Multi-dimensional data sensitivity automatic grading method and system

Abstract

The application discloses a kind of multi-dimension data sensitivity automatic grading method and system, the method includes: extracting data attribute features from data to be graded;Using compliance knowledge enhanced large model to identify the data to be graded, obtain sensitive identification result;Data attribute features and sensitive content features are matched with the rule in compliance rule base, match to mandatory classification rule, then output sensitive level;Otherwise, output compliance constraint condition;Data attribute features, compliance constraint condition and sensitive identification result are input into grading decision model, obtain prediction result;If the highest probability of prediction is inconsistent with initial sensitive level, and the probability difference is less than the set threshold, trigger compliance knowledge enhanced large model to carry out secondary semantic analysis, output calibration suggestion;According to compliance constraint condition, prediction result and calibration suggestion, output final sensitive level and its confidence degree.The application scheme can realize the accurate grading of multi-dimension data sensitivity.

Description

Technical Field

[0001] This invention relates to the field of data security, specifically to a method and system for automatic classification of multi-dimensional data sensitivity. Background Technology

[0002] Data sensitivity classification is a fundamental step in data security governance, widely applied in scenarios such as data classification and grading, data sharing and control, and security auditing. Currently, mainstream technical methods mainly include rule-based keyword matching, regular expressions, and traditional machine learning classification models.

[0003] However, existing technologies generally have the following shortcomings in practical applications: (1) Poor generalization ability and high false negative rate in complex scenarios.

[0004] Rule-based methods rely on manually preset rules and manual feature engineering, which cannot identify sensitive content such as mask variations, homophones, and split expressions. Traditional machine learning models are limited by the quality of labeled data and the ability to express features, and have limited generalization ability in cross-domain and multi-type data scenarios, which can easily lead to missed detection of highly sensitive data.

[0005] (2) The hierarchical logic is rigid and the dimensions are singular.

[0006] Existing methods typically rely solely on the presence or absence of sensitive content for binary judgment, lacking the ability to integrate multi-dimensional information such as data attributes, compliance requirements, and risk levels for decision-making. This results in classification results that are detached from actual business scenarios and compliance needs, leading to insufficient usability.

[0007] (3) High on-the-ground maintenance costs and poor industry adaptability.

[0008] Both the rule base and the model need to be re-labeled and trained for different industries and scenarios. The labeling cost is high and the iteration cycle is long, making it difficult to achieve low-cost and rapid industry adaptation.

[0009] (4) Poor interpretability, unable to meet compliance audit requirements.

[0010] Existing methods are mostly "black box" outputs, unable to provide decision-making basis for the classification results, mapping of compliance clauses, and audit evidence, making it difficult to meet the compliance requirements of laws and regulations such as the Data Security Law for the traceability and auditability of the data classification and grading process.

[0011] In summary, existing technologies struggle to simultaneously achieve the generalizability of sensitive data identification, the accuracy of classification, the low cost of industry-specific adaptation, and the interpretability of the decision-making process. Summary of the Invention

[0012] This invention provides a method and system for automatic classification of multi-dimensional data sensitivity to solve one or more of the problems mentioned above in the prior art.

[0013] Therefore, the present invention provides the following technical solution: On one hand, the present invention provides a method for automatic classification of multi-dimensional data sensitivity, the method comprising: Acquire the data to be classified, and extract data attribute features from the data to be classified; The large model is enhanced by pre-trained compliance knowledge to perform sensitive semantic understanding and identification on the data to be classified, and a structured sensitive identification result is obtained. The structured sensitive identification result includes: sensitive content features, initial sensitivity level and confidence level. The data attribute features and the sensitive content features are matched with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, the compliance constraints are output. The data attribute features, compliance constraints, and structured sensitivity identification results are input into a pre-trained hierarchical decision model to obtain prediction results; the prediction results include each sensitivity level and its prediction probability. If the highest predicted sensitivity level is inconsistent with the initial sensitivity level, and the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold, then the compliance knowledge enhancement model is triggered to perform secondary semantic analysis on the data to be classified and output calibration suggestions; and based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

[0014] Optionally, the sensitive semantic understanding and recognition of the data to be classified includes: full sensitive content recognition, context-related sensitive recognition, and initial sensitivity level recognition.

[0015] Optionally, the method further includes: Obtain the large model of the base; The compliance knowledge-enhanced large model is obtained by training the base large model using an incremental fine-tuning method.

[0016] Optionally, the process of training the compliance knowledge-enhanced large model from the base large model using an incremental fine-tuning method includes: Construct a multi-industry compliance-sensitive dataset, in which each sample contains original text, sensitive semantic annotation labels, and compliance knowledge annotation labels; Based on standardized prompt word templates for sensitive identification and classification tasks, each sample in the dataset is converted into a structured training instance. Freeze a certain number of network parameters in the base model and inject LoRA matrices only in the self-attention layer of the base model. Determine a multi-task joint loss function, which includes: an autoregressive cross-entropy loss function, a sensitivity level classification cross-entropy loss function, and a contrastive learning loss function; With the goal of minimizing the multi-task joint loss function, only the LoRA matrix of the self-attention layer is updated, while keeping the frozen parameters in the base large model unchanged. The training is iterated until convergence, and the compliance knowledge augmentation large model is obtained.

