

[0043] The traditional customer information approval process, taking a customer applying for a home loan from a bank as an example, is as follows: Figure 1 As shown, Figure 1 This is a schematic diagram of a traditional customer data approval process, including the following steps:

[0044] Step S110, Customer Application Stage. The customer needs to provide the bank with materials such as identity documents, marriage certificate, and household registration documents.

[0045] Step S120, preliminary review of bank qualifications, including preliminary review of the authenticity of materials, eligibility to purchase a house, credit report and other materials.

[0046] Step S130, bank due diligence, including manual investigation and third-party collaborative investigation.

[0047] Step S140, bank property valuation and mortgage status verification, is a key risk control point.

[0048] Property valuation process, for example, adopts Figure 2 The process shown is as follows: Figure 2 This diagram illustrates the property valuation process. After initiating the valuation process, relevant property data is obtained through a third-party valuation agency. The property type is then determined. For residential properties, the market comparison approach can be used to assess value, and on-site inspections and comparisons with surrounding transaction prices are conducted to determine the final value. For commercial / office properties, the income approach can be used to assess value, along with discounted cash flow analysis of rental income.

[0049] Step S150: Bank risk assessment and loan approval. This typically requires multiple levels of human decision-making.

[0050] Step S160: Signing of bank contract and registration of mortgage.

[0051] As can be seen from the above process, the current customer data approval process, taking a mortgage application as an example, has the following problems:

[0052] 1. Reliance on human expertise and subjective bias. The reliance on human experience in key areas such as credit report interpretation and income stability assessment leads to inconsistent risk assessment standards.

[0053] 2. Data silos and insufficient real-time performance in various systems.

[0054] 3. Passive customer interaction. In the current process, customers must passively wait for the approval result and cannot proactively optimize their application materials based on the pre-review feedback.

[0055] In summary, current customer data approval methods involve a large amount of manual processing, rely on human experience, are unreliable in accuracy and efficiency, and are time-consuming.

[0056] Figure 3 This is a flowchart illustrating a customer data approval method provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the customer data approval method provided in this embodiment includes:

[0057] Step S310: Collect the materials pending approval from the customer for the business pending approval, and perform privacy-enhanced OCR verification on the materials pending approval to obtain credible materials pending approval.

[0058] The customer data approval method provided in this application embodiment can be applied to any computer or server with processing capabilities. This computer or server is used to collect and approve personal data. The computer or server can be a computer or server deploying the software system, or a computer or server dedicated to testing the software system. This computer or server can be deployed locally or in the cloud at an organization providing services to customers.

[0059] First, the client's pending approval materials for their application are collected. These materials can be uploaded proactively by the client or collected from internal or external public databases and are related to the client or the application. Next, privacy-enhanced Optical Character Recognition (OCR) verification is performed on these materials. This verification involves two aspects: OCR recognition and privacy enhancement. OCR recognition is a text or image acquisition method based on image recognition technology. The collected client's pending approval materials can be in the form of images, documents, etc., and are generally non-text formats except for basic personal information. For subsequent processing, these materials need to be converted into formatted text data. Therefore, OCR recognition can be used to collect the text data from various pending approval materials. To ensure the accuracy of client approval, the collected materials are also verified to determine their legality, accuracy, and whether they have been tampered with. Privacy-enhanced OCR processing ensures that the personal privacy data of the verified approval materials is not leaked. Materials awaiting approval that have been verified through privacy-enhanced OCR are considered trustworthy materials, and trustworthy materials can be structured data.

[0060] In one embodiment, collecting pending approval materials from customers for pending approval business and performing privacy-enhanced OCR verification on the pending approval materials to obtain credible pending approval materials may include: collecting pending approval materials from customers for pending approval business and constructing a knowledge graph, wherein the pending approval materials uploaded by customers are processed using privacy-enhanced OCR verification; and then, based on the knowledge graph and the approval rules of the pending approval business, when it is determined that the required pending approval materials are complete and credible, credible pending approval materials are obtained.

[0061] Among them, knowledge graphs are semantic networks that describe the objective world using graph structures. By constructing knowledge graphs from the customer's pending approval materials for pending business, the correlation between various data can be intuitively demonstrated.

