A cold start method for model identification of sensitive data

By constructing a structured sensitive data knowledge base to record the sources of observations and the basis for annotation, the annotation dependency problem in the cold start phase of sensitive data identification technology is solved, enabling rapid model initialization and improved identification accuracy, and supporting continuous system optimization.

CN123115283APending Publication Date: 2026-11-17NORTHEASTERN UNIV CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611490062.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-23
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing sensitive data identification technologies rely heavily on manual annotation during the cold start phase, and different business systems lack a unified sensitive label and field naming mapping, resulting in identification differences and duplicate annotations, making it difficult to handle complex scenarios.

Method used

By constructing a structured sensitive data knowledge base, recording the sources of observations and the basis for annotation, tracing back and grouping evidence from the same source along the inheritance relationship of the annotation basis, isolating the purpose of samples, and using category knowledge state control to control the cold start of the model, the reliance on manual annotation is reduced.

Benefits of technology

It improves data standard consistency, enhances context awareness and recognition capabilities, reduces reliance on manual annotation during the cold start phase, enables rapid initialization of knowledge-driven models and interpretability of recognition results, and supports continuous system optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115283A_ABST
    Figure CN123115283A_ABST
Patent Text Reader

Abstract

This invention discloses a cold-start method for sensitive data identification, belonging to the field of data security. It involves cleaning, label normalizing, and field alignment of multi-source basic data to form a unified sensitive data standard system and standardized data objects, recording the observation source and labeling basis for each data object; and generating candidate label evidence based on a structured sensitive data knowledge base. For each sensitive category, a category knowledge state is constructed, and the prior reliability is calculated to form the category-level fusion coefficient when the category content prototype is formed; further, the prototype adoption ratio and prior retention strength during model cold start are determined. After the system is running, the affected categories are located based on the labeling basis and the inverted index of the category prototype version, and the prior knowledge is maintained, released, frozen, or re-enhanced, as well as the model is locally updated according to the category, thereby reducing the impact of duplicate contributions from the same source evidence on the cold-start model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data security technology, and specifically to a cold start method for model identification of sensitive data. Background Technology

[0002] With the digital development of government affairs, finance, healthcare, and the industrial internet, various business systems have accumulated massive amounts of structured and unstructured data, including sensitive content such as personal identity information, financial information, medical and health data, and trade secrets. To implement data classification and grading, security auditing, and leakage prevention, it is necessary to first accurately identify sensitive data in database fields, documents, and messages, and configure corresponding categories and levels.

[0003] Existing sensitive data identification technologies can be mainly divided into three categories. The first category is rule-based and feature dictionary-based schemes, which match data using regular expressions, keywords, and field name rules. These are suitable for data with fixed formats but struggle with scenarios involving non-standard field naming and complex semantics. The second category is machine learning or deep learning-based schemes, which train classification models using labeled samples. These have some generalization ability but heavily rely on large-scale, high-quality manually labeled data. The third category is knowledge organization and weakly supervised assistance-based schemes, which generate weakly labeled samples by organizing labels, rules, and dictionaries. However, existing knowledge is usually only used for sample generation and does not fully participate in model initialization.

[0004] The existing technologies still suffer from the following prominent problems: There is a lack of unified mapping for sensitive labels and field names across different business systems; historical rules and field dictionaries are scattered and lack structured association; and the cold start phase relies heavily on manual annotation. Furthermore, existing processing typically only records the data source or final label, without distinguishing where the data content was obtained or the basis for label judgment. When the same field name is used to both form field mappings and trigger field name rules, or when the same rule generates multiple pseudo-labeled samples, multiple derived results may inherit the same annotation basis, yet they are repeatedly used as independent support in candidate label judgment, sample screening, and category prior construction, causing a single basis to gain excessive contribution due to the increased propagation path or the increased number of derived samples. Summary of the Invention

[0005] In view of at least one of the above technical problems, this application provides a model cold start method and apparatus for sensitive data identification. Based on unified standards and structured knowledge organization, the observation source and annotation basis are recorded respectively, the evidence of the same source is traced back and grouped along the inheritance relationship of the annotation basis, the training purpose of the sample, the purpose of the category prototype content and the purpose of the category prior reliability are isolated, and the prior adoption and maintenance in the category content prototype construction and model cold start are controlled by using the category knowledge state.

[0006] This invention provides a cold start method for models aimed at sensitive data identification, comprising: Step 1: Obtain multi-source basic data, parse, clean, label normalize and field align the multi-source basic data to obtain a unified sensitive data standard system and a standardized data object set; set observation source identifiers for original data records, and set labeling basis identifiers, basis version identifiers and upstream basis identifiers for rules, field mappings, historical labels and manual confirmation records; Step 2: Construct a structured sensitive data knowledge base based on the unified sensitive data standard system and standardized data object set. The structured sensitive data knowledge base includes at least tag objects, data entity objects, rule objects, sample objects, observation source objects, and annotation basis objects, and establishes the observation source, basis, and inheritance relationships. Step 3: Based on the structured sensitive data knowledge base, generate candidate sensitive labels and corresponding evidence records for the data objects to be identified; associate each evidence record with the observation source identifier, the current labeling basis identifier, and the upstream labeling basis identifier, and trace back step by step along the upstream labeling basis; group the evidence that ultimately depends on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record into the same labeling basis group to obtain the evidence status corresponding to the candidate label. The evidence status includes independent supporting basis group, independent excluding basis group, and conflict status. Step 4: Based on the independent support basis group, independent exclusion basis group, and conflict state, set the model training purpose, category prototype content purpose, and category prior reliability purpose for the sample objects respectively; Step 5: For each candidate sensitive label, construct a category knowledge state vector. Calculate category content prototype The classification model is then trained to obtain the trained classification model, and the structured sensitive data knowledge base and the trained classification model are deployed to the sensitive data identification system.

[0007] Optionally, the unified sensitive data standard system in step 1 is established in the following way: For each original sensitive category record, its original label name is compared with existing standard sensitive labels in conjunction with the label description, sensitivity level, parent category, applicable business scope and standard source. When the original sensitive category and a certain standard sensitive label meet the preset matching conditions, a mapping relationship between the original sensitive category and the standard sensitive label is established. If no standard sensitive label meets the matching conditions, the original sensitive category is used as a candidate new label, and a new standard sensitive label is generated after preset rules or manual confirmation. Each standard sensitive label contains at least a standard label name and a label definition, and stores label aliases, label levels, sensitivity levels, applicable business scenarios, standard sources, and version information; multiple original sensitive labels with different names from different data sources but representing the same sensitivity category are mapped to the same standard sensitive label.

[0008] Optionally, the field alignment includes: For database fields and business message or interface data items with explicit data item names, compare them with candidate standard field concepts based on object name, object description, data type, de-identified data sample and business scenario to establish a field mapping relationship between the original data object and the standard field concept; The standard field concept is established based on the unified sensitive data standard system, historical field dictionary and confirmed field mapping results. Each standard field concept includes at least a standard field name and stores the field description, field alias, data type, typical data format and applicable business scenario.

[0009] Optionally, step 2, which involves building a structured sensitive data knowledge base, includes: The system generates label objects from standard sensitive labels in the unified sensitive data standard system, standard field objects from standard field concepts, data entity objects from standardized data objects, rule objects from historical rule records, and sample objects from historical annotation records. It also establishes observation source objects and annotation basis objects. The observation source objects record the business system, data container, record location, collection batch, and version. The annotation basis objects record the basis type, basis content or associated object, target label, supported or excluded direction, version, generation method, and upstream basis identifier. Lexical features, format features, statistical features, semantic features, and contextual features are saved as identification feature attributes for the corresponding data entity objects; candidate label evidence is associated with the data entity object, candidate label, observation source object, and annotation basis object, respectively; Establish at least one of the following relationships among knowledge objects: field mapping, category affiliation, rule pointing, applicable scenario, structural association, co-occurrence, business association, observation source, basis, and inheritance, so that candidate labels or sample labels can be traced back to their basic judgment basis along the labeling basis relationship.

[0010] Optionally, the evidence status in step 3 is generated in the following way: For the data object to be identified and candidate tags The evidence records are generated by combining rule matching, field mapping, sample matching, business scenario consistency, entity association, and exclusion conditions. The evidence record includes at least the data object identifier, candidate label identifier, evidence type, observation source identifier, current labeling basis identifier, upstream labeling basis identifier, support or rejection direction, basis version, and basis reliability. By tracing back along the upstream annotation basis of each evidence record, when two pieces of evidence ultimately rely on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record, they are classified into the same annotation basis group; when the label of one piece of evidence is directly generated by another piece of evidence, the two are classified into the same annotation basis group determined by the common upstream basis. For the same labeling basis group for the same candidate label, multiple pieces of evidence in the same direction form a group contribution, which is not directly accumulated according to the number of evidence paths or the number of derived samples; when there are both supporting and rejecting records in the same group, the conflict marker within the group is retained. Formation of Evidence State E i,k The evidence status records at least the set of observation sources, the independent supporting annotation basis group, the independent excluding annotation basis group, the reliability of each basis group, the completeness of the source record, and the conflict status.

[0011] Optionally, in step 4, the sample objects are assigned purposes for model training, category prototype content, and category prior reliability, including: Set a usage vector for each sample ,in Indicates whether participation in model training is permitted. This indicates whether the observed content is allowed to participate in the category prototype construction. The labeling criteria indicate whether adding prior reliability of the category is allowed, and each application is determined separately; When the same candidate label obtains at least two backtrackable and non-inherited supporting labeling criteria, no exclusion criteria that meet the conflict conditions appear, and it can be distinguished from competing labels, the sample is allowed to participate in model training, category prototype content construction, and category prior reliability calculation; samples that have been independently and manually confirmed form a new independent confirmation criterion group. When a sample inherits from only one set of automatically labeled criteria, if the reliability of that criterion meets the training admission criteria and there is no obvious conflict, it is allowed to participate in model training, but its observations are prohibited from entering the category prototype content candidate set and its labeling criteria are prohibited from increasing the category prior reliability; after obtaining a second independent supporting criterion or independent human confirmation, the use determination is re-executed. When supporting evidence and highly reliable exclusion evidence coexist, and multiple competing labels have independent support or uncertain sources affecting label judgment, the data object is set as a manual review object; ultimately, a training sample set T, a category prototype content candidate set P, and a category reliability evidence set V are formed respectively.