[0017] Optionally, the compliance rule base has multiple levels; The step of matching the data attribute features and the sensitive content features with rules in the compliance rule base to output compliance constraints includes: The compliance rule base is converted into a set of executable logical expressions; The data attribute features and the sensitive content features are input into the set of logical expressions for hierarchical matching, and at least one compliance constraint is output. When a conflict occurs in the matching results of multiple rules, the constraint condition corresponding to the highest priority rule is selected as the compliance constraint condition.

[0018] Optionally, the method further includes: supervising the training of a lightweight gradient boosting tree model based on a multi-industry labeled hierarchical dataset to obtain the hierarchical decision model.

[0019] Optionally, the method further includes: Generate a standardized classification decision report, which includes one or more of the following information: basic data information, a list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching status, corresponding legal and regulatory provisions, classification decision-making process and final classification basis; and / or Sensitive content in the data to be classified is highlighted.

[0020] Optionally, the method further includes: performing end-to-end evidence storage on all operations, input data, model output, rule execution status, and classification results throughout the entire hierarchical process, and generating a unique audit traceability ID.

[0021] On the other hand, the present invention also provides a multi-dimensional data sensitivity automatic classification system, the system comprising: The feature extraction module is used to acquire the data to be classified and extract data attribute features from the data to be classified. The semantic understanding and recognition module is used to enhance the large model with pre-trained compliance knowledge to perform sensitive semantic understanding and recognition on the data to be classified, and obtain a structured sensitive recognition result, which includes: sensitive content features, initial sensitivity level and confidence level; The rule matching module is used to match the data attribute features and the sensitive content features with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, the compliance constraints are output. The prediction module is used to input the data attribute features, the compliance constraints, and the structured sensitivity identification results into a pre-trained hierarchical decision model to obtain prediction results; the prediction results include each sensitivity level and its prediction probability; The inspection module is used to check whether the highest predicted sensitivity level is inconsistent with the initial sensitivity level, and whether the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold. If so, the compliance knowledge enhancement model is triggered to perform secondary semantic analysis on the data to be classified and output calibration suggestions. Based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

[0022] Optionally, the system further includes: The compliance knowledge enhancement model construction module is used to train the base model to obtain the compliance knowledge enhancement model through incremental fine-tuning. The compliance knowledge-enhanced big model construction module includes: The dataset generation unit is used to construct a multi-industry compliance-sensitive dataset. Each sample in the dataset contains the original text, sensitive semantic annotation labels, and compliance knowledge annotation labels. The instance conversion unit is used to convert each sample in the dataset into a structured training instance based on a standardized prompt word template for a sensitive identification and classification task. A matrix injection unit is used to freeze a certain number of network parameters in the base model and inject LoRA matrices only in the self-attention layer of the base model. The loss function determination unit is used to determine the multi-task joint loss function, which includes: autoregressive cross-entropy loss function, sensitivity level classification cross-entropy loss function, and contrastive learning loss function; The training unit is used to update the parameters of the LoRA matrix of the self-attention layer only, with the goal of minimizing the multi-task joint loss function, while keeping the frozen parameters in the base large model unchanged, and iteratively training until convergence to obtain the compliance knowledge augmentation large model.

[0023] Optionally, the system further includes: The report generation module is used to generate standardized hierarchical decision-making reports, which include one or more of the following information: basic data information, a list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching, corresponding legal and regulatory provisions, the hierarchical decision-making process and the basis for final classification; and / or The annotation module is used to highlight sensitive content in the data to be classified.

[0024] Optionally, the system further includes: an evidence storage module, used to store evidence for all operations, input data, model outputs, rule execution status, and classification results throughout the entire hierarchical process, and generate a unique audit traceability ID.

[0025] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when run by a processor, executes the steps of the multi-dimensional data sensitivity automatic classification method.

[0026] Compared with existing technologies, the multi-dimensional data sensitivity automatic classification method and system provided by the present invention have one or more of the following beneficial effects: 1. Significantly improved accuracy and generalization ability: Based on the deep semantic understanding capability of large models, this invention breaks through the capability bottleneck of traditional solutions, can cover the identification of sensitive data of all types and all scenarios, solves the core pain point of high false negative rate in complex scenarios of traditional solutions, and improves the recall rate of highly sensitive data by more than 20 percentage points.

[0027] 2. Significantly enhanced usability and compliance of the grading results: The multi-dimensional dynamic grading mechanism of this invention replaces the fixed mapping grading of traditional solutions. The grading results can be dynamically adjusted in combination with scenarios, permissions, and compliance requirements, which can directly support the security management of the entire data lifecycle. At the same time, the full-link compliance design fully meets the requirements of laws and regulations such as the Data Security Law and can be directly used for regulatory compliance audits.

[0028] 3. Significantly reduced implementation and maintenance costs: Based on the zero-shot / few-shot learning capability of large models, this invention shortens the industry scenario adaptation cycle from 15-30 days in traditional solutions to less than 3 days. At the same time, the iteration cost of LoRA incremental fine-tuning is only 1% of that of traditional full-scale fine-tuning solutions, which significantly reduces the implementation and maintenance costs for enterprises.

