A data sensitive information automatic grading and classification method

Abstract

The application discloses a kind of data sensitive information automatic hierarchical classification methods.The method carries out multimodal content extraction to heterogeneous data source, text block, image area and table cell are encoded into uniform intermediate representation sequence;Standardized sensitive entity set is obtained by the cascade identification of rule layer and multimodal semantic recognition layer, and is calibrated by confidence uniformity;Combining data source metadata, time-stamped time sequence blood relation diagram is constructed, and sensitivity propagation is carried out by space-time double-channel graph neural network, to obtain the sensitivity annotation result of each blood relation node;Again, with scale plate library matching and executing selective conformal risk control determination, the node sensitive grade determination result with the missed detection rate not more than preset upper limit is output;Through event-driven incremental closed loop, the real-time of data security map is maintained.The application solves the technical problems that the recognition coverage of prior art is limited to text single mode, the hierarchical result lacks provability and the dynamic updating timeliness is poor.

Description

Technical Field

[0001] This invention relates to the field of data security and privacy computing technology, specifically to a method for automatic identification and hierarchical classification of sensitive data information based on multimodal cascade recognition and spatiotemporal lineage graph risk control. This invention can be applied to the discovery of sensitive information in enterprise-level data assets, data flow lineage tracking, compliance hierarchical classification management, and data security map construction. Background Technology

[0002] With the successive implementation of data compliance both domestically and internationally, enterprises need to identify and classify sensitive information within their massive data assets. Data security governance has become a fundamental aspect of enterprise digital operations, and the automatic identification and classification of sensitive information is the most crucial and demanding technical task within this process.

[0003] In the field of data security governance, common sensitive information identification technologies mainly fall into two categories: one is the rule matching method based on regular expressions and keyword dictionaries, which scans and matches structured fields through predefined patterns such as ID card numbers, mobile phone numbers, and bank card numbers; the other is the text detection method based on the Named Entity Recognition (NER) model, which uses deep learning models to annotate and identify entities such as names, addresses, and organizations in unstructured text.

[0004] The prior art most relevant to this invention adopts the following technical principle: performing rule matching and joint scanning of text content in the data source with the NER model, identifying sensitive fields, assigning sensitivity level labels to data tables or files according to preset classification and grading standards, and updating the classification and grading results through periodic full rescanning.

[0005] However, the aforementioned existing technologies suffer from the following two technical problems: First, their recognition capabilities are limited to a single text modality, and the sensitivity level determination results lack any statistically significant risk guarantee. For scanned images, invoices, documents embedded in images, and sensitive information in tables containing complex merged cells, existing technologies suffer from severe under-detection. More importantly, their under-detection rate is immeasurable and uncontrollable, failing to provide provable evidence of the reliability of the classification results to regulatory auditors, and thus failing to meet the data compliance audit requirements for provability. Second, the classification results are static snapshots, neither tracking the structural relationships and temporal causality of data flow nor efficiently responding to dynamic data changes. Derivative data generated after data copying, processing, desensitization, and aggregation cannot automatically inherit and update sensitivity levels. Furthermore, existing technologies cannot distinguish between different processing sequences, such as desensitization before aggregation versus aggregation before desensitization, which result in drastically different sensitivity levels. When facing data changes, they can only rely on periodic full rescans, which are inefficient and have computational overhead that increases linearly with data size, failing to meet the real-time governance needs of large-scale dynamic data environments. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic classification method for sensitive data, in order to solve the technical problems in the prior art where the coverage of sensitive identification is limited to a single text modality, the false detection rate of the classification results is immeasurable and cannot meet the provability requirements of compliance audit, and the classification results cannot trace the data flow lineage and cannot efficiently respond to dynamic changes in data.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An automatic classification method for sensitive data information, applied to the classification of enterprise data assets, includes: extracting multimodal content from heterogeneous data sources, encoding the resulting text blocks, image regions, and table cells into a unified intermediate representation sequence; performing rule-layer scanning on the unified intermediate representation sequence to obtain a set of rule-hit entities, and separating undetermined content into a set of unresolved fragments; routing the set of unresolved fragments to a multimodal semantic recognition layer containing text, image, and table channels for judgment to obtain a set of semantically recognized entities; performing unified confidence calibration and sensitivity type normalization on the set of rule-hit entities and the set of semantically recognized entities to obtain a standardized set of sensitive entities; constructing a time-stamped lineage graph by combining the standardized set of sensitive entities with metadata in the data source, and propagating and fusing it through spatial and temporal channels to obtain sensitivity labeling results for each lineage node; matching the sensitivity labeling results with a compliance template library to obtain a candidate sensitivity level distribution; performing selective conformal risk control judgment on the candidate sensitivity level distribution, and outputting a node sensitivity level judgment result with a false negative rate not exceeding a preset upper limit.

[0009] In a preferred embodiment, the set of pending fragments is routed to a multimodal semantic recognition layer containing text, image, and table channels for determination to obtain a set of semantic recognition entities. This includes: inputting text fragments from the set of pending fragments into the text channel for contextual semantic recognition; inputting image regions from the set of pending fragments into the image channel for layout perception recognition in conjunction with layout features; inputting table cells from the set of pending fragments into the table channel for column semantic inference based on column name semantics and column content distribution; and fusing the channel features output from the text channel, image channel, and table channel in a shared coding space for unified determination to obtain a set of semantic recognition entities.

[0010] Further, the channel features output by the text channel, the image channel, and the table channel are fused in a shared coding space and then uniformly determined to obtain a semantic recognition entity set. This includes: mapping the channel features output by the text channel, the image channel, and the table channel to the shared coding space to obtain a unified representation vector for each modality; fusing the unified representation vectors for each modality; and performing sensitive entity determination on the fused representations to obtain a semantic recognition entity set.

[0011] In a preferred embodiment, the set of rule-hit entities and the set of semantic recognition entities are subjected to unified confidence calibration and sensitivity type normalization to obtain a standardized sensitive entity set. This includes: assigning a corresponding prior score based on whether the pattern matched by each entity in the set of rule-hit entities has a check bit verification; calibrating the original confidence score output by each entity in the set of semantic recognition entities during recognition using a temperature scaling method; mapping the prior score and the calibrated original confidence score to the same calibration probability space, and aligning all entities to a unified sensitivity type system to obtain the standardized sensitive entity set.

[0012] In a preferred embodiment, a time-stamped temporal lineage graph is constructed by combining the standardized sensitive entity set and metadata from the data source. This graph is then propagated and fused via spatial and temporal channels to obtain the sensitivity labeling results for each lineage node. The process includes: extracting time-stamped lineage relationships from the standardized sensitive entity set and the metadata; using data assets as nodes and data flow processing relationships as directed edges; recording the occurrence time and processing type of flow events on each directed edge to construct the temporal lineage graph; performing neighborhood message passing along the lineage edges of each node using the spatial channel to obtain a spatial representation vector; performing temporal encoding on the incoming event sequences of each node using the temporal channel to obtain a temporal representation vector; fusing the spatial representation vector and the temporal representation vector to obtain a spatiotemporal sensitivity representation vector for each lineage node, and writing the spatiotemporal sensitivity representation vector into the memory state of the corresponding lineage node; and inputting the spatiotemporal sensitivity representation vector into a sensitivity discrimination classification head to obtain the sensitivity labeling results for each lineage node.

[0013] Furthermore, the temporal channel performs temporal encoding on the incoming event sequence of each node to obtain a temporal representation vector, including: encoding the incoming event sequence according to the chronological order of the occurrence of each incoming event, capturing the causal pattern between processing events, and obtaining the temporal representation vector; wherein, the causal pattern is used to distinguish data flow scenarios with different processing times and different sensitivities, and the data flow scenarios include flow scenarios with desensitization before aggregation and flow scenarios with aggregation before desensitization.

[0014] In a preferred embodiment, the sensitivity labeling results are matched with a compliance template library to obtain a candidate sensitivity level distribution. Selective conformal risk control is then performed on the candidate sensitivity level distribution to output a node sensitivity level determination result where the false negative rate does not exceed a preset upper limit. This includes: matching the sensitivity labeling results with the mapping rules from sensitivity types to level standards in the compliance template library to obtain a candidate sensitivity level distribution for each node; using a selection function to select a first set of nodes from each node whose determination confidence reaches a preset selection threshold; calculating a conformal quantile threshold on a pre-held calibration dataset; constructing a level prediction set for each node in the first node set based on the conformal quantile threshold, wherein the level prediction set ensures that the false negative rate does not exceed a preset upper limit without relying on data distribution assumptions; and for nodes whose cardinality of the level prediction set does not exceed a preset cardinality threshold, taking the highest sensitivity level within the level prediction set as the final sensitivity level of that node, and outputting the node sensitivity level determination result.

[0015] In a preferred embodiment, the method further includes: routing nodes whose cardinality in the level prediction set exceeds the preset cardinality threshold, as well as nodes that were not filtered into the first node set by the selection function, to a manual review queue, where their final sensitivity level is determined by manual review.

[0016] In a preferred embodiment, the method further includes: monitoring the data change event stream; when a data change event occurs, extracting the content of the changed data and performing the rule layer scanning, the multimodal semantic recognition layer judgment, and unified confidence calibration to obtain an incremental sensitive entity set; based on the incremental sensitive entity set, determining the influence domain of the time-series lineage graph affected by the data change event, locally updating the memory state of nodes within the influence domain, and re-executing the selective conformal risk control judgment for nodes within the influence domain to obtain an updated node sensitivity level judgment result.

[0017] Further, based on the incremental sensitive entity set, the influence domain affected by the data change event in the temporal lineage graph is determined. The memory state of the nodes within the influence domain is locally updated, and the selective conformal risk control judgment is re-executed for the nodes within the influence domain to obtain an updated node sensitivity level judgment result. This includes: determining the influence domain based on the nodes associated with the incremental sensitive entity set and their downstream nodes reachable along directed edges in the temporal lineage graph; re-executing the propagation calculation of the spatial channel and the temporal channel only for the nodes within the influence domain to update the memory state of the nodes within the influence domain, while the memory state of the nodes outside the influence domain remains unchanged; re-executing the selective conformal risk control judgment only for the nodes within the influence domain to obtain an updated node sensitivity level judgment result; wherein, the computational cost of the incremental update is proportional to the size of the influence domain.

[0018] The present invention has the following beneficial technical effects:

[0019] (1) The cascaded recognition architecture includes sensitive information in images, scanned documents and complex tables in the recognition coverage, significantly reducing the false negative rate of unstructured data; at the same time, highly deterministic content is quickly determined by the rule layer, and the overall scanning throughput is greatly improved compared with the single depth model scheme.

[0020] (2) The spatiotemporal dual-channel lineage representation enables the sensitivity level of derived data to be automatically inherited and propagated along the lineage link. The temporal channel eliminates the misjudgment of the causal relationship of the processing sequence by the static graph method, and the accuracy of lineage-level sensitivity tracking is significantly improved.

[0021] (3) The classification results are accompanied by a mathematical upper limit of the missed detection rate that is independent of the distribution. Enterprises can provide the regulatory auditors with verifiable risk control evidence, thus upgrading the classification from empirical results to auditable and quantifiable compliance assets.

[0022] (4) Incremental closed loop reduces the latency from data change to security map update from hours of full rescan to seconds of event impact domain level, decoupling update computation overhead from full data scale, and supporting real-time governance of large-scale dynamic data environment.

[0023] (5) Multi-regulatory template mapping enables the same identification and lineage results to be output simultaneously as hierarchical views under multiple compliance standards. When regulatory standards change, only the mapping configuration needs to be adjusted without re-identification, which significantly reduces the operating costs of multi-jurisdictional compliance. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall process of an automatic hierarchical classification method for sensitive data according to the present invention.

