A data asset mortgage pre-audit method and system based on privacy computing

Abstract

The application discloses a kind of data asset mortgage pre-audit method and system based on privacy computing, belong to data security technical field, the present application will fund credit as first data node, will asset provider as second data node, in local trusted execution environment, original data is mapped into logical native matrix, based on Map-Reduce architecture concurrent execution integrity and uniqueness verification, Z-Score abnormal offset detection, IQR abnormal extreme value detection, self-defined extreme value boundary verification, classification enumeration legality verification and regular expression consistency verification, generate comprehensive audit result vector, then destroy local data and only return desensitization aggregation result;First data node is based on risk control tolerance threshold value and executes pre-gate determination, and through cross-node hot updating mechanism, component metadata is injected into second data node front-end container, can complete multidimensional quality audit before data asset mortgage, desensitization result return and two-way collaborative verification under the premise that original data does not go out of domain.

Description

Technical Field

[0001] This invention relates to the fields of data security and financial risk control technology, and in particular to a method and system for pre-audit of data asset collateral based on privacy computing. Background Technology

[0002] With the rapid development of the data element market, customer data, transaction data, equipment data, behavioral data, and operational data accumulated by enterprises during their operations have gradually become assessable, transferable, and usable data assets for financing and credit enhancement. In scenarios such as data asset collateralization, pledge, or credit review, lenders typically need to confirm whether the relevant data assets meet quality requirements such as completeness, uniqueness, stability, business consistency, and statistical availability before making credit decisions, approving loans, valuing assets, or granting access to joint modeling. If the data to be audited contains a large number of missing values, duplicate values, outliers, out-of-bounds values, illegal enumeration values, or inconsistent formatting records, it will lead to distorted data asset valuation, biased credit scoring, and a decline in the training effect of risk models, thereby seriously affecting the reliability of credit decisions and the healthy development of data asset financing.

[0003] Existing data asset pre-collateral audits primarily employ methods such as plaintext sampling, offline document submission, manual verification, or data holders generating their own quality reports. These methods face a core contradiction in cross-entity collaborative scenarios: asset providers need to expose raw detailed data or sample data to funding lenders or third-party auditing firms, which easily raises risks to data security, personal information protection, and trade secret leaks; simultaneously, funding lenders find it difficult to verify whether the locally generated quality reports by data providers are based on unified audit rules, unified field definitions, and unified calculation logic, resulting in insufficient credibility and verifiability of the audit results.

[0004] The aforementioned contradictions severely restrict the large-scale development of data asset-backed financing under the premise of privacy protection. There is an urgent need for a pre-audit method that allows the funding party to uniformly compile and encrypt audit rules and distribute them without the original asset details data leaving the local security boundary of the asset provider. The asset provider can then perform multi-dimensional quality verification in a local trusted execution environment or security sandbox, and only use the field-level aggregated statistical results as the data stream for the anonymized report. At the same time, there is also a need for a secure collaborative verification system that can combine the pre-audit results with subsequent joint modeling access control and cross-node front-end result display to ensure that a verifiable closed loop is formed between audit rules, execution boundaries, result feedback, result display, and process interception. Summary of the Invention

[0005] The purpose of this invention is to provide a data asset collateral pre-audit method and a secure collaborative verification system based on privacy computing, in order to solve the problems in the prior art where data asset collateral pre-audit relies on plaintext data exposure, audit rules are difficult to unify, data quality verification is difficult to adapt to the privacy collaborative computing environment, audit results are difficult to link with joint modeling access control, and cross-entity result display is difficult to verify consistently.

