Multi-source information-oriented credit data management method and system

Abstract

The invention discloses a credit data management method and system for multi-source information. The method comprises the following steps: obtaining multi-source credit data, and calculating the sensitivity of each data field based on information entropy; processing the original information according to the sensitivity difference; data fusion is carried out through a knowledge distillation algorithm, and a teacher-student network architecture is adopted to generate lightweight privacy protection features; consistency verification is carried out on the fused data, and a data relation graph is constructed to detect and process conflicts; and calculating information fidelity and privacy protection intensity, and adaptively optimizing privacy budget allocation and distillation weight. According to the method, the problems of insufficient privacy protection, poor fusion effect, lack of adaptive optimization and the like of the multi-source credit data are solved, efficient data fusion and quality optimization are realized on the premise of ensuring privacy security, and a reliable data basis is provided for a credit evaluation system.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically a credit data management method and system for multi-source information. Background Technology

[0002] Credit assessment systems need to integrate information from multiple data sources, including bank credit reporting, third-party payments, e-commerce transactions, social networks, public records, and the Internet of Things, to build a more comprehensive and accurate credit profile. However, these multi-source heterogeneous data differ significantly in data format, quality standards, and update frequency, posing a significant challenge to the effective integration and utilization of data. Furthermore, the credibility and importance of different data sources vary, making the reasonable evaluation and utilization of these differentiated data sources a key issue in the field of credit data management.

[0003] Existing credit data processing methods have significant shortcomings in privacy protection. Most rely on simple data anonymization or masking techniques, lacking differentiated protection strategies for data fields with varying levels of sensitivity. Traditional privacy protection methods often employ a one-size-fits-all approach, either over-protecting data and severely reducing its usability, or under-protecting it and leading to privacy breaches. Furthermore, existing methods lack quantitative assessment mechanisms for the sensitivity of data fields, making it impossible to formulate corresponding protection strategies based on actual risk levels and achieve an optimal balance between privacy protection and data usability. Regarding multi-source data fusion, existing technologies mainly rely on traditional data integration methods, such as data warehousing or simple data stitching. These methods are computationally inefficient when processing large-scale heterogeneous data and lack intelligent feature extraction and fusion capabilities.

[0004] Therefore, there is an urgent need for a credit data management method oriented towards multi-source information to overcome the above-mentioned shortcomings. Summary of the Invention

[0005] (1) Technical problems to be solved The purpose of this invention is to provide a credit data management method and system for multi-source information, so as to solve the problems of insufficient privacy protection, poor data fusion effect, lack of adaptive optimization capability and low computational efficiency in the existing technology.

[0006] (2) Technical solution To achieve the above objectives, the present invention provides a credit data management method for multi-source information, the method comprising the following steps: S1. Obtain multi-source credit data, including bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data, and IoT behavior data, and calculate the sensitivity of each data field based on information entropy.

[0007] S2 constructs a three-layer data processing strategy, using preset methods to perform privacy protection processing on the original multi-source credit data based on the sensitivity of the data fields.

[0008] S3: The processed multi-source credit data is fused using a knowledge distillation algorithm. The fused data is then verified for consistency. If the verification fails, the original multi-source credit data is processed to prevent conflicts before step S3 is re-executed.

[0009] S4 calculates the information fidelity and privacy protection strength of the fused data that has passed the consistency verification. If the result meets the preset requirements, output the unified credit dataset with privacy protection. Otherwise, optimize the privacy budget allocation and distillation weights until the result meets the preset requirements or reaches the maximum number of iterations.

[0010] Furthermore, the sensitivity is quantified based on the information entropy of the credit data: ;in, For the first Data fields Sensitivity For data fields Information entropy: , Represents data fields The Middle The probability of each different value occurring. This represents the total number of unique values ​​in this field. Assigning weights to different data source types: bank credit data is assigned 1.0, third-party payment data 0.9, e-commerce transaction data 0.8, social network data 0.6, public record data 0.7, and IoT behavior data 0.5. This is the risk impact factor, determined based on the field content type, with a value range of [0.5, 1.5].

[0011] Furthermore, the three-layer data processing strategy is based on sensitivity. Implement differentiated processing for privacy protection; when When k-anonymization is used, the minimum size of the anonymous set is... The value satisfies: This makes the original record and Other records cannot be distinguished by the quasi-identifier attribute.