[0012] Optionally, in step 5, a category knowledge state vector is constructed. Calculate category content prototype ,include: The tag knowledge vector is obtained by encoding and aggregating the tag name, tag definition, and tag alias. ; will point to sensitive categories The rule name, rule description, keywords, applicable fields, and textualizable rule conditions are encoded and aggregated to obtain a rule knowledge vector. Samples are read from the candidate set P of the category prototype content and grouped and aggregated according to the observation source and common annotation criteria to obtain the sample content vector. ; Read the independent supporting evidence group and the independent excluding evidence group from the category reliability evidence set V, and construct the category knowledge state vector. The It includes at least the following states: knowledge availability, independent confirmation, conflict, observation coverage, and inherited duplication. The prior reliability is calculated using the function Q. ; Based on category knowledge status Determine the category-level nonnegative fusion coefficients for label knowledge, rule knowledge, and sample content, respectively. And normalize the fusion coefficients corresponding to the available knowledge; Sensitive categories Category content prototype Represented as: ; Among them, does the increase or decrease of the number of paths or derived samples generated by the same annotation group increase? When a category content prototype is unavailable, it is marked as unusable for model priors and saved accordingly. and the corresponding version.

[0013] Optionally, step 5 involves training the classification model, including: Category content prototype The transformation T maps to the representation space of the target classification model. For newly established classification models or newly added sensitive categories, the transformation is performed using... Generate initial classification parameters for the corresponding category, where ; The training objectives include a weighted classification loss and a class prior preservation term, wherein the strength of the class prior preservation term is... This makes the same category of knowledge state Both through The knowledge composition of the control category content prototype, and through , and The extent to which the control model adopts and retains the prototype of the content in the aforementioned categories; Upon receiving changes to rules, field mappings, manual confirmations, false alarms, or omissions, the system locates the record inheriting the changed basis using the first inverse index from the annotation basis to the evidence and sample; it then locates the affected category using the second inverse index from the sample to the category prototype version. Finally, it re-groups the annotation basis, determines its usage, and... , and calculate; According to the updated , The system handles conflict states and independent manual confirmation of changes, sets prior control states for affected categories to hold, release, freeze, or re-enhance, and adjusts the prototype scale, prior hold strength, and model parameters only for the affected categories.

[0014] The beneficial effects of adopting the above technical solution are as follows: (1) Improve the consistency of data standards and knowledge management. This invention cleans, normalizes, and aligns fields of multi-source heterogeneous data, mapping the original names to unified standard labels and standard field concepts, effectively avoiding recognition differences caused by non-standard naming or different classification criteria; it transforms scattered historical rules, field dictionaries, business scenarios, and annotation results into structured knowledge objects, and establishes clear relationships such as category affiliation, field mapping, and co-occurrence association, solving the problems of scattered rules, conflicts, and untimely updates.

[0015] (2) Enhance the ability to recognize context and complex relationships. In addition to traditional field name and regular expression matching, this invention introduces semantic features, context features and entity association information; combined with the co-occurrence relationship of fields in the same data table and business scenarios, it can effectively distinguish easily confused concepts such as "account name" and "name", and greatly reduce the false alarm and false negative rates based solely on the words themselves.

[0016] (3) Significantly reduces the reliance on manual annotation during the cold start phase. This invention utilizes a knowledge base to generate candidate labels and integrates confidence scores to automatically filter high-confidence pseudo-labeled samples; it only performs targeted manual verification on data objects with uncertain identification or conflicting evidence; and it specifically constructs difficult negative samples and ordinary negative samples to add to the training set for categories with few samples and easily confused categories, which significantly improves the problem of blurred classification boundaries caused by class imbalance.

[0017] (4) Achieve rapid initialization of knowledge-driven models. This invention encodes and integrates label knowledge, rule knowledge and sample knowledge in a unified manner to generate a "sensitive category prototype" carrying prior information about the category. The classification weights or category centers of the classification model are directly initialized using this prototype, so that the model can have basic category cognition without learning from scratch under the condition of cold start with only a small number of labeled samples, which greatly shortens the model convergence cycle and improves the initial recognition accuracy.

[0018] (5) Improve the interpretability of recognition results. In the process of candidate label generation, the system records the rule support, feature coverage, and judgment evidence. During manual review, the basis for the generation of the label can be directly traced, which improves the transparency of system decision-making and review efficiency.

[0019] (6) Support closed-loop optimization for continuous system evolution. After system deployment, manual review results, false positives / false negatives, and new business requirements can be fed back to the knowledge base. The feedback information is used to update rule reliability, candidate label fusion weights, and sample screening thresholds, and to trigger the regeneration of category prototypes and incremental training of models, enabling the system to flexibly adapt to new sensitive categories and changes in business scenarios, forming a virtuous cycle of "deployment-identification-feedback-optimization". Attached Figure Description

[0020] Figure 1 This is a flowchart of the model cold start method for sensitive data identification in this embodiment; Figure 2 This is a flowchart illustrating the construction process of the structured sensitive data knowledge base in this implementation. Figure 3 This is a flowchart of the candidate label generation and cold start sample construction process in this embodiment; Figure 4 This is a flowchart of the sensitive category prototype generation and model initialization process in this embodiment; Figure 5 This is a flowchart of the feedback update process in this embodiment. Detailed Implementation

[0021] To facilitate understanding of this application, a more comprehensive description of this application will be provided below with reference to the accompanying drawings.

[0022] To address the problems of existing technologies, this invention aims to provide a cold start method for sensitive data identification, which solves the problems of inconsistent sensitive data standards, scattered historical rules and feature dictionaries, difficulty in handling contextual semantics with a single rule, insufficient initial manually labeled samples, unbalanced category distribution, and insufficient participation of existing knowledge in model initialization.

[0023] This invention establishes a unified sensitive data labeling system and standardized data objects by cleaning, aligning, and normalizing sensitive data classification and grading standards, historical rules, field dictionaries, field metadata, historical documents, and data samples. This reduces identification discrepancies between different business systems caused by inconsistencies in field names, label names, and classification criteria. Furthermore, by extracting lexical features, format features, statistical features, semantic features, contextual features, and entity association information from data objects, and uniformly organizing sensitive labels, standard fields, historical rules, data entities, business scenarios, and sample information, the identification features are saved as feature attributes of the corresponding data entities, constructing a structured knowledge base for sensitive data identification. Using this structured knowledge base, candidate labels are generated for unlabeled data, and high-confidence pseudo-labeled samples that meet preset conditions are selected. Data objects with uncertain identification results are submitted for manual review, and samples whose labels are confirmed after review are used as manually confirmed samples. The high-confidence pseudo-labeled samples, combined with a small number of existing manually labeled samples, manually confirmed samples, difficult negative samples, and ordinary negative samples, constitute a cold-start training sample set to alleviate the problems of insufficient manually labeled samples and uneven category distribution in the initial stage. Simultaneously, label knowledge and rule knowledge are formed based on label names, label definitions, label aliases, and rule descriptions. Sample knowledge is formed by combining existing historical confirmed samples, manually confirmed samples, and high-confidence pseudo-labeled samples. The category prototypes corresponding to the sensitive categories are generated through the above knowledge, and the classification weights or category centers of the sensitive data classification model are initialized using the category prototypes. This enables the model to complete initial training and cold start deployment using existing label knowledge, rule knowledge, and sample knowledge under the condition of a small number of labeled samples.

[0024] This invention primarily employs category identification of data objects under a preset sensitive data classification hierarchy. Within the same classification hierarchy, a target sensitive category is determined based on the content, characteristics, and business context of the data object to be identified. When the same original data contains multiple data contents with different sensitive categories, the original data can be divided into multiple data objects to be identified, and each object can be identified according to the method described in this invention.

[0025] Figure 1 This is a flowchart of the cold start method for sensitive data identification in this embodiment. The diagram illustrates the overall process of this invention, starting from the acquisition of multi-source basic data, through the establishment of a unified sensitive data standard system and data standardization, extraction of sensitive data identification features and entity association information, construction of a structured sensitive data knowledge base, generation of candidate sensitive labels, construction of a cold start training sample set, generation of sensitive category prototypes, and initialization and training of the classification model using the category prototypes, to form a structured knowledge base and an initial classification model.

[0026] After the classification model is trained, it is validated using an independent validation sample set. When the preset cold start deployment conditions are met, the structured sensitive data knowledge base, the initial classification model, and the corresponding recognition thresholds are deployed to the sensitive data recognition system. Upon receiving a new data object to be identified, the system obtains the recognition result through both the knowledge base recognition process and the classification model prediction process, and jointly judges the candidate labels from the knowledge base, the predicted labels from the model, the corresponding confidence levels, and the judgment evidence. If the recognition result cannot be directly determined, it undergoes manual review, and finally outputs the sensitive category and the corresponding judgment criteria.

[0027] The manual confirmation results, false alarm results, missed alarm results, as well as information on newly added rules, newly added sensitive tags, and changes in business scenarios generated during system operation, can be further fed back into the feedback update process and applied to the structured knowledge base, cold start sample screening conditions, sensitive category prototypes, and classification models, respectively. The updated knowledge base, classification models, and recognition parameters are redeployed and continue to be used for subsequent sensitive data identification, thus forming an iterative optimization process after the system's cold start deployment.

[0028] Figure 2 This is a flowchart illustrating the construction process of the structured sensitive data knowledge base in this implementation. In addition to knowledge objects such as tags, standard fields, rules, data entities, business scenarios, standard sources, and samples, this implementation further sets up observation source objects and annotation basis objects. Standardized data objects are connected to observation source objects through observation source relationships; rule hits, field mappings, historical tags, manual confirmations, and derived sample tags are connected to annotation basis objects through "based on" and "inherited from" relationships, enabling the data content source and tag judgment source to be recorded separately and traced back along the upstream basis.

[0029] After information parsing, deduplication and cleaning, format unification, label normalization and field alignment, the above multi-source data forms a unified sensitive data standard system and standardized data objects; further, lexical features, format features, statistical features, semantic features, context features and entity association information corresponding to the data objects are extracted.

[0030] The processed tags, standard fields, rules, data entities, business scenarios, standard sources, and samples form corresponding knowledge objects in the structured knowledge base. The resulting lexical features, format features, statistical features, semantic features, and contextual features are stored as identification feature attributes for the corresponding data entity objects. Knowledge relationships such as field mapping, category affiliation, rule pointing, applicable scenarios, knowledge sources, structural associations, co-occurrence, business associations, and areas of confusion are established between knowledge objects. When an identification rule hits a specific data object, the hit result further forms supporting evidence for the corresponding candidate sensitive label or exclusion evidence for excluding the corresponding candidate label.