[0029] 4. High versatility and scalability: This invention can be adapted to scenarios in all industries such as government affairs, finance, healthcare, and energy. It supports hierarchical classification of all types of data, including structured, semi-structured, and unstructured data. There is no need to develop independent solutions for different data types. At the same time, the modular architecture supports flexible functional expansion and can quickly adapt to new compliance requirements and business scenarios.

[0030] 5. Solves the compliance problem of AI model black box: The full-chain interpretability system of this invention breaks the black box attribute of traditional AI models. The classification results have complete decision-making basis and traceability, which solves the industry pain point that enterprise AI classification schemes cannot pass compliance audits. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of a multi-dimensional data sensitivity automatic classification method provided in an embodiment of the present invention; Figure 2 This is a flowchart of a compliance knowledge enhancement model generated through incremental fine-tuning in an embodiment of the present invention; Figure 3 This is a schematic diagram of a multi-dimensional data sensitivity automatic classification system provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a compliance knowledge enhancement big model construction module provided in an embodiment of the present invention; Figure 5 This is another structural schematic diagram of the multi-dimensional data sensitivity automatic classification system provided in the embodiments of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] To address the problems of existing data sensitivity classification methods, this invention provides a multi-dimensional automatic data sensitivity classification method and system. It uses a fine-tuned large language model enhanced with compliance knowledge as the core of sensitive semantic understanding, constructing a full-link architecture of "large model semantic recognition + multi-dimensional feature fusion + compliance rule engine constraints + dynamic classification decision-making + interpretability generation + human-machine collaborative iteration." This achieves end-to-end automatic sensitivity classification for all types of data, including structured, semi-structured, and unstructured data, while considering generalization, accuracy, compliance, and interpretability.

[0036] like Figure 1 The diagram shown is a flowchart of an automatic multi-dimensional data sensitivity classification method provided in an embodiment of the present invention, which includes the following steps: In step 101, the data to be classified is obtained, and data attribute features are extracted from the data to be classified.

[0037] The solution of this invention supports unified access and standardized processing of multi-source heterogeneous data. The multi-source heterogeneous data may include, but is not limited to: structured data (such as database tables, Excel, CSV format data), unstructured data (such as Word documents, PDF documents, TXT documents, emails, chat logs, audio and video transcribed text, etc.), and semi-structured data (such as JSON, XML, HTML format files, etc.). It supports multiple access methods such as local files, databases, and APIs (Application Programming Interfaces).

[0038] Extract text from files of different formats, preserving structural information such as paragraphs, headings, tables, and hierarchical levels; remove garbled characters, invalid characters, and duplicate content, and complete text normalization. For long texts exceeding the large model context window, a sliding window segmentation strategy based on semantic paragraphs can be adopted, such as a window size of 512-4096 tokens and an inter-block overlap rate of 15%, to ensure the integrity of the contextual semantics and avoid missing sensitive content across blocks.

[0039] The data attribute features can be obtained through metadata extraction, such as including but not limited to: data source, generation time, creation subject, access permissions, transmission range, file type, data volume, life cycle stage, etc., as the basic attribute features for subsequent classification.

[0040] In step 102, the pre-trained compliance knowledge-enhanced big model is used to perform sensitive semantic understanding and recognition on the data to be classified, so as to obtain a structured sensitive recognition result; the structured sensitive recognition result includes: sensitive content features, initial sensitivity level and confidence level.

[0041] The aforementioned compliance knowledge enhancement model can be obtained by fine-tuning existing large language models, enabling it to break through the upper limit of the semantic understanding ability of traditional large language models and achieve accurate identification of sensitive content in all scenarios.

[0042] It should be noted that, in order to reduce the cost of fine-tuning while retaining the general semantic capabilities of the large language model, the LoRA (Low-Rank Adaptation) incremental fine-tuning technique can be used to freeze 99% of the parameters of the selected large language model and train only the LoRA matrix of the attention layer.

[0043] like Figure 2 The diagram shown is a flowchart of an embodiment of the present invention for incrementally fine-tuning an existing large language model to generate a compliance knowledge-enhanced large model, including the following steps: Step 201: Obtain the large model of the base.

[0044] A pedestal model (also known as a base model or foundation model) refers to a large-scale deep learning model that has been pre-trained on large-scale, wide-domain data and has strong general-purpose capabilities, but has not yet been fine-tuned for a specific task.

[0045] For example, in a non-limiting embodiment, a commercially available open-source language model with hundreds of billions of parameters, such as Llama 3 70B or Qwen 2 72B, can be selected, which has powerful deep semantic understanding, contextual reasoning, and zero-shot learning capabilities.

[0046] Step 202: Construct a multi-industry compliance sensitive dataset, wherein each sample in the dataset contains the original text, sensitive semantic annotation labels, and compliance knowledge annotation labels.

[0047] The multi-industry compliance sensitive dataset can cover sensitive data labeled samples from industries such as government affairs, finance, and healthcare. It also incorporates compliance knowledge from laws and regulations such as the Data Security Law and the Personal Information Protection Law to enhance compliance knowledge and enable the model to have both sensitive semantic recognition and compliance cognition capabilities.