[0012] when At that time, calibration noise is added to the raw data: ;in, The original data values ​​without noise. These are the data values ​​after adding noise; for Privacy budget, This is the global sensitivity of the query function.

[0013] when At the same time, keep the original data unchanged and only perform format standardization processing.

[0014] Furthermore, when At that time, privacy budget Adaptive allocation based on sensitivity, total privacy budget Privacy budget for each field: .

[0015] Furthermore, the data fusion of the processed multi-source credit data using the knowledge distillation algorithm includes: implementing knowledge distillation through a combination of a teacher network and a student network. The teacher network employs a 5-layer fully connected neural network with [1024, 512, 256, 128, 64] nodes per layer, using the ReLU activation function to output a complete feature representation of the original multi-source data. The student network employs a 3-layer fully connected network with [256, 128, 64] nodes. Both the teacher and student networks receive the same input data. The student network learns by mimicking the output distribution of the teacher network and generates lightweight, privacy-preserving feature outputs through knowledge distillation.

[0016] The loss function is: ;in, For mission losses, For knowledge distillation loss, , , For student network input The predicted output, For the first The feature vector of each input sample For the output of the teacher network; For the first The true label of each sample For divergence; For knowledge distillation loss, This represents the initial distillation weight.

[0017] Furthermore, the consistency check of the fused data includes constructing a data relationship graph. ,in, Represents a set of data entities. Indicates the relationships between entities; Calculate the similarity between nodes in a data relationship graph: ,in, , They are nodes , eigenvectors, This is the weight matrix. This represents vector concatenation. It is the sigmoid activation function; when If a conflict is detected in the merged data, the original multi-source credit data is sorted by source credibility, the credit data with the lowest source credibility is deleted, and step S3 is repeated until... .

[0018] Furthermore, when At the same time, calculate the information fidelity of the fused data. and privacy protection strength , ; ;in, Represents the original data field and the actual tag value corresponding to this field mutual information, This indicates the data after privacy protection processing; dataset quality score: .

[0019] Furthermore, when the quality score When starting parameter optimization, the privacy budget allocation adjustment is represented as follows: Distillation weight adjustment is expressed as: ;in, and These represent the old and new privacy budgets for each field. and Represent the weights of the old and new distillations, respectively; Constraints: and Repeat the parameter adjustment until... Or it can reach the maximum number of iterations, 50.

[0020] Furthermore, when a new data source is added, steps S1 to S4 are re-executed.

[0021] Based on the same inventive concept, this invention also provides a credit data management system for multi-source information, used to execute the method of this invention. The system includes: a data acquisition and sensitivity analysis module, a privacy protection processing module, a knowledge distillation data fusion module, a quality assessment and optimization module, a parameter adjustment and control module, and an output interface.

[0022] Furthermore, the data acquisition and sensitivity analysis module is used to acquire multi-source credit data, including bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data, and IoT behavior data, and calculates the sensitivity of each data field based on information entropy.

[0023] Furthermore, the privacy protection processing module is used to construct a three-layer data processing strategy, and to perform privacy protection processing on the original multi-source credit data using a preset method based on the sensitivity of the data fields.

[0024] Furthermore, the knowledge distillation data fusion module includes: a teacher network unit, a student network unit, a distillation processing unit, and a consistency verification unit; the teacher network unit adopts a 5-layer fully connected neural network, with the number of nodes in each layer being [1024, 512, 256, 128, 64], and uses the ReLU activation function; the student network unit adopts a 3-layer fully connected network, with the number of nodes being [256, 128, 64]; the distillation processing unit performs data fusion on the processed multi-source credit data using a knowledge distillation algorithm; the consistency verification unit performs consistency verification on the fused data and constructs a data relationship graph for conflict detection.

[0025] Furthermore, the quality assessment and optimization module is used to calculate the information fidelity and privacy protection strength of the fused data that has passed the consistency verification. When the results do not meet the preset requirements, the privacy budget allocation and distillation weights are optimized until the results meet the preset requirements or the maximum number of iterations is reached.

[0026] Furthermore, the parameter adjustment control module is used when the quality score... At that time, parameter optimization is initiated, and privacy budget allocation and distillation weights are adjusted.

[0027] Furthermore, the output interface module is used for data quality scoring. At that time, output a unified credit dataset with privacy protection.