[0025] Figure 2This is a flowchart of the cascaded multimodal sensitive identification process for confidence routing in this invention;

[0026] Figure 3 This is a schematic diagram illustrating the spatiotemporal dual-channel lineage map construction and sensitivity propagation of the present invention;

[0027] Figure 4 This is a flowchart illustrating the selective conformal risk control determination process of the present invention.

[0028] Figure 5 This is a schematic diagram of the event-driven incremental update closed-loop mechanism of the present invention. Detailed Implementation

[0029] Example 1

[0030] like Figure 1 As shown, this embodiment provides an automatic classification and grading method for sensitive data, applied to the classification and grading of enterprise data assets, including the following steps:

[0031] S1: Extract multimodal content from heterogeneous data sources and encode the resulting text blocks, image regions, and table cells into a unified intermediate representation sequence.

[0032] In this step, the heterogeneous data sources to be addressed include, but are not limited to, various data from enterprise management processes, such as database tables, electronic documents, images, and scanned copies. For different data formats, the automatic classification and grading system for sensitive data (hereinafter referred to as the system, whose physical entity is a server cluster) employs differentiated content extraction strategies: text data is directly extracted; image data undergoes optical character recognition (OCR) and layout analysis to restore the text content and spatial layout of the image; and table data undergoes table structure recognition to restore the row and column structure and cell semantics. After the above extraction processing, a multimodal parsing result containing three types of elements—text blocks, image regions, and table cells—is obtained. Subsequently, the multimodal parsing result is uniformly serialized and encoded, with metadata such as source location, modality type, and layout coordinates appended to each parsed element. This organizes the three types of heterogeneous elements into a unified intermediate representation sequence, enabling the subsequent recognition engine to process content from different modalities in a consistent manner.

[0033] Specifically, each element in the unified intermediate representation sequence is composed of a content encoding field and a structured metadata field. The content encoding field carries the text content string of the element or its vectorized preprocessed embedded representation; the structured metadata field records information in the following dimensions: source location identifier (uniquely identifies the data source to which the element belongs and its physical address within the data source; for database fields, it records the three-level path of database name, table name, and field name; for document-type data sources, it records the file's unique identifier and page number), modality type marker (distinguishing between three modalities: text blocks, image regions, and table cells using enumerated values, each encoded with a predefined integer value), and layout coordinates (recording the bounding box information of the element in the original document layout, in the format of a two-dimensional plane coordinate pair of the upper left and lower right corners; for elements originating from database fields, the layout coordinate field is filled with null values).

[0034] During the serialization and organization phase, the three types of heterogeneous elements are arranged according to their natural spatial order in the original data source: for document-type data sources, a top-to-bottom, left-to-right page reading order is adopted; for database table-type data sources, a row-first traversal order is adopted, with fields within the same row arranged in ascending order of column index. This ordered arrangement preserves the spatial proximity between elements, allowing subsequent text channels to directly utilize the contextual information of adjacent elements in the sequence to improve judgment accuracy when performing contextual semantic recognition. At the same time, complete page coordinate information also provides a structured spatial layout prior for page-aware recognition in image and table channels, helping to achieve more accurate positioning and recognition in page-position-sensitive scenarios such as document image areas and table column boundaries.

[0035] S2: Perform a rule layer scan on the unified intermediate representation sequence to obtain the rule-hit entity set, and separate the undetermined content into a set of unresolved fragments.

[0036] In this step, a deterministic rule layer scan is first performed on the unified intermediate representation sequence obtained from S1. The rule layer utilizes a regular pattern library and a checksum algorithm to match and determine sensitive entities with strong structural patterns, such as ID card numbers, bank card numbers, and mobile phone numbers. The rule layer can identify highly deterministic entities with extremely low computational cost. After scanning, the set of rule-hit entities is output, while the content that the rule layer could not determine is separated into a set of unresolved fragments for subsequent deep-level determination.

[0037] S3: The set of unresolved fragments is routed to a multimodal semantic recognition layer containing text channels, image channels, and table channels for judgment to obtain a set of semantic recognition entities.

[0038] Specifically, the set of unresolved segments that the rule layer in S2 cannot determine are routed to the multimodal semantic recognition layer for deep judgment. The multimodal semantic recognition layer consists of three parallel channels: text, image, and table, each performing specialized recognition processing for the characteristics of different modalities. The features of the three channels are fused in a shared encoding space and then uniformly judged, outputting a set of semantically recognized entities and the corresponding original model confidence scores. By entrusting the semantic recognition layer with the processing of content that the rule layer cannot directly determine, full coverage recognition of multimodal content is achieved.

[0039] S4: Perform confidence level calibration and sensitivity type normalization on the rule-hit entity set and the semantic recognition entity set to obtain a standardized sensitive entity set.

[0040] Specifically, confidence scores are processed for both the rule-hitting entity set from the rule layer and the semantic recognition entity set from the semantic recognition layer, and the two outputs are mapped to the same calibration probability space. Rule-hitting entities are assigned prior scores based on their pattern validation strength, while the original confidence scores of semantic recognition entities are calibrated using a temperature scaling method to ensure statistical comparability of the confidence scores relied upon in subsequent risk control steps. Simultaneously, sensitivity type normalization is performed on all entities, aligning them to a unified sensitivity type system (such as personal identity, financial accounts, and healthcare), ultimately outputting a standardized set of sensitive entities with calibrated confidence scores.

[0041] S5: Combining the standardized set of sensitive entities with the metadata in the data source, construct a time-stamped lineage graph, and then propagate and fuse it through spatial and temporal channels to obtain the sensitivity labeling results of each lineage node.

[0042] The standardized sensitive entity set obtained from S4 is combined with data source metadata (including ETL task logs, SQL statement parsing results, API call records, etc.) to extract time-stamped lineage relationships. Data assets such as tables, files, and fields are used as nodes, and data flow processing relationships are used as directed edges. The occurrence time and processing type (e.g., replication, desensitization, aggregation, association) of flow events are recorded on each directed edge to construct a temporal lineage graph. A spatial-temporal dual-channel graph neural network is run on the temporal lineage graph for sensitivity propagation: the spatial channel performs neighborhood message passing along the lineage edges to capture the topological dependencies of data assets; the temporal channel performs temporal encoding on the incoming event sequences of each node in chronological order to capture the causal patterns of processing events. The representation vectors of the two channels are concatenated and fused to obtain the spatiotemporal sensitivity representation vector of each lineage node, which is written into the memory state of the corresponding node. Finally, the sensitivity labeling results of each lineage node are output through a sensitivity discrimination classification head.

[0043] S6: Match the sensitivity labeling results with the compliance template library to obtain the candidate sensitivity level distribution, perform selective conformal risk control judgment on the candidate sensitivity level distribution, and output the node sensitivity level judgment result with a false negative rate not exceeding the preset upper limit.

[0044] In this step, the sensitivity labeling results of each lineage node obtained from S5 are matched with the compliance template library. The compliance template library contains multiple sets of regulatory templates, including specific regulations, legal provisions, and industry data classification and grading guidelines. Each template defines configurable mapping rules from sensitivity types to the regulatory level standard. After matching, the candidate sensitivity level distribution of each node under the target compliance standard is obtained. Subsequently, selective conformal risk control is performed on the candidate sensitivity level distribution: high-confidence nodes are screened using a selection function, and conformal quantile thresholds are calculated on the calibration dataset. A level prediction set is constructed for each high-confidence node. This prediction set mathematically guarantees that the false negative rate does not exceed a user-preset upper limit, and this guarantee does not depend on any data distribution assumptions. For nodes whose prediction set size does not exceed a preset threshold, the highest sensitivity level is selected as the final level according to the principle of categorizing from highest to lowest. Nodes with insufficient confidence in the judgment are routed to a manual review queue. Finally, the node sensitivity level judgment result with risk assurance is output.

[0045] Taking the data asset compliance governance of a large comprehensive enterprise as a typical application scenario, this paper further illustrates the specific applicability of the automatic classification and grading method for sensitive data in different types of enterprise data. Enterprise data assets typically cover various types, including financially sensitive data, production and R&D data, and general business data. These data types differ significantly in content characteristics, sensitivity, and classification determination paths. The multimodal cascaded identification and lineage propagation mechanism can achieve differentiated and accurate classification and grading processing based on these differences.

[0046] In the scenario of hierarchical classification of enterprise financial data, the data assets to be governed encompass various forms, including electronic financial statements, scanned copies of accounting vouchers, fund transfer database tables, and bank statements. During the multimodal content extraction stage, the system performs OCR and layout-aware recognition on the scanned bank statements to restore fields such as bank account numbers, transaction amounts, and counterparty names. For the fund transfer database tables, the system performs column semantic inference on the table channel to accurately locate highly sensitive fields such as account number columns, amount columns, and payer columns. In the rule-layer scanning stage, the system uses a regular expression pattern library and the Luhn algorithm to directly detect strongly structured sensitive entities such as bank card numbers and unified social credit codes. Semantically sensitive content in accounting vouchers, such as the names of financial personnel and customer companies, is covered by contextual semantic recognition in the text channel. Through the temporal lineage graph propagation analysis, the lineage dependency relationship between the original financial statement data and downstream management dashboard reports is fully modeled. The highly sensitive attributes carried by unpublished original financial data propagate along the lineage edges to downstream derived report nodes, effectively preventing the risk of indirectly obtaining core financial information by accessing derived data. Ultimately, the field containing bank card numbers and unpublished financial statement data were identified as core data (Level 4), fund transfer records containing customer account details were identified as important data (Level 3), and archived ordinary accounting vouchers were identified as general data (Level 2). This achieves automatic and refined classification of financial data assets, directly supporting the generation of enterprise financial data access control policies and the output of compliance reports.

[0047] In the hierarchical classification scenario of production and R&D data, the data assets to be managed include product design engineering documents containing CAD screenshots, formula and process database tables, electronic documents of R&D experimental records, and scanned copies of supplier cooperation agreements. During the multimodal content extraction stage, the system performs image channel recognition on engineering documents containing CAD screenshots to extract key content such as product models and design parameters from the drawing title bars. Based on the high consistency between column name semantics (e.g., "formula ratio," "process temperature") and the numerical distribution within the columns, the table channel infers core fields in the formula and process database table as sensitive commercial secrets, covering the blind spot where the rule layer cannot recognize such unformatted commercial secret information. Furthermore, lineage graph propagation analysis reveals the lineage dependency between the core formula database table and downstream summary reports: even if the downstream report only displays the statistical mean of the formula data, the sensitive attributes of highly sensitive upstream nodes still propagate downstream along the lineage edge, assigning corresponding sensitivity levels to report data nodes, effectively preventing the omission of sensitive attribute inheritance when scanning derived data in isolation. Ultimately, the core product formula database table was identified as core data (Level 4), R&D experimental records containing key technical parameters were identified as important data (Level 3), and R&D materials such as publicly available product manuals that do not contain trade secrets were identified as public data (Level 1) or general data (Level 2). This achieves refined hierarchical management of production and R&D data assets, providing a data classification basis for enterprise intellectual property protection and authorization of access to trade secrets.

[0048] In the scenario of classifying and categorizing ordinary business data, the data assets to be governed include internal meeting minutes, employee training materials, announcements and notices, and project progress management forms. This type of data typically contains only a small amount of personal information such as scattered employee names and department names, and the number of sensitive entities in the rule-hit entity set and the semantic recognition entity set is limited. Lineage graph propagation analysis confirms that the upstream sources of the above data nodes are all non-sensitive content, and no cross-node inheritance of highly sensitive attributes has occurred. After matching with the compliance template library, this type of data is generally classified as general data (Level 2) or public data (Level 1). For personal information such as employee contact information that occasionally appears in meeting minutes, the corresponding fields are individually marked as general data (Level 2) and a desensitization suggestion is triggered, without affecting the overall classification conclusion of the document. It should be noted that because the rule layer can quickly and cost-effectively determine a large number of ordinary documents without sensitive entities, and the multimodal semantic recognition layer only performs in-depth processing on a very small number of fragments containing occasional sensitive content, the overall scanning efficiency is significantly better than the solution of performing deep model inference on all enterprise data, demonstrating a significant cost advantage in large-scale batch compliance scanning scenarios for ordinary business data.

