A system and method for secure profiling and modeling of semi-structured data

Abstract

The application discloses a semi-structured data security representation and modeling system and method, and belongs to the technical field of computer data processing. The technical problem to be solved is that the existing semi-structured data large model modeling cannot simultaneously consider privacy security, transmission efficiency and modeling accuracy: direct feeding of full data has a risk of leakage, only transmission of structural information is insufficient in modeling accuracy, the privacy protection boundary of sampling desensitization is unquantifiable, and the modeling result lacks a verification and correction mechanism of zero data exposure. The technical solution points are that a "local calculation - security coding - model generation - local verification" architecture is adopted, data structure analysis, metadata extraction and layered statistical calculation are completed locally, and a double-modal security representation is generated through differential privacy disturbance and Token adaptive compression; a large model is based on noise perception prompt output to output an initial model; local minimum exposure sampling verification is used to assume and iteratively correct, and the original data is not out of the local at all.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing technology, specifically relating to a secure representation and modeling system and method for semi-structured data. Background Technology

[0002] In fields such as water conservancy, energy, finance, and government affairs, semi-structured data often contains sensitive content such as engineering parameters, operational data, and location information. The external transmission of raw data is subject to data security regulations. Currently, semi-structured data modeling based on large language models mainly includes three technical solutions: The first type is the full data direct input solution, which converts semi-structured data into text format and then inputs it into the large language model. This solution has two technical problems: First, directly transmitting the raw data to the external large language model service poses a risk of sensitive data leakage and cannot meet the requirements for secure data transmission; second, after converting large-scale semi-structured data into text, the number of tokens is large, which cannot be fully input due to the limitation of the context window length of the large language model, and the data transmission volume is also large.

[0003] The second type is the schema-only approach, which only provides table structure information such as column names and basic data types to the large language model, without transmitting the data content. This approach has two technical problems: First, when only table structure information is provided, statistical information such as data distribution, field relationships, and business semantics is lacking, resulting in low accuracy in primary and foreign key identification, field semantic inference, and business rule extraction; second, for semi-structured data containing multi-level headers, deeply nested structures, and cross-table references, relying solely on table structure information is insufficient for structure identification.

[0004] The third type is the sampling anonymization scheme, which involves randomly sampling and generalizing anonymization of the original data before inputting the sampled data into a large language model. This scheme has three technical problems: First, the randomly sampled data may still contain identifiable sensitive records, failing to provide a quantifiable privacy protection boundary; second, the random sampling process loses long-tail distribution features and low-frequency association patterns, resulting in low-frequency business rules being unrecognizable; and third, the sampling granularity is fixed, making it impossible to dynamically adjust the sampling range according to the information requirements of the modeling task, easily leading to information redundancy or missing key information.

[0005] Existing technical solutions fail to address the following technical issues: without transmitting the original data, they cannot simultaneously meet the information density requirements, token length limitations, and modeling accuracy requirements for large language modeling. Specifically, they cannot provide sufficient modeling information for large language models without transmitting the original data; they cannot maximize information density under token length limitations; they cannot eliminate the impact of privacy noise on the inference results of large language models after adding privacy noise to statistical information; they cannot automatically identify multi-level nested structures and implicit cross-table references in semi-structured data; and they cannot verify and correct the modeling results of large language models without transmitting the full amount of data.

[0006] In view of this, the present invention is hereby proposed. Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, this invention provides a secure representation and modeling system and method for semi-structured data. This addresses the issue that existing large-scale semi-structured data modeling cannot simultaneously balance privacy, transmission efficiency, and modeling accuracy: direct transmission of full data carries the risk of leakage; transmitting only structural information results in insufficient modeling accuracy; the privacy protection boundary of sampled desensitization cannot be quantified; and the modeling results lack a verification and correction mechanism with zero data exposure.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, a secure representation and modeling system for semi-structured data includes a local processor and a large language model server communicating with the local processor; the local processor includes: The data parsing and structure recognition layer is used to receive the semi-structured data to be processed, perform structure parsing and recognition on the semi-structured data, and obtain data structure information. The metadata and statistical calculation module is used to extract structural metadata from the data structure information and calculate the statistical information corresponding to each data field in a hierarchical manner. The differential privacy perturbation module is used to perform differential privacy perturbation on the statistical information to obtain noisy statistical information; An adaptive compression coding module is used to adaptively compress and encode the structural metadata and noise statistics according to a preset token budget, generate a bimodal security representation, and send it to the large language model server. The large language model server includes: The noise perception modeling module is used to receive the dual-modal security representation and the preset noise perception prompts, generate an initial data model and return it to the local processor; The local processor also includes: The hypothesis verification and feedback module is used to extract verifiable hypotheses from the initial data model, verify the verifiable hypotheses locally based on the minimum exposure sampling strategy, generate feedback information based on the verification results and send it to the large language model server, and drive the large language model server to iteratively correct and obtain the final data model.

[0009] Furthermore, the data parsing and structure recognition layer is specifically used for: For semi-structured data such as spreadsheets, traverse all worksheets and use the outward-to-inward shrinking method to locate the effective data area of ​​each worksheet, mark merged cells and parse the formula abstract syntax tree, and replace the cell coordinates in the formula with semantic placeholders. Multi-level headers are identified by combining rule-based filtering with a lightweight classifier. A header tree structure is constructed based on the inclusion relationship of merged regions, indentation level, and font size, and the header path sequence is output. Detect cross-worksheet references and named ranges, and construct a dependency graph between worksheets; For nested semi-structured data, identify the nesting level and generate a path sequence.

[0010] Furthermore, the metadata and statistical calculation module is specifically used for: The statistical information is calculated at four levels, specifically including: Layer L1 provides basic descriptive statistics, including mean, median, standard deviation, minimum, maximum, and quartiles for numerical fields; average length, cardinality of unique values, and high-frequency patterns for text fields; time span and time granularity for date fields; and number of categories and distribution ratio for categorical fields. The L2 layer is for distribution shape statistics, including skewness, kurtosis, bin boundaries, distribution fit type, entropy, and Gini coefficient; Layer L3 is for relational statistics, including inter-column correlation coefficient matrix, mutual information, functional dependency strength, and cross-table association degree; Layer L4 provides quality and privacy statistics, including null value rate, outlier rate, duplicate record rate, sensitive field identification results, and k-anonymity estimate. The scope of statistical calculations is controlled according to the sensitivity level of the field: confidential fields are calculated only for L1 level statistics, sensitive fields are calculated for L1 level and optional L2 level statistics, and internal and lower level fields are calculated for L1-L3 level statistics.