[0028] (3) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: This invention uses an information entropy-based sensitive measurement method to accurately assess the privacy risks of different data fields and maximize data availability. It employs knowledge distillation technology and a teacher-student network architecture to achieve lightweight feature extraction and fusion of multi-source data, improving computational efficiency by about 20% compared to traditional methods. At the same time, it effectively solves the data conflict problem through consistency verification of data relationship graphs, and has broad application prospects. Attached Figure Description

[0029] Figure 1 This is a flowchart of a credit data management method for multi-source information according to the present invention; Figure 2 This is a schematic diagram of the composition of a credit data management system for multi-source information according to the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of this invention, not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] Example 1 like Figure 1 The diagram shown is a flowchart of a credit data management method for multi-source information according to the present invention. The method includes the following steps: S1. Obtain multi-source credit data, including bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data, and IoT behavior data, and calculate the sensitivity of each data field based on information entropy.

[0032] The acquisition of multi-source credit data utilizes standardized API interfaces and data transmission protocols. Bank credit data primarily includes structured data such as loan records, repayment history, and credit limits, typically in XML or JSON format. Third-party payment data includes behavioral characteristics such as transaction frequency, transaction amount, and break-even point. E-commerce transaction data covers e-commerce platform data such as purchasing preferences, spending power, and credit history. Social network data includes semi-structured data such as social relationship graphs, interaction frequency, and content preferences. Public record data includes publicly available information such as business registrations, court judgments, and administrative penalties. IoT behavioral data includes sensor data such as device usage patterns, location trajectories, and time preferences.

[0033] The sensitivity is quantified based on the information entropy of the credit data: ;in, For the first Data fields Sensitivity For data fields Information entropy: , Represents data fields The Middle The probability of each different value occurring. This represents the total number of unique values ​​in this field. Assigning weights to different data source types: bank credit data is assigned 1.0, third-party payment data 0.9, e-commerce transaction data 0.8, social network data 0.6, public record data 0.7, and IoT behavior data 0.5. This is the risk impact factor, determined based on the field content type, with a value range of [0.5, 1.5].

[0034] Specifically, risk impact factors The determination is based on the potential impact of data fields on credit assessment results. For example, fields that directly identify an individual, such as ID card numbers and bank card numbers, are set... =1.5, because the leakage of this information could lead to identity theft; financial status fields such as income level and asset status are set... =1.3, because it involves personal financial privacy; behavioral preference fields such as shopping preferences and consumption habits are set. =1.0, which is considered medium sensitivity; This setting applies to statistical summary fields such as age group and region. =0.5, because this information is relatively insensitive. Data source type weight. The settings are based on the importance and reliability of different data sources in credit assessment. Bank credit data, as the most authoritative source of credit information, is weighted at 1.0, and the weights of other data sources decrease sequentially.

[0035] S2 constructs a three-layer data processing strategy, using preset methods to perform privacy protection processing on the original multi-source credit data based on the sensitivity of the data fields.

[0036] The three-layer data processing strategy is based on sensitivity. Implement differentiated treatment for privacy protection.

[0037] when When k-anonymization is used, the minimum size of the anonymous set is... The value satisfies: This makes the original record and Other records are indistinguishable from each other based on the quasi-identifier attribute. For highly sensitive data, k-anonymization ensures that each record is associated with at least one other record. All other records are identical in their quasi-identifier attributes, making it impossible to distinguish a specific individual. For example, a record containing "Age: 28, Occupation: Software Engineer, Monthly Income: 15,000 yuan" might become "Age: 25-30, Occupation: Technical, Monthly Income: 10,000-20,000 yuan" after k-anonymization, ensuring that at least k users have the same generalized attribute value.

[0038] when At that time, calibration noise is added to the raw data: ;in, The original data values ​​without noise. These are the data values ​​after adding noise; for Privacy budget, Global sensitivity for query functions. Privacy budget. Adaptive allocation based on sensitivity, total privacy budget Privacy budget for each field: Differential privacy mechanism satisfies Differential privacy, in which The noise variance is: .

[0039] For example, for monthly income data, if the original value is 15,000 yuan, the global sensitivity is... Privacy Budget The value after adding noise is Yuan.

[0040] when At the same time, the original data remains unchanged, and only format standardization processing is performed; this includes data type conversion, missing value imputation, and outlier detection and handling. For example, dates in different formats are uniformly converted to "YYYY-MM-DD" format, amounts in different units are uniformly converted to RMB yuan, and obvious input errors (such as age as -5 years or 999 years) are marked and corrected as outliers.