[0031] The structured sensitive data knowledge base is used for subsequent candidate sensitive label generation, cold start training sample screening, sensitive category prototype generation, classification model initialization, and joint recognition by the knowledge base and classification model after system deployment. Combination Figure 1 and Figure 2 As shown, the model cold start method for sensitive data identification in this embodiment includes the following steps: Step 1: Obtain multi-source basic data, parse, clean, label normalize and field align the multi-source basic data to obtain a unified sensitive data standard system and a standardized data object set; set observation source identifiers for original data records, and set labeling basis identifiers, basis version identifiers and upstream basis identifiers for rules, field mappings, historical labels and manual confirmation records; The multi-source basic data includes category system source data, object source data, and historical rules, historical annotation results, and business scenario information that can assist in the subsequent construction of the knowledge base and the identification of sensitive data. Among them, the category system source data and object source data are the basic inputs for this method to establish a unified sensitive data standard system and form the data objects to be processed; historical rules, historical annotation results, and business scenario information can be obtained according to actual data conditions.

[0032] The category system source data includes one or more of the following: classification and grading standards that provide information such as sensitive data category names, category definitions, category levels, sensitivity levels, or applicable scope; internal enterprise data security policies; and existing sensitive data labeling systems. The internal enterprise data security policies are internal documents used by an enterprise to stipulate requirements for data classification and grading, data protection, or data security management. These policies may include sensitive data category names, category definitions, sensitivity levels, applicable business scope, or data examples. A single category system source may contain multiple sensitive data category records, and different sources may use different category names, category levels, or classification criteria. During acquisition, the category names and related descriptive information from the corresponding source are retained. The object source data is categorized according to the type of data object to be identified: database data, document data, and business messages or interface data. For database data, database fields are used as the data objects to be processed. The field name, field comments, data type, corresponding data table, corresponding business system, and one or more de-identified data samples obtained from the field can be obtained. Original structural information such as primary keys, foreign keys, and inter-table joins can be retained. For document data, text fragments, paragraphs, or other predefined text units are used as the data objects to be processed. The text content, the document it belongs to, the paragraph or chapter it belongs to, and information reflecting its contextual location can be obtained. For business messages or interface data, message data items are used as the data objects to be processed. The data item name, data item description, data type, anonymized data example, the message type it belongs to, and directly obtainable business process information can be obtained.

[0033] A data source can contain multiple data objects to be processed. For example, a database can contain multiple data tables, and a data table can contain multiple fields to be identified; a document can contain multiple text units to be identified; a business message can contain multiple data items to be identified. For database fields or business message data items, the de-identified data samples can include one or more representative data values ​​extracted from the data records corresponding to the same object to be processed and de-identified, for subsequent format feature and statistical feature extraction.

[0034] A historical field dictionary, serving as an aid to standardized data retrieval, can include one or more of the following: existing standard field names, field aliases, field descriptions, data types, typical data formats, and confirmed field mapping relationships. It is used to establish standard field concepts and perform field alignment. A historical field dictionary can contain multiple field records, each corresponding to an existing field concept or field mapping information.

[0035] Historical sensitive data identification rules serve as auxiliary knowledge data acquisition. Each rule must at least record the rule identifier or name that can identify it, as well as the rule content. It can also further record its target sensitive label, applicable data type, applicable business scenario, rule source, and historical usage information. One sensitive label can correspond to multiple historical identification rules, and the same rule can correspond to one or more target sensitive labels according to its existing configuration.

[0036] Historical manual annotation results are associated with specific data objects to be processed. Each historical annotation record corresponds to at least one specific data object and its existing sensitive tags, and can further record the annotation source, whether it has been manually verified, and its corresponding credibility status. A data source can contain multiple data objects with historical annotation results; data objects without historical annotation results can still participate in the subsequent candidate sensitive tag generation as unannotated objects.

[0037] The existing manually labeled samples refer to data objects that already have manually labeled sensitive tags before the execution of this method. Among them, existing manually labeled samples whose tags have been confirmed as valid through manual review, business confirmation, or other credible confirmation processes are called existing historical confirmed samples. Confirmed samples newly formed after manual review during the execution of this method are called manually confirmed samples.

[0038] Business scenario information is acquired as independent contextual information and may include the business scenario name, business type, affiliated business system, and applicable data scope. A single business scenario can be associated with multiple data objects to be processed, and the same data object to be processed can also be associated with one or more business scenarios depending on its source and business purpose. Business scenario information describes the business context in which the data object to be processed exists and is used for subsequent field alignment, contextual feature extraction, rule applicability determination, and candidate sensitive label identification.

[0039] When acquiring the aforementioned basic data, retain the corresponding source identifiers for each data source and the data object to be processed, and record the relevant business system, data type, and version information that can be directly obtained. The obtained multi-source basic data is saved according to the correspondence between the data source and the data object to be processed, serving as input for establishing a unified sensitive data standard system and standardizing the data.

[0040] The acquired multi-source basic data is parsed according to the correspondence between the data source and the data object to be processed, forming original sensitive category records, original data object records, historical rule records, historical annotation records, and business scenario records. Each type of record retains the corresponding data source identifier, ensuring that data from different sources maintains its correspondence with the original data source during subsequent processing.

[0041] For data sourced from a category system, sensitive data categories are extracted from each classification and grading standard, internal enterprise data security policy, or existing sensitive data labeling system. When a category system source contains multiple sensitive data categories, a separate original sensitive category record is created for each category. Each original sensitive category record must at least store the original label name and may also store information obtainable from the original source, such as label description, sensitivity level, parent category, applicable business scope, and standard source. Therefore, a category system source can correspond to multiple original sensitive category records, and the same original sensitive category record corresponds to one sensitive data category within that source.

[0042] For source data, the determined granularity of the data objects to be processed forms raw data object records. Each field to be identified in the database data forms a raw data object record, recording the field name, field comments, data type, corresponding data table, corresponding business system, one or more anonymized data samples, and directly obtainable raw structural information. Each text unit to be identified in the document data forms a raw data object record, recording the content or description of the text unit, the document it belongs to, the chapter or paragraph it belongs to, and the corresponding business scenario. Each data item to be identified in the business message or interface data forms a raw data object record, recording the data item name, data item description, data type, anonymized data sample, corresponding message type, and directly obtainable business process information.

[0043] For each acquired historical identification rule, a historical rule record is created. Each record stores the rule identifier or rule name, rule content, and, based on the information actually obtainable from the original data, the target sensitive label, rule type, applicable data type, applicable business scenario, rule source, and historical usage information. For each acquired historical manual annotation result, a historical annotation record is created. Each historical annotation record is associated with a specific data object to be processed, and information such as the existing sensitive labels, annotation source, manual confirmation status, and trust status of that data object is stored. For each acquired business scenario information, a business scenario record is created. Each business scenario record stores the business scenario name, business type, affiliated business system, and applicable data scope, and retains the correspondence between it and the corresponding data object to be processed, rule, or category system source that can be determined from the original data.

[0044] The text, identifiers, and formatting information in the above records undergoes basic cleaning and format standardization. For text information such as tag names, field names, field comments, rule descriptions, and business scenario names, garbled characters, invalid characters, and duplicate content are removed, and character encoding, whitespace characters, and symbol representations are standardized. For dates, numbers, enumeration values, and other data with fixed representation formats, they are standardized according to preset formats. For tag, rule, or object records with the same source and duplicate content, deduplication is performed based on source identifiers, object identifiers, and content consistency. This cleaning only changes the data representation format and does not alter the original data objects, original tags, or their source relationships.

[0045] After basic cleaning, the original sensitive category records are normalized by labeling. For each original sensitive category record, its original label name, along with the available label description, sensitivity level, parent category, applicable business scope, and standard source, is compared with existing standard sensitive labels. When the original sensitive category and a certain standard sensitive label meet preset matching conditions, a mapping relationship between the original sensitive category and that standard sensitive label is established; when no standard sensitive label meets the preset matching conditions, the original sensitive category is used as a candidate new label, and a new standard sensitive label is generated after preset rules or manual confirmation.

[0046] The above-described label normalization process forms a unified sensitive data standard system. This unified sensitive data standard system includes standard sensitive labels and the hierarchical relationship between them. Each standard sensitive label includes at least a standard label name and a label definition, and can further store label aliases, label levels, sensitivity levels, applicable business scenarios, standard sources, and version information. Multiple original sensitive labels from different data sources, each with a different name but representing the same sensitive category, are mapped to the same standard sensitive label, while preserving the mapping relationship between each original sensitive label and the standard sensitive label.

[0047] Each standard sensitive label contains at least a standard label name and a label definition, and stores label aliases, label levels, sensitivity levels, applicable business scenarios, standard sources, and version information; multiple original sensitive labels with different names from different data sources but representing the same sensitivity category are mapped to the same standard sensitive label.

[0048] Field alignment is performed on database fields and business message or interface data items with explicit data item names to establish a correspondence between original fields or original data items in different business systems and unified standard field concepts.

[0049] The standard field concepts are established based on a unified sensitive data standard system, a historical field dictionary, and confirmed field mapping results. Each standard field concept includes at least a standard field name and can further store field descriptions, field aliases, data types, typical data formats, and applicable business scenarios.

[0050] For each original data object requiring field alignment, the object name, description, data type, anonymized data sample, and applicable business scenario are compared with candidate standard field concepts. The object name is used to match the standard field name or field alias; the object description is used to compare with the standard field description; the data type and directly obtainable data format information are used to determine if the data representation is consistent; and the business scenario is used to determine if the original data object falls within the applicable scope of the standard field concept.

[0051] When the original data object and a certain standard field concept meet the preset mapping conditions, a field mapping relationship is established between the two, and the original data object identifier, standard field identifier, and corresponding mapping relationship are saved. For example, "phone_num", "mobile", "telephone", and "contact number" in different business systems can be mapped to the standard field concept "contact number" respectively.

[0052] For text fragments, paragraphs, or other text units in a document, since they typically do not have fixed field names identical to database fields, the aforementioned field alignment is not required. After basic cleaning, these objects are directly formed into standardized data objects based on their text content, data source, and contextual information. For documents or messages with a clearly defined field-based data item structure, corresponding standard field mappings can also be established according to the aforementioned field alignment methods.

[0053] For text fragments, paragraphs, or other text units in a document, since they typically do not have fixed field names identical to database fields, the aforementioned standard field alignment is not required. After basic cleaning, these objects are directly formed into standardized data objects based on their text content, data source, and context information. For documents or messages with a clearly defined field-based data item structure, corresponding standard field mappings can also be established according to the aforementioned field alignment methods.

[0054] After label normalization and necessary field alignment are completed, each identified data object to be processed is used as a unit to reorganize the information already obtained from the original data object and the standardization results obtained in this step, forming a standardized data object. Each standardized data object has a unique object identifier and maintains a mapping relationship with its corresponding original data object.

[0055] In one implementation, the i-th standardized data object is represented as: D i ={id i ,n i ,d i ,v i ,t i ,s i ,c i} Among them, id i A unique identifier representing a standardized data object; n i Indicates the object name; d i Represents object description; v i This indicates the de-identified data content or de-identified data sample corresponding to the object; t i Indicates data type; s i Indicates the data source; c i It represents the context information associated with the object.