[0049] Example 2

[0050] This embodiment, based on Embodiment 1, further explains the specific implementation method of routing the set of unresolved fragments to the multimodal semantic recognition layer for judgment and obtaining the set of semantic recognition entities in step S3. For example... Figure 2 As shown, this embodiment includes the following steps:

[0051] In this invention, the cascaded multimodal recognition architecture for confidence routing refers to decomposing the sensitive recognition task into a two-level sequential processing structure—a deterministic rule layer and a multimodal semantic recognition layer—based on the decidability of the content. A dynamic routing mechanism determines the processing path for each content segment between the two levels: highly deterministic content is directly determined by the first-level rule layer at a low cost; content that the rule layer cannot determine is then routed to the second-level semantic layer for deep recognition, rather than sending all content to the computationally expensive deep model. The boundary between the two levels is defined by whether the content is matched by the rule layer and verified. The routing decision itself only involves simple set classification logic and does not introduce additional computational cost.

[0052] This architecture is based on a core insight into the distribution characteristics of sensitive entities in large-scale enterprise data assets: sensitive entities with strong structured patterns (such as 18-digit ID card numbers, 13-17 digit bank card numbers, 11-digit mobile phone numbers, etc.) have clearly enumerable and highly deterministic identification rules, and constitute the majority of the total known sensitive entities in enterprise data assets; while ambiguous sensitive content requiring semantic understanding (such as names and addresses embedded in natural language descriptions, handwritten text in document images, and sensitive columns in tables with complex merged cells) is relatively limited in quantity but significantly more difficult to identify. If all content is fed into a multimodal semantic recognition model, the high computational cost of the deep model will be spread across a large amount of content that can be determined by rules at a low cost, resulting in serious redundancy of computational resources; if only the rule layer is used, sensitive information in images, scanned documents, and complex tables will form an unacceptable blind spot in recognition. The cascaded architecture achieves the optimal balance between recognition accuracy, coverage, and scanning efficiency by complementing the advantages of two layers: the rule layer completes batch recognition of highly deterministic entities with near-zero inference latency, while the semantic layer performs deep inference only on truly questionable unresolved content.

[0053] At the implementation level, the unified confidence calibration mechanism between the two layers is a key technical support for the cascaded architecture to seamlessly integrate the two outputs. The prior scores output by the rule layer and the model confidence scores output by the semantic layer have different statistical meanings in their original forms, and cannot be directly compared across layers or used in subsequent risk control judgments within the same framework. Through differential assignment of prior scores and temperature scaling calibration, the two outputs are uniformly mapped to the same [0,1] calibration probability space, so that the confidence scores of all entities have a consistent probabilistic interpretation in a statistical sense, thereby supporting the unified threshold judgment and risk guarantee calculation of the confidence scores of all entities in the subsequent selective conformal risk control link.

[0054] S3.1: Input the text fragments in the set of pending fragments into the text channel for contextual semantic recognition.

[0055] In this step, the text channel receives text-type fragments from the pending fragment set and uses a natural language processing model to perform contextual semantic recognition on these text fragments. The text channel's recognition capability covers semantically sensitive information that the rule layer cannot determine through pattern matching, such as names, address descriptions, and organization names—entity types that require contextual understanding for accurate identification. Through deep semantic analysis, the text channel outputs text-type sensitive entities and their corresponding channel feature vectors.

[0056] S3.2: Input the image regions in the set of unresolved fragments into the image channel and combine them with the layout features to perform layout perception and recognition.

[0057] In this step, the image channel receives regions belonging to the image type from the set of pending fragments and performs layout-aware recognition of sensitive content within these regions. The image channel is specifically designed for scenarios such as scanned images, invoices, and embedded document information. By fusing image content features with layout features, it identifies image morphologically sensitive information that neither rule-based nor plain text NER methods can cover. The image channel outputs image-class sensitive entities and their corresponding channel feature vectors.

[0058] S3.3: Input the table cells in the set of unresolved fragments into the table channel and perform column semantic inference based on column name semantics and column content distribution.

[0059] In this step, the table channel receives cells of table type from the set of pending fragments and performs column semantic inference based on column name semantics and column content distribution. The table channel is specifically designed to handle sensitive fields in tables containing complex merged cells that are difficult to determine at the rule level. By comprehensively analyzing the semantic meaning of column headers and the statistical distribution characteristics of data within columns, it infers potential sensitive column types, thus covering previous blind spots. The table channel outputs table-type sensitive entities and their corresponding channel feature vectors.

[0060] S3.4: After fusing the channel features output by the text channel, the image channel, and the table channel in the shared coding space, a unified determination is made to obtain a semantic recognition entity set.

[0061] In this step, the channel features output from the text, image, and table channels are uniformly mapped to a shared coding space, and cross-modal feature fusion is performed within this shared coding space. The fused multimodal joint representation is processed by a unified decision module, outputting a set of semantic recognition entities and the original model confidence scores for each entity. The design of the shared coding space ensures the comparability of features from different modalities within the same semantic space, enabling the fusion decision to fully utilize the complementary advantages of multimodal information.

[0062] Example 3

[0063] This embodiment, based on Embodiment 2, further explains the specific implementation method of fusing the three-channel features in the shared coding space and then uniformly determining to obtain the semantic recognition entity set in step S3.4. This embodiment includes the following steps:

[0064] S3.4.1: Map the channel features output by the text channel, the image channel, and the table channel to a shared coding space to obtain a unified representation vector for each modality.

[0065] In this step, the feature vectors output by the text channel, image channel, and table channel are mapped to a shared encoding space of the same dimension through their respective modal projection layers, resulting in unified representation vectors for the text modality, image modality, and table modality. The parameters of each modal projection layer are jointly optimized during the training phase to ensure that the features of different modalities have semantically aligned representations in the shared encoding space, providing a foundation for subsequent cross-modal fusion.

[0066] S3.4.2: The unified representation vectors of each modality are fused, and the fused representations are used to determine sensitive entities to obtain a semantic recognition entity set.

[0067] In this step, the unified representation vectors of each modality obtained in S3.4.1 are fused to obtain the multimodal fused representation vector.

[0068] In this embodiment of the invention, the fusion operation specifically adopts a cascaded structure of concatenation fusion and linear compression: the unified representation vectors of each modality are concatenated along the feature dimension to obtain a multimodal joint vector with a dimension three times that of the single-modal representation vector; subsequently, a linear transformation layer compresses the joint vector to the same dimension as the single-modal unified representation vector, which serves as the multimodal fusion representation vector. The parameters of the linear transformation layer are jointly optimized and learned end-to-end during the model training phase, ensuring that the compressed fusion vector retains complementary discriminative information from the three modalities.

[0069] Preferably, when dealing with scenarios where the quality of the three modal features differs significantly (e.g., low confidence in text recognition after OCR of scanned images leads to a decrease in image channel feature quality), the fusion operation can be replaced with a fusion method based on cross-modal attention: multi-head attention is performed on the unified representation vector of the three modalities using the fusion query vector of the element to be judged. The contribution of each modal feature to the current judgment is adaptively adjusted through attention weights. When the image channel feature quality is low, its weight is automatically reduced, while the influence of text and table channel features is increased, effectively suppressing misjudgment interference introduced by noisy modalities. This demonstrates more robust recognition performance in real-world scenarios with uneven multimodal information quality. The fused representation vector is input to the sensitive entity judgment module, which performs sensitivity classification on each candidate entity or fragment, outputting its sensitivity type and the corresponding original model confidence, thus forming a semantic recognition entity set. Through effective fusion of multimodal features, the judgment module can comprehensively utilize complementary information from text semantics, image layout, and table structure, significantly improving recognition accuracy, especially for ambiguous sensitive entity scenarios where single-modal information is insufficient for judgment.

[0070] Example 4

[0071] This embodiment, based on Embodiment 1, further explains the specific implementation method of performing confidence-based unified calibration and sensitivity type normalization on the rule-hitting entity set and the semantic recognition entity set in step S4 to obtain a standardized sensitive entity set. This embodiment includes the following steps:

[0072] S4.1: According to the rules, determine whether the pattern matched by each entity in the entity set has a check bit verification and assign the corresponding prior score.

[0073] For each entity in the rule-matched entity set, a prior score is assigned based on the regular expression pattern features it matches. Specifically, for strongly structured patterns with checksum verification (such as last digit checksum for ID card numbers, Luhn algorithm checksum for bank card numbers, etc.), the recognition certainty of such entities is extremely high, and a prior score close to 1 is assigned. For weakly structured patterns that only rely on format pattern matching and do not have checksum verification (such as simple patterns that only match length and character type), a relatively low prior score is assigned to reflect the possibility of mismatches. The differentiated assignment of prior scores allows for a reasonable distinction between entities of different reliability levels in the rule-matched entity set, providing rule-side input for the subsequent construction of a unified confidence space.

[0074] S4.2: The original confidence scores of each entity in the semantic recognition entity set are calibrated using a temperature scaling method during recognition.

[0075] In this step, the original confidence scores (i.e., the probability values ​​output by the classifier's softmax) for each entity in the semantic recognition entity set are post-processed and calibrated using temperature scaling. Temperature scaling introduces a learnable temperature parameter T into the softmax function, learning a T value on a reserved calibration set with the objective of minimizing the negative log-likelihood. This ensures that the model's output confidence scores are statistically consistent with the true accuracy, eliminating the overconfidence bias commonly found in deep learning models. The calibrated confidence scores have a probabilistic interpretation in a statistical sense, providing a reliable basis for subsequent unified comparisons with prior scores from the rule-based side.

[0076] It should be noted that the calibration set relied upon by the temperature scaling method is typically obtained by dividing the labeled sensitive entity sample data into certain proportions. This portion of data is pre-reserved during the training phase of the semantic recognition model and does not participate in the gradient backpropagation of model parameters. It is dedicated to learning the temperature parameters during the calibration phase to ensure the independence and reliability of the calibration results. The calibration set should cover positive and negative examples of various objectives in a unified sensitivity type system and maintain a class ratio consistent with the target data environment as much as possible to ensure the representativeness of the temperature parameter estimation. The learning of the temperature parameter T aims to minimize the negative log-likelihood loss (NLL) on the calibration set. This objective function is monotonically convex with respect to T and can be solved accurately in a finite number of iterations using an efficient one-dimensional line search algorithm. The entire calibration process does not require any modification to the weights of the original semantic recognition model.

[0077] After calibration, the T-value is fixed as a constant for the inference phase: for each candidate entity, the logit vector output by the original softmax is divided by T and then normalized again to obtain the calibrated confidence score. The additional computational cost introduced by this operation during the inference phase is negligible and does not affect the overall scanning throughput performance of the system. It should be understood that when the underlying semantic recognition model is fine-tuned or retrained, the distribution of its output confidence score may change. In this case, T-value learning should be re-performed on the calibration set to maintain the statistical validity of the calibrated confidence score and ensure that the confidence score on which subsequent selective conformal risk control decisions are based maintains a consistent correspondence with the true accuracy rate.

[0078] S4.3: Map the prior scores and the calibrated original confidence levels to the same calibration probability space, and align all entities to a unified sensitivity type system to obtain the standardized sensitive entity set.

[0079] The prior scores from the rule layer obtained in S4.1 and the calibration confidence scores from the semantic layer obtained in S4.2 are uniformly mapped to the [0,1] calibration probability space to ensure that the confidence scores of the two outputs have a unified statistical meaning, thereby supporting the consistency comparison and threshold determination of the confidence scores of all entities in the subsequent risk control process. Simultaneously, a sensitivity type normalization mapping is performed on all entities, aligning all entities from the rule layer and semantic layer to a predefined sensitivity type system (such as standard categories like personal identification information, financial account information, health and medical information, and location trajectory information), eliminating the heterogeneous differences in sensitivity type labeling between the two layers. After the above calibration and normalization processing, the final output is a standardized set of sensitive entities with calibration confidence scores, which serves as the input for subsequent lineage graph construction and hierarchical determination.