[0062] Let's continue with the example of a customer applying for a home loan at a bank. The knowledge graph construction includes the following processes: Data Collection: Only data voluntarily filled out by the customer, authorized for credit investigation, and publicly available government data is collected. Graph Modeling: Constructing three types of entities—customer, collateral, and loan—and their associated rules. Inference Engine: Implementing conditional judgments based on the Prolog rule engine.

[0063] After constructing a knowledge graph for the client's pending approval materials for their business applications, the system further assesses the suitability of these materials. This includes: Rule-driven: Employing a condition-triggered mechanism to ensure that material additions strictly match the business scenario; Legal mapping: Each rule is associated with specific legal provisions, achieving accurate translation of regulatory requirements into system logic; Dynamic response: When a client modifies key attributes (such as occupation / collateral information), the system recalculates the material list in real time, avoiding human oversight; Explainability: The front-end interaction synchronously displays the basis for the rules.

[0064] The process involves privacy-enhanced OCR verification of client-uploaded materials for approval. This includes: using multimodal verification methods to detect counterfeit materials, including physical and / or chemical feature verification; identifying logical contradictions in the OCR-recognized materials; and dynamically scoring the materials based on the counterfeit detection and logical contradiction identification results to obtain the verification outcome. The privacy-enhanced OCR verification process is performed within a hardware-isolated secure domain.

[0065] Privacy-enhanced OCR verification may include the following features: 1. Chip-level secure processing, employing a trusted execution environment: Sensitive data is processed within a hardware-isolated secure domain, ensuring runtime memory encryption and preventing unauthorized access. 2. Hardware-accelerated computing: Dedicated chips are used for critical tasks (such as optical recognition), and one-way encryption and desensitization are performed on the output data to ensure privacy and security.

[0066] OCR verification for credibility includes multimodal anti-counterfeiting detection, which can employ a combination of physical and chemical feature verification methods. Physical feature verification includes analyzing paper texture using laser diffraction to identify batch characteristics and prevent counterfeit documents. Chemical feature verification includes comparing the ink composition of the seal using near-infrared spectroscopy to ensure the authenticity of official seals.

[0067] After OCR recognition is complete, logical contradiction interception can be performed, including verifying age compliance: verifying key age thresholds to ensure business rules (such as the legal age of adulthood) are not violated, otherwise risk control interception will be triggered. Verifying time series contradictions: detecting time logic conflicts of related events (such as contradictions between document dates and key business nodes) to identify anomalies. Verifying ownership consistency: verifying the rationality of ownership relationships such as property and identity (such as missing co-ownership or abnormal proportions) to prevent concealment risks. Implementing a dynamic scoring model: allocating weights to anti-counterfeiting detection and logical contradiction identification: for example, anti-counterfeiting features account for 70% (emphasizing physical authenticity), and logical consistency accounts for 30% (emphasizing data rationality). Threshold control: when the total score is below 0.85, it automatically transfers to manual review to achieve risk-level handling.

[0068] Step S320: Input the credible materials to be approved into the comprehensive evaluation model based on federated learning to obtain the customer's comprehensive characteristic parameters.

[0069] After obtaining credible materials for approval, they are input into a federated learning-based comprehensive evaluation model to obtain the customer's comprehensive characteristic parameters. This federated learning-based comprehensive evaluation model is trained using federated learning technology, a distributed machine learning method that achieves multi-party joint modeling through encrypted parameter exchange, ensuring that training can be completed without the data leaving the local machine. Since the customer document approval process may involve personal privacy data, whether uploaded by the customer or retrieved from internal or external databases, the federated learning-based comprehensive evaluation model ensures the security of this data. The federated learning-based comprehensive evaluation model corresponds to the business to be approved. Inputting the credible materials into the model yields the customer's comprehensive characteristic parameters, which consider various key variables related to the business to be approved, outputting various characteristic parameters that can be used to rate the customer.

[0070] Among them, a distributed training method is used to train the federated learning-based comprehensive evaluation model, and the federated learning-based comprehensive evaluation model performs anonymization processing on the training data and aggregates it using a secure protocol.

[0071] Distributed training of the comprehensive evaluation model based on federated learning ensures that the data of training participants does not leave the domain, and feature calculation and model training are completed locally. Secure aggregation nodes can also be set up: deployed on hardware security modules to ensure the trusted execution of the aggregation process.