[0011] Furthermore, the differential privacy perturbation module is specifically used for: Calculate the corresponding sensitivity for different types of statistics: set the global sensitivity to 1 for count statistics, and use the smoothing sensitivity mechanism to calculate the local sensitivity for mean statistics; The noise scale is calculated based on the privacy budget parameter, and noise is sampled from the corresponding distribution and added to the statistics; Logical consistency is checked on the statistics after adding noise, and the noise value is adjusted for statistics that do not meet the monotonicity constraint using a constraint projection algorithm.

[0012] Furthermore, the differential privacy perturbation module adds noise to the quantile statistics using an exponential mechanism, specifically for: The negative of the absolute value of the difference between the target quantile rank and the candidate rank is used as the scoring function. The quantile values ​​after adding noise are obtained by sampling from the candidate quantile values ​​according to the probability distribution corresponding to the scoring function.

[0013] Furthermore, the adaptive compression coding module is specifically used for: A fixed number of tokens are reserved from the total token budget for system prompts and output format constraints, thus obtaining the available token budget. The available token budget is allocated to structural metadata, L1 layer statistics, semantic and relational metadata, and L2-L3 layer statistics according to a preset priority ratio, with L2-L3 layer statistics selected in order of information gain. Compact encoding is performed on each part of the allocated content, including column name semantic compression, statistical precision truncation, sparse matrix coordinate format storage, dynamic truncation of enumeration values ​​according to Token budget, and header path point encoding. Generate a bimodal security representation of MDS-Code in JSON format, including metadata version, privacy budget, token budget, structure information of each worksheet, noise statistics, inter-table relationship information, and global statistics.

[0014] Furthermore, the noise perception modeling module is specifically used for: Noise perception inference is performed based on noise perception cues: For statistics marked as noisy in the bimodal security representation, the corresponding confidence interval is calculated according to the noise scale, and interval estimation is used instead of point estimation; the confidence of the corresponding inference result is adjusted by a confidence attenuation function according to the ratio of the noise scale to the magnitude of the statistics. The output includes a structured initial data model containing a conceptual model, a logical model, business rules, and uncertainties. Each inference result is accompanied by a confidence score.

[0015] Furthermore, the hypothesis verification and feedback module is specifically used for: Extract different types of verifiable hypotheses from the initial data model, including constraint hypotheses, relational hypotheses, and rule hypotheses, and calculate the minimum sample size based on the hypothesis type and confidence requirements; For different types of hypotheses, corresponding targeted sampling strategies are adopted: constraint hypotheses use outlier priority sampling, relation hypotheses use hash bucketing collision detection sampling, and rule hypotheses use residual sorting sampling. Validate the sampled data locally to obtain validation results and anomaly pattern summaries.

[0016] Furthermore, the hypothesis verification and feedback module is also used for: The validation results are encoded into structured feedback information containing the hypothesis content, validation results, pass rate, abnormal evidence, and correction suggestions, and sent to the large language model server. The iteration termination is controlled based on the verification pass rate and the number of iteration rounds: when the verification pass rate of all verifiable hypotheses is greater than a preset threshold or the maximum number of iteration rounds is reached, the iteration stops and the final data model is obtained. After each round of verification is completed, the sample data obtained from sampling is destroyed immediately, and only the verification result statistics are retained.

[0017] Furthermore, the data parsing and structure recognition layer is also used for primary-foreign key relationship inference, specifically for: Calculate the unique value rate and null value rate of each column, mark the column with a unique value rate of 1 and a null value rate of 0 as candidate primary keys, and confirm the primary key by combining the column name semantic matching results; Calculate the value overlap rate and column name semantic similarity across columns, and combine them with cross-table reference frequency to mark column pairs with both value overlap rate and name similarity greater than a preset threshold as candidate foreign keys.

[0018] Secondly, a secure representation and modeling method for semi-structured data includes: S1. The local processor receives the semi-structured data to be processed, performs structure parsing and recognition on the semi-structured data, and obtains data structure information; S2. The local processor extracts structural metadata from the parsed semi-structured data and calculates the statistical information corresponding to each data field in a hierarchical manner. S3. The local processor performs differential privacy perturbation on the calculated statistical information to obtain noisy statistical information; S4. The local processor performs adaptive compression encoding on the structural metadata and noise statistics information according to the preset Token budget, generates a bimodal security representation, and sends it to the large language model server. S5. The large language model server receives the bimodal security representation and the preset noise perception prompt, generates the initial data model and returns it to the local processor; S6. The local processor extracts verifiable hypotheses from the initial data model, verifies the verifiable hypotheses locally based on the minimum exposure sampling strategy, generates feedback information based on the verification results and sends it to the large language model server, driving the large language model server to iteratively correct and obtain the final data model.

[0019] The present invention has at least the following beneficial effects: (1) Ensure data privacy and security: The original semi-structured data is processed only on the local processor side throughout the process, and only aggregated statistical information protected by differential privacy is transmitted. This provides a quantifiable and provable privacy protection boundary, meets the compliance requirements of zero outbound sensitive data, and avoids the risk of privacy leakage caused by direct transmission of full data. (2) Improve token utilization efficiency: Use a dual-modal representation that combines structured metadata with hierarchical statistical information to replace the original data. Combined with the token adaptive compression coding strategy, maximize information density within a limited token budget, significantly reduce data transmission volume and inference cost, and break through the length limit of the context window of a large language model. (3) Ensure modeling accuracy: supplement the missing business semantics and data distribution characteristics of the table structure by hierarchical statistical information, combine the noise perception mechanism to calibrate the inference bias caused by differential privacy noise, and then pass the local minimum exposure sampling verification and iterative correction closed loop to ensure modeling accuracy under the premise of zero original data out of the domain, and solve the problem of poor modeling effect caused by only transmitting schema information. (4) Adapt to complex semi-structured scenarios: Supports automatic recognition and structured representation of non-standard semi-structured structures such as multiple worksheets, merged cells, multi-level headers, cross-sheet formula references, and deep nesting, solving the shortcomings of existing solutions in handling complex semi-structured data, and is applicable to a wider range of scenarios; (5) Enhance the interpretability of results: Each inference output by the large language model is accompanied by a confidence score and uncertainty description, clearly marking the risk level of high-noise fields, making it easier for users to judge the credibility of the results and providing a clear decision basis for subsequent data governance. Attached Figure Description