[0080] Taking a typical scenario of data asset compliance scanning for a company as an example, the overall operational effect of the above cascaded recognition architecture is illustrated: A scanning task covers multiple data formats, including employee file database tables, electronic contract documents, and scanned expense receipts. The rule layer directly identifies and verifies structured sensitive entities (ID numbers, bank card numbers, mobile phone numbers, etc.), accounting for the vast majority of all identified sensitive entities in this scan. The semantic recognition layer only performs in-depth processing on unresolved content that the rule layer cannot determine. This includes handwritten names in expense receipt images, address descriptions in electronic contract texts, and non-standard column names in sensitive fields in employee file tables—areas completely uncovered by the rule layer. Through this cascaded processing, while significantly reducing the number of deep model inference calls, sensitive information in expense receipt images and complex tables is included in the recognition coverage. The overall false negative rate is significantly reduced compared to a pure rule-based solution, and the overall scanning throughput is significantly improved compared to a single deep model full-scale inference solution, balancing the effectiveness and completeness requirements of large-scale data asset scanning scenarios.

[0081] Example 5

[0082] This embodiment, based on Embodiment 1, further explains the specific implementation method of constructing a time-stamped kinship map in step S5 and fusing it through spatiotemporal dual-channel propagation to obtain the sensitivity labeling results of each kinship node. For example... Figure 3 As shown, this embodiment includes the following steps:

[0083] In this invention, the spatiotemporal dual-channel lineage graph representation learning method refers to: designing a dual-channel graph neural network on a timestamped directed graph of data lineage, consisting of parallel spatial and temporal channels, to independently model the topological features and temporal event features of the lineage graph, and then fusing the output representation vectors of the two channels to obtain a spatiotemporal sensitivity representation vector that simultaneously encodes the "position" (spatial dimension) and "processing history" (temporal dimension) of the data asset in the lineage network, used for propagating and inheriting sensitive attributes. The core difference of this method is that: traditional static graph neural networks only encode the topological structure of the lineage graph, ignoring the temporal sequence information of events on the edges; pure temporal modeling methods only focus on the sequence of processing events, ignoring the topological propagation contribution of upstream nodes in the lineage graph; the spatiotemporal dual-channel design of this invention, by modeling the two types of information separately in independent channels and then fusing them, enables the model to simultaneously perceive the lineage topological dependence of the data asset and the causal sequence of processing events, solving the problem of mutual interference when the two types of information are mixed and encoded.

[0084] This method or embodiment can be understood from the following two dimensions: First, in the spatial dimension, the sensitivity of a data asset node depends not only on the data content it directly holds, but also on the sensitivity attributes of all upstream source nodes in the lineage topology. The neighborhood message passing mechanism of the graph neural network iteratively aggregates the sensitive entity features of upstream neighbor nodes, implicitly encoding the sensitivity attributes of all upstream assets reachable in the lineage topology in the spatial representation vector of each node. This enables the automatic inheritance of the sensitivity attributes of source data by derived data, eliminating the need for manual tracing of lineage links one by one. Second, in the temporal dimension, the same set of data assets can produce output data with drastically different sensitivities through processing operations in different temporal sequences. This temporal causal relationship cannot be expressed in the directed edge topology of the lineage graph and must be captured by explicitly modeling the temporal order of the event sequence on the edge. The temporal channel encodes the temporal order of the incoming event sequence of each node, enabling the model to correctly distinguish data flow scenarios with essential differences in sensitivity due to different processing times, fundamentally eliminating the systematic misjudgment of the causal order of processing by the static topology graph method.

[0085] S5.1: Combine the standardized sensitive entity set with the metadata to extract the lineage with timestamps, use data assets as nodes and data flow processing relationships as directed edges, and record the occurrence time and processing type of the flow event on each directed edge to construct the time-series lineage graph.

[0086] In this embodiment of the invention, the standardized sensitive entity set obtained in step S4 of embodiment 1 is combined with data source metadata to extract lineage relationships. Metadata sources include ETL task logs, SQL statement parsing results, API call records, etc., from which the flow and processing relationships between data assets are extracted. A lineage graph is constructed using data assets such as tables, files, and fields as graph nodes and data flow and processing relationships as directed edges. Each directed edge records not only its semantics (i.e., processing type, including replication, desensitization, aggregation, association, etc.) but also the precise occurrence time of the corresponding flow event, thus embedding the processing time sequence information into the edge attributes of the lineage graph, constructing a time-stamped time-series lineage graph. The construction of the time-series lineage graph allows for a complete expression of the dependencies, processing semantics, and temporal causality between data assets, laying a structural foundation for subsequent spatiotemporal dual-channel sensitivity propagation.

[0087] S5.2: The spatial representation vector is obtained by performing neighborhood message passing on the sensitive entity features of each node along the bloodline edge of the spatial channel.

[0088] In this step, the spatial channel, based on the topology of a temporal lineage graph, employs a message-passing mechanism from a graph neural network to perform neighborhood aggregation operations on the sensitive entity features of each node along the directed edges of the lineage. For each node in the lineage graph, the spatial channel aggregates the sensitive entity features held by its upstream neighbor nodes according to edge weights to the current node. After iterative multi-hop propagation, the spatial representation vector of each node incorporates the sensitive attribute information of its upstream data assets reachable in the lineage topology. Through the message passing of the spatial channel, derived data nodes can automatically perceive the sensitive content in their source data, realizing the topological propagation and inheritance of sensitive attributes along the lineage links.

[0089] S5.3: The time sequence channel is used to perform time sequence encoding on the incoming event sequence of each node to obtain the time sequence representation vector.

[0090] In this step, the temporal channel performs temporal modeling on the incoming event sequence of each node in the temporal lineage graph. After sorting the incoming event sequences by the occurrence time of each event, they are processed by a temporal coding model (such as a recurrent neural network or a Transformer temporal encoder) to capture the causal patterns between processing events and output a temporal representation vector. The key value of the temporal channel lies in the fact that different processing event sequences have a fundamental impact on data sensitivity—for example, even involving the operations of desensitization and aggregation, the data generated by desensitizing first and then aggregating may have drastically different final sensitivity levels compared to the data generated by aggregating first and then desensitizing. By explicitly modeling the event sequence, the temporal channel enables the system to correctly distinguish data flow scenarios with different processing sequences and varying sensitivities, fundamentally eliminating the misjudgment of causal order in static topology graph methods.

[0091] S5.4: The spatial representation vector and the temporal representation vector are fused to obtain the spatiotemporal sensitivity representation vector of each bloodline node, and the spatiotemporal sensitivity representation vector is written into the memory state of the corresponding bloodline node.

[0092] In this step, the spatial representation vector obtained in S5.2 and the temporal representation vector obtained in S5.3 are concatenated and fused to obtain a spatiotemporal sensitivity representation vector that simultaneously encodes the topological dependence of the flow and the causal relationship of the processing time. The spatiotemporal sensitivity representation vector integrates the positional characteristics (spatial dimension) of the data asset in the lineage graph with its historical processing event sequence characteristics (temporal dimension), and is a comprehensive expression of the current sensitive state of the node. The fused spatiotemporal sensitivity representation vector is written into the memory state of the corresponding lineage node and persistently stored in the node as a state carrier for memory reuse during subsequent incremental updates, avoiding repeated calculations of the entire graph when data changes.

[0093] Furthermore, the memory state of each lineage node is a persistent storage unit maintained by each lineage node. Its stored content is the spatiotemporal sensitivity representation vector output by the node's most recent spatiotemporal dual-channel propagation calculation. The memory state is initialized when the node is first added to the lineage graph. The initial value can be set to a zero vector or a local representation vector based on the node's local sensitive entity features. After each full propagation calculation or incremental local update, the newly calculated spatiotemporal sensitivity representation vector is overwritten, updating the node's memory state. This persistent design of the memory state allows the system to directly reuse the stored historical memory states of lineage nodes unrelated to data changes, without needing to recalculate their spatiotemporal representations, in order to participate in the neighborhood aggregation operation of adjacent nodes within the influence domain. This is the key data structure foundation for decoupling computational cost and full graph scale in the subsequent incremental update mechanism.

[0094] S5.5: Input the spatiotemporal sensitivity representation vector into the sensitivity discrimination classification head to obtain the sensitivity labeling results of each bloodline node.

[0095] Specifically, the spatiotemporal sensitivity representation vectors of each lineage node obtained in S5.4 are input into the sensitivity discrimination classification head. The classification head performs classification inference on the spatiotemporal fusion representation of each node, outputting the sensitivity type label and sensitivity propagation intensity of the node, forming the sensitivity labeling results of the entire link node. The sensitivity propagation intensity reflects the credibility of the sensitivity attributes inherited by the current node from the upstream lineage link, providing a reference for confidence estimation in subsequent classification determination.

[0096] Using the user data processing workflow of an internet company as a specific case, this paper illustrates the complete operation process of the aforementioned time-series lineage graph construction and spatiotemporal dual-channel propagation mechanism. The company's data processing workflow involves the following data asset nodes: the original user information database table (node ​​A, containing user name, ID number, mobile phone number, and home address fields); the marketing analysis wide table after ETL anonymization (node ​​B, where the ID number and mobile phone number fields have been replaced with anonymous identifiers using irreversible hashing); the user profile data table (node ​​C) aggregated from node B and user purchase behavior records; and the user group statistical analysis report (node ​​D) generated from node C. In addition, the lineage graph also contains a supply chain inventory database table (node ​​E) completely independent of the above four nodes, and its downstream procurement statistical report (node ​​F, derived from node E).

[0097] During the time-series lineage graph construction phase, the system extracts processing events between the aforementioned nodes from the ETL task logs and SQL statement parsing results: Directed edge A→B records a de-identification processing event, labeled "De-identification," with an occurrence time of T1; directed edge B→C records a convergence processing event, labeled "convergence," with an occurrence time of T2 (T2 is approximately 72 hours later than T1); directed edge C→D records a convergence processing event, labeled "convergence," with an occurrence time of T3 (T3 is approximately 24 hours later than T2); and directed edge E→F in the independent subgraph records a convergence processing event, occurring at T4, which is unrelated to the processing time sequence of the personal information subgraph. Each directed edge is fully accompanied by a processing type label and a precise occurrence time, and the time-series lineage graph fully expresses the topological structure and causal relationships of the data processing flow.

[0098] During the spatiotemporal dual-channel propagation phase, the spatial channel performs multi-hop neighborhood message passing to each node along the directed edge of bloodline. Node A identifies four types of highly sensitive entities—ID card number, mobile phone number, name, and address—through step S4. Its local sensitive entity feature vector carries high-confidence personal identity information and location trajectory information annotations. Node B aggregates the highly sensitive features of upstream node A through the desensitization processing weight of edge A→B. The spatial representation vector retains semantic traces of bloodline-related personal identity information, enabling the system to identify the bloodline relationship between node B's data and the highly sensitive source data, rather than treating it as a completely independent new data asset. Nodes C and D further inherit the upstream sensitive features from node A via node B through multi-hop propagation. The spatial channel thus automatically captures the sensitive attribute inheritance relationship along the entire data flow link. The temporal channel performs temporal encoding on the incoming event sequence of each node: the incoming event sequence of node B contains only the single-step desensitization event at time T1, and the temporal channel outputs a temporal representation vector encoding the "single-step desensitization" processing mode; the incoming event of node C contains only the convergence event at time T2, and the temporal channel captures the causal order of "desensitization first, convergence later", forming a distinguishable vector difference with the temporal representation generated by "convergence first, desensitization later"; the temporal channels of nodes E and F operate independently and are not affected by any node in the personal information subgraph.