[0006] To achieve the above objectives, the technical solution provided by the present invention includes: The system of this invention includes a first data node and a second data node, which form a privacy-cooperative network through a secure communication channel. The first data node is the host of the funding credit provider, including a risk control orchestration module, a rule serialization and public key encryption module, a network communication module, a front-end gating judgment module, and a component metadata configuration module. The second data node is the host of the asset provider, including a decryption and deserialization module, a network communication module, a trusted execution environment or security sandbox, a result encapsulation module, a front-end container, a front-end visualization component, and a result management page. During the rule issuance phase, the risk control orchestration module of the first data node constructs a pre-audit rule set containing multiple dimensions of data quality based on the asset-backed risk control model. This audit rule set includes at least an extreme value boundary dictionary, a classification enumeration dictionary, a regular expression matching template, and verification parameters related to missing, duplicate, anomaly, range, enumeration, and consistency. The rule serialization and public key encryption module of the first data node serializes the audit rule set into a structured configuration string and encrypts it using the public key of the second data node to generate an audit configuration ciphertext package. The network communication module of the first data node sends the workflow scheduling instructions for the pre-audit task and the audit configuration ciphertext package to the second data node through a secure communication channel.

[0007] During the local mapping phase, the decryption and deserialization module of the second data node receives the audit configuration ciphertext packet, decrypts it using the local private key, and calls the deserialization function to parse out the structured audit rule set. The second data node mounts the original asset detail dataset within a locally allocated trusted execution environment or security sandbox. The underlying privacy computing engine declares the dataset as a distributed virtual data frame object. The logical native matrix mapping module of the second data node maps and loads the distributed virtual data frame object into a logical native matrix under the distributed computing framework, provided the data process does not leave the local security sandbox. This logical native matrix is ​​physically distributed across multiple working computing nodes within the local security sandbox. Each working node carries one or more row-oriented local shards, and each shard corresponds to an independent working unit in the MapReduce architecture.

[0008] During the statistical verification phase, the MapReduce concurrent computing engine of the second data node, targeting the j-th target feature column vector in the logical native matrix (where j is an integer from 1 to m), extracts the corresponding configuration parameters from the structured audit rule set and concurrently performs multi-dimensional data quality verification based on multiple worker nodes within the local security sandbox. This multi-dimensional data quality verification includes integrity and uniqueness verification, anomaly offset detection based on the Z-Score algorithm, anomaly extreme value detection based on the IQR algorithm, custom extreme value boundary verification, classification enumeration legality verification, and regular expression consistency verification.

[0009] The integrity and uniqueness verification is performed in the Map phase by having each worker node scan its local shard in parallel, calculating the number of local null values ​​and the set of local non-null independent values. In the Reduce phase, the aggregation node summarizes the local state values ​​of each worker node, calculating the global missing value ratio and the global duplicate value ratio. The global missing value ratio is the total number of null values ​​divided by the total number of samples, and the global duplicate value ratio is the cardinality of the complement of the set of non-null independent values ​​divided by the total number of samples.

[0010] The Z-Score-based outlier detection process executes three rounds of asynchronous Map-Reduce collaborative computation: In the first round, each worker node outputs a local scalar sum from the Map function, and the aggregation node outputs the global mean from the Reduce function; in the second round, each worker node calculates the local centered variance sum in parallel based on the global mean, and the aggregation node calculates the global standard deviation from the Reduce function; in the third round, each worker node extracts a threshold from the rule set based on the global mean and the global standard deviation, locally counts the number of outliers that meet the deviation conditions, and the aggregation node summarizes and outputs the global outlier ratio from the Reduce function.

[0011] The outlier detection based on the IQR algorithm calls a distributed approximate quantile algorithm to avoid the communication overhead caused by distributed full sorting: each worker node constructs a local data sketch in the Map phase, and the aggregation node constructs a global sketch distribution model through the merge operator in the Reduce phase, based on which the global lower quartile and upper quartile are approximately calculated, the interquartile range is calculated and the lower and upper limits of the system boundary are constructed, and the aggregation node summarizes the number of out-of-bounds elements and calculates the global outlier ratio.