[0048] It should be noted that the embodiments of the present invention do not limit the specific method of injecting compliance knowledge. For example, in a non-limiting embodiment, a "triple structure" training sample can be constructed: fine-grained annotation of the original text is performed to identify the sensitive data types and their sensitivity levels contained therein, and each annotated segment is associated with corresponding legal and regulatory clauses, and key constituent elements of the clauses are extracted. The above results are then encapsulated into the following structure as training data for fine-tuning the model: Input: Raw text; The compliance reasoning chain (part of the model output) includes "identified sensitive elements → triggered legal clauses → logical deduction process → judgment and basis for compliance".

[0049] Alignment constraint: forces the model to generate a complete legal reasoning clause before outputting a judgment conclusion.

[0050] Step 203: Based on the standardized prompt word template for the sensitive identification and classification task, each sample in the dataset is converted into a structured training instance.

[0051] The standardized prompt word template may include, but is not limited to, information such as: task definition, sensitive category definition, compliance basis, and output format constraints. This standardized prompt word template can guide large models to accurately complete recognition tasks.

[0052] For example, a standardized prompt word template in a non-limiting embodiment is as follows: Task: You are a professional data security compliance expert who needs to perform full-scale sensitive information identification and sensitivity analysis on input text.

[0053] Sensitivity level definition (based on the Data Security Law): 1. Top Secret: State secrets, core trade secrets, and classified information whose disclosure could lead to major security incidents; 2. Confidential Level: Important trade secrets, core sensitive personal information, undisclosed major business / government decisions; 3. Confidential Level: General trade secrets, ordinary personal information, and internal, non-public work information; 4. Internal level: Non-sensitive work information that can only be circulated within the organization; 5. Public Level: Contains no sensitive information and can be publicly released.

[0054] Compliance basis: Data Security Law of the People's Republic of China and Personal Information Protection Law of the People's Republic of China.

[0055] Output requirements: 1. Identify all sensitive content in the text, and label the content, start and end positions, sensitivity category, and potential harm from disclosure; 2. Analyze the contextual relationships across paragraphs to determine whether combining them increases the sensitivity level; 3. Output the initial sensitivity level prediction and confidence level of the text; 4. Strictly adhere to the JSON format for output; no extra content is allowed.

[0056] Input text: {text}.

[0057] Output format: {"sensitive_content_list": [{"content": "xxx", "start_pos":x, "end_pos": x, "sensitive_category": "xxx", "leak_harm": "xxx"}], "context_sensitive_analysis": "xxx", "initial_sensitive_level": "xxx", "confidence": "xxx"}.

[0058] Step 204: Freeze a certain number of network parameters in the base model and inject the LoRA matrix only in the self-attention layer of the base model.

[0059] Step 205: Determine the multi-task joint loss function, which includes: autoregressive cross-entropy loss function, sensitivity level classification cross-entropy loss function, and contrastive learning loss function.

[0060] To address the shortcomings of traditional fine-tuning methods that only focus on output fit, this invention designs a multi-task joint loss function for sensitive hierarchical tasks, thereby improving the model's sensitive semantic discrimination and hierarchical accuracy. The formula is as follows: ; The meanings of each parameter are as follows: This is the autoregressive cross-entropy loss function, used to ensure the accuracy of the model output format and consistency with semantic understanding; The cross-entropy loss function for sensitivity level classification is used to optimize the model's prediction accuracy for sensitivity levels; The contrastive learning loss function, used to improve the model's semantic discriminative power for texts with different sensitivity levels, is formulated as follows: ; In the formula, This is the semantic embedding vector of the input text. For positive example text embedding vectors of the same sensitivity level, For negative example text embedding vectors with different sensitivity levels, The cosine similarity function is used. Temperature coefficient (e.g.) The value ranges from 0.05 to 0.2. N The number of positive samples. M The number of negative samples; α ,β , γ The value is a weighting coefficient. For example, in a non-limiting embodiment, it can be 0.4, 0.35, or 0.25, and the specific value can be dynamically adjusted according to the scenario.

[0061] Step 206: With the goal of minimizing the multi-task joint loss function, only the LoRA matrix of the self-attention layer is updated, while keeping the frozen parameters in the base large model unchanged. The training is iterated until convergence to obtain the compliance knowledge augmentation large model.

[0062] The compliance knowledge enhancement model provided in this embodiment of the invention needs to complete three core tasks on the input text: Full-scale sensitive content identification: It can identify not only fixed-format sensitive information, but also masked variations, homophonic puns, and split expressions of sensitive content, as well as trade secrets and classified information without fixed formats; Contextual Sensitivity Analysis: Semantically associate text across paragraphs and blocks to identify scenarios where "individual content is not sensitive, but when combined, it forms highly sensitive content," thus avoiding missed detection of combined sensitive information; Initial sensitivity level prediction: Based on the recognition results and compliance requirements, output the initial sensitivity level and confidence level of the overall text.

[0063] It should be noted that in step 102, when using compliance knowledge to enhance the large model for sensitive semantic understanding and recognition of the graded data, a general prompt word template can be used to output structured sensitive recognition results. For example, a non-limiting example of a general prompt word template includes the following aspects: [Character Setting] You are a sensitive information analysis expert with advanced compliance reasoning skills. Your core competencies include: identifying variant / implicit sensitive expressions, cross-fragment semantic correlation analysis, and data classification based on compliance knowledge.

