A method for sensitive data identification based on graph search enhanced generation

Abstract

This invention discloses a sensitive data identification method based on graph retrieval enhancement, belonging to the fields of data security and artificial intelligence. The main steps of this method are: collecting sensitive data identification rules and specifications, and obtaining structured rule data through preprocessing; constructing a sensitive data identification rule knowledge graph, forming a structured association network through entity extraction and relationship mining; mining the core entity association information of the samples to be identified, and enhancing the context by combining vector retrieval and graph feature filtering; combining sensitive rules and historical cases, injecting them into a large model for classification to obtain whether it is sensitive, the sensitive type, and the judgment criteria; integrating the output results of the large model, and visually presenting the logic and graph of sensitive data determination; and manually updating the graph and the large model through closed-loop feedback. This invention combines knowledge graphs and retrieval enhancement generation capabilities to identify implicit sensitive information, and the basis for the judgment results is traceable, making it suitable for assisting in the security management of sensitive data in multiple fields.

Description

Technical Field

[0001] This invention relates to a sensitive data identification method based on graph retrieval enhancement, belonging to the fields of data security and artificial intelligence technology. Background Technology

[0002] With the acceleration of global digitalization, the amount of data processed by enterprises and organizations is exploding, including a wide range of sensitive data such as personally identifiable information, financial records, health status, and trade secrets. Effective identification, classification, and protection of sensitive data has become a legal responsibility and a core requirement for business security that all organizations must fulfill.

[0003] However, current mainstream sensitive data identification technologies have limitations when dealing with complex and dynamic data environments. Traditional identification methods, such as matching techniques based on regular expressions and keyword databases, are simple to deploy and fast, but due to their mechanical pattern matching mechanism, they cannot understand the specific context of the data, resulting in limited identification accuracy and a high likelihood of generating a large number of false positives (e.g., misidentifying a string of ordinary order numbers as ID card numbers) and false negatives.

[0004] To enhance semantic understanding, the industry has introduced machine learning or deep learning-based models, such as Named Entity Recognition (NER) and text classifiers. While these methods improve accuracy to some extent, they typically require significant costs for manual data annotation to train the models and struggle to generalize to novel or hidden sensitive data. More importantly, both traditional and machine learning methods largely treat data as isolated entities for analysis, failing to effectively capture and utilize the complex and multi-layered intrinsic relationships between data. The sensitivity of data is often not determined by the data itself, but by the "relationship network" it belongs to. For example, a single name and address may not be sensitive, but when they are linked through a clear relationship to the same person, this data combination constitutes highly sensitive personal information.

[0005] While the recently emerging Retrieval Augmentation (RAG) technology enhances the model's responsiveness by leveraging external knowledge bases, its retrieval mechanism largely relies on vector similarity, essentially remaining a "flat" text block retrieval method that fails to fully utilize the structured graph relationships between data entities. Therefore, existing technologies generally lack an effective means to deeply mine the intrinsic connections between data and profoundly understand the context, resulting in a lack of capability in identifying sensitive data defined by relationships and hidden within complex networks. A more intelligent and accurate technological solution is urgently needed. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing technologies and provide a sensitive data identification method based on graph retrieval enhancement. This method constructs a sensitive data identification rule knowledge graph, utilizes graph-enhanced contextual retrieval, and combines the reasoning capabilities of a large language model to improve the automation level and accuracy of sensitive data identification.

[0007] The technical solution of this invention is: This invention provides a sensitive data identification method based on graph retrieval enhancement, the method comprising:

[0008] Collect sensitive data identification rules and specifications, and obtain structured rule data through preprocessing;

[0009] Construct a knowledge graph of sensitive data identification rules, and form a structured association network through entity extraction and relationship mining;

[0010] Mine the core entity association information of the sample to be identified, and enhance the context by combining vector retrieval and graph feature filtering;

[0011] By combining sensitive rules with historical cases, a large model is used to classify whether something is sensitive, the type of sensitivity, and the criteria for judgment.

[0012] Integrate the output results of the large model and visualize the decision-sensitive logic and graph;

[0013] Human feedback loop updates the graph and large model.

[0014] Furthermore, the method includes the following steps:

[0015] Step 1. Based on the enterprise's business scenarios and compliance requirements, collect sensitive data to identify relevant rules and regulations, extract core elements, including sensitivity types, judgment conditions, and legal basis, and preprocess them into structured rule data;

[0016] Step 2. Using the structured rule data collected and preprocessed in Step 1 as input, firstly, a named entity recognition model that integrates rule keywords is used to automatically extract the core entities in the rules; then, the extracted entities are classified through a large model, and the inherent relationships between entities are mined to generate entity-relationship pairs; subsequently, the labeled entities, entity attributes, and entity-relationship pairs are imported into a graph database; a structured knowledge graph of sensitive data recognition rules is constructed.

[0017] Step 3. Receive query requests for sensitive data to be identified, extract the core entity from the query request, and locate the node corresponding to the core entity in the sensitive data identification rule knowledge graph; based on the core entity node, use a graph traversal algorithm to mine its associated nodes and associated paths to obtain the structured association information of the core entity; convert the structured association information and the text content of the data to be identified into vector representations, and combine vector similarity calculation to retrieve relevant context fragments in the preprocessed text database; simultaneously, introduce graph structure features such as the degree of nodes and the length of associated paths in the graph, weight and sort the relevance of the search results, filter out the context information associated with the core entity, and form enhanced search results including structured association information and text context;

[0018] Step 4. Use the enhanced retrieval results, including structured association information and textual context, generated by graph retrieval enhancement as input to the large model, and inject sensitive data judgment rules and historical sensitive data judgment case library into the large model; based on the input context information and judgment rules, the large model first classifies the data into sensitive types, and then combines the entity association relationships in the knowledge graph to evaluate whether the data belongs to sensitive data, while identifying and eliminating false positives and false negatives that are prone to occur in traditional methods; output intelligent recognition results;

[0019] Step 5. Structure and integrate the results output by the large model, mapping the scattered sensitive data annotation information with the associated nodes and paths in the sensitive data knowledge graph to form a unified sensitive data identification result report; present the judgment logic of sensitive data in a graphical form through visualization tools, and convert the integrated results into an exportable format to allow users to directly obtain the sensitive data list; and design a visual interactive interface to allow users to click on sensitive data items to view the corresponding graph association information and judgment logic chain.

[0020] Step Six. Build a human feedback interaction platform that allows users to review the results, mark misjudged cases, missed cases, and cases with inaccurate judgment criteria, and enter feedback opinions; classify and organize the cases of human feedback and convert them into model update datasets; set regular update cycles to automatically trigger the update process of knowledge graph and model; form a closed-loop mechanism of human feedback - data processing - model / graph update.