[0012] The custom extreme value boundary verification extracts the business range for the feature column from the rule set. During the Map phase, each worker node calculates in parallel the number of failed records exceeding the range limit locally, and the aggregation node calculates the global failed record ratio. The classification enumeration legality verification extracts the set of valid enumeration items for the feature column from the rule set. During the Map phase, each worker node performs set subordination tests row by row, counting the number of locally illegal classification records, and the aggregation node calculates the global illegal classification record ratio. The regular expression consistency verification extracts the preset pattern matching string from the rule set. Each worker node calls the regular expression matching engine to concurrently perform local operations, counting the number of dirty data that does not conform to the format specification locally, and the aggregation node calculates the global dirty data ratio.

[0013] During the aggregation output phase, the second data node sequentially concatenates the statistical ratios calculated for the j-th target feature column vector to construct an evaluation component vector for a single column feature. After traversing all asset feature columns of the logical native matrix, the evaluation component vectors of each column are aggregated to generate a comprehensive audit result vector. The second data node triggers a memory snapshot destruction protocol within its local sandbox, erasing the logical native matrix and its derived detailed records, including temporary fragments, local statistical intermediates, and derived detailed records that could potentially restore the original asset details. The result encapsulation module of the second data node encapsulates the comprehensive audit result vector into a de-identified report data stream, which is then returned to the first data node via the secure communication channel.

[0014] During the pre-interception and collaborative verification phase, the pre-gating judgment module of the first data node receives the comprehensive audit result vector and performs pre-gating judgment based on the risk control tolerance threshold: if the assessment fails to meet the standard, entry into the subsequent joint modeling process is blocked; if the assessment meets the standard, entry into the subsequent joint modeling process is allowed. The component metadata configuration module of the first data node extracts the component metadata configuration file compiled and packaged by this node, which contains a list of operator front-end capabilities. The system injects the component metadata configuration file into the front-end container of the second data node through a cross-node hot update mechanism. This front-end container is independent of the underlying computing sandbox. The second data node uses the updated front-end visualization component to parse and render the comprehensive audit result vector, and displays a data quality inspection report on the result management page, realizing bidirectional collaborative verification between the first data node and the second data node.

[0015] Furthermore, the secure communication channel of this invention achieves physical and logical dual isolation based on the TLS / SSL protocol, employs an asymmetric encryption algorithm for key negotiation, and supports two-way authentication and data integrity verification. The trusted execution environment supports at least one of Intel SGX, ARM TrustZone, or AMDSEV. The MapReduce concurrent computing engine further includes aggregation nodes for summarizing the local state variables of each worker node and generating field-level statistical ratios; the front-end container is independent of the underlying computing sandbox, and the cross-node hot update mechanism updates the front-end display components without affecting the underlying computing tasks.

[0016] Beneficial effects: Compared with the prior art, the present invention has the following characteristics: The original asset details data are always kept in the local trusted execution environment or security sandbox of the second data node, while the first data node only receives the de-identified aggregation results, which reduces the risk of privacy leakage during the data asset collateral review process; Audit rules are uniformly constructed, serialized, and encrypted by the first data node, ensuring consistent rule and field definitions for cross-entity data quality audits. The local mapping of distributed virtual data frames to logical native matrices enables encrypted objects in the privacy-coordinated computing environment to enter the local Map-Reduce concurrent statistical process, improving the execution efficiency of large-scale asset data quality verification. The multi-dimensional verification of integrity, uniqueness, abnormal offset, abnormal extreme value, business scope, classification enumeration and format consistency constitutes the pre-audit indicator system for data asset collateral, which improves the ability of audit results to express asset availability and model adaptability. By destroying memory snapshots, sending back de-identified reports, implementing pre-gating judgment, and cross-node front-end hot updates, a secure closed loop is formed for data quality auditing, result display, and joint modeling access control, thereby improving the automation, verifiability, and security of data asset financing scenarios.