[0064]

Core Task 1: Full-scale Sensitive Content Identification

[0065] Output format: List of identified sensitive segments and sensitive types.

[0066] [Core Task Two: Full-Scale Sensitive Content Identification] Perform semantic association analysis across paragraphs, sentences, and blocks of text, focusing on identifying: 1. Individual content is not sensitive, but when combined, it forms highly sensitive information; 2. Scattered signage information in different locations is pieced together to form a profile of an individual or entity; 3. Sensitive leaks caused by semantic inference and background knowledge completion.

[0067] Output format: Association analysis conclusions, combined sensitive paths, and sensitivity level after combination.

[0068]

Core Task 3: Full-scale Sensitive Content Identification

[0069] 2. Confidence level (percentage).

[0070] 3. The main basis for rating determination (referencing compliance knowledge dimensions: legal requirements / business hazards / privacy risks).

[0071] [Overall Output Format] === Full Sensitive Content Recognition === === Context-sensitive analysis === === Initial Sensitivity Level Prediction === It should be noted that the above general prompt template is only an example. In actual implementation, it can be set according to the actual application requirements.

[0072] Continue to refer to Figure 1 In step 103, the data attribute features and the sensitive content features are matched with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, compliance constraints are output.

[0073] By matching with rules in the compliance rule base, hard compliance requirements can be incorporated, the uncertainty of large model outputs can be compensated for, and the compliance of the classification results can be guaranteed.

[0074] In one non-limiting embodiment, the compliance rule base can be configured as multi-level, for example, constructing a three-level compliance rule base, namely: National level: Sensitive data classification requirements and mandatory classification rules stipulated in laws and regulations such as the Data Security Law and the Law on Guarding State Secrets; Industry-specific data grading standards and regulatory requirements for sectors such as finance, healthcare, and government. Enterprise level: Internal data classification and grading standards, access control rules, and data flow rules.

[0075] To facilitate matching, the compliance rule base can be converted into an executable logical expression. The data attribute features (such as scenario and data flow) and the sensitive content features (such as ID card number) are substituted into the logical expression to determine whether a violation occurs and output compliance constraints.

[0076] For example: "ID number found and not desensitized → Desensitization must be performed".

[0077] When multiple rules have different requirements for the same behavior (such as corporate rules allowing it but national laws prohibiting it), the rules are automatically adjudicated according to their hierarchy to ensure that the output never violates the highest priority rule.

[0078] If a mandatory classification rule is matched, that is, if the data to be classified conforms to the mandatory classification rule, the sensitivity level can be directly locked. For example, data containing state secret information can be directly classified as top secret without entering the subsequent model reasoning stage.

[0079] In some embodiments, the present invention can further support visual editing and version management of the compliance rule base, allowing for rapid iteration when compliance requirements are updated without the need to retrain the model.

[0080] In step 104, the data attribute features and sensitive content features are input into the pre-trained hierarchical decision model to obtain the prediction results; the prediction results include each sensitivity level and its prediction probability.

[0081] Specifically, the Lightweight Gradient Boosting Machine (LightGBM) model can be trained under supervision using a multi-industry labeled hierarchical dataset to obtain the hierarchical decision model. Its input is a multi-dimensional feature vector, and its output is the predicted probability for five levels: Top Secret, Confidential, Secret, Internal, and Public.

[0082] During training, a weighted loss function can be used to apply higher penalty weights to classification errors with high sensitivity levels in order to reduce the false negative rate of highly sensitive data.

[0083] In step 105, determine whether the highest predicted sensitivity level is inconsistent with the initial sensitivity level and whether the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold; if so, proceed to step 106; otherwise, proceed to step 108.

[0084] In step 106, the compliance knowledge enhancement model is triggered to perform secondary semantic analysis on the data to be classified and output calibration suggestions; and based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

[0085] It should be noted that although the semantic analysis of the data to be classified in step 106 and step 102 uses the same large model, it differs from the initial semantic analysis in step 102. The secondary semantic analysis in step 106 mainly allows the large model to perform inference and calibration separately for data with relatively low prediction probabilities. Therefore, targeted prompts are used, which at least include: fragments of the original text (i.e., the data to be classified), the sensitivity identification results obtained from the first large model identification, the highest sensitivity level and probability predicted by the classification decision model, and the probability difference between the two levels. The task given by the prompts has less context and less misleading information, enabling the large model to make more accurate judgments.

[0086] For example, in a non-limiting embodiment, targeted prompts include: ## Task - Data Sensitivity Level Calibration and In-Depth Analysis: You are a senior data security compliance expert with over 10 years of experience, and you now need to perform final calibration on a piece of data with a classification dispute. ## Known Information: 1. Original text segment: {disputed_text_segment}; 2. All sensitive content identified by the first large model: {first_sensitive_content_list}; 3. Initial prediction level of the first large model: {first_model_level}, confidence level: {first_confidence}; 4. The hierarchical decision model predicts the highest probability level: {ml_model_level}, with a probability of {ml_probability}; 5. The probability difference between the two levels: {probability_diff} (less than 0.2, indicating a sample with ambiguous boundaries).