[0041] S3: The processed multi-source credit data is fused using a knowledge distillation algorithm. The fused data is then verified for consistency. If the verification fails, the original multi-source credit data is processed to prevent conflicts before step S3 is re-executed.

[0042] The data fusion of the processed multi-source credit data using the knowledge distillation algorithm includes: implementing knowledge distillation through a combination of a teacher network and a student network. The teacher network employs a 5-layer fully connected neural network with [1024, 512, 256, 128, 64] nodes per layer, using the ReLU activation function to output a complete feature representation of the original multi-source data. The student network employs a 3-layer fully connected network with [256, 128, 64] nodes. Both the teacher and student networks receive the same input data. The student network learns by mimicking the output distribution of the teacher network and generates lightweight, privacy-preserving feature outputs through knowledge distillation.

[0043] Teacher networks, as large-capacity deep neural networks, are capable of learning complex feature representations and potential correlations from raw multi-source data; the forward propagation process of the network is as follows: , ,...,in and The first Layer weight matrix and bias vector; the output of the teacher network is processed by temperature parameters. Soften: ,in For the first The logits output for each category, The value is typically set between 3 and 5. The student network structure is relatively simple, and it learns knowledge by imitating the soft target probability distribution of the teacher network. The distillation loss uses KL divergence to measure the difference between the two probability distributions. During training, the student network not only needs to fit the hard targets of the real labels, but also needs to learn the soft target distribution output by the teacher network.

[0044] The loss function is: ;in, For mission losses, For knowledge distillation loss, , , For student network input The predicted output, For the first The feature vector of each input sample For the output of the teacher network; For the first The true label of each sample For divergence; For knowledge distillation loss, This represents the initial distillation weight.

[0045] S4 calculates the information fidelity and privacy protection strength of the fused data that has passed the consistency verification. If the result meets the preset requirements, output the unified credit dataset with privacy protection. Otherwise, optimize the privacy budget allocation and distillation weights until the result meets the preset requirements or reaches the maximum number of iterations.

[0046] The consistency check of the merged data includes constructing a data relationship graph. ,in, Represents a set of data entities. It indicates the relationship between entities.

[0047] Calculate the similarity between nodes in a data relationship graph: ,in, , They are nodes , eigenvectors, This is the weight matrix. This represents vector concatenation. It is the sigmoid activation function.

[0048] when If a conflict is detected in the merged data, the original multi-source credit data is sorted by source credibility, the credit data with the lowest source credibility is deleted, and step S3 is repeated until... .

[0049] when At the same time, calculate the information fidelity of the fused data. and privacy protection strength , ; ;in, Represents the original data field and the actual tag value corresponding to this field mutual information, This indicates data that has undergone privacy protection processing; Sensitivity The number of all data fields, including: highly sensitive fields processed with k-anonymization ( ) and the medium sensitivity field with added Laplace noise ( ).

[0050] For example, if the mutual information between the original income data and the credit rating is 1.2, and it decreases to 1.0 after noise is added, then the information fidelity of this field is... .

[0051] Dataset quality score: .

[0052] When quality score When starting parameter optimization, the privacy budget allocation adjustment is represented as follows: Distillation weight adjustment is expressed as: ;in, and These represent the old and new privacy budgets for each field. and Represent the weights of the old and new distillations, respectively; Constraints: and .

[0053] Repeat the parameter adjustment until Or it can reach the maximum number of iterations, 50.

[0054] When a new data source is added, repeat steps S1 to S4.

[0055] Taking a bank's credit card approval system as an example for practical application testing, the system needs to integrate user information from six data sources, including the People's Bank of China Credit Reference Center, Alipay, Taobao, WeChat, the State Administration for Industry and Commerce, and the Internet of Vehicles, involving approximately 5 million user records, with each user containing an average of 120 data fields.

[0056] During the data acquisition phase, the system collected multi-dimensional data including basic personal information, credit records, consumption behavior, social relationships, corporate background, and travel patterns. Through information entropy calculation, the sensitivity of the ID card number was determined to be [value missing]. The sensitivity to monthly income is The sensitivity of shopping preferences is The sensitivity of the age group is .

[0057] Based on a three-tier processing strategy, the ID card number field is processed using... Anonymization processing; Laplace noise was added to the monthly income field; privacy budget allocation was... Fields sensitive to age groups are only standardized.