[0017] This invention is applicable to scenarios involving data asset financing, data asset valuation, data element circulation, joint risk modeling, and privacy-focused collaborative computing, where multiple stakeholders, including lenders, asset appraisers, data asset providers, or data holders, participate. The method of this invention can be deployed in privacy-focused computing systems that support multi-party secure computation, trusted execution environments, security sandboxes, or equivalent security isolation mechanisms. It can also be used as a pre-processing computational module for data quality assessment before data assets enter the subsequent joint modeling process. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of the data asset collateral pre-audit system of the present invention; Figure 2This is a schematic diagram of the process of performing local dimensionality reduction mapping and multidimensional quality verification of dense objects in the second data node according to the present invention. Figure 3 This is a schematic diagram illustrating the principle of joint modeling pre-interception and bidirectional collaborative verification of the present invention. Detailed Implementation

[0019] The following provides a detailed description of a data asset collateral pre-audit method and a secure collaborative verification system based on privacy computing, as provided in the embodiments of the present invention.

[0020] The embodiments of the present invention employ the following technical solution: a data asset collateral pre-audit method and a secure collaborative verification system based on privacy computing, the execution process of which includes the following steps: The first phase involves the initialization of the privacy computing network and the issuance of audit rules. This phase aims to establish a secure communication network between heterogeneous nodes and issue risk control rules. S1: The system initializes a privacy-focused collaborative network based on multi-party secure computation (MPC). The funding provider's host acts as the first data node. The asset provider's host acts as the second data node. Based on an asymmetric encryption algorithm, public and private key pairs are negotiated and generated for each node. Let the public key of the first data node be denoted as . The public key of the second data node is .like Figure 1 As shown in the overall interaction architecture, the first data node With the second data node They are physically and logically isolated from each other through a secure communication channel based on TLS / SSL.

[0021] S2: First Data Node The risk control orchestration module is based on an asset-backed model and constructs a system that includes... A set of pre-audit rules for data quality across multiple dimensions The set It includes an extreme value boundary dictionary, a category enumeration dictionary, and regular expression matching templates.

[0022] S3: First Data Node rule set The audit configuration string is serialized into JSON format and utilized using a second data node. public key Encrypt it to generate an audit configuration ciphertext package. The calculation formula is:

[0023] S4: First Data Node The network communication module will send the workflow scheduling instructions for the pre-audit task along with the configuration ciphertext. It is sent to the second data node via a secure communication channel. .

[0024] Figure 1 In the middle, the first data node The primary host for the lending party, and the secondary data node. The asset provider host and the data node are physically and logically isolated via a secure communication channel based on TLS / SSL. Generate an audit rule set through the risk control orchestration module. The audit configuration ciphertext package is generated by the rule serialization / public key encryption module. The network communication module transmits workflow scheduling instructions and Send to the second data node Second data node The set of structured audit rules is obtained by parsing through the decryption / deserialization module. The original asset detail dataset is then sequentially generated within a trusted execution environment or security sandbox. Distributed virtual data frames Logical native matrix MapReduce statistical verification results and comprehensive audit result vector Second data node The result encapsulation module will Encapsulate the data into a de-identified report data stream and return it to the first data node. First data node The pre-gating judgment module determines whether to proceed to the subsequent joint modeling process based on the risk control tolerance threshold. Figure 1 The dashed arrow in the middle indicates that the component metadata configuration file is generated by the first data node. Cross-node hot update to the second data node The front-end container.

[0025] The second stage is the dimensionality reduction mapping step for dense objects. This stage aims to transform the virtual data frame into a native computation matrix without leaving the domain of the original data. S5: Combined with appendix Figure 2 A schematic diagram of local dimensionality reduction for dense-state objects, second data node. Received encrypted packet Then, the internal decryption module uses the local private key. Decrypt it and call the deserialization function to parse out the structured set of audit rules. .