[0020] Figure 1 This is an architecture diagram of a secure representation and modeling system for semi-structured data provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0023] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0024] Example 1 See Figure 1 , Figure 1 This is an architecture diagram of a secure representation and modeling system for semi-structured data proposed in this invention. It executes structured parsing, statistical computation, differential privacy perturbation, token adaptive compression, and bimodal fusion encoding on a local processor. Without recording the original data locally, it transmits a compact and complete bimodal representation to a large language model, enabling the large model to generate a high-accuracy data model. Furthermore, it automatically corrects model biases through local hypothesis testing closure, thus simultaneously meeting the requirements of data privacy protection, token efficiency, and modeling accuracy. Specifically, it may include: M1, the local processor, specifically includes: M11, the data parsing and structure recognition layer, is used to receive the semi-structured data to be processed, perform structure parsing and recognition on the semi-structured data, and obtain data structure information; This layer, deployed on the local processor, is used to perform structure parsing and feature recognition of multi-format semi-structured data, providing basic structural information for subsequent metadata extraction and statistical calculations. This layer supports parsing common semi-structured data formats such as Excel (.xlsx, .xls format), CSV (comma-separated, tab-separated), JSON (containing nested objects and arrays), and XML. Corresponding parsing strategies are adopted for different data formats: for spreadsheet-like data, an outward-to-inward shrinking method is used to locate the effective data area and mark merged cells; for JSON-like data, the nesting level is automatically flattened and the path expressions of each field are recorded; for CSV-like data, the delimiter, quotation mark rules, and header row position are automatically inferred.

[0025] The data parsing for Excel format documents includes the following sub-algorithms: The M111 multi-sheet (sheet) valid area detection algorithm, executed by the local processor, includes the following steps: M1111: Traverse all worksheets in the workbook, including hidden worksheets and chart worksheets, and identify the valid data boundaries of each worksheet one by one. M1112. Use the outside-to-inward contraction method for boundary positioning: Scan from the four corners of the worksheet towards the center to locate the position of the first non-empty cell, and determine the coordinate range of the valid data area accordingly. ,in These are the minimum and maximum row numbers of the valid region, respectively. These are the minimum and maximum column numbers of the valid region, respectively; M1113. Mark the merged cells within the valid area and record the coordinate range and merge type of each merged area. The merge types include three categories: header merge, data merge, and format merge. M1114. Extract formula information from the worksheet, parse the abstract syntax tree (AST) of the formula, retain the structural information of the formula, and replace the specific cell coordinates in the formula with the corresponding semantic placeholders.

[0026] The M112 multi-level header recognition algorithm, executed by the local processor, includes the following steps: M1121, Rule-based filtering: Detects cell font boldness, background color fill characteristics, border style, position characteristics (whether it is in the first row / first column), and data type abruptness characteristics with the cells below to filter and obtain candidate header cells; M1122, Model Layer Confirmation: For the selected candidate header cells, extract feature vectors [font weight, background color entropy, position index, type consistency below, merge marker], and input the feature vectors into the lightweight gradient boosting tree ( The classifier outputs the probability value that the cell is a header. ; M1123. Hierarchical Tree Construction: For cells identified as header cells, construct the header tree structure based on the inclusion relationship of merged areas, cell indentation level, and font size differences. , where the set of nodes Each node in For a given header cell, the edge set Each edge in This indicates the hierarchical relationship between header cells; M1124. Output header path sequence: Generate the header path sequence corresponding to each field based on the header tree structure, which is used for subsequent semantic enhancement processing.

[0027] The M113 implicit relation detection algorithm, executed on the local processor, includes the following steps: M1131, Cross-Sheet Reference Detection: Parses cross-worksheet reference nodes in the formula abstract syntax tree and constructs an inter-worksheet dependency graph. , where the set of nodes For all worksheets, edge set This refers to the reference relationships between worksheets; M1132, Named Range Parsing: Extracts named ranges at the workbook and worksheet levels in Excel, and records the name, reference range, and scope of each named range.

[0028] M114, the primary key and foreign key inference algorithm, is executed by the local processor and consists of two parts: primary key inference and foreign key inference. M1141, Primary Key Inference includes: Calculate the unique value rate for each column The specific expression is: ; If the unique value rate of a column And the null value rate If so, then mark that column as a candidate primary key; Perform semantic verification on candidate primary keys: if the column name matches the preset regular expression pattern This will increase the primary key confidence of that column.

[0029] M1142, Foreign Key Inference includes: Calculate the overlap rate of values ​​across column pairs in a worksheet The specific expression is:

[0030] in, These are the sets of duplicate values ​​for the two columns, respectively. The name similarity S between two columns is calculated using the following expression:

[0031] Embedding is generated through a pre-trained business terminology vector model; If the value overlap rate And name similarity If so, then mark the column pair as a candidate foreign key; Verify by considering cross-worksheet reference frequency: If the worksheet There are many formula references in the worksheet. If a certain column is selected, the confidence level of the foreign key determination for that column pair is strengthened.

[0032] The M115 hierarchical / nested structure recognition algorithm is executed on a local processor. It uses a rule-based filtering system (indentation features, bracket matching features, keyword position features) combined with a lightweight classifier to confirm the structure and construct a tree structure for the corresponding data. For JSON objects... Generate the path sequence corresponding to each field. For spreadsheets with multi-level headers, a header tree is constructed to output the path sequence for subsequent semantic enhancement processing.