[0099] After the spatiotemporal representation vectors are fused, the spatiotemporal sensitivity representation vectors of each node are written into the corresponding memory state. The sensitivity discrimination classification head outputs sensitivity labeling results based on the spatiotemporal representations of each node: Node A is labeled as highly sensitive (personal identity information and location trajectory information); Node B is labeled as moderately sensitive because the direct field sensitivity is reduced by the desensitization operation, but the bloodline relationship attribute still exists; Node C maintains the moderately sensitive labeling because the aggregation operation may partially restore the identifiability of the user profile; Node D is labeled as moderately sensitive due to the anonymization effect of multi-step statistical aggregation; Nodes E and F are independently determined based on the supply chain inventory data attributes and are not affected by the personal information subgraph. After selective conformal risk control classification in step S6, Node A is determined to be core data (level four), Nodes B and C are determined to be important data (level three), and Node D is determined to be general data (level two). The above case fully demonstrates the ability of the spatiotemporal dual-channel lineage graph representation learning method to automatically and accurately determine the causal order of the inheritance and processing of sensitive attributes of derived data in multi-path data flow scenarios, as well as the isolation characteristic that the hierarchical results between two completely independent lineage subgraphs do not interfere with each other.

[0100] Example 6

[0101] This embodiment, based on embodiment 5, further explains the specific implementation method of temporal encoding of the incoming event sequence by the temporal channel in step S5.3 to obtain the temporal representation vector. This embodiment includes:

[0102] The incoming event sequence is encoded according to the order of occurrence of each incoming event to capture the causal pattern between processing events and obtain the temporal representation vector; wherein, the causal pattern is used to distinguish data flow scenarios with different processing times and different sensitivities, and the data flow scenarios include flow scenarios with desensitization before aggregation and flow scenarios with aggregation before desensitization.

[0103] In this step, for a target node in the temporal lineage graph, its incoming edge event sequence is formed by arranging the processing events corresponding to all directed edges directly pointing to that node in chronological order of their occurrence. The temporal channel inputs this event sequence into the temporal coding model, using the absolute timestamp of the event occurrence time and the relative time interval to form the temporal position code, and the embedding vector of the processing type as the event content code. After the two are superimposed, they are processed step by step by the temporal coding model to obtain a temporal representation vector that captures the causal relationship between the events.

[0104] Optionally, the construction of the time position embedding is as follows: For absolute timestamps, the Unix timestamp of the event occurrence is logarithmically scaled to a fixed interval after taking the base-10 logarithm, and then converted into a time feature vector of the same dimension as the processing type embedding vector through a learnable scalar time projection layer; for relative time intervals, the time difference (in seconds) between the current event and the previous event in the sequence is calculated, and similarly, after logarithmic transformation to compress the magnitude difference of large-span intervals, it is projected into an interval feature vector of the same dimension. The absolute time feature vector and the relative interval feature vector are added element-wise to obtain the time position embedding vector of the current event, and then added to the processing type embedding vector to serve as the complete input representation of the corresponding time step of the time-series coding model. The introduction of logarithmic transformation effectively compresses the magnitude difference between minute-level intervals and month-level intervals, avoiding numerical instability in the time features of long-span processing event sequences.

[0105] Regarding the architecture selection of temporal coding models, the following principles can be considered: When the length of the incoming event sequence of a lineage node is short (generally no more than a few dozen events) and system computing resources are limited, lightweight recurrent neural networks such as gated recurrent units (GRUs) are preferred as the backbone network of the temporal channel. Their hidden states are naturally adapted to sequential causal modeling of sequences, and inference latency is low. When the historical processing links of lineage nodes are long, the event temporal span is large, and it is necessary to capture long-range temporal dependencies (such as the long-range impact of early desensitization operations on later convergence sensitivity), a temporal encoder based on the Transformer architecture is preferred. Its multi-head self-attention mechanism can establish direct temporal associations between pairs at any position in the sequence, which has a significant advantage in capturing long-range causal relationships. In engineering deployment, a suitable architecture can be pre-selected based on the topological statistical characteristics of the business lineage graph (such as the average number of incoming events and the longest lineage path depth), or a lightweight architecture can be used as the default option, and nodes exceeding the sequence length threshold can be automatically switched to the Transformer encoder to achieve a dynamic balance between computational efficiency and modeling capability.

[0106] To illustrate the necessity of distinguishing between temporal and causal factors, consider a typical scenario: If the original data table A contains name and ID number fields (highly sensitive), and the processing path is "first desensitize (remove ID number) → then aggregate (merge with address data)," the final sensitivity level of the aggregated result will be low. However, if the processing path is "first aggregate (merge with address data to form name-ID-address association records) → then desensitize (only remove ID number)," highly sensitive association records have already been generated during the aggregation stage. Even if partial desensitization is performed subsequently, this data node should still maintain a high sensitivity level. The temporal channel explicitly distinguishes and encodes the two event sequences of "desensitize first, then aggregate" and "aggregate first, then desensitize," enabling the classification head to output sensitivity level judgments that match the semantics of actual data processing. This fundamentally solves the misjudgment problem caused by the static topology graph method ignoring temporal and causal factors.

[0107] Taking the temporal encoding process of node C in the above case as a specific example, the operational details of the temporal position embedding construction and temporal encoding model processing are further explained. Assume node C has two incoming edges: one is edge B→C, corresponding to the convergence processing event, with a Unix timestamp of t_BC (corresponding to 14:00:00 on a certain day, on the order of approximately 1.70 × 10⁻⁶). 9 The first is edge G→C (node ​​G is a user order database table), corresponding to the associated processing event, with a Unix timestamp of t_GC (occurring approximately 3600 seconds earlier than edge B→C, i.e., 13:00:00 on the same day). The sequence of incoming edge events for node C, sorted by occurrence time, is: event 1 ("association" operation, time t_GC), event 2 ("convergence" operation, time t_BC).

[0108] In the temporal embedding construction stage, for event 1, its absolute temporal feature is obtained by taking the logarithm of t_GC to base 10 (approximately 9.23) and then linearly scaling and projecting it into a d-dimensional absolute temporal feature vector τ1_abs. Since event 1 is the first event in the sequence, its relative time interval is defined as 0. After logarithmic transformation, it is mapped to the endpoint of a predefined minimum interval and projected into a d-dimensional relative interval feature vector τ1_rel. The two are added element-wise to obtain the temporal embedding e1_time of event 1, which is then superimposed with the "association" operation type embedding vector to obtain the complete input representation x1 of event 1. For event 2, its absolute temporal feature vector τ2_abs is obtained by linearly scaling and projecting the logarithm of t_BC (also of the same magnitude, approximately 9.23). The relative time interval between event 2 and event 1 is 3600 seconds, which is approximately 3.56 after taking the logarithm to base 10 and projected into a d-dimensional relative interval feature vector τ2_rel. The two are superimposed to obtain the temporal embedding e2_time, which is then superimposed with the "convergence" operation type embedding vector to obtain the complete input representation x2 of event 2. It can be observed that the logarithmic transformation effectively compresses the numerical span between time intervals of different orders of magnitude—a 3600-second interval corresponds to a numerical value of approximately 3.56, a 86400-second interval corresponds to approximately 4.94, and a monthly interval corresponds to approximately 2.6 × 10⁻⁶. 6 The value corresponding to a second is approximately 6.41. The intervals between each order of magnitude are uniformly distributed in the logarithmic space to approximately 1 to 1.5 units. This allows the time-series coding model to maintain stable feature representation within a similar numerical range when dealing with processing time intervals ranging from minutes to months, fundamentally avoiding the problem of unstable time feature values ​​in long-term processing chains.

[0109] In the processing stage of the temporal coding model, taking the gated recurrent unit (GRU) as an example, the temporal encoder receives x1 and x2 in sequence: When processing x1, the update gate and reset gate of the GRU update the hidden state h1 based on the content of x1 ("association" operation type embedding superimposed with the time position embedding at the corresponding moment). h1 implicitly encodes the processing mode information of "the sequence starts with the association operation". When processing x2, the GRU uses h1 as the historical state basis and combines it with x2 ("convergence" operation type embedding superimposed with the time position embedding at an interval of about 3600 seconds) to update and obtain the final hidden state h2. h2 also encodes the temporal processing causal pattern of "association operation first, convergence operation is performed after an interval of about one hour", which is output as the temporal representation vector of node C. It should be understood that if the processing events of two incoming edges are swapped (i.e., the convergence operation occurs earlier than the association operation), the type embedding vectors and relative time interval features corresponding to event 1 and event 2 will both change. The hidden state output by GRU will encode a different temporal pattern of "convergence first, association later," producing a distinguishable vector difference from the above description. It is this difference that enables the temporal channel to provide the sensitivity discrimination classification head with feature inputs reflecting the essential differences in processing temporal order, supporting the system to make correct sensitivity discrimination judgments in data flow scenarios with identical topological structures but different temporal orders. The timestamp values ​​and embedding dimensions mentioned above are illustrative parameters. In actual deployment, the scaling parameters should be calibrated accordingly based on the temporal span characteristics of the business lineage graph.

[0110] Example 7

[0111] This embodiment, based on Embodiment 1, further explains the specific implementation method of matching the sensitivity labeling results with the compliance template library and performing selective conformal risk control judgment in step S6, outputting the node sensitivity level judgment result with a false negative rate not exceeding a preset upper limit. In this invention, the selective conformal risk control grading judgment mechanism refers to: combining the selective prediction framework and the conformal prediction framework to implement two-stage risk control for the sensitivity level judgment of data asset nodes. The first stage uses a selection function to filter out nodes with insufficient overall judgment confidence and imports them into the manual review channel; the second stage only uses conformal prediction technology to construct a level prediction set for the selected nodes, and without relying on any data distribution assumptions, makes the false negative rate of the prediction set—that is, the probability that the real high sensitivity level is mistakenly excluded from the prediction set—have a mathematical upper limit that can be pre-configured by the user. Compared with standard conformal prediction, which constructs a prediction set indiscriminately for all samples, the "selectivity" of this invention is reflected in the fact that by pre-screening low-confidence nodes, a conformal prediction set is constructed only for nodes that the model has a certain degree of confidence in. Under the same calibration dataset size, a prediction set with a smaller cardinality (more accurate judgment) is obtained, which improves the coverage ratio of nodes that can be automatically judged. At the same time, nodes with insufficient confidence in judgment are covered by manual review, forming a complementary closed loop of automatic judgment and manual review.

[0112] The "false negative rate" is specifically defined as the proportion of nodes whose actual sensitivity level is high, but which are ultimately classified as low by the automatic judgment mechanism; that is, the rate of misclassification of high-sensitivity nodes. In data compliance scenarios, the compliance penalty risk of failing to detect highly sensitive data far outweighs the additional processing costs of falsely reporting low-sensitivity data. Therefore, the false negative rate is a core metric for the compliance reliability of the hierarchical classification system. By setting the upper limit of the false negative rate α to a regulatory-acceptable risk threshold (e.g., α=0.01 corresponds to a false negative rate of no more than 1%), enterprises can mathematically provide a quantifiable and verifiable risk control commitment for their automatic hierarchical classification process, upgrading the hierarchical classification results from a "soft commitment" based on engineering experience to a "hard commitment" with statistical assurance, thus meeting the data compliance audit requirements for the verifiability of results.