[0021] Furthermore, step two specifically includes:

[0022] Using the structured rule data output from step one as input, a named entity recognition model that integrates rule keywords is employed to automatically identify and extract core entities from the rules, including: sensitive type entities, judgment condition entities, legal basis entities, and sensitivity level entities. The extracted entities are then processed through a large model: core entities are categorized according to semantic attributes; the inherent relationships between different categories of entities are identified, generating entity-relationship pairs; the processed entities and relationships are imported into a graph database to form a structured graph: attribute information is added to each entity, and through a graph data query language, entities, attributes, and entity-relationship pairs are imported into the graph database to construct an initial rule knowledge graph; ensuring the uniqueness of entities in the graph, a structured sensitive data identification rule knowledge graph is ultimately formed.

[0023] Furthermore, step three specifically includes:

[0024] Data to be identified A named entity recognition model that integrates rule keywords is used to... Perform entity recognition and extract the core entity set. ,in For the first There are three core entities; the sensitive data identification rule knowledge graph built in S2 locates the nodes corresponding to the core entities by matching entity names and attributes, thus obtaining the node set. ;

[0025] A graph search algorithm is used to mine related nodes, and a traversal depth threshold is set. With core nodes Starting from the first hop, the first hop associated node is the set of directly connected nodes. , For nodes, For the relationship, For a set of nodes, For a set of relations; the second hop associated node is The directly connected nodes, i.e. And so on, until the traversal depth reaches... This yields the total set of associated nodes. Record the path from the core node to each associated node. Represented as a sequence of nodes and relations: ,in This represents the path length, i.e., the number of hops. Indicates the first The relationship of jumping, Indicates the first Jump nodes; integrate all core nodes' associated nodes, associated paths, and relationship types to form structured associated information. ;

[0026] Structured related information The text content of the data to be identified is converted into a vector representation: a knowledge graph embedding model is used to convert entity nodes and relationships into low-dimensional vectors, with each node... Mapped to vector Each relationship Mapped to vector ;path The vector is the weighted sum of the vectors of each node and the relation vector: , For the first The node vector of the jump, For the first The vector mapped from the jump nodes. For the first The vector after the jump relation mapping; the vector representation of the final structured association information is as follows: Mean indicates averaging; the text content of the data to be identified is Using a pre-trained model Convert to text vector Concatenate the structured vector and the text vector into a joint vector. , This indicates vector concatenation, used as the query vector for retrieval;

[0027] Based on joint vector retrieval of relevant context fragments, and combined with weighted ranking based on graph structure features, enhanced retrieval results are generated from the preprocessed text database. Extracting a collection of text fragments A text embedding method is used to generate a vector for each segment. ; Calculate the query vector With fragment vector Cosine similarity:

[0028]

[0029] Filtering with similarity greater than a threshold The fragments form the initial search results. ;right Each segment Extract the structural features of the entities contained in the fragment within the graph: the nodes corresponding to the entities in the fragment. degree This refers to the number of associated edges; the average path length between entities and core nodes in a fragment. ,in for The path length to the nearest core node; the final relevance score of the segment is calculated by combining similarity and graph structure features:

[0030]

[0031] in, For similarity weights, , The maximum degree of the node. This represents the maximum average path length.

[0032] according to sorted in descending order Take the front Each fragment, associated with structured information This integration creates enhanced search results that include structured relational information and textual context. .

[0033] Furthermore, step four specifically includes:

[0034] Enhanced search results that include structured relational information and textual context The structured association information and association paths in the input text are converted into structured text descriptions while retaining text context fragments to form the basic input text. Define a sensitive data determination rule base. The rules clearly define the legal basis and judgment logic, and the rule base is converted into text format. ; Screening typical historical sensitive data to determine case libraries Each case includes the data to be judged, related information, judgment result, and basis, converted into text format. ; Will , , By logically piecing together the context, rules, and cases, the complete input text for the large model is formed. And add a command prompt;

[0035] Define sensitive types , For the first For each sensitive type, the large model uses semantic understanding and rule matching to output the probability that the data to be identified belongs to each type. ( ); Set type determination threshold If there exists a unique Then the sensitive type of the data to be identified is determined to be... If there are multiple Then, combining the entity relationships in the knowledge graph, the type with the highest matching degree with the relationship is selected as the final result, denoted as... The large model, based on decision rules and historical cases, outputs the probability that the data to be identified belongs to sensitive data. Set a sensitivity threshold ,like If the data is sensitive, it is considered sensitive; otherwise, it is considered non-sensitive. Simultaneously, knowledge graph entity association weights are introduced. Correcting the sensitivity probability:

[0036]

[0037] in, and Indicates weight; if If the data is positive, it is considered sensitive data; otherwise, it is considered non-sensitive data. For scenarios prone to false positives using traditional methods, the large model compares rule matching with historical cases to output the false positive probability. ,like False alarms are excluded if the value exceeds a set threshold; for scenarios prone to missed alarms, the large model mines entity relationships and outputs the missed alarm correction probability. ,like If the value exceeds the set threshold, it will be classified as sensitive data.

[0038] Large model combined with sensitive types Based on entity relationships and regulatory requirements, assess the level of sensitive data. The specific process is as follows: Determine whether the data is sensitive. , Indicates sensitivity, Indicate non-sensitive; and define the corresponding assessment indicators and weights for whether it is sensitive: regulatory severity. Entity association strength Scope of the leak ,satisfy The large model scores each indicator; if the score is... , , The total score for sensitivity level is:

[0039]

[0040] in, The score indicates the severity of the regulation. Indicates the entity association strength score. The score indicates the scope of the leak's impact.

[0041] Set a score threshold range, and output the sensitivity level based on the sensitivity corresponding to the threshold. ;

[0042] The large model outputs complete intelligent recognition results and performs structured processing to clarify key information. The specific operations are as follows: The large model outputs results in a preset format, containing four types of core information: Sensitivity determination conclusion: "belongs to sensitive data" or "does not belong to sensitive data"; Key annotations: specific location and sensitivity type of the sensitive data. Sensitivity Level Judgment criteria: cited judgment rules, knowledge graph relationships, and historical cases; Result structuring: converting the unstructured text output of the large model into a structured data format, performing format validation on the structured results to ensure the integrity and accuracy of the output information, and ultimately forming intelligent recognition results directly usable for subsequent result integration. .

[0043] Furthermore, step five specifically includes:

[0044] The intelligent recognition results output in step four The scattered sensitive data annotation information is precisely mapped to the associated nodes and paths in the sensitive data identification rule knowledge graph constructed in step two, establishing a relationship between the judgment result and the graph structure. The basic information of the sensitive data, the graph association information, and the judgment logic are integrated to form a structured, unified identification result report, including: basic information of the sensitive data, including data content, location index, and sensitive type. Sensitivity Level The system includes: graph association information, such as associated node IDs and names, associated path IDs and sequences, node degrees and path lengths; a judgment logic chain, including referenced rules, cases, and reasoning processes for excluding false positives / false negatives; the original text of the data to be identified and context fragments; a visualization tool to present the judgment logic of sensitive data in a graphical form, intuitively showing the complete chain including entity association, rule matching, and level determination; and a user-interactive visualization interface designed to allow users to view detailed information about sensitive data through clicks, filtering, and other operations.