[0072] The federated learning-based comprehensive evaluation model employs further privacy protection techniques, including gradient encryption: homomorphic encryption is used to protect transmission parameters and prevent leakage of original data. Differential privacy: controllable noise is added to key features to meet strict privacy budget constraints. Feature desensitization is performed on the training data input to the model; for structured data, k-anonymization + differential privacy processing is used. For unstructured data, neural network embedding is used for extraction. The data transmission security of the federated learning-based comprehensive evaluation model can be achieved by combining financial-grade data transmission encryption with hardware-level key management. Table 1 illustrates one data privacy protection architecture.

[0073] Table 1

[0074]

[0075] The training protocol for a federated learning-based comprehensive evaluation model can include secure aggregation and federated optimization strategies. The secure aggregation strategy includes: an improved multi-party secure computation protocol, double masking technology + SSS (Shamir's Secret Sharing), and a Hardware Security Module (HSM) guarantee. The federated optimization includes: dynamic participant weight adjustment, a privacy-utility balance algorithm, and an automatic convergence detection mechanism.

[0076] Step S330: Based on the customer's comprehensive characteristic parameters, conduct an approval rating of the customer's pending business and obtain the approval rating result.

[0077] Different pending approval items can have different approval ratings. Based on the customer's comprehensive characteristic parameters and the approval rating status of the pending items, an approval rating result can be obtained. The approval rating of pending items is categorized according to the needs of the approval process. For example, some pending items only have two approval results: pass or fail. Other pending items, even if approved, may include multiple different levels of approval ratings. Finally, the approval rating results for the customer's pending items can be provided to the customer.

[0078] For the approval rating results, for example, a multi-entity credit rating system can be established, as shown in Table 2. Table 2 is a schematic diagram of a credit rating system.

[0079] Table 2

[0080]

[0081] Intelligent decision-making engines can be introduced into the approval rating of loan applications. The intelligent decision-making engine's approval rate model uses a logistic regression model to assess the probability of loan application approval, taking into account key variables such as: customer credit score: based on central bank credit reporting, historical repayment records, etc., reflecting the customer's overall creditworthiness; collateral value stability: combining market transaction data and government-guided price volatility to assess collateral risk; guarantor qualifications: including net asset coverage ratio, occupational stability, etc., to measure the guarantee's credit enhancement ability; market funding costs: adjusting dynamic thresholds with reference to interbank lending rates (such as the Shanghai Interbank Offered Rate, SHIBOR) to ensure that funding returns cover risks; and risk control policy constraints: dynamically adjusting approval standards based on current non-performing loan tolerance to ensure compliance with regulatory requirements.

[0082] The model optimizes the weights of each variable using Maximum Likelihood Estimate (MLE) and employs L2 regularization to prevent overfitting. All input features are standardized using Z-scores to ensure model stability and fairness. The decision threshold can be dynamically adjusted according to business strategies, and the Area Under Curve (AUC-ROC) curve is periodically used for validation to ensure the model's discriminative power.

[0083] For loan applicants, a repayment prediction model can be introduced, employing a survival analysis model to assess customer repayment risk. This model comprehensively considers the following key variables: customer fundamental characteristics: age, income stability, etc., reflecting long-term repayment ability; debt stress indicators: the ratio of monthly repayment to income, measuring short-term debt repayment pressure; collateral liquidity: assessing asset liquidity based on market transaction data; and guarantor credit dynamics: historical credit change trends, predicting potential default risk. The model quantifies the impact weight of each variable on default risk using statistical learning methods and dynamically adjusts parameters to adapt to market changes. All features are standardized to ensure model fairness and interpretability.

[0084] In one embodiment, for approval ratings of rejection, structured rejection reasons and actionable optimization suggestions are generated. The approval rating results can incorporate an interpretable rejection suggestion module, providing clients with structured rejection reasons and actionable optimization suggestions based on risk control model decisions. This ensures transparency in rejection decisions and helps clients improve their qualifications for reapplication. A dual-drive model of rule engine + machine learning is adopted to ensure suggestions conform to business logic and possess dynamic adaptability. Rule interpretability means that each suggestion is associated with specific risk control rules, and the legal basis is displayed on the front end. Decision tree visualization technology is used to show clients the rejection logic path (e.g., "Insufficient property score → Excessive regional vacancy rate"). Dynamic adaptability: Suggestion parameters (e.g., down payment ratio, debt ratio threshold) are automatically updated with market policies (e.g., connecting to the central bank's benchmark interest rate application programming interface (API)). The effectiveness of suggestions is optimized through reinforcement learning, prioritizing the solutions with the highest historical success rates. Compliance assurance: The rejection reason database is reviewed by legal counsel to avoid discriminatory statements (e.g., not involving sensitive factors such as gender or region). All suggestion generation processes are traceable and comply with relevant regulations. Table 3 shows examples of reasons for approval rejection and optimization suggestions.