[0087] ## Official rating standards (strictly adhered to): 1. Top Secret: Contains state secrets and core trade secrets; disclosure of such information would cause particularly serious damage to national security, major public interests, or the core interests of the enterprise. 2. Confidential: Contains important trade secrets, core sensitive personal information (biometrics, medical and health information, financial accounts), and undisclosed major business / government decisions; disclosure of such information would cause serious damage. 3. Confidential: Contains general trade secrets, ordinary personal information, and internal non-public work information; disclosure of such information would cause general harm. 4. Internal level: Non-sensitive work information that can only be circulated within the organization, and its disclosure to external parties will not cause substantial harm; 5. Public Level: Contains no sensitive information and can be publicly released.

[0088] ## Compliance basis: "Data Security Law of the People's Republic of China" and "Personal Information Protection Law of the People's Republic of China"; ## Your task: 1. Based on all the known information above, The key focus is on comparing and analyzing the applicable boundaries of the {first_model_level} and {ml_model_level} models. ; 2. List the core reasons for supporting the {first_model_level} level, one by one; 3. List the core reasons for supporting the {ml_model_level} level, one by one; 4. Based on the severity of the harm, scope of impact, and compliance requirements of the leakage of sensitive content, provide a final calibration level recommendation; 5. Clearly identify any potential judgment biases in the initial analysis or hierarchical decision-making model; 6. Give your calibration confidence level (0-100%).

[0089] ## Output Requirements: - Strictly adhere to the JSON format below for output; no additional explanations or content are permitted. - The reasons must be specific and verifiable, and cannot be vague. - Only one of the above 5 calibration levels can be selected; - The confidence level must be an integer: { "support_first_level_reasons": [ "reason 1", "reason 2" ], "support_ml_level_reasons": [ "reason 1", "reason 2" ], "final_calibration_level": "xxx", "judgment_bias_analysis": "xxx", "calibration_confidence": xxx}.

[0090] In step 107, the sensitivity level with the highest predicted probability is taken as the sensitivity level of the data to be classified.

[0091] To meet the needs of different applications, the sensitivity level results output in this embodiment of the invention can support multiple formats, such as JSON, XML, CSV, PDF compliance reports, and other formats, which can better meet the needs of different scenarios.

[0092] In some embodiments, a standardized API interface can be further provided: a RESTful (representational state delivery) standard API interface is provided to support the connection with third-party systems such as DLP (Data Loss Prevention System), data security governance platform, document management system, OA (Office Automation) system, and cloud storage system, so as to synchronize the classification results to the downstream system and realize security policies such as access control, encryption, de-identification, and transmission control based on the sensitivity level; In some embodiments, a visual management interface may also be provided: providing a visual display of the overall data classification, distribution of sensitive data, compliance risk statistics, and iterative optimization status to support management personnel in decision-making.

[0093] In some embodiments, the method of the present invention may further include any one or more of the following steps: (1) Generation of interpretable reports for hierarchical decision-making: Based on the semantic generation capability of the large model and combined with the full-link hierarchical data, a standardized hierarchical decision-making report is automatically generated. The content may include, but is not limited to, any one or more of the following: basic data information, list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching status, corresponding legal and regulatory clauses, hierarchical decision-making process and final classification basis. (2) Full-link audit evidence storage: All operations, input data, model output, rule execution status and classification results of the entire hierarchical process are stored in the full-link evidence storage, generating a unique audit traceability ID to support subsequent compliance audits and problem backtracking. All stored evidence data is tamper-proof. (3) Visualization of sensitive content: Highlight sensitive content in the original text (i.e., data to be classified) to support quick location, verification and desensitization.

[0094] In some embodiments, human-machine collaborative iterative optimization can also be supported, the specific process of which is as follows: 1. Manual verification entry: Provides a visual interface for verifying classification results, supporting security experts to correct classification results and supplementary annotations for sensitive identification results; 2. Feedback Data Processing: The manually corrected feedback data is cleaned, deduplicated, and converted in format, and divided into two categories: rule optimization data and model optimization data; 3. Rule and Model Iteration: For rule defects, rule optimization suggestions are automatically generated and submitted to the administrator for review and update; for model identification errors, an incremental learning approach is adopted to incrementally fine-tune the LoRA weights of large models and incrementally train hierarchical decision models without the need for full retraining. 4. Performance Evaluation and Deployment: After each iteration, test set validation is automatically performed, and core metrics such as accuracy, recall, and F1 score are output. Only model versions that reach the preset threshold can be deployed and released to ensure continuous improvement in the tiered performance.

[0095] Accordingly, the multi-dimensional data sensitivity automatic classification system provided in the embodiments of the present invention, such as Figure 3 The diagram shown is a structural schematic of the system.