[0045] Furthermore, step six specifically includes:

[0046] Build an interactive platform that supports user review and feedback input: The platform consists of 3 core modules, each corresponding to the needs of the entire feedback process:

[0047] The module for displaying pending review results is used to connect to the sensitive data identification result report output in step five. It displays the core information of each result in a list format, including data content, sensitivity type, level, and judgment basis. It supports filtering by unreviewed / reviewed and sensitivity type.

[0048] Case labeling module: Provides visual labeling controls, allowing users to select the labeling type for a single result, including false positive cases: non-sensitive data is judged as sensitive;

[0049] Missed case: Sensitive data was not identified;

[0050] Case of inaccurate judgment criteria: The type / level is correct, but the criteria used are incorrect;

[0051] And input feedback through the text box;

[0052] Feedback Record Management Module: Automatically records key information for each feedback, forming a feedback log. Supports querying historical feedback by time / tag type; sets hierarchical permissions; adds feedback verification rules; performs structured processing on feedback cases collected by the platform, converting them into datasets that can be directly used for knowledge graph and model updates; sets automatic update cycles to trigger the update process of knowledge graph and core model.

[0053] The present invention also provides a sensitive data identification system based on graph retrieval enhancement, the system including a module for executing the sensitive data identification method based on graph retrieval enhancement.

[0054] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned sensitive data identification method based on graph retrieval enhancement.

[0055] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned sensitive data identification method based on graph retrieval enhancement.

[0056] The beneficial effects of this invention are that by introducing a knowledge graph of sensitive data identification rules, isolated data points are connected into a graph, providing rich context for sensitive data identification. This enables the discovery of implicit sensitive information generated by data associations that traditional methods cannot identify, thereby improving the accuracy of identification. Simultaneously, leveraging the reasoning capabilities of large models, this method can deeply understand the complex semantics of natural language text, handle various unstructured and formatted data, and overcome the limitations of regular expressions and traditional machine learning methods. The identification result is not merely a label, but also includes detailed reasoning and a visualized chain of evidence, making auditing and tracing simple and intuitive, meeting compliance requirements. Furthermore, through a feedback update mechanism, the knowledge graph and model can continuously learn new sensitive data patterns and business knowledge, allowing for continuous iterative optimization of system performance to adapt to emerging sensitive data types and business changes. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0058] Figure 1This is a flowchart illustrating a sensitive data identification method based on graph retrieval enhancement provided by the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be systematically and completely described below with reference to the accompanying drawings. Obviously, the presented embodiments are only partial examples of the present invention and are not exhaustive. Based on the content disclosed in this invention, any other embodiments obtained by those skilled in the art without creative effort fall within the protection scope of this invention.

[0060] Unless otherwise specified, the various embodiments and technical features of this application can be combined arbitrarily without conflict. Furthermore, the executing entity of the embodiments of this invention can be any terminal device with text processing capabilities, including but not limited to: desktop computers, laptops, smartphones, tablets, wearable devices, personal digital assistants (PDAs), and mobile internet devices (MIDs), etc., and this invention does not limit this.

[0061] Example 1: As Figure 1 As shown in the figure, this invention provides a sensitive data identification method based on graph retrieval enhancement. Taking a sensitive data identification scenario of a municipal government service platform as an example, this platform needs to process government data submitted by citizens and identify core sensitive data. The specific steps include:

[0062] S1: Collect sensitive data identification rules and specifications, and obtain structured rule data through preprocessing;

[0063] S1 includes collecting sensitive data to identify relevant rules and regulations for enterprise business scenarios and compliance requirements, extracting core elements, including sensitive types, judgment conditions, and legal basis, and preprocessing them into structured rule data.

[0064] S2: Construct a knowledge graph of sensitive data identification rules, and form a structured association network through entity extraction and relationship mining;

[0065] S2 includes taking the structured rule data collected and preprocessed in step S1 as input, firstly using a named entity recognition model that integrates rule keywords to automatically extract the core entities in the rules; then classifying the extracted entities through a large model and mining the inherent relationships between entities to generate entity-relationship pairs; subsequently importing the labeled entities, entity attributes, and entity-relationship pairs into a graph database; and constructing a structured sensitive data recognition rule knowledge graph.

[0066] S3: Mine the core entity association information of the sample to be identified, and enhance the context by combining vector retrieval and graph feature filtering;

[0067] S3 includes receiving a query request for sensitive data to be identified, extracting the core entity from the query request, and locating the node corresponding to the core entity in the sensitive data identification rule knowledge graph; based on the core entity node, using a graph traversal algorithm to mine its associated nodes and associated paths to obtain the structured association information of the core entity; converting the structured association information and the text content of the data to be identified into vector representations, and retrieving relevant context fragments in the preprocessed text database by combining vector similarity calculation; simultaneously introducing graph structure features such as the degree of nodes and the length of associated paths in the graph, weighting and sorting the relevance of the search results, filtering out the context information associated with the core entity, and forming enhanced search results including structured association information and text context;

[0068] S4: Combining sensitive rules and historical cases, inject them into a large model to classify whether it is sensitive, the type of sensitivity, and the basis for judgment;

[0069] S4 includes taking the enhanced retrieval results, which include structured association information and text context, generated by graph retrieval enhancement as input to the large model, and injecting sensitive data judgment rules and a historical sensitive data judgment case library into the large model; based on the input context information and judgment rules, the large model first classifies the data into sensitive types, and then combines the entity association relationships in the knowledge graph to evaluate whether the data belongs to sensitive data, while identifying and eliminating false positives and false negatives that are prone to occur in traditional methods; and outputting intelligent recognition results.

[0070] S5: Integrate the output results of large models and visualize the decision-sensitive logic and graphs;

[0071] S5 includes: structurally integrating the results output by the large model; mapping the scattered sensitive data annotation information with the associated nodes and paths in the sensitive data knowledge graph to form a unified sensitive data identification result report; presenting the judgment logic of sensitive data in a graphical form through visualization tools; converting the integrated results into an exportable format to support users in directly obtaining the sensitive data list; and designing a visual interactive interface that allows users to click on sensitive data items to view the corresponding graph association information and judgment logic chain.

[0072] S6: Human feedback closed-loop update of the graph and large model.

[0073] S6 includes: building a human feedback interaction platform, allowing users to review the results, mark misjudged cases, missed cases, and cases with inaccurate judgment criteria, and enter feedback opinions; classifying and organizing the cases of human feedback and converting them into model update datasets; setting a regular update cycle to automatically trigger the update process of knowledge graph and model; and forming a closed-loop mechanism of human feedback - data processing - model / graph update.