[0033] M12, the metadata and statistical calculation module, is used to extract structural metadata from the parsed semi-structured data and calculate the statistical information corresponding to each data field in a hierarchical manner. M121. This module is deployed on the local processor. Based on the output of the data parsing and structure recognition layer, it extracts five types of structured metadata from semi-structured data, providing basic metadata support for subsequent statistical information calculations and secure compression encoding. The extracted metadata categories and their corresponding contents are as follows: M1211, Basic Structure Metadata: Includes information such as data source type, table / collection name, field list, estimated record count range, structure tree depth, valid region coordinates, and header tree path; Data structure example: .

[0034] M1212, Data Type Metadata: Includes the inferred data type of each field, specifically including numeric, text, date, boolean, fractional, currency, percentage, object, and array types; Data structure example: .

[0035] M1213, Constraint Metadata: Includes field null constraint, uniqueness flag, value range, enumerated value list (taking the top K most frequent values), and other constraint information; Data structure example: .

[0036] M1214. Semantic Metadata: Includes standard business semantic tags, field unit information, and time granularity information obtained through matching from a business terminology mapping library; Data structure example: .

[0037] M1215, Relational Metadata: Includes cross-table / cross-file relationships, nested paths, inter-column dependencies, primary and foreign key relationships, formula / reference dependency graphs, named region mappings, and other relationship information; Data Structure Example: .

[0038] M122. This module uses semantic enhancement technology to achieve semantic tag matching and unit identification of fields, specifically including the following steps: M1221. Construct a business terminology mapping library. The storage structure of the mapping library is as follows: That is, each standard label corresponds to a set of synonyms, such as a thesaurus: Corresponding standard label ; M1222, Column Name Vectorization Similarity Calculation: Using a pre-trained domain word vector model, column names are mapped to 128-dimensional vectors. The cosine similarity between the column name vector and each standard term vector in the mapping library is calculated, and the top-1 standard label with the highest similarity is taken as the semantic label of the field. M1223, Unit Recognition for Numeric Fields: Matches unit patterns in field values ​​using a pre-defined regular expression library, for example... g, combined with the semantic tags of the field, confirms the unit of measurement for that field.

[0039] M13, Differential Privacy Perturbation Module, is used to perform differential privacy perturbation on the calculated statistical information to obtain noisy statistical information; This module is deployed on the local processor. Based on the output of the metadata and statistical computing module, it calculates statistical information in four levels according to the gradient of information density and computing cost. At the same time, it performs calculation optimization in combination with the field sensitivity level, so as to control the consumption of local computing resources and the degree of exposure of statistical information while ensuring the validity of the information required for modeling.

[0040] Statistical information is categorized into the following four levels: M131, L1 Basic Descriptive Statistics (Required Layer): These are basic statistical items that are mandatory for calculation across all fields, and they calculate corresponding basic indicators for fields of different data types. Numeric fields: Calculate mean, median, standard deviation, minimum, maximum, and quartiles; Text fields: Calculate average length, cardinality of unique values, and high-frequency patterns; Date field: Calculate time span and time granularity; Categorization field: Calculates the number of categories and the distribution ratio of each category.

[0041] M132 and L2 distribution morphology statistics (optional layer): optional statistical items dynamically triggered by the Token budget, including skewness, kurtosis, bin boundaries, distribution fit type (normal distribution, log-normal distribution, power-law distribution), entropy, and Gini coefficient.

[0042] M133 and L3 correlation statistics (optional layer): optional statistical items that are dynamically triggered by the complexity of data relationships, including the inter-column correlation coefficient matrix (stored in a sparse manner), mutual information, functional dependency strength, and cross-table correlation degree.

[0043] M134 and L4 Quality and Privacy Statistics (Security Layer): Statistical items used for data quality assessment and privacy risk assessment, including null value rate, outlier rate, duplicate record rate; sensitive field identification results (identifying sensitive fields such as ID card number, mobile phone number, and bank card number through regular expression matching); k-anonymity estimate.

[0044] The sensitivity-level driven computation optimization mechanism includes: Before performing L2 and L3 level statistical calculations, this module first determines the calculation range based on the sensitivity level of the field. If the field is classified as confidential, the statistical calculations at the L2 and L3 levels are skipped, and only the L1 mandatory layer statistics are performed to reduce the consumption of computing resources and the exposure of sensitive information. If the field is sensitive, L2 level statistics will be selectively calculated based on the current token budget, and L3 level statistics will not be executed. If the field is at the internal level or below, the L1-L3 level statistics can be fully calculated based on the budget and relational complexity.

[0045] M14, an adaptive compression coding module, is used to adaptively compress and encode the structural metadata and noise statistics according to a preset Token budget, generate a bimodal security representation, and send it to the large language model server. This module, deployed on the local processor, is the core privacy protection and transmission optimization module. On one hand, it performs differential privacy perturbations on the statistics output by the statistical information calculation module, providing a provable and quantifiable privacy protection boundary. On the other hand, it performs adaptive compression encoding on the structural metadata and noisy statistical information according to a preset token budget, generating a compact bimodal secure representation that maximizes the information density of the representation within a limited token budget. Specifically, it includes: M141, the differential privacy statistical perturbation algorithm, is executed on the local processor. It applies corresponding noise addition methods to different types of statistics while ensuring the logical consistency of the statistics after noise addition. Specifically, it includes the following steps: M1411. Sensitivity Calculation: Calculate the corresponding sensitivity for different types of statistics. For count-type statistics (such as the number of rows or the number of unique values), apply global sensitivity. ; For mean-like statistics, smoothing sensitivity is used. Mechanism for calculating locality sensitivity: After sorting the data, calculate the sensitivity for deleting the [number]th [item]. The difference between the mean of the dataset after each set of data and the mean of the original dataset. The smoothing sensitivity S is calculated using the following formula:

[0046] in, The smoothing parameter has a default value of 0.1; noise scale. ,in Privacy budget parameters (user-configurable, default) ).