[0085] Table 3

[0086]

[0087] This application embodiment also introduces a dynamic property value assessment system, the assessment model of which is as follows:

[0088]

[0089] Property_Score is the overall evaluation score of the target object, used to quantify its value or risk level. n is the total number of evaluation dimensions, which can be adjusted according to business needs (e.g., 5 dimensions, 7 dimensions, etc.). The weight coefficient for the i-th dimension reflects the contribution of that dimension to the total score, and is usually determined through data-driven optimization methods. For the scoring function of the i-th dimension, input the original data. It outputs a standardized score (e.g., 0~1 or 0~100). The original observation data for the i-th dimension can come from a structured database, API interface, or third-party data source.

[0090] The design principles for assessment dimensions include: the selection of assessment dimensions must meet the requirements of independence, quantifiability, and business relevance. Specifically, independence means that each dimension should minimize correlation to avoid double counting; quantifiability means that data should be objectively obtainable to avoid subjective judgment; and business relevance means that dimensions must be strongly related to the assessment objectives (e.g., real estate assessments should cover economic, legal, and market aspects).

[0091] The methods for setting threshold rules include: the threshold is used to filter the raw data. Mapped to rating The setting of these thresholds must adhere to the following principles: Data-driven: Risk levels are classified based on historical data distribution (e.g., quantiles, standard deviations). Business logic: Key thresholds are defined based on industry experience (e.g., red-line indicators stipulated by policies). Dynamic adjustment: Thresholds can be updated periodically as the market changes (e.g., calculated using a rolling time window). Common threshold types: Qualitative thresholds: such as "existence / non-existence" (e.g., seizure records directly trigger downgrading). Quantitative thresholds: such as interval segmentation (dividing continuous variables into "low risk / medium risk / high risk").

[0092] Compliance Explanation for Weight Training: Weight Source: Training is conducted using publicly available algorithms (such as linear regression and gradient boosting trees). Interpretability: The reasonableness of the weights is analyzed using general interpretability methods such as Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), without relying on black-box models. Data Compliance: Training data is ensured not to involve user privacy or protected commercial data.

[0093] The model's output can be applied in the following scenarios: Risk Management: Identifying high-risk individuals (such as real estate, businesses, and credit clients). Resource Allocation: Optimizing investment or policy priorities (such as allocating regional development funds). Dynamic Monitoring: Adjusting scores in real time to respond to market changes.

[0094] The system also features a matching early warning mechanism. It monitors loan risk in real time using a dynamic assessment model, primarily based on the following key indicators: Loan-to-Value Ratio (LTV): measures the degree of matching between the loan amount and the value of the collateral; Customer Comprehensive Risk Score: reflects the borrower's creditworthiness and repayment ability. When the system detects that a preset risk threshold has been exceeded, it will automatically trigger the following risk control measures: initiating an on-site inspection process to verify the true condition of the collateral; and transferring the process to a manual review stage for a secondary assessment by professional risk control personnel. The early warning mechanism is parameterized and can flexibly adjust the threshold standards according to different business scenarios and regulatory requirements. All decision-making processes retain complete audit logs to ensure compliance.

[0095] Finally, end-to-end compliance assurance includes: data authorization and privacy protection; federated learning feature engineering: adopting a standards-compliant federated learning framework to ensure data is "usable but not visible," preventing the leakage of raw data; and customer self-authorization management: using blockchain technology to achieve authorization and notarization, ensuring that customer data usage is traceable and tamper-proof throughout the entire process.

[0096] Regulatory compliance and risk control sandbox, including real-time credit reporting integration: The system automatically connects to the People's Bank of China's credit reporting system to ensure data source compliance and update credit assessment results in real time. Regulatory sandbox testing: Different risk control strategies are simulated in a closed environment to ensure that new models meet regulatory requirements before going live.

[0097] Model Transparency and Auditability: SHAP Explainability Report: Employing the SHAP method to quantify the basis of model decisions, ensuring that risk control logic is transparent and auditable. Dynamic Parameter Monitoring: Regularly verifying model parameters to ensure they comply with regulatory policies and business risk control standards.