[0074] Furthermore, S2 specifically includes:

[0075] Using the structured rule data output by S1 as input, a named entity recognition model that integrates rule keywords is adopted to automatically identify and extract core entities in the rules, including: sensitive type entities, judgment condition entities, legal basis entities, and sensitive level entities. The extracted entities are processed through a large model: core entities are classified according to semantic attributes; the inherent relationships between different categories of entities are identified to generate entity-relation pairs; the processed entities and relations are imported into a graph database to form a structured graph: attribute information is added to each entity, and the entity, attributes, and entity-relation pairs are imported into the graph database through a graph data query language to construct an initial rule knowledge graph; ensuring that the entities in the graph are unique, a structured sensitive data identification rule knowledge graph is finally formed.

[0076] Furthermore, S3 specifically includes:

[0077] Data to be identified A named entity recognition model that integrates rule keywords is used to... Perform entity recognition and extract the core entity set. ,in For the first There are three core entities; the sensitive data identification rule knowledge graph built in S2 locates the nodes corresponding to the core entities by matching entity names and attributes, thus obtaining the node set. ;

[0078] A graph search algorithm is used to mine related nodes, and a traversal depth threshold is set. With core nodes Starting from the first hop, the first hop associated node is the set of directly connected nodes. , For nodes, For the relationship, For a set of nodes, For a set of relations; the second hop associated node is The directly connected nodes, i.e. And so on, until the traversal depth reaches... This yields the total set of associated nodes. Record the path from the core node to each associated node. Represented as a sequence of nodes and relations: ,in This represents the path length, i.e., the number of hops. Indicates the first The relationship of jumping, Indicates the first Jump nodes; integrate all core nodes' associated nodes, associated paths, and relationship types to form structured associated information. ;

[0079] Structured related information The text content of the data to be identified is converted into a vector representation: a knowledge graph embedding model is used to convert entity nodes and relationships into low-dimensional vectors, with each node... Mapped to vector Each relationship Mapped to vector ;path The vector is the weighted sum of the vectors of each node and the relation vector: , For the first The node vector of the jump, For the first The vector mapped from the jump nodes. For the first The vector after the jump relation mapping; the vector representation of the final structured association information is as follows: Mean indicates averaging; the text content of the data to be identified is Using a pre-trained model Convert to text vector Concatenate the structured vector and the text vector into a joint vector. , This indicates vector concatenation, used as the query vector for retrieval;

[0080] Based on joint vector retrieval of relevant context fragments, and combined with weighted ranking based on graph structure features, enhanced retrieval results are generated from the preprocessed text database. Extracting a collection of text fragments A text embedding method is used to generate a vector for each segment. ; Calculate the query vector With fragment vector Cosine similarity:

[0081]

[0082] Filtering with similarity greater than a threshold The fragments form the initial search results. ;right Each segment Extract the structural features of the entities contained in the fragment within the graph: the nodes corresponding to the entities in the fragment. degree This refers to the number of associated edges; the average path length between entities and core nodes in a fragment. ,in for The path length to the nearest core node; the final relevance score of the segment is calculated by combining similarity and graph structure features:

[0083]

[0084] in, For similarity weights, , The maximum degree of the node. This represents the maximum average path length.

[0085] according to sorted in descending order Take the front Each fragment, associated with structured information This integration creates enhanced search results that include structured relational information and textual context. .

[0086] Furthermore, S4 specifically includes:

[0087] Enhanced search results that include structured relational information and textual context The structured association information and association paths in the input text are converted into structured text descriptions while retaining text context fragments to form the basic input text. Define a sensitive data determination rule base. The rules clearly define the legal basis and judgment logic, and the rule base is converted into text format. ; Screening typical historical sensitive data to determine case libraries Each case includes the data to be judged, related information, judgment result, and basis, converted into text format. ; Will , , By logically piecing together the context, rules, and cases, the complete input text for the large model is formed. And add a command prompt;

[0088] Define sensitive types , For the first For each sensitive type, the large model uses semantic understanding and rule matching to output the probability that the data to be identified belongs to each type. ( ); Set type determination threshold If there exists a unique Then the sensitive type of the data to be identified is determined to be... If there are multiple Then, combining the entity relationships in the knowledge graph, the type with the highest matching degree with the relationship is selected as the final result, denoted as... The large model, based on decision rules and historical cases, outputs the probability that the data to be identified belongs to sensitive data. Set a sensitivity threshold ,like If the data is sensitive, it is considered sensitive; otherwise, it is considered non-sensitive. Simultaneously, knowledge graph entity association weights are introduced. Correcting the sensitivity probability:

[0089]

[0090] in, and Indicates weight; if If the data is positive, it is considered sensitive data; otherwise, it is considered non-sensitive data. For scenarios prone to false positives using traditional methods, the large model compares rule matching with historical cases to output the false positive probability. ,like False alarms are excluded if the value exceeds a set threshold; for scenarios prone to missed alarms, the large model mines entity relationships and outputs the missed alarm correction probability. ,like If the value exceeds the set threshold, it will be classified as sensitive data.

[0091] Large model combined with sensitive types Based on entity relationships and regulatory requirements, assess the level of sensitive data. The specific process is as follows: Determine whether the data is sensitive. , Indicates sensitivity, Indicate non-sensitive; and define the corresponding assessment indicators and weights for whether it is sensitive: regulatory severity. Entity association strength Scope of the leak ,satisfy The large model scores each indicator; if the score is... , , The total score for sensitivity level is:

[0092]

[0093] in, The score indicates the severity of the regulation. Indicates the entity association strength score. The score indicates the scope of the leak's impact.

[0094] Set a score threshold range, and output the sensitivity level based on the sensitivity corresponding to the threshold. ;

[0095] The large model outputs complete intelligent recognition results and performs structured processing to clarify key information. The specific operations are as follows: The large model outputs results in a preset format, containing four types of core information: Sensitivity determination conclusion: "belongs to sensitive data" or "does not belong to sensitive data"; Key annotations: specific location and sensitivity type of the sensitive data. Sensitivity Level Judgment criteria: cited judgment rules, knowledge graph relationships, and historical cases; Result structuring: converting the unstructured text output of the large model into a structured data format, performing format validation on the structured results to ensure the integrity and accuracy of the output information, and ultimately forming intelligent recognition results directly usable for subsequent result integration. .