[0026] Figure 2In the middle, the second data node Received audit configuration encrypted package Then, the private key is used within a trusted execution environment or security sandbox. Decrypt and deserialize to obtain a set of structured audit rules. Meanwhile, the original asset details dataset Declared as a distributed virtual data frame within the local security boundary Then mapped to a logical native matrix For matrices The first in Target feature column vectors The system divides it into Part of the film Each of these is then assigned to a corresponding worker node for concurrent computation. Figure 2 The Map-Reduce concurrent computation area displays six types of verification branches, including integrity and uniqueness verification, Z-Score outlier detection, IQR outlier detection, custom extreme value boundary verification, classification enumeration validity verification, and regular expression consistency verification. Each branch outputs the results after the Map and Reduce phases. , , , , , and And spliced ​​together to form the first Column feature evaluation components:

[0027] Traverse all After the first feature column, the second data node Forming a comprehensive audit result vector And trigger a local memory snapshot destruction to erase the logical native matrix. And derived detailed records, and then the anonymized report data stream is returned to the first data node. . Figure 2 Solid arrows indicate data processing and statistical aggregation paths, dashed boundaries indicate trusted execution environments or security sandboxes, and dashed arrows indicate memory snapshot destruction paths.

[0028] S6: As Figure 2 As shown by the dashed boundary, the second data node Mount the original asset detail dataset within the locally allocated Trusted Execution Environment (TEE / Security Sandbox). The underlying privacy computing engine declares the dataset as a distributed virtual data frame (VDataFrame) object. .

[0029] S7: As Figure 2 As shown in the dimensionality reduction mapping logic block, under the condition that the data process does not leave the local security sandbox, the second data node Distributed Virtual DataFrame (VDataFrame) objects Mapped and loaded as a logical native matrix under a distributed computing framework ,in This represents the number of rows in the asset sample. The matrix represents the number of columns for asset characteristics. Physically, it is distributed across multiple working computation nodes within the local security sandbox; logically, it is divided into rows. Each region is divided into shards, and each shard corresponds to an independent work unit in the MapReduce architecture.

[0030] The third stage is a distributed, multidimensional mathematical and statistical verification step based on the MapReduce architecture. For example... Figure 2 The concurrent computation branch flow diagram at the bottom shows the logic native matrix. The first in Target feature column vectors ,in Considering the large-scale nature of financial-grade asset data, the system leverages the multi-task concurrency capabilities within a local security sandbox. Based on the row-oriented local partitioning in S7, each worker node directly reads the sub-vector of the corresponding column from the relevant partition as... Local data fragments , of which Each shard is carried by a specific worker node and contains There are 100 samples, satisfying the following total global sample size:

[0031] Second data node From the set of rules After extracting the corresponding configuration parameters, the following distributed verification is executed concurrently: S8: Integrity and uniqueness verification. During the Map phase, each worker node scans its local shard in parallel. Calculate the number of local null values With locally nonempty independent value sets ,in Indicator function:

[0032]

[0033] During the Reduce phase, the aggregation node summarizes the local state variables of each worker node and calculates the global missing value ratio for that feature column. The global duplicate value ratio is calculated using the set union operator. :

[0034]

[0035] S9: Anomaly offset detection based on the Z-Score algorithm. It extracts non-empty subsets from column vectors and performs multiple rounds of asynchronous MapReduce collaborative computation to obtain distributed statistical parameters.

[0036] Round 1 (Calculating the global mean): Each worker node's Map outputs a local scalar sum. The aggregation node's Reduce output is the global mean. And it was distributed, including:

[0037] Second round (calculating global standard deviation): Each working node is based on the global mean. Parallel computation of local centralized variance :

[0038] The aggregation node's Reduce calculation yields the global standard deviation. :

[0039] Third round (outlier identification): Each working node based on the issued... and Extract rule set Threshold in The number of outliers that meet the deviation criteria in local statistics. :

[0040] The aggregation node's Reduce function summarizes the global outlier ratio. :