[0056] All information in the standardized data object originates from the original data already obtained, rather than being regenerated through label normalization or field alignment processes. For database fields, the object name can be determined based on the cleaned field name, the object description can be determined based on the field comments, the anonymized data content or anonymized data sample can be obtained from the data record corresponding to the field, the data type can be determined based on the field metadata, the data source can be determined based on the business system, database, and data table, and the context information can include the business scenario, the data table, and related field information.

[0057] For a document text unit, the object name can be determined based on the text unit identifier or name, the object description and desensitized data content can be determined based on the document description and desensitized text content, the data source is used to record the document and its source, and the context information can include the chapter, paragraph, preceding and following text positions, and the business scenario to which it belongs.

[0058] For business messages or interface data items, the object name can be determined based on the data item name, the object description can be determined based on the data item description, the de-identified data content or de-identified data example can be determined based on the corresponding message content, the data type can be determined based on the interface definition or message structure, the data source can be determined based on the business system, interface or message type, and the context information can include the business process node, message structure and business scenario.

[0059] When a certain type of data object to be processed does not contain any of the above-mentioned information, the corresponding attribute can be empty, which does not affect the participation of other obtained information in the standardized data object in subsequent processing. For database fields or business message data items containing multiple de-identified data samples, these multiple de-identified data samples are saved as a sample set corresponding to the standardized data object for the extraction of format features and statistical features.

[0060] The final results include: a unified sensitive data standard system, a set of standard field concepts, a set of standardized data objects, and historical rule records, historical annotation records, and business scenario records that have undergone cleaning and source association. Specifically, a tag mapping relationship is maintained between standard sensitive labels and original sensitive categories; a field mapping relationship is maintained between standardized data objects requiring field alignment and standard field concepts; and an object correspondence relationship is maintained between historical annotation records and corresponding standardized data objects. These results collectively serve as the foundational data for feature and entity association information extraction and the construction of a structured sensitive data knowledge base.

[0061] Step 2: Construct a structured sensitive data knowledge base based on the unified sensitive data standard system and standardized data object set. The structured sensitive data knowledge base includes at least tag objects, data entity objects, rule objects, sample objects, observation source objects, and annotation basis objects, and establishes the observation source, basis, and inheritance relationships. Figure 2 This document illustrates the construction process of the structured sensitive data knowledge base in this embodiment. The unified sensitive data standard system, standardized data object set, historical rules, historical annotation results, and business scenario information obtained in step 1 are used as inputs for knowledge base construction. The standardized data objects undergo feature extraction and entity association analysis, and the corresponding data is converted into knowledge objects with unified identifiers and clear relationships.

[0062] For each standardized data object, lexical features, format features, statistical features, semantic features, and contextual features are extracted for sensitive data identification. Entity association information between different data objects is determined based on the data source, structural context, and business context. Different types of data objects can choose appropriate feature extraction methods based on their actual available information; it is not required that every data object possess all of the above features.

[0063] For standardized data objects in database fields, keywords, synonyms, abbreviations, prefixes, and suffixes are extracted based on field names, field comments, and information that can be converted into text to form lexical features; format features are formed based on the character composition, character length, fixed prefixes and suffixes, combination of numbers and characters, and available format validation rules of the anonymized data samples; when there are multiple anonymized data samples or statistical information is available, the cases of null values, unique values, length distribution, character distribution, and high-frequency value distribution are statistically analyzed to form statistical features.

[0064] For document text units, lexical features, semantic features, and contextual features are extracted based on the anonymized text content, the chapter, the paragraph, and the preset text range before and after it. The preset text range can be set according to paragraphs, sentences, character ranges, or document structure.

[0065] For business messages or interface data items, extract lexical features, format features, semantic features, and context features based on the data item name, data item description, data type, de-identified data sample, message type, business process node, and business scenario.

[0066] For standardized data objects containing textual semantic information, the object name, object description, desensitized text content, or other information that can reflect the semantics of the object are input into the text encoding model to obtain the corresponding semantic representation, and the semantic representation is used as the semantic feature of the corresponding standardized data object.

[0067] For database fields, contextual features are formed based on the data table to which they belong, the business scenario to which they belong, and related fields that are in the same data table as the target field or have primary keys, foreign keys, or inter-table joins. Structural associations are established based on primary keys, foreign keys, and inter-table joins, and co-occurrence associations are established based on the co-occurrence of multiple fields in the same data table.

[0068] For document text units, co-occurrence associations are established based on the common occurrence of different text objects in the same paragraph or within a preset text range; for business messages or interface data items, business associations are established based on the corresponding relationships of different data items in the same message structure, business process node, or business scenario.

[0069] After the above processing, identification feature information corresponding to the unique identifier of each standardized data object, as well as entity association information between different standardized data objects, are formed. This identification feature information and entity association information, together with the standardized data objects, serve as inputs for constructing the structured sensitive data knowledge base.

[0070] Constructing the structured sensitive data knowledge base includes: generating tag objects from standard sensitive tags in the unified sensitive data standard system, generating data entity objects from standardized data objects, generating rule objects from historical rule records, generating sample objects from historical annotation records, and establishing observation source objects and annotation basis objects respectively. Depending on the actual needs of knowledge base construction, standard field objects, business scenario objects, and standard source objects can also be established.

[0071] Each tag object records at least the standard tag name and tag definition, and can further record tag aliases, tag level, sensitivity level, applicable business scenarios, standard source and version information.

[0072] Each standard field object records at least the standard field name and can further record the field description, field alias, data type, typical data format, and applicable business scenarios. The field mapping relationship between the original data objects and standard field concepts established in Step 1 is converted into the field mapping relationship between the corresponding data entity objects and standard field objects in the structured sensitive data knowledge base.

[0073] Each standardized data object corresponds to a data entity object. Database fields correspond to database field entities, document text units correspond to document text entities, and business messages or interface data items correspond to message data item entities. The data entity object stores the object name, object description, anonymized data content or anonymized data sample, data type, data source, and business context, and retains the correspondence between these elements and the unique identifier of the standardized data object. Lexical features, format features, statistical features, semantic features, and contextual features are stored as identification feature attributes for the corresponding data entity object.

[0074] Each historical rule record is converted into a rule object. A rule object records at least the rule identifier or rule name and rule content, and can further record the rule type, target sensitive tags, applicable data types, applicable business scenarios, rule source, rule version, exclusion conditions, and historical usage information. When a rule has been configured with target sensitive tags, a rule-pointing relationship is established between the rule object and the corresponding tag object.

[0075] Each historical annotation record is used to generate an initial sample object. Each sample object is associated with a corresponding data entity object, and existing sensitive labels, annotation sources, manual verification status, and trust status are recorded. When the historical annotation results have confirmed the sensitive category to which the data entity belongs, the label association is saved between the sample object and the corresponding label object.

[0076] The observation source object is used to record the source of data content, and includes at least the business system, data container, recording location, collection batch, and version information. Data entity objects are associated with their corresponding observation source objects through observation source relationships, so that the source of data content can be queried and traced back.

[0077] The labeling criteria are used to record the source of judgment for candidate labels or sample labels, and include at least the criteria type, criteria content or related objects, target sensitive labels, support or rejection direction, criteria version, generation method and upstream criteria identifier.

[0078] Rule hits, field mappings, historical labels, manual confirmations, and derived sample labels are associated with the corresponding labeling reference objects through either a relation or an inheritance relation, so that label judgment can be traced back level by level along the upstream labeling reference.

[0079] At least one of the following relationships is established between the knowledge objects: field mapping, category affiliation, rule pointing, applicable scenario, structural association, co-occurrence, business association, observation source, based on, and inherited from. The category affiliation relationship between standard field objects and tag objects can be determined based on a unified sensitive data standard system, historical field dictionaries, existing manual annotation results, or confirmed correspondences between fields and sensitive tags.

[0080] For existing manually confirmed samples, when the data entities corresponding to two sensitive labels have similar field names, similar data formats, or similar semantic features, but the manual confirmation results belong to different sensitive categories, a confusing relationship can be established between the corresponding label objects for subsequent candidate label conflict judgment and difficult negative sample selection.

[0081] The structured sensitive data knowledge base also sets up a data structure for candidate label evidence records, used to record evidence such as rule hits, field mappings, sample matching, business scenario consistency, entity associations, and exclusion conditions formed during subsequent identification processes. The data structure includes at least a data object identifier, candidate label identifier, evidence type, observation source identifier, current annotation basis identifier, upstream annotation basis identifier, and support or exclusion direction. This step establishes the data structure and relationships of the evidence records; candidate label evidence records for specific data objects to be identified are generated in step 3.

[0082] Through the above-mentioned knowledge object transformation, attribute writing, and relationship establishment, standard sensitive labels, standard fields, data entities, historical rules, business scenarios, historical annotations, standard sources, observation sources, annotation basis, identification features, and entity association information are organized into a structured sensitive data knowledge base with unified identification and clear association relationships.

[0083] The structured sensitive data knowledge base can be implemented using graph databases, relational databases, document databases, or a combination of these. When using a graph database, each knowledge object can be represented as a node, and the relationships between objects can be represented as connecting edges. When using a relational database, knowledge object tables, knowledge relationship tables, observation source tables, and annotation basis tables can be established respectively, and corresponding records can be associated through unique object identifiers.

[0084] Step 3: Based on the structured sensitive data knowledge base, generate candidate sensitive labels and corresponding evidence records for the data objects to be identified; associate each evidence record with the observation source identifier, the current labeling basis identifier, and the upstream labeling basis identifier, and trace back step by step along the upstream labeling basis; group the evidence that ultimately depends on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record into the same labeling basis group to obtain the evidence status corresponding to the candidate label. The evidence status includes independent supporting basis group, independent excluding basis group, and conflict status. Combination Figure 3 For the i-th data object D to be identified i Data standardization is performed as described in step 1, and identification features are obtained as described in step 2. The structured sensitive data knowledge base is queried to obtain standard field objects, rule objects, historical sample objects, business scenario objects, and associated data entities related to the data object Di to be identified, and a candidate sensitive label set Yi is formed within a preset classification level. For the k-th candidate sensitive label in the candidate sensitive label set Yi The evidence records are generated by combining rule matching, field mapping, sample matching, business scenario consistency, entity association, and exclusion conditions. Where j represents the data object to be identified. and candidate sensitive tags The serial number of the evidence record formed. The evidence record. It should include at least the data object identifier, candidate label identifier, evidence type, observation source identifier, current labeling basis identifier, upstream labeling basis identifier, support or rejection direction, basis version, and basis reliability.