[0096] Furthermore, S5 specifically includes:

[0097] The intelligent recognition results output in step four The scattered sensitive data annotation information is precisely mapped to the associated nodes and paths in the sensitive data identification rule knowledge graph constructed in step two, establishing a relationship between the judgment result and the graph structure. The basic information of the sensitive data, the graph association information, and the judgment logic are integrated to form a structured, unified identification result report, including: basic information of the sensitive data, including data content, location index, and sensitive type. Sensitivity Level The system includes: graph association information, such as associated node IDs and names, associated path IDs and sequences, node degrees and path lengths; a judgment logic chain, including referenced rules, cases, and reasoning processes for excluding false positives / false negatives; the original text of the data to be identified and context fragments; a visualization tool to present the judgment logic of sensitive data in a graphical form, intuitively showing the complete chain including entity association, rule matching, and level determination; and a user-interactive visualization interface designed to allow users to view detailed information about sensitive data through clicks, filtering, and other operations.

[0098] Furthermore, S6 specifically includes:

[0099] Build an interactive platform that supports user review and feedback input: The platform consists of 3 core modules, each corresponding to the needs of the entire feedback process:

[0100] The module for displaying pending review results is used to connect to the sensitive data identification result report output in step five. It displays the core information of each result in a list format, including data content, sensitivity type, level, and judgment basis. It supports filtering by unreviewed / reviewed and sensitivity type.

[0101] Case labeling module: Provides visual labeling controls, allowing users to select the labeling type for a single result, including false positive cases: non-sensitive data is judged as sensitive;

[0102] Missed case: Sensitive data was not identified;

[0103] Case of inaccurate judgment criteria: The type / level is correct, but the criteria used are incorrect;

[0104] And input feedback through the text box;

[0105] Feedback Record Management Module: Automatically records key information for each feedback, forming a feedback log. Supports querying historical feedback by time / tag type; sets hierarchical permissions; adds feedback verification rules; performs structured processing on feedback cases collected by the platform, converting them into datasets that can be directly used for knowledge graph and model updates; sets automatic update cycles to trigger the update process of knowledge graph and core model.

[0106] The present invention also provides a sensitive data identification system based on graph retrieval enhancement, the system including a module for executing the sensitive data identification method based on graph retrieval enhancement.

[0107] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned sensitive data identification method based on graph retrieval enhancement.

[0108] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned sensitive data identification method based on graph retrieval enhancement.

[0109] Example 2: A sensitive data identification method based on graph retrieval enhancement, the method comprising the following steps:

[0110] S1: Collect sensitive data identification rules and specifications, and obtain structured rule data through preprocessing;

[0111] S11 specifies the system's data collection rules for the three core business scenarios of the government service platform: "citizen service data processing," "enterprise government approval," and "cross-departmental data sharing."

[0112] S12 extracts three core elements from the rules and specifications in S11, including:

[0113] Sensitive data types, such as: sensitive personal identification data, sensitive government business data, and secret-related data (such as "code names of major municipal infrastructure projects" and "code names of undisclosed livelihood project budgets").

[0114] Judgment criteria: For example, the combination of "ID number + bank card number + subsidy amount" is judged as high-sensitivity government data; "citizen's name / address associated with low-income family identification code" is judged as sensitive data; "city-level major infrastructure project code (including 'urban rail / people's livelihood' keywords + year-number format)" is judged as confidential data; "individual citizen's name / company name (without associated sensitive identifiers)" is judged as non-sensitive data.

[0115] Legal basis element, that is, which law or regulation is used to determine the sensitive data.

[0116] S13 Preprocessing of Sensitive Government Data Rules: The core elements extracted are cleaned and structured: Duplicate elements are removed: identical expressions in different documents are deleted; ambiguous expressions are corrected, unifying "highly sensitive government data" and "sensitive-level government data" as "highly sensitive government data," and clarifying "confidential data" as "secret government data"; missing information is supplemented, adding the level "medium sensitive government data" to "enterprise credit punishment records" (which are only marked "sensitive"), and adding "cross-departmental data sharing" as the applicable scenario. Data structuring: Following the logical association of "sensitivity type - judgment criteria - legal basis - applicable scenario," the data is organized into a tabular format, forming structured rule data. This provides input for subsequent knowledge graph construction.

[0117] S2: Construct a knowledge graph of sensitive data identification rules, and form a structured association network through entity extraction and relationship mining;

[0118] S21 government affairs rules core entity extraction, using the structured rule data output by S1. For input, a Named Entity Recognition (NER) model that integrates "sensitive government keywords" (e.g., incorporating "ID number," "low-income family identification code," and "Urban Rail 2024-01" as domain dictionaries into a pre-trained BERT model) is used to automatically extract core entities, mainly including: sensitive type entities, including sensitive personal identity data, sensitive government business data, and secret-related data; judgment condition entities, including individual ID numbers, name + ID number combinations, low-income family identification codes, and Urban Rail 2024-01 (project code); legal basis entities, including various legal and regulatory clauses; and sensitive level entities, including highly sensitive government data, moderately sensitive government data, and confidential government data.

[0119] S22 Government Entity Classification and Relationship Mining uses a customized government model (fine-tuned based on DeepSeek, incorporating government data scenario knowledge) to process extracted entities: Entity classification: "ID card number" and "social security account number" are classified as "Personal Identity Sensitive Data Entities"; "Low-income family identification code" and "Government procurement bid number" are classified as "Government Business Sensitive Data Entities"; "Urban Rail 2024-01" and "Livelihood Project Budget Code 2024-05" are classified as "State Secret Related Data Entities"; and the relevant legal provisions for each data are classified as "Legal Basis Entities". Relationship mining: Entity-relationship pairs are generated, as shown in the following examples: (Personal Identity Sensitive Data, must meet the following criteria: individual ID card number); (Government Business Sensitive Data, must meet the following criteria: citizen address + low-income family identification code combination); (Low-income family identification code, corresponding level, medium-sensitivity government data).

[0120] S23 Construction and Import of Knowledge Graph for Sensitive Government Data Rules. Entity attributes are labeled, for example: For "Sensitive Personal Identity Data," the attributes are labeled: Applicable Scenario: Citizen Service Data Processing; Update Time: March 2024; Impact of Leakage: Infringement of Citizen Privacy; For "Confidential Government Data," the attributes are labeled: Access Permission: Limited to Core Municipal Government Departments; Encryption Requirements: AES-256 Encryption; Neo4j graph database is used, and entities, attributes, and relationships are imported using Cypher statements.

[0121] S3: Mine the core entity association information of the sample to be identified, and enhance the context by combining vector retrieval and graph feature filtering;

[0122] Enhanced contextual retrieval of sensitive government data: The data to be identified includes "The municipal government service platform received application data from citizen 'Zhang San, ID number 430102********1234, applying for the second quarter of 2024 minimum living allowance subsidy, family identification code LB2024-0567', and sensitive data needs to be identified within it." Perform the following steps:

[0123] S31 core government entity extraction and graph node localization uses the NER model from S21, which incorporates government keywords, for... Perform entity recognition and extract the core entity set:

[0124]

[0125] Knowledge graph of government rules built in S2 In the example, the node is located by matching entity name and attribute: "430102********1234 (ID number)" matches the "judgment condition entity" node. (Name: Individual ID Number, Format Requirement: 18 digits); Match "LB2024-0567 (Low-Income Family Identification Code)" with the "Judgment Criteria Entity" node. (Name: Low-income family identification code; Applicable scenario: application for livelihood subsidies); Locate the node set corresponding to the core entity. .