[0041] S10: Outlier detection based on the IQR algorithm. To avoid the significant communication overhead of distributed full sorting, the system calls a distributed approximate quantile algorithm. Each worker node constructs a local data sketch (DataSketch) during the Map phase. Aggregation nodes use merge operators during the Reduce phase. Construct a global sketch distribution model Based on this, the global lower quartile was approximately calculated. and the upper quartile Calculate the interquartile range :

[0042] Extraction rule set Extreme value proportion threshold Construct the lower bound of the system boundary and upper limit Each working node counts the number of locally out-of-bounds elements in parallel. :

[0043] The final aggregation node's Reduce calculation yields the global outlier ratio. :

[0044] S11: Custom extreme value boundary verification. Extract the rule set. The business range for this feature column During the Map phase, each worker node calculates in parallel the number of failure records that locally exceed the limit of that interval. :

[0045] The aggregation node calculates the global failure record ratio during the Reduce phase. :

[0046] S12: Classification enumeration validity verification. Extract rule set. The set of valid enumeration items for this feature column. During the Map phase, each worker node performs a set dependency test row by row, counting the number of locally invalid classification records. :

[0047] The aggregation node calculates the global illegal classification record ratio during the Reduce phase. :

[0048] S13: Regular expression consistency check. Extract rule set. Preset pattern matching string Each worker node calls the regular expression matching engine to concurrently perform local operations, counting the number of dirty data that does not conform to the format specifications. ,in For pattern matching functions:

[0049] The aggregation node summarizes and calculates the global dirty data ratio during the Reduce phase. :

[0050] The fourth stage involves the destruction of memory snapshots and the aggregation of output steps: S14: Second Data Node In steps S8 to S13, the first Target feature column vectors The calculated statistical ratios are concatenated in order to construct the evaluation component vector of a single-column feature. :

[0051] Traversing the original logical matrix All After listing the asset characteristics, the evaluation component vectors of each column are aggregated to generate a comprehensive audit result vector. :

[0052] S15: As Figure 2 As shown in the feedback route, the second data node After extraction and construction Then, the memory snapshot destruction protocol within the local sandbox is triggered, erasing the logical native matrix. And its derived detailed records.

[0053] S16: Second Data Node Call the underlying data connection module to vectorize the comprehensive audit results. Encapsulated as a de-identified report data stream, such as Figure 1 The data is returned to the first data node across the secure communication channel. .

[0054] The fifth stage consists of joint modeling pre-interception and two-way collaborative verification steps: S17: First Data Node Receive the vector of comprehensive audit results Pre-gate control is performed based on the risk control tolerance threshold; if the assessment fails to meet the standard, the batch of asset data is blocked from entering the subsequent joint modeling process.

[0055] S18: First Data Node Extract the compiled and packaged component metadata configuration file for this node. This configuration file contains a list of operator front-end capabilities.

[0056] S19: Combined with appendix Figure 3The collaborative verification principle diagram shows that the system injects the metadata configuration file into the second data node through a cross-node hot update mechanism. Within the front-end container, which is independent of the underlying computing sandbox; the second data node Using the updated front-end visualization components, parse and render the comprehensive audit result vector. .

[0057] Figure 3 In the middle, the first data node Receive the vector of comprehensive audit results Then, the pre-gating judgment module performs a judgment based on the risk control tolerance threshold; when If the corresponding audit results fail to meet the standards, the batch of asset data will be blocked from entering the subsequent joint modeling process. Once the audit results meet the standards, entry into the subsequent joint modeling process will be permitted. First Data Node It also extracts the component metadata configuration file, which contains a list of operator front-end capabilities, and injects it into the second data node through a cross-node hot update mechanism. The front-end container. Second data node. The front-end container is independent of the underlying computing sandbox, and uses the updated front-end visualization components to parse the comprehensive audit result vector. The system then renders a data quality inspection report, which is displayed on the results management page. Figure 3 The solid arrows in the text represent the vector of comprehensive audit results. The usage path is indicated by the dashed arrow, which represents the cross-node hot update path of component metadata, and the dashed box represents the isolation boundary between the underlying computing sandbox and the front-end container.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for pre-audit of data asset collateral based on privacy computing, characterized in that, It includes the following five steps: Step 1. Initialize the privacy computing network and issue audit rules, establish a secure communication network between heterogeneous nodes and issue risk control rules; Step 2. Dimensionality reduction mapping of dense objects: under the condition that the original data does not leave the domain, the virtual data frame is transformed into a local native computation matrix; Step 3. Distributed multidimensional mathematical statistical verification based on MapReduce architecture; Step 4. Destroy and aggregate memory snapshots; Step 5. Joint modeling for pre-interception and two-way collaborative verification.