[0098] Customer notification and appeal mechanism: Standardized loan rejection notice: Provide customers with a clear and compliant explanation of the reasons for loan rejection, including a comparison of key indicators. Optimization suggestions: Provide feasible credit enhancement solutions based on risk control rules.

[0099] The customer data approval method provided in this embodiment verifies the materials to be approved using privacy-enhanced OCR to obtain credible materials to be approved, and processes the credible materials to be approved using a comprehensive evaluation model based on federated learning to finally obtain the approval rating result. Under the premise of ensuring the security of customers' personal privacy data, it greatly reduces the problems of low approval efficiency and low accuracy of approval results caused by manual participation.

[0100] Figure 4 This is a flowchart illustrating another customer data approval method provided in an embodiment of the present invention, such as... Figure 4 As shown, the customer data approval method provided in this embodiment includes:

[0101] Step S410: Collect the materials pending approval from the customer for the pending business, and perform privacy-enhanced OCR verification on the materials pending approval to obtain credible materials pending approval.

[0102] Step S420: Input the credible materials to be approved into the comprehensive evaluation model based on federated learning to obtain the customer's comprehensive characteristic parameters.

[0103] Step S430: Based on the customer's comprehensive characteristic parameters, conduct an approval rating of the customer's pending business and obtain the approval rating result.

[0104] Step S440: Collect multi-source heterogeneous data related to the business to be approved.

[0105] The customer data approval method provided in this embodiment is... Figure 1 Based on the illustrated embodiment, the federated learning-based comprehensive evaluation model can also be dynamically optimized. The execution order of steps S440 and S450 in this embodiment is not... Figure 4 As a limit, optimization of the federated learning-based comprehensive evaluation model can be performed at any time.

[0106] First, multi-source heterogeneous data related to the pending approval process is collected. This multi-source heterogeneous data can be obtained from external public databases, and all data is related to the pending approval process. Then, privacy protection processing is applied to the multi-source heterogeneous data. Different weights can also be assigned to various multi-source heterogeneous data based on the credibility of their sources.

[0107] Multi-source heterogeneous data source classification can include:

[0108] Transaction data: By connecting to real estate transaction platforms through compliant APIs, we collect indicators such as regional transaction prices, listing volume, and sales cycle.

[0109] Policy data: Real-time capture of publicly available government documents (such as purchase restriction policies and tax adjustments), and extraction of key clauses using Natural Language Processing (NLP).

[0110] Information and data: Based on open-source sentiment analysis models, social media / news are processed to identify regional economic risk signals.

[0111] Privacy protection design for multi-source heterogeneous data:

[0112] 1. Data De-identification (k-anonymization, k≥N). Purpose: To ensure that each individual transaction record is indistinguishable from at least N other records in terms of key attributes (such as age, region, transaction amount), preventing individual identification. Technical Implementation: Generalize the dataset (e.g., replace precise age with age range) and suppress (e.g., remove rare transaction features) to ensure that each equivalence class has ≥N records.

[0113] 2. Differential Privacy Processing (Laplace Noise, ε=0.*). Purpose: To add controllable noise to statistical indicators (such as average housing price in a neighborhood), preventing external attackers from inferring the original data from the query results. The strength of privacy protection is determined by ε (privacy budget); ε=0.1: indicates extremely strong privacy protection (higher noise, slightly reduced data availability); the smaller ε, the stronger the privacy protection, but the lower the data accuracy. Technical Implementation: Add random noise conforming to a Laplace distribution (mean=0, scale=Δf / ε) to the aggregated query results, where Δf is the query sensitivity.

[0114] For multi-source heterogeneous data, data quality verification can also be performed, including: logical verification: detecting abnormal fluctuations through time series analysis (such as triggering manual review by a single-day price change of ±30%); source credibility scoring: government data sources: default weight 1.0; third-party platforms: dynamically adjusted according to historical accuracy (0.6~0.9); related information data: reduced to 0.3~0.5 after sentiment polarity analysis.

[0115] Step S450: Periodically optimize the federated learning-based comprehensive evaluation model based on multi-source heterogeneous data.