[0126] The S32 government affairs graph is traversed to mine related nodes and paths using a breadth-first search (BFS) algorithm, with a traversal depth threshold set. (Only related nodes within 2 hops are mined to avoid redundant information).

[0127] First hop associated node: The direct associated nodes for (a unique ID number) are:

[0128] ;

[0129] The direct associated node of the (low-income family identification code) is:

[0130] ;

[0131] Second hop related node: ( Exclude directly related nodes. ), ( Exclude directly related nodes. );

[0132] Total set of associated nodes: , .

[0133] Record the path from the core node to the related nodes, as shown in the example below:

[0134] ; ; .

[0135] form:

[0136]

[0137] The S33 government affairs structure and text information vector conversion uses the TransE knowledge graph embedding model to convert nodes and relationships into 128-dimensional vectors: (Individual ID Number) Vector ; (Personally identifiable sensitive data) vector The relation "must satisfy" vector ;path vector ;

[0138]

[0139] Vector representation of structured association information :

[0140] .

[0141] Sentence-BERT (a slightly modified version for government scenarios) will recognize the text to be identified. ("Zhang San, ID number 430102********1234, applying for the second quarter of 2024 minimum living allowance, family identification code LB2024-0567") converted to vector By concatenating the structured vector and the text vector, the query vector is obtained. .

[0142] S34 Government Affairs Context Retrieval and Graph Feature Weighted Sorting, from the "Municipal Government Affairs Historical Service Data Text Database" preprocessed in S1. Extracting a collection of text fragments (e.g., "Citizen Li Si, ID number 310104********5678, applying for 2023 minimum living allowance, family identification code LB2023-0345", "Company A submits government procurement application, bid number CG2024-012", etc.), generate vectors for each segment. ;calculate and Cosine similarity, setting a threshold Filter out the initial search results (Example:) For "Citizen Li Si, ID number 310104********5678, applying for the 2023 minimum living allowance subsidy, family identification code LB2023-0345", Graph structure feature extraction, node degree: The node corresponding to "ID number 310104********5678" degree (Connecting 4 relation edges), maximum degree of the graph Average path length Medium entity and core node Average path length Maximum average path length Weighted score calculation: Setting (Similarity weight) (Degree weight) (Path length weight), calculate score:

[0143]

[0144] The top 3 segments are selected in descending order of score. ),and Integration to create enhanced search results .

[0145] S4: Combining sensitive rules and historical cases, inject them into a large model to classify whether it is sensitive, the type of sensitivity, and the basis for judgment;

[0146] S41 Large Model Input Data Integration and Formatting: Basic Input Text Judgment rules Historical Cases Enter the complete text = Add the instruction: "Please identify sensitive data in the government data to be identified, mark the type, level, location and judgment basis, and exclude false alarms / omissions."

[0147] The S42 large-scale model's sensitive type classification and determination uses a government-customized version of the large-scale model (fine-tuned based on DeepSeek R1 70B) to perform the analysis and output the probabilities of each type:

[0148] ,

[0149] ,

[0150] ;

[0151] Set threshold Combining the graph association ("ID number associated with sensitive personal identity data, low-income assistance identification code associated with sensitive government affairs data"), the final sensitive types are... :

[0152]

[0153] Initial sensitivity probability Correlation weight The ID number and sensitive personal identity data are linked in a single hop. The low-income assistance identification code and sensitive government affairs data are linked in a single hop. ),average Corrected probability ( ), was identified as sensitive data; false negatives and missed reports were eliminated: (The combination of "ID number + low-income assistance identification code" complies with rule 3, and there are no false positives.) (No unidentified sensitive entities) No corrections required.

[0154] S43 Large Model Sensitivity Level Assessment Settings (Severity of regulations) (Entity association strength) (Scope of impact of the leak). ID number: (Explicit protection) (1-hop related sensitive type) (Disclosure violates privacy). Score :

[0155] ;

[0156] Low-income family identification code: (Internal details are clearly defined) (1-hop association) (The leak affected eligibility for minimum living allowance) .

[0157] Threshold range For political affairs with high sensitivity, The government information is classified as "medium sensitive"; therefore, the ID number is "highly sensitive" and the low-income family identification code is "medium sensitive". The final level is... :

[0158] .

[0159] S44 intelligent recognition results output: "Belongs to sensitive government affairs data", "Personal identity sensitive data", "Government affairs highly sensitive", "Rule 1 (Article 13 of the Personal Information Protection Law), Rule 2 (Article 5.2 of the Implementation Rules for the Classification and Grading of Municipal Government Affairs Data), Rule 3 ('ID number + low-income identification code' combination is highly sensitive)", "430102********1234 - Determination must meet - Personal identity sensitive data; LB2024-0567 - Determination must meet - Government affairs business sensitive data (1 jump)" "Case 1 ('Wang Wu + ID number + low-income identification code' is determined to be highly sensitive / medium sensitive)".

[0160] S5: Integrate the output results of large models and visualize the decision-sensitive logic and graphs;

[0161] S51 structured results and map association mapping:

[0162] Entity-node mapping:

[0163] ;

[0164] Path-relationship mapping:

[0165] ;

[0166] Feature mapping: .

[0167] S52 Unified Government Affairs Identification Result Report is generated; the "Government Affairs Sensitive Data Identification Report (20240520-001)" is generated, which includes:

[0168] 1. Basic Information: Source of data to be identified (Minimum Living Allowance Application Module of Government Service Platform), Identification Time (2024-05-20 14:30), Number of Sensitive Entities (2);

[0169] 2. Graph association information: associated node ID, path sequence, node degree;

[0170] 3. Judgment logic chain: Extract entities → Graph association → Rule matching (rule 1+2+3) → Level evaluation (0.87 points → high sensitivity, 0.77 points → medium sensitivity) → Eliminate false negatives and false negatives;

[0171] 4. Raw data: text to be recognized and Context fragment.

[0172] S54 Government Affairs Visualization Interactive Interface Design: Design the interactive interface for the "Municipal Government Affairs Sensitive Data Identification Platform".

[0173] S6: Human feedback closed-loop update of the graph and large model.

[0174] The S61 government affairs manual feedback interaction platform was established; a "municipal-level government affairs data identification and feedback platform" was built, comprising 3 modules:

[0175] The module for displaying pending review results connects to reports output by S5 and displays them according to business categories such as "Minimum Living Allowance Application", "Enterprise Approval", and "Project Management". It supports batch filtering of "Unreviewed" results.