2. The data asset collateral pre-audit method based on privacy computing according to claim 1, characterized in that, Step one includes the following specific implementation steps: S1. Initialize a privacy-cooperative network based on multi-party secure computation, with the funding lender's host as the first data node and the asset provider's host as the second data node; negotiate and generate public and private key pairs for each node based on an asymmetric encryption algorithm; the first data node and the second data node are physically and logically isolated from each other through a secure communication channel; S2. The risk control orchestration module of the first data node is based on the asset mortgage model and constructs a set of pre-audit rules for data quality. This set of rules includes an extreme value boundary dictionary, a classification enumeration dictionary and a regular expression matching template. S3. The first data node serializes the rule set into an audit configuration string and encrypts it using the public key of the second data node to generate an audit configuration ciphertext package; S4. The network communication module of the first data node sends the workflow scheduling instruction of the pre-audit task, along with the audit configuration ciphertext packet, to the second data node via a secure communication channel.

3. The data asset collateral pre-audit method based on privacy computing according to claim 1, characterized in that, Step two includes the following specific implementation steps: S5. After the second data node receives the audit configuration ciphertext packet, the internal decryption module uses the local private key to decrypt it and calls the deserialization function to parse out the structured audit rule set; S6. The second data node mounts the original asset detail dataset within the locally allocated trusted execution environment, and the underlying privacy computing engine declares the original asset detail dataset as a distributed virtual data frame object; S7. Under the condition that the data process does not leave the local security sandbox, the second data node maps and loads the distributed virtual data frame object into a logical native matrix under the distributed computing framework. This logical native matrix is ​​physically distributed in multiple working computing nodes within the local security sandbox, and logically divided into multiple local partitions by rows. Each partition corresponds to an independent working unit in the Map-Reduce architecture.

4. The data asset collateral pre-audit method based on privacy computing according to claim 1, characterized in that, Step three includes the following specific implementation steps: S8. Integrity and uniqueness verification: In the mapping phase, each working node scans its local fragment in parallel and calculates the number of local null values ​​and the set of local non-null independent values; in the reduction phase, the aggregation node summarizes the local state variables of each working node, calculates the global missing value ratio, and calculates the global duplicate value ratio through the set union operator. S9. Anomaly Detection Based on Z-Score Algorithm: Extract non-empty subsets from column vectors and perform multiple rounds of MapReduce asynchronous collaborative computation to obtain distributed statistical parameters. Round 1: Calculate the global mean: Each worker node's Map outputs a local scalar; the aggregation node's Reduce outputs the global mean and distributes it to each worker node; Round 2: Calculate the global standard deviation: Each working node calculates the local centralized variance in parallel based on the global mean, and the aggregation node reduces the calculation to obtain the global standard deviation; Third round: Outlier identification: Each working node extracts the threshold from the rule set based on the issued mean and standard deviation, and locally counts the number of outliers that meet the deviation conditions; the aggregation node reduces and outputs the global outlier ratio. S10. Outlier Detection Based on IQR Algorithm: A distributed approximate quantile algorithm is invoked. Each worker node constructs a local data sketch during the Map phase; the aggregation node constructs a global sketch distribution model through merging operators during the Reduce phase, and approximates the global lower and upper quartiles based on this model, calculating the interquartile range; the extreme value ratio threshold is extracted from the rule set, constructing lower and upper bounds; each worker node counts the number of locally out-of-bounds elements in parallel; the aggregation node calculates the global outlier ratio during the Reduce phase. S11. Custom extreme value boundary verification: Extract the business range for this feature column from the rule set; each worker node calculates the number of failure records that locally exceed the limit of this range in parallel during the Map phase; the aggregation node calculates the global failure record ratio during the Reduce phase; S12. Classification Enumeration Legality Verification: Extract the set of legal enumeration items for this feature column from the rule set; each worker node performs set subordination tests row by row in the Map phase and counts the number of locally illegal classification records; the aggregation node calculates the global illegal classification record ratio in the Reduce phase; S13. Regular Expression Consistency Verification: Extract the preset pattern matching string from the rule set; each worker node calls the regular expression matching engine to concurrently execute local operations and count the number of dirty data that does not conform to the format specification; the aggregation node summarizes and calculates the global dirty data ratio in the Reduce phase.