[0058] During the knowledge distillation and fusion phase, the teacher network achieved an AUC of 0.876 on the test set after training convergence, while the student network achieved an AUC of 0.851 after distillation training, representing a 3.4% improvement compared to a directly trained network of the same size (AUC=0.823). During data fusion, 13,247 conflicting records were detected, and 8,156 low-confidence records were automatically removed based on source credibility ranking, resulting in a final similarity score of [missing data]. It passed the consistency verification.

[0059] The fidelity of the unified credit dataset output by the final system Privacy protection strength Quality score It meets the preset requirements.

[0060] Compared with traditional methods, this invention improves data availability by approximately 15% while maintaining the same level of privacy protection; and improves privacy protection strength by approximately 23% while maintaining the same data availability.

[0061] The computational efficiency of the entire data processing flow is improved by about 20% compared with the traditional method, and the time is shortened from 45 minutes to 36 minutes, which fully verifies the technical advantages and practical value of the method of the present invention.

[0062] Example 2 like Figure 2 The diagram shown illustrates the composition of a credit data management system for multi-source information according to the present invention. The system includes: a data acquisition and sensitivity analysis module, a privacy protection processing module, a knowledge distillation data fusion module, a quality assessment and optimization module, a parameter adjustment and control module, and an output interface.

[0063] The data acquisition and sensitivity analysis module is used to acquire multi-source credit data, including bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data, and IoT behavior data, and calculates the sensitivity of each data field based on information entropy.

[0064] The privacy protection processing module is used to construct a three-layer data processing strategy, and to perform privacy protection processing on the original multi-source credit data using preset methods based on the sensitivity of the data fields.

[0065] The knowledge distillation data fusion module includes: a teacher network unit, a student network unit, a distillation processing unit, and a consistency verification unit. The teacher network unit adopts a 5-layer fully connected neural network with the number of nodes in each layer being [1024, 512, 256, 128, 64], and uses the ReLU activation function. The student network unit adopts a 3-layer fully connected network with the number of nodes being [256, 128, 64]. The distillation processing unit performs data fusion on the processed multi-source credit data using a knowledge distillation algorithm. The consistency verification unit performs consistency verification on the fused data and constructs a data relationship graph for conflict detection.

[0066] The quality assessment and optimization module is used to calculate the information fidelity and privacy protection strength of the fused data that has passed the consistency verification. When the results do not meet the preset requirements, the privacy budget allocation and distillation weights are optimized until the results meet the preset requirements or the maximum number of iterations is reached.

[0067] The parameter adjustment control module is used to adjust the quality score. At that time, parameter optimization is initiated, and privacy budget allocation and distillation weights are adjusted.

[0068] The output interface module is used for data quality scoring. At that time, output a unified credit dataset with privacy protection.

[0069] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A credit data management method for multi-source information, characterized by, The method comprises the following steps: S1, acquiring multi-source credit data, the multi-source credit data comprising bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data and Internet of Things behavior data, and calculating the sensitivity of each data field based on information entropy; S2, constructing a three-layer data processing strategy, and adopting a preset method to perform privacy protection processing on the original multi-source credit data according to the sensitivity of the data field; S3, performing data fusion on the processed multi-source credit data through a knowledge distillation algorithm, and performing consistency verification on the fused data, if the consistency verification fails, performing anti-conflict processing on the original multi-source credit data, and then re-executing step S3; S4, calculating the information fidelity and privacy protection strength of the fused data that passes the consistency verification, if the result meets the preset requirement, outputting the unified credit data set with privacy protection, otherwise, optimizing the privacy budget allocation and distillation weight until the result meets the preset requirement or the maximum iteration number is reached.

2. The method of claim 1, wherein, The sensitivity is quantified according to the information entropy of the credit data: ; wherein, is the sensitivity of the first data field, is the information entropy of the data field , represents the probability of the first different value appearing in the data field , is the total number of non-repeated values of the field; is the data source type weight, 1.0 for bank credit data, 0.9 for third-party payment data, 0.8 for e-commerce transaction data, 0.6 for social network data, 0.7 for public record data, and 0.5 for Internet of Things behavior data; is the risk impact factor, which is determined according to the field content type, and the value range is [0.5, 1.5].​​ 3. The method of claim 2, wherein, The three-layer data processing strategy is based on sensitivity Implementing privacy protection differentiated processing; When k-anonymity is used, the minimum size of the anonymous set is chosen to satisfy: so that the original record cannot be distinguished from other records on the quasi-identifier attributes. When a calibration noise is added to the original data: ; wherein, is the original data value without noise, is the data value after adding noise; is a privacy budget, is the global sensitivity of the query function; When The original data is kept unchanged, only format standardization is performed.