[0096] In this embodiment, the multi-dimensional data sensitivity automatic classification system 300 includes: Feature extraction module 301 is used to acquire data to be classified and extract data attribute features from the data to be classified. The semantic understanding and recognition module 302 is used to enhance the large model with pre-trained compliance knowledge to perform sensitive semantic understanding and recognition on the data to be classified, and obtain a structured sensitive recognition result, which includes: sensitive content features, initial sensitivity level and confidence level; The rule matching module 303 is used to match the data attribute features and the sensitive content features with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, compliance constraints are output. Prediction module 304 is used to input the data attribute features, the compliance constraints, and the structured sensitivity identification results into a pre-trained hierarchical decision model to obtain prediction results; the prediction results include each sensitivity level and its prediction probability; The inspection module 305 is used to check whether the highest predicted sensitivity level is inconsistent with the initial sensitivity level, and whether the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold. If so, the compliance knowledge enhancement model is triggered to perform secondary semantic analysis and output calibration suggestions. Based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

[0097] The aforementioned compliance knowledge enhancement model can be trained from the base model using an incremental fine-tuning method by the corresponding compliance knowledge enhancement model construction module. The compliance knowledge enhancement model construction module can be part of the system of this invention, or it can be independent of the system of this invention; there is no limitation on this.

[0098] like Figure 4The diagram shown is a structural schematic of a compliance knowledge enhancement big model construction module in an embodiment of the present invention.

[0099] This compliance knowledge-enhanced large model building module 400 includes: Dataset generation unit 401 is used to construct a multi-industry compliance-sensitive dataset, wherein each sample in the dataset contains original text, sensitive semantic annotation labels, and compliance knowledge annotation labels; The instance conversion unit 402 is used to convert each sample in the dataset into a structured training instance based on a standardized prompt word template for a sensitive identification and classification task. The matrix injection unit 403 is used to freeze a certain number of network parameters in the base model and inject LoRA matrices only in the self-attention layer of the base model. The loss function determination unit 404 is used to determine the multi-task joint loss function, which includes: autoregressive cross-entropy loss function, sensitivity level classification cross-entropy loss function, and contrastive learning loss function; Training unit 405 is used to update the parameters of the LoRA matrix of the self-attention layer only, with the goal of minimizing the multi-task joint loss function, while keeping the frozen parameters in the base large model unchanged, and iteratively training until convergence to obtain the compliance knowledge augmentation large model.

[0100] like Figure 5 The diagram shown is another structural schematic of the multi-dimensional data sensitivity automatic classification system provided in an embodiment of the present invention.

[0101] and Figure 3 The difference in the illustrated embodiment is that, Figure 5 In the illustrated embodiment, the multi-dimensional data sensitivity automatic classification system 300 further includes: a report generation module 501 and / or a labeling module 502. Wherein: The report generation module 501 is used to generate a standardized hierarchical decision report. This standardized decision report may include, but is not limited to, any one or more of the following information: basic data information, a list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching, corresponding legal and regulatory clauses, the hierarchical decision-making process, and the basis for final classification; and / or The annotation module 502 is used to highlight sensitive content in the data to be classified.

[0102] In some embodiments, the multi-dimensional data sensitivity automatic classification system 300 may further include: an evidence storage module (not shown), used to perform end-to-end evidence storage of all operations, input data, model output, rule execution status, and classification results throughout the classification process, and generate a unique audit traceability ID.

[0103] The specific implementation methods of the above modules can be referred to the description in the previous embodiments of the present invention, and will not be repeated here.

[0104] The multi-dimensional data sensitivity automatic classification method and system provided in this invention constructs a sensitive semantic recognition system based on a large language model enhanced with compliance knowledge. It designs a multi-task joint loss function for sensitivity classification tasks, breaking through the capability limits of traditional rules and small models. This achieves accurate identification of sensitive content in complex scenarios such as variations, cryptic language, and contextual associations, significantly reducing the false negative rate of highly sensitive data. Furthermore, it proposes a multi-dimensional dynamic classification decision mechanism, integrating four dimensions: sensitive content, data attributes, compliance attributes, and risk attributes. This constructs a fusion decision architecture of "compliance rule engine + lightweight classification decision model + large model semantic calibration," solving the problems of rigid classification logic and single dimensions in traditional solutions. The classification results fully align with business and compliance requirements. Moreover, a dual-track classification architecture of large model + rule engine is designed. This leverages the generalization ability of the large model to cover long-tail complex scenarios while using the hard constraints of the compliance rule engine to address the uncertainty of the large model's output, simultaneously considering the generalization, accuracy, and compliance of the classification scheme.

[0105] Furthermore, the present invention also constructs a hierarchical end-to-end interpretability and audit traceability system, which automatically outputs complete hierarchical decision-making basis and compliance reports based on the large model generation capability, realizing the end-to-end tamper-proof evidence storage of the hierarchical process, and fully meeting the compliance audit requirements of data security laws and regulations.

[0106] Furthermore, the present invention also designs a human-machine collaborative incremental iterative optimization closed loop. Based on LoRA incremental fine-tuning technology, it realizes low-cost and rapid model iteration, while supporting rapid adaptation to industry scenarios with zero or few samples, which greatly reduces the implementation and maintenance costs of the solution and improves the adaptability across all industries.

[0107] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0108] To further verify the effectiveness of the present invention, the present invention underwent thorough testing.