5. The data asset collateral pre-audit method based on privacy computing according to claim 1, characterized in that, Step four includes the following specific implementation steps: S14. The second data node will sequentially concatenate the statistical ratios calculated for a certain target feature column vector to construct the evaluation component vector of a single column feature; after traversing all asset feature columns of the logical native matrix, the evaluation component vectors of each column will be aggregated to generate a comprehensive audit result vector. S15. After the second data node extracts and constructs the comprehensive audit result vector, it triggers the memory snapshot destruction protocol in the local sandbox to erase the logical native matrix and its derived detailed records. S16. The second data node calls the underlying data connection module to encapsulate the comprehensive audit result vector into a de-identified report data stream and return it to the first data node via a secure communication channel.

6. The data asset collateral pre-audit method based on privacy computing according to claim 1, characterized in that, Step five includes the following specific implementation steps: S17. The first data node receives the comprehensive audit result vector and performs a pre-gating judgment based on the risk control tolerance threshold; If the assessment fails to meet the standards, the process will be blocked from proceeding to the subsequent joint modeling stage. S18. The first data node extracts the compiled and packaged component metadata configuration file; S19. Through the cross-node hot update mechanism, the metadata configuration file is injected into the front-end container of the second data node. The second data node uses the updated front-end visualization component to parse and render the comprehensive audit result vector.

7. A data asset collateral pre-audit security collaborative verification system based on privacy computing, characterized in that, It includes a funding provider host as the first data node and an asset provider host as the second data node, and the first data node and the second data node form a privacy-cooperative network through a secure communication channel; The first data node is the host of the funding credit provider, which includes a risk control orchestration module, a rule serialization and public key encryption module, a network communication module, a front-end gating judgment module, and a component metadata configuration module; The second data node is the asset provider host, which includes a decryption and deserialization module, a trusted execution environment or security sandbox, a logical native matrix mapping module, a MapReduce concurrent computing engine, a result encapsulation module, a front-end container, a front-end visualization component, and a result management page.

8. The system according to claim 7, characterized in that, The secure communication channel is based on the TLS / SSL protocol to achieve both physical and logical isolation, uses an asymmetric encryption algorithm for key negotiation, and supports two-way authentication and transmission data integrity verification.

9. The system according to claim 7, characterized in that, The risk control orchestration module is used to construct a set of pre-audit rules for data quality based on the asset-backed risk control model. The set of audit rules includes an extreme value boundary dictionary, a classification enumeration dictionary, a regular expression matching template, and verification parameters related to missing, duplicate, anomaly, range, enumeration, and consistency.

10. The system according to claim 7, characterized in that, The pre-gating judgment module is used to receive the comprehensive audit result vector and perform pre-gating judgment based on the risk control tolerance threshold. If the assessment fails to meet the standards, the process will be blocked from proceeding to the subsequent joint modeling stage. If the assessment is satisfactory, the process will proceed to the next stage of joint modeling.

Images