[0176] Case Marking Module: When reviewing the above results, the government data security specialist found that "the business name of 'Second Quarter 2024 Low-Income Subsidy' for Citizen Zhang San was not misjudged and does not need to be marked"; when reviewing "Company B's procurement application data (including 'Bid Number CG2024-012')," it was found that the system did not recognize "CG2024-012," and marked it as a "Missed Case." The feedback was entered as: "Bid Number CG2024-012 meets the requirements of Article 5.2 of the 'Implementation Rules for the Classification and Grading of Municipal Government Data,' and should be judged as highly sensitive government data."

[0177] Feedback record management module: Records feedback logs and sets permissions (only data security specialists can mark "missed judgment", and business specialists can only mark "inaccurate basis").

[0178] S62 Government Feedback Case Classification and Dataset Conversion: In the case set of missed judgments, add "CG2024-012 (base number), Sensitive type: Sensitive data of government business, Level: High sensitivity of government affairs, Basis: Article 5.2 of the "Implementation Rules for Classification and Grading of Municipal Government Data""; Case set based on amendments: None (no case with incorrect basis in this case); Case set of misjudged judgments: None (no misjudged case in this case).

[0179] The S63 government knowledge graph and model are updated regularly, with a set update cycle. Today, due to a missed judgment in this case (bid number), which falls under "core sensitive data in government procurement," the number of high-priority cases has reached [number missing]. This triggers an immediate update, instantly updating the knowledge graph constructed by S2.

[0180] Table 1 shows the advantages and disadvantages of the method of the present invention compared with other methods.

[0181]

[0182] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0183] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A sensitive data identification method based on graph retrieval enhancement, characterized in that: include: Collect sensitive data identification rules and specifications, and obtain structured rule data through preprocessing; Construct a knowledge graph of sensitive data identification rules, and form a structured association network through entity extraction and relationship mining; Mine the core entity association information of the sample to be identified, and enhance the context by combining vector retrieval and graph feature filtering; By combining sensitive rules with historical cases, a large model is used to classify whether something is sensitive, the type of sensitivity, and the criteria for judgment. Integrate the output results of the large model and visualize the decision-sensitive logic and graph; Human feedback loop updates the graph and large model.

2. The sensitive data identification method based on graph retrieval enhancement according to claim 1, characterized in that: The method includes the following steps: Step 1. Based on the enterprise's business scenarios and compliance requirements, collect sensitive data, identify relevant rules and regulations, extract core elements, including sensitivity types, judgment conditions, and legal basis, and preprocess them into structured rule data; Step 2. Using the structured rule data collected and preprocessed in Step 1 as input, firstly, a named entity recognition model that integrates rule keywords is used to automatically extract the core entities in the rules; then, the extracted entities are classified through a large model, and the inherent relationships between entities are mined to generate entity-relationship pairs; subsequently, the labeled entities, entity attributes, and entity-relationship pairs are imported into a graph database; a structured knowledge graph of sensitive data recognition rules is constructed. Step 3. Receive the query request for sensitive data to be identified, extract the core entity from the query request, and locate the node corresponding to the core entity in the knowledge graph of sensitive data identification rules; based on the core entity node, use the graph traversal algorithm to mine its associated nodes and associated paths to obtain the structured association information of the core entity. The structured association information and the text content of the data to be identified are converted into vector representations. The relevant context fragments are retrieved from the preprocessed text database by combining vector similarity calculation. At the same time, graph structure features such as the degree of nodes and the length of associated paths in the graph are introduced to weight and sort the relevance of the retrieval results, and the context information associated with the core entity is filtered out to form enhanced retrieval results including structured association information and text context. Step 4. Use the enhanced retrieval results, including structured association information and textual context, generated by graph retrieval enhancement as input to the large model, and inject sensitive data judgment rules and historical sensitive data judgment case library into the large model; based on the input context information and judgment rules, the large model first classifies the data into sensitive types, and then combines the entity association relationships in the knowledge graph to evaluate whether the data belongs to sensitive data, while identifying and eliminating false positives and false negatives that are prone to occur in traditional methods; output intelligent recognition results; Step 5. Structure and integrate the results output by the large model, and map the scattered sensitive data annotation information with the associated nodes and paths in the sensitive data knowledge graph to form a unified sensitive data identification result report; The logic for determining sensitive data is presented in a graphical form through visualization tools, and the integrated results are converted into an exportable format, allowing users to directly obtain a list of sensitive data. A visual interactive interface was designed, allowing users to click on sensitive data entries to view the corresponding graph association information and judgment logic chain; Step 6. Build a human feedback interaction platform that allows users to review the results, mark misjudged cases, missed cases, and cases with inaccurate judgment criteria, and enter feedback opinions; classify and organize the cases of human feedback and convert them into model update datasets; set regular update cycles to automatically trigger the knowledge graph and model update process; This establishes a closed-loop mechanism of human feedback, data processing, and model / map updates.

3. The sensitive data identification method based on graph retrieval enhancement according to claim 2, characterized in that: Step two specifically includes: Using the structured rule data output from step one as input, a named entity recognition model that integrates rule keywords is employed to automatically identify and extract core entities from the rules, including: sensitive type entities, judgment condition entities, legal basis entities, and sensitivity level entities. The extracted entities are then processed through a large model: core entities are categorized according to semantic attributes; the inherent relationships between different categories of entities are identified, generating entity-relationship pairs; the processed entities and relationships are imported into a graph database to form a structured graph: attribute information is added to each entity, and through a graph data query language, entities, attributes, and entity-relationship pairs are imported into the graph database to construct an initial rule knowledge graph; ensuring the uniqueness of entities in the graph, a structured sensitive data identification rule knowledge graph is ultimately formed.