[0085] Tracing back along the upstream annotation basis of each evidence record, when two pieces of evidence ultimately rely on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record, they are grouped into the same annotation basis group; when the label of one piece of evidence is directly generated by another piece of evidence, the two are grouped into the same annotation basis group determined by the common upstream basis; for evidence with incomplete source records but whose inheritance relationship can be confirmed, they are grouped according to the most recent common upstream basis; for evidence that cannot be traced back and whose dependency relationship cannot be ruled out, an undetermined source mark is set, it is not included in the number of independent annotation basis, and it is reserved for manual review.

[0086] For the same labeling basis group for the same candidate label, multiple pieces of evidence in the same direction form a group contribution, which is not directly accumulated according to the number of evidence paths or the number of derived samples; when there are both supporting and rejecting records in the same group, the conflict marker within the group is retained. Formation of evidence status The evidence status records at least the set of observation sources, the independent supporting annotation basis group, the independent excluding annotation basis group, the reliability of each basis group, the completeness of the source record, and the conflict status.

[0087] After completing the backtracking of annotation criteria and grouping of evidence from the same source, the annotation criteria group can be used as the unit of fusion calculation to assist in scoring candidate sensitive labels. Multiple pieces of evidence in the same direction within the same annotation criteria group only constitute one group-level contribution and are not repeatedly accumulated due to an increase in the number of evidence paths or derived samples.

[0088] For any given data object to be identified Each candidate sensitive label in the candidate sensitive label set Yi is then scored using auxiliary methods. For simplicity, the index i of the data object to be identified is omitted in the following scoring process. , , , , and Representing the data objects to be identified Sensitive label for the k-th candidate The scores are based on rule support, feature coverage, semantic consistency, contextual consistency, evidence conflict, and fusion confidence.

[0089] The rule support Based on the candidate sensitive tags The independent rule annotation is determined based on the group and the reliability of each group. Multiple rule hits of the same rule version or inherited from the same upstream annotation basis only constitute one group-level contribution.

[0090] The feature coverage Based on the data object to be identified Satisfied identification features and candidate sensitive labels The degree of matching between corresponding reference identification features is determined. Multiple feature matching results that depend on the same field mapping record or the same basic fact do not overlap and form independent contributions.

[0091] semantic consistency S k The semantic similarity between the semantic representation of the data object to be identified and the semantic representation of the candidate labels is determined. The semantic representation of the data object to be identified uses the obtained semantic features. The semantic representation of the candidate labels is obtained from information such as label names, label definitions, and label aliases stored in the knowledge base, using a text encoding method that is the same as or output to the same semantic representation space. Semantic consistency is determined by calculating the semantic similarity between the two; the closer the semantic representation of the data object to be identified is to the semantic representation of the candidate labels, the higher the semantic consistency.

[0092] Context consistency U k Based on the data table, relevant fields, document context, and business scenario or process to which the data object to be identified belongs, and the candidate tags... y k The matching between the applicable business scenarios and related entities is determined; among them, the relevant fields and document context scope are determined in the manner described above. For example, if the field name or data format of the field to be identified is insufficient to determine the sensitive category, but the data table in which it is located, the relevant fields, or the related business entities have a stable association with a certain candidate tag, the corresponding context information can improve the contextual consistency of the candidate tag.

[0093] The degree of conflict of evidence Q k Based on supporting candidate tags y k The degree of conflict between knowledge evidence and knowledge evidence that refutes the candidate label is determined. When a field name, rule, or semantic information supports a candidate label, but the exclusion rule, data format, business scenario, or related entity information is significantly inconsistent with the label, the degree of conflict of evidence for that candidate label is increased. When multiple types of evidence, such as rules, fields, semantics, and context, point to the same conclusion, the degree of conflict of evidence is correspondingly decreased. For exclusion rules with high reliability or evidence that clearly contradicts the current business scenario, a larger contribution to the conflict can be assigned.

[0094] Before fusion computation, rule support, feature coverage, semantic consistency, contextual consistency, and evidence conflict degree are mapped to a unified preset numerical range. In one implementation, the above indicators can be normalized to the [0,1] interval to reduce the impact of different original value ranges of different types of identification evidence on the fusion result.

[0095] Candidate sensitive labels y k Fusion confidence score C k Represented as: ; in, w 1 to w 5 represents the corresponding non-negative fusion weight, the , , , , and Both are for the currently given data object to be identified. Sure.

[0096] The fusion weights are determined based on the effectiveness of different identification evidence in judging sensitive categories. In the cold start phase, where sufficient historical confirmation samples are lacking, a preset initial weight can be used. As the number of manually confirmed samples increases, the corresponding weights can be adjusted based on the identification results of various types of identification evidence in the confirmed samples. For example, if a certain type of rule has a high correct identification rate in manually confirmed samples, the fusion weight corresponding to the rule's support can be increased; if a certain type of evidence frequently causes false alarms, its impact on the candidate label score can be reduced.

[0097] When a candidate label lacks partial knowledge sources, a fusion confidence score is calculated based on the currently available identification evidence. For evidence components that cannot be calculated due to the data source itself not providing the corresponding information, these components do not participate in the current fusion calculation, and the weights of the positive evidence actually participating in the fusion are normalized. If the corresponding information exists but does not support the candidate label, the corresponding supporting component takes a lower value or zero. For example, if a data object lacks available business context information, contextual consistency may not participate in the current score; if a business context exists but does not match the applicable scenario of the candidate label, this situation is treated as lower contextual consistency or conflicting evidence.

[0098] After scoring each candidate label, they are sorted from highest to lowest fusion confidence score, and the candidate labels, corresponding fusion confidence scores, and knowledge evidence used for judgment are output. Within the same classification level, the candidate label with the highest score is denoted as y1, and its fusion confidence score is denoted as C1. When there are two or more valid candidate labels in the candidate label set, the candidate label with the second highest score is denoted as y2, and its fusion confidence score is denoted as C2. For data objects that form only one valid candidate label, y1, C1, and their supporting and excluding evidence are used as inputs to step 4; for data objects that form two or more valid candidate labels, y1, C1, y2, C2, and their corresponding judgment evidence are used as inputs to step 4. When there are no rules, standard field mappings, sample matching, or other knowledge evidence in the knowledge base that can form valid candidate sensitive labels, it is recorded that the current data object has not formed a valid candidate label, and this result is passed to step 4 for processing as temporarily unlabeled.

[0099] The above-mentioned weighted fusion method based on rule support, feature coverage, semantic consistency, contextual consistency and evidence conflict is one implementation method. Weighted voting, logistic regression, tree model or other methods that can integrate multiple identification evidence can also be used to determine the comprehensive score of candidate sensitive labels.

[0100] Step 4: Based on the independent support criteria group, independent exclusion criteria group, and conflict state, assign model training purpose, category prototype content purpose, and category prior reliability purpose to the sample objects, including: Based on the candidate labels and evidence status, the purpose of model training, the purpose of category prototype content, and the purpose of category prior reliability are determined for each data object to be processed, and a purpose vector is set. Each application is determined independently, and the magnitude of the training contribution is determined by the training weights. To be indicated separately.

[0101] When a candidate label receives at least two backtracking supporting annotation sets that are not inherited, no conflicting exclusionary criteria appear, and it can be distinguished from competing labels, the sample is allowed to participate in model training, category prototype content construction, and category prior reliability calculation. Samples that have been independently verified by humans form new sets of independent verification criteria.

[0102] When a sample comes from a new observation record but its label is inherited from only one set of automatic labeling criteria, the sample can be allowed to participate in weighted training as long as the criteria meet the training admission conditions and there is no obvious conflict. However, its observation content is prohibited from entering the category prototype content candidate set, and its labeling criteria are prohibited from increasing the category prior reliability. Only after obtaining a second independent supporting criterion or independent human confirmation will the usage determination be re-executed.

[0103] When independent supporting annotation criteria groups and high-reliability independent exclusion annotation criteria groups coexist, multiple competing labels all have independent support, or the source uncertainty affects the candidate label judgment, the data object to be identified is set as a manually reviewed object. Before the manual review is completed, the data object to be identified is prohibited from being used for model training, category prototype content, or category prior reliability.

[0104] The final result is a training sample set T, a candidate set of category prototype content P, and a set of category reliability criteria V. T contains only samples permitted for training purposes, P contains only samples permitted for prototype content purposes, and V stores only independent annotation criteria groups permitted for reliability purposes. Difficult negative samples and ordinary negative samples can be added to T according to the configuration, but they are not used as positive prototype content or reliability support for the target category.

[0105] If the data object to be identified does not match a valid rule, and its fields, semantics, and contextual information have a low degree of matching with each candidate sensitive label, thus failing to form a candidate label with clear identification basis, then the data object will not be assigned a sensitive label for the time being, nor will it be directly added to the cold start training sample set. The data object can be retained as data to be processed and re-identified after the knowledge base, identification rules, or manually confirmed samples are updated.

[0106] To enhance the model's ability to distinguish between easily confused categories, difficult negative samples are constructed based on "easily confused" relationships in the structured knowledge base, as well as existing historical and human-confirmed samples. When a data object has high similarity to the target sensitive category in terms of field name, data format, semantic representation, or contextual features, but its human-confirmed result or the sensitive label corresponding to an existing historical confirmed sample belongs to another category, this data object is considered a difficult negative sample corresponding to the target sensitive category. For example, the "Account Name" field may be similar to a natural person's name in character form and field value, but it represents an account identifier in field comments, the data table it belongs to, and the business context; therefore, it can be considered a difficult negative sample for the "Name" category.

[0107] Ordinary negative samples are selected from data objects with clearly defined category labels. For the target sensitive category y... k When the manual annotation results, manual confirmation results, or existing historical confirmation labels of a certain data object do not belong to the target sensitive category y kWhen a data object has a low degree of matching with the target category in terms of rules, features, semantics, and context, it can be used as a regular negative sample corresponding to the target sensitive category. Hard negative samples are used to represent data with features similar to the target category but different from the actual category, while regular negative samples are used to represent data with significant differences from the target category and whose actual category has been determined. Both types of negative samples are used together to assist the model in establishing category boundaries. The final cold-start training sample set includes high-confidence pseudo-labeled samples, existing manually labeled samples, and manually confirmed samples formed after manual review. Hard negative samples and regular negative samples can be added according to the actual data conditions. Data objects awaiting manual review but whose labels have not yet been confirmed are not directly added to the model training sample set.

[0108] Each training sample record must at least save the corresponding data object, training label, and sample source; high-confidence pseudo-labeled samples must simultaneously save the candidate label fusion confidence and judgment evidence; manually confirmed samples must save the verified confirmation label and verification result; difficult negative samples and ordinary negative samples must both record their corresponding target category and actual category. This information is used in subsequent model training to distinguish sample sources and determine the training weights for different types of samples.