[0047] M1412, Noise injection: Sample noise from the corresponding distribution and add it to the statistics. For count and mean statistics, from the Laplace distribution Medium sampling noise Adding the original statistic yields a noisy result, such as the mean after adding noise. ; For quantile statistics, an exponential mechanism is used instead of direct noise addition: the negative of the absolute value of the difference between the target quantile rank and the candidate rank is used as the scoring function. The noisy quantile values ​​are obtained by sampling from the candidate quantile values ​​according to the probability distribution corresponding to the scoring function, satisfying the following conditions: Differential privacy requirements.

[0048] M1413, Noise Consistency Check: After injecting noise, perform a logical consistency check on the noise-added statistics, such as checking the monotonicity of quantiles. If the verification fails, the noise value is adjusted using the constraint projection algorithm to make the noise statistics satisfy the monotonicity constraint.

[0049] M142. Tiered Anonymization Strategy: This module performs tiered anonymization based on the sensitivity level of the fields, matching different privacy budgets and statistical information disclosure scopes. See Table 1 for details: Table 1

[0050] The M143 Token Adaptive Compression Coding Algorithm (MDS-Code) is executed on the local processor. It dynamically allocates the encoding length according to the Token budget and uses a multi-strategy compression to generate a structured bimodal representation. Specifically, it includes the following steps: M1431, Token Budget Allocation: Receives the total token budget configured by the user. Alternatively, the total budget can be automatically calculated based on the context window of the large language model; Reserve a fixed number of tokens Used for system prompts and output format constraints to obtain the available token budget. ; Allocate available budget according to preset priority ratios: Structure metadata (basic structure + data types + constraints): allocation ; L1 layer statistics: allocation ; Semantic and relational metadata: allocation ; L2-L3 layer statistics: Allocation of remaining resources If the budget is insufficient, select statistical items by sorting them by information gain.

[0051] M1432, compact coding, performs multi-strategy compression coding on each part of the allocated content: Column name semantic encoding: Replace long column names with corresponding phrase semantic tags through a business term mapping library; Precision truncation of statistics: Numerical statistics should retain 3 significant figures, and the date format should be YYYY-MM-DD. Sparse matrix compression: The correlation coefficient matrix is ​​stored in COO coordinate format, retaining only the sparse matrix. Triplet; Enumeration value truncation: Categorical variables retain only the top K most frequent enumeration values. The value of K is dynamically calculated based on the token budget allocated to the enumeration value and the average number of tokens per value. Header path compression: Encodes the header tree path into a dotted string format.

[0052] Generates M1433 and MDS-Code formats, producing a JSON-structured bimodal security representation, including fields such as metadata version, privacy budget, token budget, worksheet structure information, noise statistics, inter-table relationships, and global statistics; the specific code expression is as follows:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] M15, Hypothesis Validation and Feedback Module, is used to extract verifiable hypotheses from the initial data model, validate the verifiable hypotheses locally based on the minimum exposure sampling strategy, generate feedback information based on the validation results and send it to the large language model server, and drive the large language model server to iteratively correct and obtain the final data model. This layer, deployed on the local processor, is the closed-loop validation module of the modeling process. It is responsible for extracting verifiable hypotheses from the initial data model output by the large language model, performing validation locally using a minimum exposure sampling strategy, and feeding the structured encoded validation results back to the large language model for iterative correction. This improves the final modeling accuracy without exposing the entire original data. Specifically, it includes: M151. Hypothesis Extraction and Encoding: This involves parsing the business rules and constraint information output from the large language model using a local processor to extract a set of verifiable hypotheses. Each hypothesis contains the following information: Assumption types: constraint type (such as non-empty, unique, range), relation type (such as functional dependency, correlation), rule type (such as formula relation); Validate the required data characteristics: For example, to verify the functional dependency of X→Y, it is necessary to check whether Y is consistent under the same value of X; Minimum sample size estimation: Based on the hypothesis type and confidence level requirements, the minimum sample size is calculated through statistical power analysis. .

[0098] M152, the minimum exposure sampling algorithm, executes the following steps via the local processor: M1521. Targeted sampling: Employing corresponding sampling strategies for different hypothesis types. For constraint-type assumptions (such as amount > 0): outlier priority sampling is adopted: locate records that may violate the constraints from the L1 statistics (such as outliers with amount ≤ 0) and sample these records first; For relational assumptions (such as customer ID being the primary key): use conflict detection sampling: use hash bucketing (such as hash(customer ID)%1000), extract one record from each bucket, and check for conflicts within the bucket; For rule-based assumptions (such as Amount = Quantity × Unit Price): Residual sorting sampling is used: calculate the residual res = |Amount - Quantity × Unit Price| for each record, and extract the Top-K records in descending order of residuals.

[0099] M1522. Minimum sample size calculation and proportion estimation (e.g., outlier proportion), the specific formula is as follows:

[0100] in, The quantiles corresponding to the confidence level (1.96 for 95%). To estimate the proportion, Allowable error; The specific formula for mean estimation is as follows:

[0101] σ is obtained from L1 statistics.

[0102] M1523, Verification Execution and Feedback Coding: Performs verification on the sampled data locally and generates a Boolean result. ; Calculate validation statistics, such as validation pass rate. Summary of abnormal patterns (e.g., "3 records deviated due to a discount of 0.8"); M153. Encode the verification results into structured feedback, the specific expression of which is:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] M154, Iterative correction closed loop, the iterative correction process is executed through the local processor: M1541. Attach the structured feedback information to the context of the next round of dialogue and send it to the large language model server; M1542, the large language model corrects the data model output based on feedback information and returns the corrected model; M1543, Determine the iteration termination condition: When the verification pass rate of all verifiable hypotheses is greater than a preset threshold ( When the maximum number of iterations is reached (default 3 rounds), the iteration stops and the final data model is output. M1544. After each round of verification is completed, the sample data obtained from this sampling should be destroyed immediately, and only the verification result statistics should be retained.