4. The sensitive data identification method based on graph retrieval enhancement according to claim 2, characterized in that: Step three specifically includes: Data to be identified A named entity recognition model that integrates rule keywords is used to... Perform entity recognition and extract the core entity set. ,in For the first There are three core entities; the sensitive data identification rule knowledge graph built in S2 locates the nodes corresponding to the core entities by matching entity names and attributes, thus obtaining the node set. ; A graph search algorithm is used to mine related nodes, and a traversal depth threshold is set. With core nodes Starting from the first hop, the first hop associated node is the set of directly connected nodes. , For nodes, For the relationship, For a set of nodes, For a set of relations; the second hop associated node is The directly connected nodes, i.e. And so on, until the traversal depth reaches... This yields the total set of associated nodes. Record the path from the core node to each associated node. Represented as a sequence of nodes and relations: ,in This represents the path length, i.e., the number of hops. Indicates the first The relationship of jumping, Indicates the first Jump nodes; integrate all core nodes' associated nodes, associated paths, and relationship types to form structured associated information. ; Structured related information The text content of the data to be identified is converted into a vector representation: a knowledge graph embedding model is used to convert entity nodes and relationships into low-dimensional vectors, with each node... Mapped to vector Each relationship Mapped to vector ;path The vector is the weighted sum of the vectors of each node and the relation vector: , For the first The node vector of the jump, For the first The vector mapped from the jump nodes. For the first The vector after the jump relation mapping; the vector representation of the final structured association information is as follows: Mean indicates averaging; the text content of the data to be identified is Using a pre-trained model Convert to text vector Concatenate the structured vector and the text vector into a joint vector. , This indicates vector concatenation, used as the query vector for retrieval; Based on joint vector retrieval of relevant context fragments, and combined with weighted ranking based on graph structure features, enhanced retrieval results are generated from the preprocessed text database. Extracting a collection of text fragments A text embedding method is used to generate a vector for each segment. ; Calculate the query vector With fragment vector Cosine similarity: ； Filtering with similarity greater than a threshold The fragments form the initial search results. ;right Each segment Extract the structural features of the entities contained in the fragment within the graph: the nodes corresponding to the entities in the fragment. degree This refers to the number of associated edges; the average path length between entities and core nodes in a fragment. ,in for The path length to the nearest core node; the final relevance score of the segment is calculated by combining similarity and graph structure features: ； in, For similarity weights, , The maximum degree of the node. This represents the maximum average path length. according to sorted in descending order Take the front Each fragment, associated with structured information This integration creates enhanced search results that include structured relational information and textual context. .

5. The sensitive data identification method based on graph retrieval enhancement according to claim 2, characterized in that: Step four specifically includes: Enhanced search results that include structured relational information and textual context The structured association information and association paths in the input text are converted into structured text descriptions while retaining text context fragments to form the basic input text. Define a sensitive data determination rule base. The rules clearly define the legal basis and judgment logic, and the rule base is converted into text format. ; Screening typical historical sensitive data to determine case libraries Each case includes the data to be judged, related information, judgment result, and basis, converted into text format. ; Will , , By logically piecing together the context, rules, and cases, the complete input text for the large model is formed. And add a command prompt; Define sensitive types , For the first For each sensitive type, the large model uses semantic understanding and rule matching to output the probability that the data to be identified belongs to each type. ( ); Set type determination threshold If there exists a unique Then the sensitive type of the data to be identified is determined to be... If there are multiple Then, combining the entity relationships in the knowledge graph, the type with the highest matching degree with the relationship is selected as the final result, denoted as... The large model, based on decision rules and historical cases, outputs the probability that the data to be identified belongs to sensitive data. Set a sensitivity threshold ,like If the data is sensitive, it is considered sensitive; otherwise, it is considered non-sensitive. Simultaneously, knowledge graph entity association weights are introduced. Correcting the sensitivity probability: ； in, and Indicates weight; if If the data is positive, it is considered sensitive data; otherwise, it is considered non-sensitive data. For scenarios prone to false positives using traditional methods, the large model compares rule matching with historical cases to output the false positive probability. ,like False alarms are excluded if the value exceeds a set threshold; for scenarios prone to missed alarms, the large model mines entity relationships and outputs the missed alarm correction probability. ,like If the value exceeds the set threshold, it will be classified as sensitive data. Large model combined with sensitive types Based on entity relationships and regulatory requirements, assess the level of sensitive data. The specific process is as follows: Determine whether the data is sensitive. , Indicates sensitivity, Indicate non-sensitive; and define the corresponding assessment indicators and weights for whether it is sensitive: regulatory severity. Entity association strength Scope of the leak ,satisfy The large model scores each indicator; if the score is... , , The total score for sensitivity level is: ； in, The score indicates the severity of the regulation. Indicates the entity association strength score. The score indicates the scope of the leak's impact. Set a score threshold range, and output the sensitivity level based on the sensitivity corresponding to the threshold. ; The large model outputs complete intelligent recognition results and performs structured processing to clarify key information. The specific operations are as follows: The large model outputs results in a preset format, containing four types of core information: Sensitivity judgment conclusion: "belongs to sensitive data" or "does not belong to sensitive data"; Key annotations: specific location and sensitivity type of sensitive data. Sensitivity Level Judgment criteria: cited judgment rules, knowledge graph relationships, and historical cases; Result structuring: converting the unstructured text output of the large model into a structured data format, performing format validation on the structured results to ensure the integrity and accuracy of the output information, and ultimately forming intelligent recognition results directly usable for subsequent result integration. .

6. The sensitive data identification method based on graph retrieval enhancement according to claim 2, characterized in that: Step five specifically includes: The intelligent recognition results output in step four The scattered sensitive data annotation information is precisely mapped to the associated nodes and paths in the sensitive data identification rule knowledge graph constructed in step two, establishing a relationship between the judgment result and the graph structure. The basic information of the sensitive data, the graph association information, and the judgment logic are integrated to form a structured, unified identification result report, including: basic information of the sensitive data, including data content, location index, and sensitive type. Sensitivity Level The system includes: graph association information, such as associated node IDs and names, associated path IDs and sequences, node degrees and path lengths; a judgment logic chain, including referenced rules, cases, and reasoning processes for excluding false positives / false negatives; the original text of the data to be identified and context fragments; a visualization tool to present the judgment logic of sensitive data in a graphical form, intuitively showing the complete chain including entity association, rule matching, and level determination; and a user-interactive visualization interface designed to allow users to view detailed information about sensitive data through clicks, filtering, and other operations.

7. The sensitive data identification method based on graph retrieval enhancement according to claim 2, characterized in that: Step six specifically includes: Build an interactive platform that supports user review and feedback input: The platform consists of 3 core modules, each corresponding to the needs of the entire feedback process: The module for displaying pending review results is used to connect to the sensitive data identification result report output in step five. It displays the core information of each result in a list format, including data content, sensitivity type, level, and judgment basis. It supports filtering by unreviewed / reviewed and sensitivity type. Case labeling module: Provides visual labeling controls, allowing users to select the labeling type for a single result, including false positive cases: non-sensitive data is judged as sensitive; Missed case: Sensitive data was not identified; Case of inaccurate judgment criteria: The type / level is correct, but the criteria used are incorrect; And input feedback through the text box; Feedback Record Management Module: Automatically records key information for each feedback, forming a feedback log. Supports querying historical feedback by time / tag type; sets hierarchical permissions; adds feedback verification rules; performs structured processing on feedback cases collected by the platform, converting them into datasets that can be directly used for knowledge graph and model updates; sets automatic update cycles to trigger the update process of knowledge graph and core model.

8. A sensitive data identification system based on graph retrieval enhancement, characterized in that, The system includes a module for performing a sensitive data identification method based on graph retrieval enhancement as described in any one of claims 1-7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a sensitive data identification method based on graph retrieval enhancement as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a sensitive data identification method based on graph retrieval enhancement as described in any one of claims 1-7.

Images