[0109] Step 5: For each candidate sensitive label, construct a category knowledge state vector. Calculate category content prototype The classification model is then trained to obtain the trained classification model. The structured sensitive data knowledge base and the trained classification model are then deployed to the sensitive data identification system, including: Figure 4 The category knowledge state is shown. Prior reliability Category Content Prototype And the relationship between the model cold start. For the k-th sensitive category Each will form a tag knowledge vector. Rule knowledge vector and sample content vector And construct the category knowledge state based on the category reliability criterion set V. .

[0110] Sensitive categories The tag name, tag definition, and tag alias are organized into tag description text according to a preset format, and the corresponding tag semantic representation is obtained through a text encoding model; when there are multiple tag description texts, the tag semantic representations are aggregated to obtain a tag knowledge vector. .

[0111] Will point to sensitive categories The rule name, rule description, keywords, applicable fields, and text-based rule conditions are organized into rule description text, and the corresponding rule semantic representation is obtained through a text encoding model. Rule records that depend on the same rule version or the same upstream annotation basis are grouped together, ensuring that multiple rule records within the same annotation basis group do not duplicate each other and form independent reliability contributions as the number of records increases. The usable rule semantic representations after grouping are aggregated to obtain the rule knowledge vector. .

[0112] Read the sensitive categories from the candidate set P of category prototype content. The corresponding sample. For the i-th sample, it is only if its prototype content purpose satisfies... Only when certain conditions are met are their observations allowed to participate in category prototype construction. The admission samples are grouped and aggregated according to their observation source and common annotation criteria to obtain sample content vectors. High-confidence pseudo-labeled samples can only participate in the construction of the sample content vector after the purpose of the prototype content has been determined.

[0113] Label knowledge, rule knowledge, and sample content are encoded using a shared parameter method to obtain corresponding vectors, ensuring that different knowledge vectors reside in the same category representation space. When different knowledge sources use different encoding methods, they are transformed into the category representation space of the target classification model through a pre-defined representation mapping, making each knowledge vector compatible in terms of dimension and semantic representation.

[0114] Read the sensitive categories from the category reliability criteria set V. The corresponding independent support annotation basis group and independent exclusion annotation basis group are combined with knowledge availability status, independent confirmation status, conflict status, observation coverage status, source record completeness status, and inheritance duplication status to construct a category knowledge state vector. .

[0115] From category knowledge state vector The prior reliability is calculated using the function Q. ,Right now: =Q

[0116] The prior reliability This indicates the usability of the current category content prototype as a model prior. Regardless of the number of evidence paths or derived samples generated by the same annotation basis group, it only constitutes one contribution to the reliability of the category prior; new independent observations can change the sample content vector. However, the prior reliability does not automatically increase with the increase in the number of observation records. .

[0117] Based on category knowledge state vector Determine the tag knowledge vector respectively Rule knowledge vector and sample content vector Corresponding category-level nonnegative fusion coefficient , and ,Right now: ; For available knowledge sources, the category-level fusion coefficients satisfy: , , ; + + =1; When a certain sensitive category lacks corresponding rule knowledge, the fusion coefficient corresponding to the rule knowledge will be used. Set to zero; when no sample meets the category prototype content usage admission criteria, the fusion coefficient corresponding to the sample content is set to zero. Set to zero; when other knowledge sources are unavailable, set their corresponding fusion coefficients to zero, and renormalize the fusion coefficients corresponding to the remaining available knowledge sources.

[0118] Sensitive categories Category content prototype Represented as: ; in, This indicates that the category representations obtained through fusion are normalized to make the category content prototypes more accurate. It is compatible with the category representation used by the target classification model.

[0119] As the number of independent observations that meet the criteria for inclusion in the category prototype content usage increases, and as the number of independent manually confirmed or independently supported annotation groups increases, the updated category knowledge status will be updated accordingly. Adjust the category-level fusion coefficients corresponding to the sample content An increase in the number of evidence paths or derived samples generated from the same annotation group does not directly improve the category-level fusion coefficient. or prior reliability .

[0120] When the label knowledge vector, rule knowledge vector, and sample content vector are all unavailable, or when the current category knowledge state reaches the condition that prohibits the use of model priors, the corresponding category content prototype is marked as unavailable for model priors.

[0121] Finally, the sensitive category identifier, category knowledge state vector, prior reliability, category content prototype, and corresponding version are associated and stored, namely: ( ; in, Indicates sensitive categories The current category knowledge state, prior reliability, and version of the category content prototype.

[0122] In one implementation, function Q comprehensively calculates prior reliability based on independent support labeling basis group reliability, independent confirmation status, independent exclusion basis, competing label conflicts, observation source integrity, and inherited duplication status. The function F calculates the category-level fusion coefficients corresponding to label knowledge, rule knowledge, and sample content based on the availability of label knowledge, the reliability of rule knowledge, the independent confirmation of samples, and the observation coverage.

[0123] When a newly added sensitive category only has a label name and label definition, an initial category prototype can be formed based on the label knowledge vector; as subsequent rule knowledge and confirmed samples increase, the category prototype is updated according to the aforementioned fusion method.

[0124] When a certain sensitive category lacks sufficient samples, the label hierarchy in a structured knowledge base can be used to supplement the category prototype. If category y k For a subcategory of a higher-level sensitive label, the category prototype corresponding to its higher-level label is read and fused with the current category prototype according to a preset ratio. This allows subcategories with fewer samples to inherit some stable semantic information from their higher-level counterparts. As the number of manually confirmed samples for a subcategory increases, the weight of the higher-level category prototype in the fusion result is gradually reduced.

[0125] Based on the data representation method of the target classification model, the category content prototype is transformed by T. Mapping to the model representation space. For newly established classification models or new sensitive categories that need to be added, generate the initial classification parameters for the corresponding categories using the adapted category content prototypes.

[0126] In one implementation, ,in Based on prior reliability and category knowledge status Jointly determined; when the category content prototype is unavailable or meets the freezing conditions, Set to zero.

[0127] The model is trained using the training sample set T obtained in step 4. The training objective may include a weighted classification loss and a class prior preservation term. ,in It is determined by the knowledge state of the same category and its prior reliability.

[0128] Therefore, the same category of knowledge status Both through Control Category Content Prototype The knowledge composition, and through , and The control model adopts and maintains the prototype to a certain extent; for existing classes that have been trained and are in operation, the initial parameter settings are not repeated when feedback is received, but the prototype reference, sample weights, or prior maintenance strength are updated.

[0129] For existing sensitive categories that have been trained and are in operation, the initial parameter settings are not repeated when receiving feedback. Instead, the prototype reference, sample training weights, or prior retention strength are adjusted based on the updated category knowledge state, category content prototype, and prior reliability.

[0130] During training, sample training weights are set based on sample source, intended use, and label credibility. Manually labeled samples and samples that have been manually verified and whose intended use for model training is permitted have clear sensitive labels and can be assigned higher training weights.

[0131] For high-confidence pseudo-labeled samples permitted for model training, corresponding training weights are determined based on the fusion confidence of candidate sensitive labels, allowing samples with higher fusion confidence to have a greater impact on model parameter updates. High-confidence pseudo-labeled samples not permitted for model training are not added to the training sample set.

[0132] Difficult negative samples are used in training according to the recorded target category and the actual category to enhance the model's ability to distinguish easily confused sensitive categories; ordinary negative samples are used to supplement non-target data that differ significantly from the target sensitive category.

[0133] After the model training is completed, samples that did not participate in the training of the current model parameters are selected from the existing manually labeled samples or manually confirmed samples to form an independent validation sample set, and the classification model is validated using the independent validation sample set.

[0134] Based on the sensitive category identification results of the classification model on the independent validation sample set, and the false positives and false negatives of different sensitive categories, it is determined whether the current classification model has met the preset cold start deployment conditions. When the cold start deployment conditions are met, the trained classification model is determined as the initial classification model of the sensitive data identification system and deployed together with the structured sensitive data knowledge base.

[0135] After deploying the structured sensitive data knowledge base and the initial classification model to the sensitive data identification system, this invention further utilizes the knowledge base identification process and the classification model prediction process to identify the data objects to be identified, and combines the candidate labels from the knowledge base and the predicted labels from the model to determine the final identification result, including: When the knowledge base does not form effective candidate sensitive labels, if the predicted label output by the classification model meets the corresponding model recognition conditions, the predicted label of the model will be used as the sensitive recognition result of the current data object; if the prediction result of the classification model does not meet the corresponding model recognition conditions, the data object will be marked as an object to be manually reviewed. When the knowledge base generates valid candidate sensitive labels that match the predicted labels output by the classification model, if at least one of the fusion confidence of the candidate labels in the knowledge base and the model prediction confidence reaches its corresponding recognition condition (i.e., one type of recognition result reaches the corresponding recognition condition, while the other type of recognition result does not, but both point to the same sensitive label and there is no clear exclusion evidence or other conflicting evidence), then the sensitive label that both types of results point to can be determined as the sensitive recognition result of the current data object based on the type of result that reaches the recognition condition; if neither type of result reaches the corresponding recognition condition, then it is marked as an object awaiting manual review. When the candidate labels in the knowledge base are inconsistent with the labels predicted by the model, the fusion confidence of the candidate labels in the knowledge base and the confidence of the model prediction are judged to see if they meet their respective recognition conditions, and the processing result is determined by combining the judgment evidence of the candidate labels; if neither of the two results meets the corresponding recognition conditions, the manual review process is initiated. The candidate labels in the knowledge base and the prediction results of the model are each assigned a corresponding recognition threshold to determine whether their respective recognition results are credible enough to be directly involved in the final decision.

[0136] When a candidate label from the knowledge base meets the corresponding recognition conditions and there is supporting evidence such as highly reliable rules, clear field mapping relationships, or evidence consistent with the current business scenario, but the model prediction result does not meet the corresponding recognition conditions, the candidate label from the knowledge base can be used. When the model prediction result meets the corresponding recognition conditions, but the candidate label from the knowledge base does not meet the corresponding recognition conditions or is mainly formed by low-reliability rules and weak matching evidence, the model prediction label can be used.

[0137] When both the knowledge base results and the model results meet their respective recognition criteria but point to different sensitive categories, the final label is not determined solely by their confidence scores. Instead, a conflict assessment is conducted by combining supporting evidence, excluding evidence, field mapping results, contextual information, and model prediction results corresponding to the candidate labels. If a clear conclusion still cannot be reached, the data object is marked as awaiting manual review. If neither result meets the corresponding recognition criteria, the process also proceeds to manual review.