[0109] M2, Large Language Model Server: This module is deployed on the large language model server, receiving bimodal security representations and noise perception cues transmitted from the local processor. Based on the noise perception mechanism, it performs data modeling and outputs a structured initial data model. Its core function is to enable the large language model to proactively perceive differential privacy noise in statistical information, calibrate inference confidence, and reduce the impact of noise on modeling accuracy. Specifically, it includes: M21, a noise perception modeling module, is used to receive the bimodal security representation and preset noise perception prompts, generate an initial data model, and return it to the local processor; specifically including: M211, Noise Perception Cueing Project: A dedicated system cues are built by the local processor and input into the large language model along with the bimodal representation. The cues include noise annotations, inference correction rules, task requirements, and output format constraints. Noise annotation explanation: This indicates that some statistics in the representation have been injected with differential privacy noise, and this is annotated. The field is the noise field, and the corresponding noise scale is given; Inference correction rule: requires that interval estimation be used instead of point estimation for noisy statistics, and the confidence level of the corresponding inference be reduced according to the relative magnitude of the noise; Task requirements: The modeling tasks should be clearly defined, including business entity and attribute identification, logical model design, primary and foreign key relationship and association rule identification, data quality suggestion generation, and potential business rule inference. Output format constraints: Results must be output strictly according to the specified JSON structure, and each inference result must be accompanied by a confidence score and an explanation of uncertainty.

[0110] M212, Noise-Aware Modeling Algorithm: After receiving the bimodal representation, the large language model performs the following noise-aware inference steps: M2121, Interval Estimation Alternative to Point Estimation: For statistics marked as noisy, calculate the corresponding 95% confidence interval based on the noise scale b, where the width of the 95% confidence interval for Laplace noise is approximately... Inferences are made based on confidence intervals rather than directly using noisy point estimates; M2122, Confidence Attenuation Calibration: Calculating Noise Scale With the magnitude of the statistical measure ratio Through the confidence decay function Adjust the base confidence level of the corresponding inference result; the larger the relative magnitude of the noise, the lower the confidence level of the inference. M2123 Uncertainty Labeling: For high-noise fields (noise scale greater than 10% of the statistical mean), explicitly label the uncertainty level in the output and suggest that the local system perform supplementary verification to trigger subsequent verification loop.

[0111] M213, the large language model is forced to output a structured data model in JSON format. The structured output specification forces the large model to output the following JSON structure:

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] Example 2 This invention proposes a secure representation and modeling method for semi-structured data, specifically including: S1. The local processor receives the semi-structured data to be processed, performs structure parsing and recognition on the semi-structured data, and obtains data structure information; S2. The local processor extracts structural metadata from the parsed semi-structured data and calculates the statistical information corresponding to each data field in a hierarchical manner. S3. The local processor performs differential privacy perturbation on the calculated statistical information to obtain noisy statistical information; S4. The local processor performs adaptive compression encoding on the structural metadata and noise statistics information according to the preset Token budget, generates a bimodal security representation, and sends it to the large language model server. S5. The large language model server receives the bimodal security representation and the preset noise perception prompt, generates the initial data model and returns it to the local processor; S6. The local processor extracts verifiable hypotheses from the initial data model, verifies the verifiable hypotheses locally based on the minimum exposure sampling strategy, generates feedback information based on the verification results and sends it to the large language model server, driving the large language model server to iteratively correct and obtain the final data model.

[0122] Example 3 This embodiment uses a semi-structured data security representation and modeling system and method proposed in this invention, taking the security representation and modeling of hydropower engineering geological survey data as the application scenario. The original data is geological survey data in Excel format containing multiple worksheets, totaling about 20,000 rows, containing sensitive information such as dam site coordinates, borehole parameters, and geotechnical mechanics parameters. The original data is prohibited from being transmitted out of the intranet environment.

[0123] The Excel data in this example contains three worksheets, and the information for each worksheet is as follows: Sheet1: Basic Drilling Information The data includes the following fields: borehole ID (primary key), latitude and longitude coordinates (GCJ-02 encrypted coordinates), borehole elevation (unit: m), borehole depth (unit: m), groundwater level depth (unit: m), and exploration date; a total of 500 rows of data; among which the coordinate and elevation fields are classified as confidential.

[0124] Sheet2: Soil and Rock Layering and Physical and Mechanical Parameters The data includes the following fields: borehole ID (foreign key), layer number, soil / rock name, top layer depth (m), bottom layer depth (m), natural density (g / cm³), saturated uniaxial compressive strength (MPa), shear strength (friction angle φ / cohesion c), and deformation modulus (GPa). It contains approximately 15,000 rows of data (each borehole corresponds to multiple layers). All parameter fields are classified as sensitive, with a privacy budget ε=0.5.

[0125] Sheet3: Field Test Record (Heavy Dynamic Penetration Test N63.5) The data includes the following fields: borehole ID, test depth (in meters), actual number of hammer blows, and corrected number of hammer blows; approximately 4,500 rows of data in total; it is an internal data set with a privacy budget of ε=2.0.

[0126] This embodiment is executed through the cooperation of a local processor and a large language model server. The specific steps are as follows: B1. After receiving the Excel data, the local processor performs structure parsing: Identify valid data ranges in the three worksheets; Detecting cross-worksheet references: The drill ID columns of Sheet2 and Sheet3 form an implicit foreign key relationship with the drill ID column of Sheet1. The calculated overlap rate is 100% and the column name similarity is 1.0. Based on semantic inference, it is automatically marked as a foreign key relationship. Identify nested structures: In Sheet2, the soil and rock stratification is a one-to-many nested structure (one borehole corresponds to multiple stratifications). Construct an implicit nested relationship from borehole ID to the stratification list.

[0127] B2. The local processor extracts five types of metadata from the parsed data: Semantic metadata: The unit for saturated uniaxial compressive strength is MPa, the unit for natural density is g / cm³, and the semantic tags for the coordinate field are domestic encrypted coordinates; Constraint metadata: Identify field constraints where the top layer depth is less than the bottom layer depth and the compressive strength is a positive value.

[0128] B3. Local processors calculate statistical information based on sensitivity levels: Statistics for the L1 required layer: the mean saturated uniaxial compressive strength is 45.3 MPa, the median is 38.2 MPa, the standard deviation is 22.1 MPa, the minimum is 5.2 MPa, the maximum is 120.5 MPa, and the void rate is 0%; the mean borehole depth is 85.6 m. L2 selection layer statistics: The compressive strength is a sensitive level, and the distribution pattern statistics are allowed to be calculated. The skewness is 1.8 and the kurtosis is 4.2, which conforms to the log-normal distribution characteristics. L4 security layer statistics: latitude, longitude, elevation, compressive strength, etc. are marked as sensitive fields; the k-anonymity of latitude and longitude fields is low. For confidential latitude, longitude and elevation fields, L2 and L3 level statistical calculations are skipped, and only the basic count and null value rate of L1 layer are retained.