[0113] The core technical principle of this mechanism originates from the finite sample permutation theory: Suppose the calibration dataset contains n samples with true grade labels. For each calibration sample, calculate its inconsistency score s(x,y) (characterizing the degree of inconsistency between the model prediction and the true grade), and use α to determine the ⌈(1−α)×(n+1)⌉th smallest inconsistency score as the conformal threshold τ. For a new target node, include all candidate grades with inconsistency scores not exceeding τ into the prediction set. According to permutation statistics theory, the n calibration samples and the new sample are considered as any permutation of n+1 commutative samples. The ranking of the new sample's inconsistency score among these n+1 values ​​is uniformly distributed in units of 1 / (n+1), thus strictly ensuring that the probability of the new sample's true grade being excluded from the prediction set does not exceed α. The commutability assumption only requires that the calibration samples and the new samples come from the same data generation process, without requiring any parameterized distribution form. Therefore, this guarantee is equally valid even in complex real-world situations such as data distribution shifts, class imbalances, and long-tailed distributions, and is the mathematical basis for the provability of the classification results of this invention.

[0114] like Figure 4 As shown, this embodiment includes the following steps:

[0115] S6.1: Match the sensitivity labeling results with the mapping rules from sensitivity types to level standards in the compliance template library to obtain the candidate sensitivity level distribution of each node.

[0116] In this step, the compliance template library contains multiple sets of regulatory templates, such as a certain law, a certain protection regulation, a certain legal provision, and industry data classification and grading guidelines. Each set of regulatory templates defines the correspondence between sensitive types and the regulatory level standard in the form of configurable mapping rules.

[0117] In this embodiment of the invention, the data structure design of the compliance template library is as follows: Each set of regulatory templates consists of two parts: template metadata and a set of mapping rules. The template metadata includes fields such as regulatory identifier, full name of the regulatory, applicable region, effective date, and template version number, which are used for the unique identification of the template and historical version tracking management. The set of mapping rules consists of several atomic mapping rules. Each atomic rule is in the format of a triple (source sensitive type identifier, target level identifier, condition modifier), where the source sensitive type identifier references the standard category in the unified sensitive type system (such as "personal identity information", "financial account information", "health and medical information", "location trajectory information", etc.), the target level identifier corresponds to the specific level defined by the regulation, and the condition modifier is an optional field used to describe the level upgrade or downgrade rules under special scenarios (for example, when location trajectory information appears in the same data asset together with personal identity information, it should be automatically upgraded to a higher protection level according to the regulation). The compliance template library supports parallel mapping calculations for multiple templates: the same sensitivity labeling result can be mapped and matched with multiple active regulatory templates simultaneously, and the hierarchical views under each regulatory standard can be output in parallel; when a specific regulatory standard is revised, the system only needs to update the mapping rule set of the corresponding regulatory template in the compliance template library, and re-execute the mapping matching operation of this step for the nodes that have completed the lineage analysis, without having to re-trigger upstream steps with high computational costs such as multimodal content extraction, cascade recognition and lineage graph propagation.

[0118] The system matches the sensitivity labeling results (including sensitivity type and propagation intensity) of each lineage node obtained in step S5 of Example 1 with the mapping rules under the user-specified target compliance standard, generating a candidate sensitivity level distribution for each node, that is, a distribution representation of the various sensitivity levels that the node may belong to under the current compliance standard and the corresponding probabilities. The multi-regulation built-in design of the compliance template library supports one-time identification and multi-standard classification. When the regulatory standards change, only the mapping configuration needs to be adjusted, without having to re-execute the identification and lineage analysis process.

[0119] S6.2: Use a selection function to select the first set of nodes whose confidence level reaches the preset selection threshold from each node.

[0120] In this step, a selection function is used to evaluate the confidence level of the candidate sensitivity level distribution of all nodes, and nodes whose confidence level meets the preset selection threshold are selected to form the first node set. A typical form of the selection function is the maximum confidence selection function, which means that when the confidence level of the highest probability level in the candidate sensitivity level distribution of a node exceeds the preset selection threshold, the node is included in the first node set; otherwise, it is considered that the current model has insufficient confidence in the judgment of that node, and it is not included in the first node set. The selection threshold is a user-configurable parameter, and its value determines the balance between automatic judgment coverage and judgment quality.

[0121] S6.3: Calculate the conformal quantile threshold on the pre-held calibration dataset, and construct a grade prediction set for each node in the first node set based on the conformal quantile threshold. The grade prediction set ensures that the false negative rate does not exceed a preset upper limit without relying on the data distribution assumption.

[0122] In this step, a conformal prediction framework is used to construct a statistically guaranteed set of grade predictions for each node in the first set of nodes. Specifically, an independent calibration dataset is held in advance before system deployment, in which each sample has a known sensitivity grade label.

[0123] In this invention, for the sensitivity level classification task, the inconsistency score is defined as follows: For a sample x in the calibration dataset with a true sensitivity level of y, its inconsistency score s(x,y) is defined as 1 minus the predicted probability of the model's output for the candidate level y, i.e., s(x,y) = 1 − f_y(x), where f_y(x) is the probability output value corresponding to that level after temperature scaling calibration. Under the above definition, a higher inconsistency score indicates a lower confidence level in the model's prediction of the true level of the sample; a lower inconsistency score indicates a more consistent prediction of the true level of the sample with the actual situation.

[0124] Specifically, the calculation process for the conformal quantile threshold τ is as follows: First, calculate the inconsistency score for each of the n samples in the calibration dataset, obtaining an ordered set of inconsistency scores {s_1, s_2, …, s_n}; second, determine the value of τ according to the following quantile levels: τ is the ⌈(1−α)×(n+1)⌉th smallest inconsistency score (⌈·⌉ represents rounding up, and α is the user-preset upper limit of the false negative rate). For each target node in the first node set, include all candidate sensitivity levels y that satisfy s(x,y)≤τ (i.e., f_y(x)≥1−τ) into the level prediction set of that node. Conformal prediction theory proves that, under the condition that the calibration samples and new samples satisfy the exchangeability assumption—which essentially only requires that the samples come from the same data generation process, without requiring a specific parameterized distribution form—the probability that the level prediction set excludes the true high sensitivity level (i.e., the false negative rate) strictly does not exceed α under finite sample conditions, without making any assumptions about the data distribution. It should be understood that when the data distribution shifts significantly, the calibration dataset should be updated accordingly and the conformal quantile threshold should be recalculated to ensure the continued effectiveness of the false negative rate guarantee.

[0125] S6.4: For nodes whose base number in the level prediction set does not exceed a preset base number threshold, the highest sensitivity level in the level prediction set is taken as the final sensitivity level of the nodes whose base number in the level prediction set does not exceed the preset base number threshold, and the node sensitivity level determination result is output.

[0126] In this step, nodes whose level prediction sets have been constructed in S6.3 are classified according to the cardinality of the prediction set. The cardinality of the level prediction set (i.e., the number of candidate levels contained in the set) reflects the certainty of the model's judgment: a cardinality of 1 means that the model definitively points to a single level while ensuring the false negative rate; a larger cardinality means that even while ensuring the false negative rate, there are still multiple levels to be determined, and the judgment is not certain enough. For nodes whose level prediction set cardinality does not exceed the preset cardinality threshold, the highest sensitivity level in the node's level prediction set is taken as its final sensitivity level according to the principle of returning to the highest level, and the node's sensitivity level judgment result is output. The principle of returning to the highest level ensures that in the uncertain situation of multiple levels coexisting, a more conservative sensitivity level is prioritized to minimize the compliance risk of underreporting highly sensitive data.

[0127] The complete operation of the above mechanism is illustrated using a specific numerical scenario: Assume the calibration dataset contains 200 node samples, and the user sets the upper limit of the false negative rate α = 0.05 (i.e., 5%). Then the conformal threshold τ is the value corresponding to the 191st smallest inconsistency score (⌈0.95×201⌉), which is assumed to be 0.80. For a target node A in the first node set filtered by the selection function, assuming its predicted probabilities after temperature scaling calibration at the four sensitivity levels (levels 1 to 4) are [0.05, 0.10, 0.70, 0.15], the inconsistency scores corresponding to each level are [0.95, 0.90, 0.30, 0.85]. There are only three candidate levels with an inconsistency score not exceeding τ = 0.80 (inconsistency score of 0.30). Therefore, the predicted level set for node A is {level 3}, with a cardinality of 1. Following the principle of categorizing by the highest level, level 3 is directly taken as the final sensitivity level output, and this result comes with a statistical guarantee that the false negative rate does not exceed α. For another node B, assuming its inconsistency scores for the four levels are [0.78, 0.62, 0.71, 0.76], then the inconsistency scores for all four levels do not exceed τ=0.80, and the cardinality of the level prediction set is 4, exceeding the preset cardinality threshold. Node B is automatically routed to the manual review queue, where professionals determine the final level based on the business context, forming a double-safety loop of automatic judgment and manual review. Through the above two-stage processing, the system provides a strict statistical guarantee that the false negative rate of the automatically judgeable node set within the calibration distribution does not exceed the preset upper limit α. It is worth noting that the above values ​​are illustrative assumptions, and the actual conformal threshold should be calculated based on real calibration data during formal deployment.

[0128] In one embodiment of the present invention, the compliance template library-driven multi-regulatory configurable hierarchical mapping explicitly decouples the two stages of "sensitive entity identification and lineage propagation" and "compliance hierarchical standard adaptation" at the system architecture level. The compliance template library serves as the configurable interface layer between the two. The identification and lineage analysis stage outputs a regulatory-neutral unified sensitive type labeling result. The hierarchical stage reads the mapping rules corresponding to the target regulation from the compliance template library and converts the unified sensitive type labeling result into a specific level under that regulatory standard. This allows the same identification and lineage analysis result to output a hierarchical view under any built-in regulatory template without re-executing the identification process. Different regulations highly overlap in the objects of sensitive information identification—all targeting common sensitive categories such as personal identification information, financial account information, and health and medical information—but there are significant differences in the hierarchical classification methods, the protection levels corresponding to various types of sensitive information, and the joint level calculation rules in special scenarios. By introducing a unified sensitive type system as a common semantic interface between the identification layer and the hierarchical layer, the identification layer only needs to maintain a type system that is independent of regulations, while the hierarchical layer independently adapts to various regulatory standards through configurable template mapping rules. The evolution of the two is independent of each other, which significantly reduces the maintenance complexity of the system in multi-jurisdictional compliance scenarios.

[0129] In implementing parallel hierarchical classification across multiple regulations, after completing the sensitivity labeling of lineage nodes, the system simultaneously performs mapping and matching of the labeling results with all active regulatory templates in the compliance template library, generating multiple sets of candidate sensitivity level distributions for each node in parallel. Each set of classification results is stored in the data security map as an independent hierarchical view. Enterprise operators can retrieve the corresponding hierarchical view under the relevant regulatory standard as needed based on the target object of the current report (such as regulatory agencies, industry associations, internal compliance committees, etc. in different regions), realizing customized output of compliance reports on demand. Taking a typical multi-jurisdictional compliance scenario of a multinational financial institution as an example: After the financial business data of the same batch of domestic and foreign customers are identified and analyzed by this invention, the sensitivity labeling results include three unified sensitive types: "personal identity information," "financial account information," and "location trajectory information." Based on a domestic data security classification and hierarchical compliance template, "personal identity information" is mapped as important data (level 3), and "financial account information" is mapped as core data (level 4); while based on a compliance template of an international data protection regulation, the same "personal identity information" and "financial account information" are both mapped as the highest level special category data. Through parallel mapping of the compliance template library, the organization can simultaneously output complete hierarchical reports under two sets of regulatory standards based on a single identification process, meeting the need to submit compliance materials required by regulatory agencies in different jurisdictions. When a revision of a regulatory provision causes an adjustment in the hierarchical mapping relationship, it is only necessary to update the mapping rule configuration of the corresponding regulatory template in the compliance template library, and trigger a partial recalculation of the affected nodes only in the mapping matching step S6.1. The completion time of the entire adaptation process is significantly reduced compared to the traditional full-scale re-scanning solution, enabling enterprises to cope with the continuous dynamic changes in regulatory standards at extremely low operating costs.