[0138] Both the knowledge base candidate labels and the model prediction results are assigned corresponding recognition thresholds to determine whether their respective recognition results reach a level of credibility suitable for direct participation in the final decision. For knowledge base candidate labels, the knowledge base recognition threshold can be set based on the fusion confidence level; for the classification model, the model recognition threshold can be set based on the model prediction confidence level. Both the knowledge base and model sides can apply a uniform threshold to all sensitive categories, or they can set category-specific thresholds based on the sample distribution of different sensitive categories, category confusion, and the impact of false positives and false negatives. These thresholds can be determined based on existing manually confirmed or verified samples, with preset initial values ​​used during the initial cold start phase when sufficient samples are lacking.

[0139] The system saves the final identification labels corresponding to the data objects to be identified, determines the sensitivity level of the data object based on the sensitivity level recorded for the corresponding label object, and records the candidate labels in the knowledge base and their fusion confidence, model prediction results, hit rules, and other judgment evidence. For data objects that have been manually reviewed, the manual confirmation results are also saved as feedback information for subsequent updates to the knowledge base and classification model.

[0140] Figure 5 This is a flowchart of the feedback update process based on the inheritance relationship of annotation criteria in this embodiment. The flowchart shows the processing steps of feedback classification, reverse index positioning, recalculation of annotation criteria group, re-determination of sample usage, recalculation of category knowledge state and prior reliability, category content prototype update, prior control state transition, and local model update.

[0141] The system categorizes operational feedback into new observations, independent confirmations, non-independent confirmations, false alarms, missed alarms, rule modifications, new tags, and changes in business scenarios, and updates the observation source object, annotation basis object, conflict status, or knowledge object version accordingly.

[0142] Establish a first inverse index from the labeling basis to the evidence record and sample record, and a second inverse index from the sample record to the category prototype version; when the rule version, field mapping, historical labels or manual confirmation results change, locate the evidence and sample that inherit the basis of the change through the first inverse index, and then locate the affected category through the second inverse index.

[0143] For affected records, the annotation criteria based on grouping and sample usage determination are re-executed. Prototype content usage or reliability usage that no longer meets the conditions is revoked, and the sample content vectors of the affected categories are recalculated. Category knowledge status Prior reliability and category content prototype .

[0144] For each sensitive category, set prior control states, including at least hold, release, freeze, and re-enhancement. A new independent, manually confirmed statement that aligns with the current category knowledge can enter the release state; a new, highly reliable conflicting statement, whose source cannot be traced back and affects label judgment, or whose independent confirmation result significantly contradicts the current category knowledge, can enter the freeze state; and a new, enhanced state can enter after subsequent conflict resolution and restoration of independent support.

[0145] The state transition applies only to the affected categories. In the release state, the category-level prior retention strength is progressively reduced. Maintain the current state while keeping it in place. In the frozen state, the current category prototype version is retained for retrospection, but its prior holdovers are suspended from participating in model updates; in the re-enhanced state, the model is updated based on the restored version. and Recalculate the target retention strength.

[0146] For manually verified samples, they are associated with corresponding data entities and sensitive tags, and added to the structured sensitive data knowledge base in the form of the sample objects. When the manual verification result is inconsistent with the original rule judgment result, the reliability, applicable business scenarios, or exclusion conditions of the corresponding rules are adjusted according to the historical hit rate of the rules and the manual verification result. Newly added identification rules are added to the knowledge base in the form of rule objects, and their target sensitive tags and applicable scope are recorded.

[0147] For newly added sensitive tags, their tag name, tag definition, tag level, sensitivity level, applicable business scenario, and standard source are determined according to the tag normalization method described above. The confirmed new tags are then added to the unified sensitive data standard system. Subsequently, corresponding tag objects are generated according to the described method and written into the structured sensitive data knowledge base. When a new tag temporarily lacks rule knowledge or sample knowledge, an initial category prototype is formed based on the currently available tag knowledge, according to the missing knowledge handling method described above, and is gradually updated as subsequent rules and confirmed samples are added.

[0148] When new sensitive labels need to be included in the online identification scope of the sensitive data identification system, the category representation of the model is expanded accordingly based on the discrimination structure of the classification model adopted. For classification models that use classification weights for category discrimination, a category output unit corresponding to the new sensitive label is added to the model's classification layer, and the classification weights corresponding to the new category are initialized using the category prototype of the new sensitive label. For prototype classification models that use category centers for discrimination, the category prototype of the new sensitive label is added to the existing category prototype set as a new category center. After completing the category expansion, incremental training is performed using newly added manually confirmed samples, high-confidence pseudo-labeled samples, or other training samples that are permitted for model training purposes.

[0149] When business scenarios change, update the corresponding business scenario object and the "applicable" relationship between it and tags, rules, and standard fields, and recalculate the context recognition basis for the affected candidate tags accordingly.

[0150] As the number of independently verified samples increases, the recognition effect of different types of identification evidence on manually verified samples can be statistically analyzed according to the annotation basis group. The reliability of the corresponding evidence type or annotation basis group and the candidate label scoring parameters can be adjusted according to the statistical results. Multiple pieces of evidence generated by the same annotation basis group will not be counted repeatedly due to the increase in the number of evidence paths.

[0151] High-confidence pseudo-labeled samples are manually sampled for verification or subsequent results are validated. The first threshold, second threshold, or the screening threshold corresponding to the category are adjusted based on the confirmation accuracy, mislabeling situation, and confidence distribution of different sensitive categories. If a certain category has a large number of mislabeled samples, the corresponding screening conditions can be increased; if a certain category has a long-term shortage of effective samples, the corresponding screening conditions can be appropriately adjusted based on the existing manual confirmation results and mislabeling situation of that category to supplement the candidate samples that can be used for training.

[0152] When the label definition, label level, rule knowledge, or sample knowledge changes, the affected sensitive category prototype is regenerated in the manner described above; after independent manual confirmation or changes in the independent supporting labeling basis group, the category knowledge state G is updated. k The category-level fusion coefficients are then redefined using F(Gk). The updated category prototypes are used to update the classification weights or category centers of the corresponding categories.

[0153] For newly added manually confirmed samples and high-confidence pseudo-labeled samples that have been re-screened or updated, the classification model can be incrementally trained or periodically retrained in the manner described above. Based on the updated validation results, the recognition thresholds corresponding to the candidate labels in the knowledge base and the model prediction results can also be adjusted.

[0154] Through the aforementioned feedback updates, manual confirmation results, rule changes, and new business data can be sequentially applied to the knowledge base, candidate label generation, training sample screening, category prototypes, and classification models, forming a closed-loop optimization process after the sensitive data identification system is deployed. The feedback updates are a preferred implementation method and are not necessarily a necessary step for completing the initial structured knowledge base construction and model cold start.

[0155] Specific Implementation: In one exemplary implementation, the data to be processed is a customer information table in business system A, with the target field name being "mobile". The system encapsulates this field as a data object. The data is recorded separately, including the observation source identifier, field mapping, rule hit, and the labeling basis for the label judgment.

[0156] The field dictionary maps "mobile" to the standard field "contact number", following the field naming rules. The string "mobile" also matches the "personal contact information" tag. The system traces back along the upstream basis of both; if both field mapping and rule matching use the same field name as the starting point for tag judgment, they are grouped into the same tagging basis group. It is not calculated as two separate supports. The de-identified field values ​​further independently satisfy the mobile phone number format verification rules. At that time, another set of annotations is formed based on field value observations. .

[0157] when and When all conditions are met and there is no highly reliable exclusionary evidence or competing label conflict, It can be set for training purposes, prototype content purposes, and reliability contribution purposes. If subsequent contact number field records are obtained from different observation sources, their labels are all governed by the same rule. If automatically generated, these records can participate in model training according to restricted training weights, but they will not enter the category prototype content candidate set or increase the category prior reliability until a second independent supporting evidence or independent human confirmation is obtained.

[0158] For fields such as "account name" that may be confused with natural persons' names, when different candidate labels have independent supporting evidence, they are set as objects for manual review. After the reviewers independently confirm based on the original data, label definitions and business context, new manual confirmation labeling evidence is formed and the evidence grouping and sample usage determination are re-executed.

[0159] The system generates sample content vectors based on samples allowed to participate in prototype content construction, and constructs category knowledge states based on independent support, independent confirmation, conflict criterion, observation coverage, and inheritance / repetition. ;Depend on calculate It also determines the fusion coefficients of tag knowledge, rule knowledge, and sample content to form a category content prototype. The aforementioned Parameter initialization used for creating a category for the first time or adding a new category. and The initialization ratio and the prior maintenance strength during training are jointly determined.

[0160] After the system is running, if the rules Modified or revoked, inheritance is located through the first reverse index. Evidence and samples are used to locate the affected category through a second inverted index, re-determine the purpose of the samples, and update the data. , , And the corresponding prior control states and model parameters.

[0161] The key point of this invention is: First, the sources of data content and label judgments are recorded separately. Standardized data objects save the observation source identifier, while candidate label evidence and sample labels save the labeling basis identifier, basis version, and upstream basis, enabling the data content source and label judgment source to be traced back independently.

[0162] Second, by tracing back the labeling basis inheritance chain along the field mapping, rule hit, historical label, sample label and manual confirmation relationship, evidence that ultimately depends on the same basic facts, the same rule version or the same upstream label record is grouped into the same labeling basis group. Evidence in the same group will not repeatedly form independent reliability contributions due to the increase in the number of paths or derived samples.

[0163] Third, the samples are set for model training purposes, category prototype content purposes, and category prior reliability purposes respectively; when only new observations are added without new independent annotation basis, their contribution to training content can be retained, but they will not be automatically upgraded to prototype content or added reliability support.

[0164] Fourth, construct category knowledge states for each sensitive category. Prior reliability and category content prototype , and by Determine the category-level fusion coefficients for the three types of knowledge, so that the category prototype content and prior reliability adopt different data relationships and counting methods.

[0165] Fifth, utilize the same category of knowledge state to simultaneously control the cold start of category content prototypes and models; pass Decide The knowledge composition, and through , and The control model adopts the initialization method of the prototype and the degree of training retention.

[0166] Sixth, by labeling the basis to evidence and samples, and the reverse index from samples to category prototype versions, the affected objects are located when feedback arrives, the basis is regrouped, the sample usage is determined, the category status is calculated and the model is updated, and the prior is maintained, released, frozen or re-enhanced according to the category.

[0167] The advantages achieved by this invention are: First, by separately preserving the observation sources and labeling criteria, the source, attribution, and inheritance relationships between data objects, evidence, samples, and categories can be queried and traced back.

[0168] Second, by grouping the rule hits, field mappings, evidence paths, and derived samples that share upstream basis into the same label basis group, the class prior reliability of the same label judgment basis is limited from being repeatedly increased due to propagation path or sample replication.