4. The method of claim 3, wherein, When the privacy budget is adaptively allocated based on the sensitivity, the total privacy budget is allocated to each field .

5. The method of claim 4, wherein, The data fusion on the processed multi-source credit data through the knowledge distillation algorithm comprises: knowledge distillation is realized through a combination of a teacher network and a student network, the teacher network adopts a 5-layer fully connected neural network, the number of nodes of each layer is [1024, 512, 256, 128, 64] respectively, a ReLU activation function is used, and complete feature representation of original multi-source data is output; the student network adopts a 3-layer fully connected network, the number of nodes is [256, 128, 64], and the teacher network and the student network receive the same input data The student network learns by imitating the output distribution of the teacher network, and generates a lightweight privacy protection feature output through knowledge distillation. The loss function is: ;in, For mission losses, For knowledge distillation loss, , , For student network input The predicted output, For the first The feature vector of each input sample For the output of the teacher network; For the first The true label of each sample For divergence; For knowledge distillation loss, This represents the initial distillation weight.

6. The method of claim 5, wherein, The consistency check on the fused data includes constructing a data relationship graph wherein, denotes a set of data entities, denotes a relationship between the entities; Computing similarity between nodes of the data relationship graph: wherein, , are feature vectors of nodes , respectively, is a weight matrix, denotes vector concatenation, is a sigmoid activation function; When a conflict exists in the fused data, the original multi-source credit data is sorted according to source credibility, the credit data with the lowest source credibility is deleted, and step S3 is re-executed until .

7. The method of claim 6, wherein, When the information fidelity of the fused data is calculated and the privacy protection strength , ; ; wherein denotes the mutual information of the original data field and the true label value corresponding to the field , denotes the data after implementing the privacy protection processing; the data set quality score: .

8. The method of claim 7, wherein, When the quality score , the parameter optimization is initiated, the privacy budget allocation adjustment is represented as: ; the distillation weight adjustment is represented as: ; wherein, and represent the new and old privacy budgets of each field respectively, and represent the new and old distillation weights respectively; the constraint conditions are: and ; The parameter adjustment is repeatedly performed until or a maximum number of iterations of 50 times is reached.

9. The method according to any one of claims 1 to 8, characterized in that, When a new data source is added, steps S1 to S4 are re-executed.

10. A multi-source information oriented credit data management system for performing the method of any one of claims 1 to 9, characterized in that, The system comprises a data acquisition and sensitivity analysis module, a privacy protection processing module, a knowledge distillation data fusion module, a quality evaluation and optimization module, a parameter adjustment control module and an output interface; The data acquisition and sensitivity analysis module is configured to acquire multi-source credit data, the multi-source credit data comprising bank credit data, third-party payment data, e-commerce transaction data, social network data, public record data and Internet of Things behavior data, and calculate the sensitivity of each data field based on information entropy; The privacy protection processing module is configured to construct a three-layer data processing strategy, and adopt a preset method to perform privacy protection processing on the original multi-source credit data according to the sensitivity of the data field; The knowledge distillation data fusion module comprises a teacher network unit, a student network unit, a distillation processing unit and a consistency verification unit; The teacher network unit adopts a 5-layer fully connected neural network, the number of nodes of each layer being [1024, 512, 256, 128, 64], and using a ReLU activation function; The student network unit adopts a 3-layer fully connected network, the number of nodes being [256, 128, 64]; The distillation processing unit performs data fusion on the processed multi-source credit data through a knowledge distillation algorithm; The consistency verification unit performs consistency verification on the fused data, and constructs a data relationship graph for conflict detection; The quality evaluation and optimization module is configured to calculate the information fidelity and privacy protection strength of the fused data that passes the consistency verification, when the result does not meet the preset requirement, optimize the privacy budget allocation and distillation weight until the result meets the preset requirement or the maximum iteration number is reached. The parameter adjustment control module is configured to initiate parameter optimization to adjust the privacy budget allocation and the distillation weight when the quality score is less than the threshold. The output interface module is configured to output the output privacy-protected unified credit dataset when the data quality score is above a threshold.

Images