[0130] Example 8

[0131] This embodiment, based on Embodiment 7, further explains the specific implementation method of routing nodes with insufficient certainty of judgment to the manual review queue. This embodiment also includes the following steps in addition to Embodiment 7:

[0132] Nodes whose cardinality in the predicted level set exceeds the preset cardinality threshold, as well as nodes that were not selected by the selection function to enter the first node set, are routed to the manual review queue, where their final sensitivity level is determined by manual review.

[0133] In this step, the following two types of nodes will be automatically routed to the manual review queue: The first type is nodes whose grade prediction set exceeds a preset threshold. Although these nodes have passed the confidence selection in S6.2, their grade prediction set still contains too many candidate grades while meeting the false negative rate guarantee, indicating that the current model cannot provide sufficiently accurate single grade judgments while maintaining risk assurance. The second type is low-confidence nodes that did not pass the S6.2 selection function screening. The judgment confidence of these nodes did not reach the preset selection threshold, and the model lacks sufficient understanding of their overall judgment. Both types of nodes enter the manual review queue, where professionals conduct manual reviews based on business context, data content, and compliance requirements. Their sensitivity level is ultimately determined and fed back to the system, forming a double-safety loop of automatic judgment and manual review. The results of the manual review can also serve as new calibration samples, continuously enriching the calibration dataset, improving the coverage and accuracy of subsequent automatic judgments, and achieving continuous self-optimization of the system.

[0134] Example 9

[0135] This embodiment, based on embodiment 5, further illustrates the specific implementation method of maintaining the real-time performance of the data security map through event-driven incremental closed loop.

[0136] In this invention, the node-memorized event-driven incremental closed-loop update mechanism refers to: using the persistent memory state of each node in the lineage graph as a reuse carrier for intermediate calculation results, combined with an event-driven triggering mechanism, precisely limiting the processing cost of data change events to the local subgraph (i.e., the influence domain) on the lineage graph that is only affected by the change. This completely decouples the update calculation cost of the data security map from the global node scale of the lineage graph and makes it an incremental update architecture that is only proportional to the size of the influence domain. The core technical contribution of this mechanism is that it changes the maintenance paradigm of the data security map from "timed full reconstruction" to "event-triggered local correction," compressing the update latency from hours of full rescanning to seconds of event influence domain processing, fundamentally meeting the real-time governance needs of large-scale dynamic data environments.

[0137] The technical principle of this decoupling mechanism for computational costs is based on the following decomposability: the spatiotemporal sensitivity representation vector of a node in the lineage graph is determined by three types of inputs—the node's local sensitive entity features, the current memory state of all its direct upstream neighbors, and the node's incoming edge event sequence. If none of these three types of inputs to a node change during the current data change event, the result of the node's spatiotemporal sensitivity representation vector in the new round of propagation computation is completely consistent with the historical memory state, and no recalculation is required. The node memory state is persistently stored by storing the spatiotemporal representation vector output by each node in the most recent propagation, providing intermediate results that can be directly read at any time for the neighborhood aggregation operation of other nodes. When the data change event only affects a few source nodes, the three types of inputs of all nodes outside the affected domain do not change, and their memory states can be safely reused. The entire update process only performs propagation and hierarchical judgment calculations within the affected domain. Therefore, the computational cost is strictly proportional to the number of nodes in the affected domain and is completely independent of the number of nodes outside the affected domain.

[0138] like Figure 5 As shown, this embodiment adds the following steps to embodiment 5:

[0139] Step A': Listen to the data change event stream.

[0140] In this step, the system establishes a continuous monitoring mechanism for data change event streams. These streams originate from various sources, including database change logs (such as MySQL binlog and PostgreSQL WAL), file system change notifications, and data platform API pushes. The system uses an event-driven approach to detect data additions, modifications, and deletions in real time. Whenever a change event occurs, it triggers subsequent incremental processing, replacing the traditional periodic full scan mechanism and fundamentally eliminating the time delay of scheduled scans.

[0141] Step B': When a data change event occurs, the changed data is extracted, and after scanning by the rule layer, judgment by the multimodal semantic recognition layer, and unified calibration of confidence, an incremental sensitive entity set is obtained.

[0142] In this step, for the data change event detected in step A', the system only extracts partial content from the changed data. A unified intermediate representation sequence of the changed data is obtained through the multimodal content extraction process described in step S1 of Example 1. Subsequently, following the rule layer scanning described in step S2 of Example 1, the multimodal semantic recognition layer determination described in step S3 of Example 1, and the confidence level unified calibration process described in step S4 of Example 1, an incremental sensitive entity set is obtained, focusing only on the changed content. The design of partial content extraction and incremental recognition ensures that the processing cost of a single change event is only related to the range of the changed data, without involving a rescan of the entire dataset, significantly reducing the computational overhead of real-time updates.

[0143] Step C': Based on the incremental sensitive entity set, determine the influence domain of the data change event in the temporal lineage graph, locally update the memory state of the nodes within the influence domain, and re-execute the selective conformal risk control judgment on the nodes within the influence domain to obtain the updated node sensitivity level judgment result.

[0144] In this step, based on the incremental sensitive entity set obtained in step B', the changed data asset nodes are located in the temporal lineage graph, and reachability analysis is performed along the directed edges of the lineage to determine the influence domain affected by this change event. Only nodes within the influence domain are recalculated using the spatiotemporal dual-channel propagation mechanism described in Example 5 to update their memory states; the memory states of nodes outside the influence domain remain unchanged and do not require recalculation. After completing the local update of the memory states, the selective conformal risk control judgment described in Example 7 is re-executed only for nodes within the influence domain to obtain the updated node sensitivity level judgment results, and the level heatmap display of the affected area in the data security map is updated accordingly. Since the node memory states preserve historical spatiotemporal representations, incremental updates only need to be performed within the influence domain. The update computation cost is proportional to the size of the influence domain and independent of the overall map size, thereby reducing the update latency of the data security map from hours for a full rescan to seconds for the event influence domain, supporting the real-time governance needs of large-scale dynamic data environments.

[0145] Taking the daily operation scenario of a certain enterprise data platform as an example, the complete operation process of the node memory event-driven incremental closed-loop update mechanism is explained. The enterprise data platform has completed full initialization analysis, and the time-series lineage graph includes the six nodes (nodes A to F) involved in the case of Example 5 above. Each node has persistently stored the corresponding spatiotemporal sensitivity representation vector as a memory state. The initial classification results of each node in the data security map are as follows: Node A is core data (level 4), Nodes B and C are important data (level 3), Node D is general data (level 2), and Nodes E and F are determined according to the supply chain data attributes (both are general data (level 2)).

[0146] One day, the system detected a structural change in the original user information database table corresponding to node A by continuously monitoring the MySQL binlog event stream (step A'): ​​due to new business needs, the company added an "emergency contact phone number" field to the table and wrote the corresponding data into the existing user records in batches. The above binlog change event was detected by the system in real time and immediately triggered the incremental identification process in step B'. The system only performed partial content extraction on the newly added "emergency contact phone number" field data to obtain a unified intermediate representation sequence containing only the content of the newly added field; in the rule layer scanning stage, the phone number data in the newly added field all conformed to the eleven-digit format and passed the format validation, and were directly hit as phone number sensitive entities, and all were included in the incremental sensitive entity set; due to the high degree of structure of the newly added field content, the number of calls to the multimodal semantic recognition layer in this incremental identification was extremely limited, and the overall time consumption of incremental identification was much lower than the time cost of rescanning the full data.

[0147] In step C', the system performs forward reachability analysis along the directed edge A→B→C→D in the temporal lineage graph based on node A associated with the incremental sensitive entity set. This determines that the impact domain of this change is four nodes: {node A, node B, node C, node D}. Since nodes E and F have no lineage connection with node A, their memory states remain unchanged throughout the process and they do not participate in any recalculation. The hierarchical view related to the supply chain data is unaffected throughout the entire change event handling process. For nodes A to D within the influence domain, the spatiotemporal dual-channel propagation mechanism recalculates their spatiotemporal sensitivity representation vectors and updates their memory states. Subsequently, the selective conformal risk control judgment is re-executed for each node within the influence domain: Node A, due to the addition of a sensitive mobile phone number field locally, has a further increased confidence in identifying sensitive attributes of personal identity information, and its final sensitivity level remains unchanged as core data (level four); Nodes B and C, due to the propagation of newly added sensitive attributes upstream through kinship edges, have seen an increase in the weight of personal identity information-related features in their spatiotemporal representation vectors, and their classification results remain as important data (level three), but their corresponding conformal confidence has significantly increased; Node D, under the influence of multi-hop propagation upstream, experiences a shift in its candidate sensitivity level distribution. Through conformal prediction set construction and classification from higher levels, its final sensitivity level is upgraded from general data (level two) to important data (level three). The updated classification results are synchronized to the data security map in real time. The end-to-end processing latency from the time the binlog event is perceived by the system to the time the data security map is refreshed is much lower than that of the timed full rescan scheme, effectively supporting the need for real-time perception and continuous governance of the sensitive status of data assets in a large-scale dynamic data environment.

[0148] Example 10

[0149] This embodiment, based on embodiment 9, further explains the specific implementation methods of determining the influence domain, locally updating the memory state, and locally re-determining in step C'. This embodiment includes the following steps:

[0150] Step C'.1: Determine the influence domain based on the nodes associated with the incremental sensitive entity set and their downstream nodes reachable along directed edges in the temporal lineage graph.

[0151] In this step, the influence domain is determined using forward reachability analysis on the lineage graph. First, based on the data assets to which each sensitive entity in the incremental sensitive entity set belongs, the corresponding set of source nodes is located in the time-series lineage graph, and these source nodes are marked as the influencing starting points. Then, starting from the influencing starting point, a breadth-first search (BFS) or depth-first search (DFS) is performed along the directed edges of the lineage, traversing all downstream nodes reachable along the directed edges. The set of source nodes and all their downstream reachable nodes together constitute the influence domain of this data change event. The scope of the influence domain precisely defines the propagation boundary of this change in the lineage topology, ensuring that subsequent updates are performed only on nodes truly affected by the change, while absolutely guaranteeing that nodes unrelated to the change are not included in the update scope.

[0152] Step C'.2: Only for nodes within the influence domain, re-execute the propagation calculation of the spatial channel and the temporal channel to update the memory state of the nodes within the influence domain, while the memory state of the nodes outside the influence domain remains unchanged.

[0153] In this step, for each node within the influence domain determined in step C.1, the current memory state of its neighboring nodes is used (the memory state of nodes outside the influence domain can be directly reused as it remains unchanged). The spatial channel neighbor message passing described in step S5.2 of Example 5 and the temporal channel temporal coding described in step S5.3 are then re-executed to obtain the updated spatial representation vector and temporal representation vector for each node within the influence domain. The updated spatiotemporal sensitivity representation vector is obtained through the fusion operation in step S5.4 and written into the memory state of the corresponding node, overwriting the historical state. The memory state of nodes outside the influence domain remains unchanged throughout the process and does not participate in any recalculation, thus achieving an incremental update characteristic where the computational cost is strictly proportional to the size of the influence domain.

[0154] Step C'.3: Re-execute the selective conformal risk control determination only for nodes within the influence domain to obtain the updated node sensitivity level determination result; wherein, the computational cost of incremental update is proportional to the size of the influence domain.

[0155] In this step, the updated spatiotemporal sensitivity representation vector obtained in step C.2 is used to re-execute the entire selective conformal risk control judgment process described in Example 7 for nodes within the influence domain. This includes mapping and matching with the compliance template library, selection function confidence screening, conformal prediction set construction, and judgment from high-level nodes, resulting in updated node sensitivity level judgment results. Simultaneously, the manual review routing mechanism described in Example 8 is also applicable in incremental updates, automatically routing nodes with insufficient confidence in their judgment within the influence domain to the manual review queue. Sensitivity level judgment results for nodes outside the influence domain do not need to be recalculated; historical results are directly used. The updated node sensitivity level judgment results are synchronized to the data security map in real time, refreshing the corresponding node's level heat map distribution and flow identifier. Combining all the processing logic from steps C.1 to C.3, the computational load of each step in the incremental update closed loop is strictly proportional to the node size of the influence domain, completely decoupled from the size of nodes outside the influence domain in the entire graph. Even in large-scale dynamic data governance scenarios with continuously growing data asset scale, stable real-time update response capabilities can still be maintained.