[0169] Third, by setting separate purposes for training, prototype content, and reliability contribution, the field values, context, and business scenario information in different observation records are preserved, while preventing samples inherited from the same label from being mistakenly identified as new independent reliability support.

[0170] Fourth, through category knowledge states Simultaneously controlling the category prototype fusion coefficient and the degree of model prior adoption, the category content representation and category prior reliability are calculated separately, avoiding the determination of the credibility of the category prototype solely based on the number of samples or knowledge items.

[0171] Fifth, by using reverse indexes and prior control states by category, only the affected categories are located and updated when rules, manual confirmation, false alarms, false negatives, or changes in business data change, thus avoiding the need to perform a complete recalculation for unrelated categories.

[0172] Sixth, for the first establishment of a category or the addition of a new category, the category content prototype provides an initial category representation, and the category knowledge state and prior reliability control its adoption ratio and training constraints; as independent confirmation and conflict change, the system can perform hold, release, freeze or re-enhancement by category, so that the prior knowledge control can be adjusted locally and reversibly.

[0173] Seventh, improve the interpretability of cold start sample screening and identification results. In the preferred implementation using fusion confidence calculation, the system can comprehensively screen pseudo-labeled samples based on rule support, feature coverage, semantic consistency, contextual consistency, and evidence conflict, and record candidate labels, hit rules, supporting features, and judgment criteria to provide a reference for manual review.

[0174] Eighth, support subsequent optimization after the system completes cold start deployment. In a preferred embodiment, the results of manual review, false positives, and false negatives can be used to adjust rule reliability, supplement training samples, regenerate category prototypes, and update model parameters, enabling the system to adapt to new business data and new sensitive categories.

[0175] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. For any equivalent substitutions or modifications that can be made by those skilled in the art based on the disclosure of the present invention, the scope of protection of the patent right of the present invention shall be determined by the scope defined in the claims.

Claims

1. A cold start method for models oriented towards sensitive data identification, characterized in that, include: Step 1: Obtain multi-source basic data, parse, clean, label normalize and field align the multi-source basic data to obtain a unified sensitive data standard system and a standardized data object set; set observation source identifiers for original data records, and set labeling basis identifiers, basis version identifiers and upstream basis identifiers for rules, field mappings, historical labels and manual confirmation records; Step 2: Construct a structured sensitive data knowledge base based on the unified sensitive data standard system and standardized data object set. The structured sensitive data knowledge base includes at least tag objects, data entity objects, rule objects, sample objects, observation source objects, and annotation basis objects, and establishes the observation source, basis, and inheritance relationships. Step 3: Based on the structured sensitive data knowledge base, generate candidate sensitive labels and corresponding evidence records for the data objects to be identified; associate each evidence record with the observation source identifier, the current labeling basis identifier, and the upstream labeling basis identifier, and trace back step by step along the upstream labeling basis; group the evidence that ultimately depends on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record into the same labeling basis group to obtain the evidence status corresponding to the candidate label. The evidence status includes independent supporting basis group, independent excluding basis group, and conflict status. Step 4: Based on the independent support basis group, independent exclusion basis group, and conflict state, set the model training purpose, category prototype content purpose, and category prior reliability purpose for the sample objects respectively; Step 5: For each candidate sensitive label, construct a category knowledge state vector. Calculate category content prototype The classification model is then trained to obtain the trained classification model, and the structured sensitive data knowledge base and the trained classification model are deployed to the sensitive data identification system.

2. The cold start method for model identification of sensitive data according to claim 1, characterized in that, The unified sensitive data standard system in step 1 is established in the following way: For each original sensitive category record, compare it with existing standard sensitive labels based on its original label name and in combination with label description, sensitivity level, parent category, applicable business scope and standard source. When the original sensitive category and a certain standard sensitive label meet the preset matching conditions, a mapping relationship between the original sensitive category and the standard sensitive label is established; when there is no standard sensitive label that meets the matching conditions, the original sensitive category is used as a candidate new label, and a new standard sensitive label is generated after preset rules or manual confirmation. Each standard sensitive label contains at least a standard label name and a label definition, and stores label aliases, label levels, sensitivity levels, applicable business scenarios, standard sources, and version information; multiple original sensitive labels with different names from different data sources but representing the same sensitivity category are mapped to the same standard sensitive label.

3. The cold start method for model identification of sensitive data according to claim 1, characterized in that, The field alignment includes: For database fields and business message or interface data items with explicit data item names, compare them with candidate standard field concepts based on object name, object description, data type, de-identified data sample and business scenario to establish a field mapping relationship between the original data object and the standard field concept; The standard field concept is established based on the unified sensitive data standard system, historical field dictionary and confirmed field mapping results. Each standard field concept includes at least a standard field name and stores the field description, field alias, data type, typical data format and applicable business scenario.

4. The cold start method for model identification of sensitive data according to claim 1, characterized in that, Step 2, which involves building a structured sensitive data knowledge base, includes: The system generates label objects from standard sensitive labels in the unified sensitive data standard system, standard field objects from standard field concepts, data entity objects from standardized data objects, rule objects from historical rule records, and sample objects from historical annotation records. It also establishes observation source objects and annotation basis objects. The observation source objects record the business system, data container, record location, collection batch, and version. The annotation basis objects record the basis type, basis content or associated object, target label, supported or excluded direction, version, generation method, and upstream basis identifier. Lexical features, format features, statistical features, semantic features, and contextual features are saved as identification feature attributes for the corresponding data entity objects; candidate label evidence is associated with the data entity object, candidate label, observation source object, and annotation basis object, respectively; Establish at least one of the following relationships among knowledge objects: field mapping, category affiliation, rule pointing, applicable scenario, structural association, co-occurrence, business association, observation source, basis, and inheritance, so that candidate labels or sample labels can be traced back to their basic judgment basis along the labeling basis relationship.

5. The cold start method for model identification of sensitive data according to claim 1, characterized in that, The evidence status mentioned in step 3 is generated in the following manner: For the data object to be identified and candidate tags The evidence records are generated by combining rule matching, field mapping, sample matching, business scenario consistency, entity association, and exclusion conditions. The evidence record includes at least the data object identifier, candidate label identifier, evidence type, observation source identifier, current labeling basis identifier, upstream labeling basis identifier, support or rejection direction, basis version, and basis reliability. By tracing back along the upstream annotation basis of each evidence record, when two pieces of evidence ultimately rely on the same basic facts, the same rule version, the same field mapping record, the same historical label record, or the same manual confirmation record, they are classified into the same annotation basis group; when the label of one piece of evidence is directly generated by another piece of evidence, the two are classified into the same annotation basis group determined by the common upstream basis. For the same labeling basis group for the same candidate label, multiple pieces of evidence in the same direction form a group contribution, which is not directly accumulated according to the number of evidence paths or the number of derived samples; when there are both supporting and rejecting records in the same group, the conflict marker within the group is retained. Formation of Evidence State E i,k The evidence status records at least the set of observation sources, the independent supporting annotation basis group, the independent excluding annotation basis group, the reliability of each basis group, the completeness of the source record, and the conflict status.

6. The cold start method for model identification of sensitive data according to claim 5, characterized in that, Step 4 sets the purpose of model training, category prototype content, and category prior reliability for the sample objects, including: Set a usage vector for each sample ,in Indicates whether participation in model training is permitted. This indicates whether the observed content is allowed to participate in the category prototype construction. The labeling criteria indicate whether adding prior reliability of the category is allowed, and each application is determined separately; When the same candidate label obtains at least two backtrackable and non-inherited supporting labeling criteria, no exclusion criteria that meet the conflict conditions appear, and it can be distinguished from competing labels, the sample is allowed to participate in model training, category prototype content construction, and category prior reliability calculation; samples that have been independently and manually confirmed form a new independent confirmation criterion group. When a sample inherits from only one set of automatically labeled criteria, if the reliability of that criterion meets the training admission criteria and there is no obvious conflict, it is allowed to participate in model training, but its observations are prohibited from entering the category prototype content candidate set and its labeling criteria are prohibited from increasing the category prior reliability; after obtaining a second independent supporting criterion or independent human confirmation, the use determination is re-executed. When supporting evidence and highly reliable exclusion evidence coexist, and multiple competing labels have independent support or uncertain sources affecting label judgment, the data object is set as a manual review object; ultimately, a training sample set T, a category prototype content candidate set P, and a category reliability evidence set V are formed respectively.

7. The cold start method for model identification of sensitive data according to claim 1, characterized in that, In step 5, construct the category knowledge state vector. Calculate category content prototype ,include: The tag knowledge vector is obtained by encoding and aggregating the tag name, tag definition, and tag alias. ; will point to sensitive categories The rule name, rule description, keywords, applicable fields, and textualizable rule conditions are encoded and aggregated to obtain a rule knowledge vector. Samples are read from the candidate set P of the category prototype content and grouped and aggregated according to the observation source and common annotation criteria to obtain the sample content vector. ; Read the independent supporting evidence group and the independent excluding evidence group from the category reliability evidence set V, and construct the category knowledge state vector. The It includes at least the following states: knowledge availability, independent confirmation, conflict, observation coverage, and inherited duplication. The prior reliability is calculated using the function Q. ; Based on category knowledge status Determine the category-level nonnegative fusion coefficients for label knowledge, rule knowledge, and sample content, respectively. And normalize the fusion coefficients corresponding to the available knowledge; Sensitive categories Category content prototype Represented as: ; Among them, does the increase or decrease of the number of paths or derived samples generated by the same annotation group increase? When a category content prototype is unavailable, it is marked as unusable for model priors and saved accordingly. and the corresponding version.

8. The cold start method for model identification of sensitive data according to claim 1, characterized in that, Step 5 involves training the classification model, including: Category content prototype The transformation T maps to the representation space of the target classification model. For newly established classification models or newly added sensitive categories, the transformation is performed using... Generate initial classification parameters for the corresponding category, where ; The training objectives include a weighted classification loss and a class prior preservation term, wherein the strength of the class prior preservation term is... This makes the same category of knowledge state Both through The knowledge composition of the control category content prototype, and through , and The extent to which the control model adopts and retains the prototype of the content in the aforementioned categories; Upon receiving changes to rules, field mappings, manual confirmations, false alarms, or omissions, the system locates the record inheriting the changed basis using the first inverse index from the annotation basis to the evidence and sample; it then locates the affected category using the second inverse index from the sample to the category prototype version. Finally, it re-groups the annotation basis, determines its usage, and... , and calculate; According to the updated , The system handles conflict states and independent manual confirmation of changes, sets prior control states for affected categories to hold, release, freeze, or re-enhance, and adjusts the prototype scale, prior hold strength, and model parameters only for the affected categories.