[0129] B4. Local processor performs hierarchical desensitization and adaptive token compression: B41. Graded Desensitization Treatment The drill ID field is set to public level, privacy budget. Fully disclose the base number of the fields; The compressive strength field is classified as sensitive and subject to privacy budget. Laplace noise is injected into the mean, and the noise scale is calculated by smoothing sensitivity and privacy budget. The original mean is 45.3 MPa, and the noise is added to 45.5 MPa. Only the coarse-grained statistics of the L1 layer (mean, median, standard deviation, quartiles) are disclosed, and the minimum, maximum and L2-L3 layer statistics are not disclosed. The borehole depth field is an internal level field, with a privacy budget. The statistics for L1-L2 layers are disclosed, while only the correlation direction is retained for L3 layer. The latitude and longitude fields are classified as confidential, privacy budget. No numerical statistics are disclosed; only data types and encrypted coordinate semantic tags are disclosed.

[0130] B42, Token Adaptive Compression The total token budget is set at 5000, with 500 tokens reserved for system prompts and output format constraints, leaving 4500 tokens available. Since sensitive and confidential fields do not disclose L2-L3 layer statistics, 10% of the budget originally allocated to L2-L3 layers (approximately 450 tokens) has been reallocated to L1 layer statistics and relational metadata. The final allocation is as follows: structural metadata 2000 tokens (approximately 44.4%), noisy L1 layer statistics 2200 tokens (approximately 48.9%), and relational and semantic metadata 800 tokens (approximately 17.8%).

[0131] The compression process includes: retaining three significant figures for compressive strength statistics; and compressing the long column name "Saturated Uniaxial Compressive Strength" into an industry-standard abbreviation. Finally, a JSON-formatted MDS-Code bimodal security representation is generated, with an example fragment as follows:

[0132]

[0133]

[0134]

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] B5. Noise-Aware Modeling for Large Language Models: The server of the large language model receives the MDS-Code bimodal representation and noise-aware prompts, and performs noise-aware modeling. Noise perception inference: The large model identifies noise in the mean compressive strength. Based on the noise scale, the 95% confidence interval is approximately 45.5 ± 3.5 MPa. Interval estimation is used when inferring the range of rock strength. Output: Generates a list containing... The logical model of the three data tables has a clear primary and foreign key relationship; it identifies an empirical formula relationship between the N63.5 hammer blow count and the degree of rock weathering and deformation modulus, with a corresponding confidence level of 0.70; the output uncertainty statement suggests that the statistical noise of compressive strength should be added locally to verify its empirical relationship with deformation modulus, and also indicates that the lack of statistical information on borehole coordinates makes spatial autocorrelation analysis impossible.

[0144] B6. The local processor performs closed-loop verification and iterative correction, specifically including: B61. Hypothesis Extraction: Extract constraint-type hypothesis from the output of the large model: "When the soil and rock name is weakly weathered granite, its saturated uniaxial compressive strength should be greater than 30MPa".

[0145] B62. Minimum Exposure Sampling: This hypothesis is a range-bound hypothesis. An outlier-first sampling strategy is adopted. A total of 46 records with compressive strength less than 30 MPa are located from the L1 statistics and all of them are sampled for validation. The sample size accounts for 0.3% of the total data, which is lower than the sample size required for statistical power analysis.

[0146] B63. Local verification: Perform verification on the sampling records to check whether the records with soil and rock name of weakly weathered granite meet the requirement of compressive strength greater than 30MPa. The verification results are as follows: Of the 46 low-intensity records, 44 correspond to completely weathered or strongly weathered rocks, which does not violate the hypothesis; 2 correspond to weakly weathered granite, which violates the hypothesis; the hypothesis pass rate is approximately 0.956.

[0147] B64. Feedback Coding: Encode the validation results into structured feedback information, including the hypothesis, validation results, pass rate, outlier evidence, and suggested corrections. An example is shown below:

[0148]

[0149]

[0150]

[0151]

[0152] B65. Iterative Correction: Input structured feedback into the large language model, the large model corrects the expression of business rules, and the second round of verification shows that the pass rate of all hypotheses reaches 100%, the iteration stops, and the final data model is output. The 46 original sampling records were destroyed immediately after verification.

[0153] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A secure representation and modeling system for semi-structured data, characterized in that, It includes a local processor and a large language model server that communicates with the local processor; the local processor includes: The data parsing and structure recognition layer is used to receive the semi-structured data to be processed, perform structure parsing and recognition on the semi-structured data, and obtain data structure information. The metadata and statistical calculation module is used to extract structural metadata from the data structure information and calculate the statistical information corresponding to each data field in a hierarchical manner. The differential privacy perturbation module is used to perform differential privacy perturbation on the statistical information to obtain noisy statistical information; An adaptive compression coding module is used to adaptively compress and encode the structural metadata and noise statistics according to a preset token budget, generate a bimodal security representation, and send it to the large language model server. The large language model server includes: The noise perception modeling module is used to receive the dual-modal security representation and the preset noise perception prompts, generate an initial data model and return it to the local processor; The local processor also includes: The hypothesis verification and feedback module is used to extract verifiable hypotheses from the initial data model, verify the verifiable hypotheses locally based on the minimum exposure sampling strategy, generate feedback information based on the verification results and send it to the large language model server, and drive the large language model server to iteratively correct and obtain the final data model.

2. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The data parsing and structure recognition layer is specifically used for: For semi-structured data such as spreadsheets, traverse all worksheets and use the outward-to-inward shrinking method to locate the effective data area of ​​each worksheet, mark merged cells and parse the formula abstract syntax tree, and replace the cell coordinates in the formula with semantic placeholders. Multi-level headers are identified by combining rule-based filtering with a lightweight classifier. A header tree structure is constructed based on the inclusion relationship of merged regions, indentation level, and font size, and the header path sequence is output. Detect cross-worksheet references and named ranges, and construct a dependency graph between worksheets; For nested semi-structured data, identify the nesting level and generate a path sequence.

3. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The metadata and statistical calculation module is specifically used for: The statistical information is calculated at four levels, specifically including: Layer L1 provides basic descriptive statistics, including mean, median, standard deviation, minimum, maximum, and quartiles for numerical fields; average length, cardinality of unique values, and high-frequency patterns for text fields; time span and time granularity for date fields; and number of categories and distribution ratio for categorical fields. The L2 layer is for distribution shape statistics, including skewness, kurtosis, bin boundaries, distribution fit type, entropy, and Gini coefficient; Layer L3 is for relational statistics, including inter-column correlation coefficient matrix, mutual information, functional dependency strength, and cross-table association degree; Layer L4 provides quality and privacy statistics, including null value rate, outlier rate, duplicate record rate, sensitive field identification results, and k-anonymity estimate. The scope of statistical calculations is controlled according to the sensitivity level of the field: confidential fields are calculated only for L1 level statistics, sensitive fields are calculated for L1 level and optional L2 level statistics, and internal and lower level fields are calculated for L1-L3 level statistics.

4. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The differential privacy perturbation module is specifically used for: Calculate the corresponding sensitivity for different types of statistics: set the global sensitivity to 1 for count statistics, and use the smoothing sensitivity mechanism to calculate the local sensitivity for mean statistics; The noise scale is calculated based on the privacy budget parameter, and noise is sampled from the corresponding distribution and added to the statistics; Logical consistency is checked on the statistics after adding noise, and the noise value is adjusted for statistics that do not meet the monotonicity constraint using a constraint projection algorithm.

5. The secure representation and modeling system for semi-structured data according to claim 4, characterized in that, The differential privacy perturbation module adds noise to quantile statistics using an exponential mechanism, specifically for: The negative of the absolute value of the difference between the target quantile rank and the candidate rank is used as the scoring function. The quantile values ​​after adding noise are obtained by sampling from the candidate quantile values ​​according to the probability distribution corresponding to the scoring function.

6. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The adaptive compression coding module is specifically used for: A fixed number of tokens are reserved from the total token budget for system prompts and output format constraints, thus obtaining the available token budget. The available token budget is allocated to structural metadata, L1 layer statistics, semantic and relational metadata, and L2-L3 layer statistics according to a preset priority ratio, with L2-L3 layer statistics selected in order of information gain. Compact encoding is performed on each part of the allocated content, including column name semantic compression, statistical precision truncation, sparse matrix coordinate format storage, dynamic truncation of enumeration values ​​according to Token budget, and header path point encoding. Generate a bimodal security representation of MDS-Code in JSON format, including metadata version, privacy budget, token budget, structure information of each worksheet, noise statistics, inter-table relationship information, and global statistics.

7. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The noise perception modeling module is specifically used for: Noise perception inference is performed based on noise perception cues: For statistics marked as noisy in the bimodal security representation, the corresponding confidence interval is calculated according to the noise scale, and interval estimation is used instead of point estimation; the confidence of the corresponding inference result is adjusted by a confidence attenuation function according to the ratio of the noise scale to the magnitude of the statistics. The output includes a structured initial data model containing a conceptual model, a logical model, business rules, and uncertainties. Each inference result is accompanied by a confidence score.

8. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The hypothesis verification and feedback module is specifically used for: Extract different types of verifiable hypotheses from the initial data model, including constraint hypotheses, relational hypotheses, and rule hypotheses, and calculate the minimum sample size based on the hypothesis type and confidence requirements; For different types of hypotheses, corresponding targeted sampling strategies are adopted: constraint hypotheses use outlier priority sampling, relation hypotheses use hash bucketing collision detection sampling, and rule hypotheses use residual sorting sampling. Validate the sampled data locally to obtain validation results and anomaly pattern summaries.

9. The secure representation and modeling system for semi-structured data according to claim 8, characterized in that, The hypothesis verification and feedback module is also used for: The validation results are encoded into structured feedback information containing the hypothesis content, validation results, pass rate, abnormal evidence, and correction suggestions, and sent to the large language model server. The iteration termination is controlled based on the verification pass rate and the number of iteration rounds: when the verification pass rate of all verifiable hypotheses is greater than a preset threshold or the maximum number of iteration rounds is reached, the iteration stops and the final data model is obtained. After each round of verification is completed, the sample data obtained from sampling is destroyed immediately, and only the verification result statistics are retained.

10. The secure representation and modeling system for semi-structured data according to claim 1, characterized in that, The data parsing and structure recognition layer is also used for primary-foreign key relationship inference, specifically for: Calculate the unique value rate and null value rate of each column, mark the column with a unique value rate of 1 and a null value rate of 0 as candidate primary keys, and confirm the primary key by combining the column name semantic matching results; Calculate the value overlap rate and column name semantic similarity across columns, and combine them with cross-table reference frequency to mark column pairs with both value overlap rate and name similarity greater than a preset threshold as candidate foreign keys.

11. A secure representation and modeling method for semi-structured data, characterized in that, include: S1. The local processor receives the semi-structured data to be processed, performs structure parsing and recognition on the semi-structured data, and obtains data structure information; S2. The local processor extracts structural metadata from the parsed semi-structured data and calculates the statistical information corresponding to each data field in a hierarchical manner. S3. The local processor performs differential privacy perturbation on the calculated statistical information to obtain noisy statistical information; S4. The local processor performs adaptive compression encoding on the structural metadata and noise statistics information according to the preset Token budget, generates a bimodal security representation, and sends it to the large language model server. S5. The large language model server receives the bimodal security representation and the preset noise perception prompt, generates the initial data model and returns it to the local processor; S6. The local processor extracts verifiable hypotheses from the initial data model, verifies the verifiable hypotheses locally based on the minimum exposure sampling strategy, generates feedback information based on the verification results and sends it to the large language model server, driving the large language model server to iteratively correct and obtain the final data model.

Images