[0116] Each pending approval process may change at any time due to the influence of policies, markets, and various internal or external factors. Therefore, the federated learning-based comprehensive evaluation model can be dynamically optimized periodically based on collected multi-source heterogeneous data. This means retraining the federated learning-based comprehensive evaluation model at regular intervals using the collected multi-source heterogeneous data, or retraining it when the collected multi-source heterogeneous data reaches a preset threshold or meets preset conditions. Retraining the federated learning-based comprehensive evaluation model can be done incrementally. Furthermore, retraining can be performed only when the deviation in customer document approval exceeds a preset deviation threshold to ensure the stability of the federated learning-based comprehensive evaluation model.

[0117] The iterative optimization of the federated learning-based comprehensive evaluation model includes: model bias detection: for example, calculating the Kullback-Leibler divergence (KL) between the actual and predicted default rates monthly, triggering retraining if the bias exceeds 15%; and using adversarial validation to identify feature distribution shifts. Incremental learning framework: based on the Follow-the-Regularized-Leader (FTRL) online learning algorithm, supporting real-time fine-tuning of model parameters; training data isolation: new data must pass compliance verification (e.g., excluding data from litigated clients) before being added to the training set.

[0118] Finally, an audit trail solution can be provided, recording end-to-end logs, including model input / output, market data versions, and parameter adjustment records, stored in a tamper-proof database (such as a Merkle tree-based blockchain structure); log fields comply with relevant requirements. A regulatory interface is provided, offering a standardized API for regulatory agencies to query model versions and training data overviews (returning only aggregated statistics).

[0119] To conduct risk stress testing on the customer information approval method provided in this application embodiment, a scenario library is first constructed, including: macroeconomic shocks: simulating the depreciation pressure of collateral when the GDP growth rate falls to 3%; policy changes: testing the impact of stricter purchase restrictions on customer qualifications. Capital adequacy ratio verification is then performed: the risk reserve ratio is dynamically adjusted based on the test results to ensure compliance with relevant requirements.

[0120] The federated learning-based comprehensive evaluation model can provide market dashboards: generate regional risk heat maps (based on GeoHash Geocoding System (GeoHash) grid aggregation data) to assist in credit allocation; provide model iteration reports: output monthly SHAP feature importance change analysis to ensure continuous interpretability; and have automated early warning: when collateral liquidity indicators deteriorate (such as depletion period > 12 months), automatically reduce the corresponding region's LTV limit.

[0121] The customer data approval method provided in this application significantly improves the efficiency of the customer data approval process, particularly in the material review and risk control decision-making stages, where efficiency is significantly improved compared to manual decision-making. Secondly, the multi-level intelligent approval system also significantly improves the accuracy of risk identification compared to traditional risk control models. Finally, the use of privacy-enhanced OCR verification technology greatly reduces the risk of sensitive information leakage, and the federated learning-based comprehensive evaluation model improves credit assessment accuracy while ensuring data isolation.

[0122] Figure 5 This is a schematic diagram of a customer data approval device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the customer data approval device provided in this embodiment includes:

[0123] The data collection module 51 is used to collect the pending approval materials of customers for pending approval business and to perform privacy-enhanced OCR verification on the pending approval materials to obtain credible pending approval materials; the model evaluation module 52 is used to input the credible pending approval materials into a comprehensive evaluation model based on federated learning to obtain the customer's comprehensive feature parameters; the approval rating module 53 is used to rate the customer's pending approval business according to the customer's comprehensive feature parameters to obtain the approval rating result.

[0124] The customer data approval device provided in this embodiment is used to achieve... Figure 3 The customer data approval method provided in the illustrated embodiment has a similar implementation principle and technical effect, and will not be described in detail here.

[0125] In one embodiment, the customer data approval device further includes a model optimization module for collecting multi-source heterogeneous data related to the business to be approved; and periodically optimizing the comprehensive evaluation model based on federated learning based on the multi-source heterogeneous data.

[0126] In one embodiment, the data collection module 51 is specifically used to collect the pending approval materials from customers for pending approval business and construct a knowledge graph. The pending approval materials uploaded by customers are verified using privacy-enhanced OCR. When the required pending approval materials are determined to be complete and credible according to the knowledge graph and the approval rules of the pending approval business, credible pending approval materials are obtained.