[0156] It should be noted that when large-scale batch data changes occur (such as data architecture reorganization leading to the simultaneous migration of a large number of data tables), the affected domain may cover most nodes of the lineage graph. In this case, the computational cost of incremental updates approaches that of a full rescan. The system can set an early warning threshold for the affected domain size. When the affected domain of a single change event exceeds this threshold, it is automatically merged into a batch full update task for unified scheduling. This avoids a large number of local update requests in a short period of time, which could cause system load fluctuations. It ensures that the system can maintain the real-time performance of the data security map at a near-optimal computational cost under various change scales.

[0157] Based on the "emergency contact mobile phone number" field addition event described in Embodiment 9 above, the complete execution process of steps C.1 to C.3 will be further explained in detail with specific operation details.

[0158] In the influence domain determination stage (step C.1), the system performs a breadth-first search starting from node A: starting from node A, it follows the directed edge A→B to reach node B, adding node B to the influence domain; then, starting from node B, it follows the directed edge B→C to reach node C, adding node C to the influence domain; starting from node C, it follows the directed edge C→D to reach node D, adding node D to the influence domain; node D has no outgoing edges, so the BFS traversal terminates. The final influence domain is determined to be four nodes: {node A, node B, node C, node D}. During the traversal, nodes E and F are never visited, and the influence domain clearly excludes these two nodes, ensuring that subsequent update calculations are strictly limited to within the influence domain.

[0159] During the local update phase of the memory state (step C.2), the system sequentially performs spatiotemporal dual-channel propagation calculations from upstream to downstream according to the lineage topology to ensure that the memory states of each node's direct upstream neighbors have been updated in this round when performing neighborhood aggregation. First, node A is processed: its spatial channel fuses the incremental sensitive entity feature vector corresponding to the newly added "emergency contact phone number" field with the historical local feature vector to update the spatial representation vector; node A's incoming edge event sequence has no new events, and the temporal representation vector directly reuses historical results; after spatiotemporal fusion, the newly calculated spatiotemporal representation vector overwrites node A's memory state. Next, node B is processed: during neighborhood aggregation, the spatial channel reads the newly updated memory state (containing the spatiotemporal representation vector of the newly added phone number sensitive feature) from node A, and after desensitization processing and weight aggregation of edge A→B, updates node B's spatial representation vector; the temporal representation vector reuses historical results; after spatiotemporal fusion, it overwrites node B's memory state. This process continues, sequentially updating the memory states of nodes C and D. It is worth noting that the memory states of nodes E and F are not read or written during the entire update process. If the neighborhood aggregation of a node in the domain needs to be affected and node E or node F needs to be referenced as an upstream neighbor (such situations exist in other lineage topology scenarios), its historical memory state can be read directly without triggering any additional calculations. This fully demonstrates the support capability of the node memory state persistence design for cross-domain computation reuse.

[0160] In the local re-determination stage (step C.3), the updated spatiotemporal sensitivity representation vectors obtained in step C.2 are used to re-execute the mapping matching and conformal risk control determination with the compliance template library for nodes A to D in the influence domain. Taking node D as an example: After the update, the probability of level 2 in its candidate sensitivity level distribution decreases from 0.78 to 0.52, and the probability of level 3 increases from 0.18 to 0.41. Assuming the current conformal quantile threshold τ of the system is 0.80, both level 2 (inconsistency score 1 − 0.52 = 0.48) and level 3 (inconsistency score 1 − 0.41 = 0.59) satisfy the condition of not exceeding τ. The level prediction set of node D expands from the historical {level 2} to {level 2, level 3}, and the cardinality increases from 1 to 2. Assuming the preset cardinality threshold is 2, and cardinality 2 does not exceed the threshold, level 3 is taken as the final sensitivity level of node D according to the principle of grading from high to low, and it is upgraded from general data (level 2) to important data (level 3). The update result is synchronized to the data security map in real time and triggers the automatic adjustment of relevant access control policies. The classification results of nodes E and F outside the domain do not need to be recalculated, and the historical judgment results are used throughout the process. Combining all the processing steps C.1 to C.3, the computational cost of this incremental update closed loop is strictly limited to four influence domain nodes, completely decoupled from the scale of other nodes in the lineage graph. Even in large-scale data governance scenarios with continuously expanding data assets, it maintains stable real-time update response capabilities, fully verifying the core design goal that the computational cost of the incremental update mechanism is strictly proportional to the size of the influence domain. The above values ​​are illustrative assumptions; in actual deployment, the conformal threshold and cardinality threshold should be calibrated accordingly based on the distribution characteristics of real business data.

[0161] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the above embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically classifying and grading sensitive data information, characterized in that, Applications include the classification and grading of enterprise data assets, including: Multimodal content extraction is performed on heterogeneous data sources, and the resulting text blocks, image regions, and table cells are encoded into a unified intermediate representation sequence; Perform a rule layer scan on the unified intermediate representation sequence to obtain a set of rule-hit entities, and separate the undetermined content into a set of unresolved fragments; The set of unresolved fragments is routed to a multimodal semantic recognition layer containing text, image, and table channels for determination, resulting in a set of semantically recognized entities. The set of entities that hit the rule and the set of entities that are semantically recognized are subjected to unified confidence calibration and sensitivity type normalization to obtain a standardized sensitive entity set. By combining the standardized set of sensitive entities with the metadata in the data source, a time-stamped lineage graph is constructed. After propagation and fusion through spatial and temporal channels, the sensitivity labeling results of each lineage node are obtained. The sensitivity labeling results are matched with the compliance template library to obtain the candidate sensitivity level distribution. Selective conformal risk control judgment is performed on the candidate sensitivity level distribution, and the node sensitivity level judgment result with a false negative rate not exceeding a preset upper limit is output.

2. The method according to claim 1, characterized in that, The process of routing the set of unresolved fragments to a multimodal semantic recognition layer containing text, image, and table channels for determination yields a set of semantically recognized entities, including: Text fragments from the set of pending fragments are input into the text channel for contextual semantic recognition; The image regions in the set of unresolved fragments are input into the image channel and combined with the layout features for layout perception and recognition. Input the table cells in the set of unresolved fragments into the table channel to perform column semantic inference based on column name semantics and column content distribution; The channel features output from the text channel, the image channel, and the table channel are fused in a shared coding space and then uniformly determined to obtain a set of semantically recognized entities.

3. The method according to claim 2, characterized in that, The process involves fusing the channel features output from the text channel, the image channel, and the table channel in a shared coding space and then uniformly determining them to obtain a semantic recognition entity set, including: The channel features output from the text channel, the image channel, and the table channel are mapped to a shared coding space to obtain a unified representation vector for each modality; The unified representation vectors of each modality are fused, and the fused representations are used to determine sensitive entities to obtain a semantic recognition entity set.

4. The method according to claim 1, characterized in that, The step of performing confidence-based unified calibration and sensitivity type normalization on the rule-hit entity set and the semantic recognition entity set to obtain a standardized sensitive entity set includes: According to the rules, the pattern matched by each entity in the entity set is checked for check bits, and a corresponding prior score is assigned. The original confidence scores output by each entity in the semantic recognition entity set during recognition are calibrated using a temperature scaling method. The prior scores and the calibrated original confidence levels are uniformly mapped to the same calibration probability space, and all entities are aligned to a unified sensitivity type system to obtain the standardized sensitive entity set.

5. The method according to claim 1, characterized in that, The process involves combining the standardized set of sensitive entities with metadata from the data source to construct a time-stamped lineage graph. This graph is then propagated and fused via spatial and temporal channels to obtain the sensitivity labeling results for each lineage node, including: By combining the standardized sensitive entity set with the metadata to extract time-stamped lineage relationships, using data assets as nodes and data flow and processing relationships as directed edges, and recording the occurrence time and processing type of flow events on each directed edge, the time-series lineage graph is constructed. The spatial representation vector is obtained by performing neighborhood message passing on the sensitive entity features of each node along the bloodline edge through the spatial channel; The time-series channel is used to perform time-series encoding on the incoming event sequence of each node to obtain a time-series representation vector; The spatial representation vector and the temporal representation vector are fused to obtain the spatiotemporal sensitivity representation vector of each bloodline node, and the spatiotemporal sensitivity representation vector is written into the memory state of the corresponding bloodline node. The spatiotemporal sensitivity representation vector is input into the sensitivity discrimination classification head to obtain the sensitivity labeling results of each bloodline node.

6. The method according to claim 5, characterized in that, The temporal encoding of the incoming event sequence of each node by the temporal channel to obtain a temporal representation vector includes: The sequence of incoming edge events is encoded according to the order in which they occur, capturing the causal pattern between the processing events to obtain the temporal representation vector. The sequential causal pattern is used to distinguish data flow scenarios with different processing times and different sensitivities. The data flow scenarios include data flow scenarios where data is desensitized before aggregation and data flow scenarios where data is aggregated before desensitization.

7. The method according to claim 1, characterized in that, The process of matching the sensitivity labeling results with the compliance template library to obtain a candidate sensitivity level distribution, performing selective conformal risk control judgment on the candidate sensitivity level distribution, and outputting a node sensitivity level judgment result with a false negative rate not exceeding a preset upper limit includes: The sensitivity labeling results are matched with the mapping rules from sensitivity types to level standards in the compliance template library to obtain the candidate sensitivity level distribution of each node; The selection function is used to filter out the first set of nodes whose confidence level reaches the preset selection threshold from each node; Calculate the conformal quantile threshold on the pre-held calibration dataset, and construct a grade prediction set for each node in the first node set based on the conformal quantile threshold. The grade prediction set ensures that the false negative rate does not exceed a preset upper limit without relying on data distribution assumptions. For nodes whose base number in the level prediction set does not exceed a preset base number threshold, the highest sensitivity level in the level prediction set is taken as the final sensitivity level of the nodes whose base number in the level prediction set does not exceed the preset base number threshold, and the node sensitivity level determination result is output.

8. The method according to claim 7, characterized in that, The method further includes: Nodes whose cardinality in the predicted level set exceeds the preset cardinality threshold, as well as nodes that were not selected by the selection function to enter the first node set, are routed to the manual review queue, where their final sensitivity level is determined by manual review.

9. The method according to claim 5, characterized in that, The method further includes: Listen for data change event streams; When a data change event occurs, the changed data is extracted and then subjected to the rule layer scanning, the multimodal semantic recognition layer judgment, and the confidence level unified calibration to obtain an incremental sensitive entity set. Based on the incremental sensitive entity set, the influence domain affected by the data change event in the temporal lineage graph is determined, the memory state of the nodes within the influence domain is locally updated, and the selective conformal risk control judgment is re-executed for the nodes within the influence domain to obtain the updated node sensitivity level judgment result.

10. The method according to claim 9, characterized in that, The step involves determining the influence domain of the data change event in the temporal lineage graph based on the incremental sensitive entity set, locally updating the memory state of nodes within the influence domain, and re-executing the selective conformal risk control judgment on the nodes within the influence domain to obtain an updated node sensitivity level judgment result, including: The influence domain is determined based on the nodes associated with the incremental sensitive entity set and their downstream nodes reachable along directed edges in the temporal lineage graph; The propagation calculation of the spatial channel and the temporal channel is re-executed only for nodes within the influence domain to update the memory state of nodes within the influence domain, while the memory state of nodes outside the influence domain remains unchanged. The selective conformal risk control determination is re-executed only for nodes within the influence domain to obtain an updated node sensitivity level determination result. The computational cost of incremental updates is proportional to the size of the influence domain.

Images