[0127] In one embodiment, the data collection module 51 is specifically used to perform anti-counterfeiting detection on the materials to be approved uploaded by the customer using a multimodal verification method, which includes physical feature verification and / or chemical feature verification; to perform logical contradiction identification on the materials to be approved obtained by OCR recognition; and to dynamically score the materials to be approved based on the detection results of anti-counterfeiting detection and the identification results of logical contradiction identification to obtain the verification result; wherein, the privacy-enhanced OCR verification is performed in a hardware-isolated security domain.

[0128] In one embodiment, the customer data approval device further includes a model training module, which is used to train a federated learning-based comprehensive evaluation model using a distributed training method. The federated learning-based comprehensive evaluation model performs de-identification processing on the training data and aggregates it using a security protocol.

[0129] In one embodiment, the model training module is further configured to generate structured rejection reasons and executable optimization suggestions for approval rating results with a rejection rating.

[0130] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device may include: a transceiver 61, a processor 62, and a memory 63.

[0131] Processor 62 executes computer execution instructions stored in memory, causing processor 62 to perform the scheme in the above embodiments. Processor 62 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0132] The memory 63 is connected to the processor 62 via the system bus and completes communication between them. The memory 63 is used to store computer program instructions.

[0133] Transceiver 61 can be used to obtain the task to be run and its configuration information.

[0134] The system bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0135] The electronic device provided in this application embodiment may be a computer or server deployed in the customer data approval device of the above embodiments.

[0136] This application also provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the technical solution of the customer data approval method described above.

[0137] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solution of the customer data approval method in the above embodiments.

[0138] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0139] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for approving customer information, characterized in that, include: Collect the materials pending approval from customers for pending business applications, and perform privacy-enhanced OCR verification on the materials to obtain credible materials pending approval; The credible materials to be approved are input into a comprehensive evaluation model based on federated learning to obtain the comprehensive characteristic parameters of the customer; The approval rating is determined based on the customer's comprehensive characteristic parameters to obtain the approval rating result.

2. The method according to claim 1, characterized in that, The method further includes: Collect multi-source heterogeneous data related to the pending approval business; The federated learning-based comprehensive evaluation model is periodically optimized based on the multi-source heterogeneous data.

3. The method according to claim 1, characterized in that, The process involves collecting pending approval materials from customers for business transactions and performing privacy-enhanced OCR verification on these materials to obtain credible pending approval materials, including: Collect pending approval materials from customers for pending business and build a knowledge graph. Among them, privacy-enhanced OCR verification is used for pending approval materials uploaded by customers. When the required materials for approval are determined to be complete and reliable based on the knowledge graph and the approval rules of the business to be approved, reliable materials for approval are obtained.

4. The method according to claim 3, characterized in that, The process of verifying the pending approval materials uploaded by customers using privacy-enhanced OCR includes: Multimodal verification methods are used to detect counterfeit materials uploaded by customers for approval. These multimodal verification methods include physical feature verification and / or chemical feature verification. Logical contradiction identification is performed on the materials to be approved obtained from OCR recognition. Based on the detection results of the anti-counterfeiting detection and the identification results of the logical contradiction identification, the materials to be approved are dynamically scored to obtain the verification results; The privacy-enhanced OCR verification process is performed within a hardware-isolated secure domain.

5. The method according to claim 1, characterized in that, The method further includes: The federated learning-based comprehensive evaluation model is trained using a distributed training method. The federated learning-based comprehensive evaluation model performs anonymization processing on the training data and aggregates it using a secure protocol.

6. The method according to claim 1, characterized in that, After obtaining the approval rating result by rating the customer's pending business based on the customer's comprehensive characteristic parameters, the process further includes: For approval ratings that are rejected, generate structured reasons for rejection and actionable optimization suggestions.

7. A customer data approval device, characterized in that, include: The data collection module is used to collect the materials pending approval from customers for pending approval business, and to perform privacy-enhanced OCR verification on the materials pending approval to obtain credible materials pending approval. The model evaluation module is used to input the credible materials to be approved into a comprehensive evaluation model based on federated learning to obtain the comprehensive characteristic parameters of the customer. The approval rating module is used to rate the customer's pending business based on the customer's comprehensive characteristic parameters and obtain the approval rating result.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the customer data approval method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the customer data approval method as described in any one of claims 1-6.

10. A computer program product, characterized in that, It includes a computer program stored in a computer-readable storage medium, at least one processor can read the computer program from the computer-readable storage medium, and the at least one processor, when executing the computer program, can implement the customer data approval method as described in any one of claims 1 to 6.

Images