[0109] 1. Experimental Environment and Setup Hardware environment: CPU is Intel Xeon 8375C, GPU is NVIDIA A100 80G 4. 512GB RAM; Software environment: Ubuntu 22.04 operating system, CUDA 12.1, PyTorch 2.2.0, large model base is Qwen 2 72B, LoRA fine-tuning parameters r=8, α=16; Test dataset: A multi-industry mixed test set was constructed, with a total of 10,000 data entries, including 2,000 structured data entries, 5,000 unstructured documents, and 3,000 chat / email text entries, covering 5 sensitivity levels. Among them, variant sensitive content and context-sensitive content accounted for 40% of the samples. The annotation results were cross-verified by 3 senior data security experts as the gold standard.

[0110] 2. Experimental results show that: (1) Accuracy and recall: The accuracy of the proposed solution is 96.7%, and the recall rate of highly sensitive data is 98.2%, which greatly reduces the risk of missing highly sensitive data. The core of this solution is the deep semantic understanding capability of the large model.

[0111] (2) Adaptability to complex scenarios: The present invention has a recognition rate of over 94% for variants and context-sensitive content, which is far higher than the existing technology, effectively solving the core pain point of poor generalization ability of traditional solutions.

[0112] (3) Performance and cost: The solution of this invention takes 45ms to process a single data item, which fully meets the needs of enterprise-level batch data processing. The performance can be linearly scaled through distributed deployment. The industry adaptation cycle is shortened to less than 3 days, and the implementation and maintenance costs are greatly reduced.

[0113] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 or Figure 2 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0114] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0115] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and apparatus of the present invention, and are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically classifying multi-dimensional data sensitivity, characterized in that, The method includes: Acquire the data to be classified, and extract data attribute features from the data to be classified; The large model is enhanced by pre-trained compliance knowledge to perform sensitive semantic understanding and identification on the data to be classified, and a structured sensitive identification result is obtained. The structured sensitive identification result includes: sensitive content features, initial sensitivity level and confidence level. The data attribute features and the sensitive content features are matched with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, the compliance constraints are output. The data attribute features, compliance constraints, and structured sensitivity identification results are input into a pre-trained hierarchical decision model to obtain prediction results; the prediction results include each sensitivity level and its prediction probability. If the highest predicted sensitivity level is inconsistent with the initial sensitivity level, and the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold, then the compliance knowledge enhancement model is triggered to perform secondary semantic analysis on the data to be classified and output calibration suggestions; and based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

2. The multi-dimensional data sensitivity automatic classification method according to claim 1, characterized in that, The process of understanding and recognizing sensitive semantics in the data to be classified includes: full-volume sensitive content recognition, context-related sensitive recognition, and initial sensitivity level recognition.

3. The multi-dimensional data sensitivity automatic classification method according to claim 2, characterized in that, The method further includes: Obtain the large model of the base; The compliance knowledge-enhanced large model is obtained by training the base large model using an incremental fine-tuning method.

4. The multi-dimensional data sensitivity automatic classification method according to claim 3, characterized in that, The compliance knowledge enhancement model obtained by training the base model using an incremental fine-tuning method includes: Construct a multi-industry compliance-sensitive dataset, in which each sample contains original text, sensitive semantic annotation labels, and compliance knowledge annotation labels; Based on standardized prompt word templates for sensitive identification and classification tasks, each sample in the dataset is converted into a structured training instance. Freeze a certain number of network parameters in the base model and inject LoRA matrices only in the self-attention layer of the base model. Determine a multi-task joint loss function, which includes: an autoregressive cross-entropy loss function, a sensitivity level classification cross-entropy loss function, and a contrastive learning loss function; With the goal of minimizing the multi-task joint loss function, only the LoRA matrix of the self-attention layer is updated, while keeping the frozen parameters in the base large model unchanged. The training is iterated until convergence, and the compliance knowledge augmentation large model is obtained.

5. The multi-dimensional data sensitivity automatic classification method according to claim 1, characterized in that, The compliance rule base has multiple levels; The step of matching the data attribute features and the sensitive content features with rules in the compliance rule base to output compliance constraints includes: The compliance rule base is converted into a set of executable logical expressions; The data attribute features and the sensitive content features are input into the set of logical expressions for hierarchical matching, and at least one compliance constraint is output. When a conflict occurs in the matching results of multiple rules, the constraint condition corresponding to the highest priority rule is selected as the compliance constraint condition.

6. The multi-dimensional data sensitivity automatic classification method according to claim 1, characterized in that, The method further includes: The lightweight gradient boosting tree model is trained under supervision using a multi-industry labeled hierarchical dataset to obtain the hierarchical decision model.

7. The multi-dimensional data sensitivity automatic classification method according to any one of claims 1 to 6, characterized in that, The method further includes: Generate a standardized classification decision report, which includes one or more of the following information: basic data information, a list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching status, corresponding legal and regulatory provisions, classification decision-making process and final classification basis; and / or Sensitive content in the data to be classified is highlighted.

8. The multi-dimensional data sensitivity automatic classification method according to claim 7, characterized in that, The method further includes: All operations, input data, model outputs, rule execution status, and classification results throughout the entire grading process are stored and verified at the end of the chain, generating a unique audit traceability ID.

9. A multi-dimensional data sensitivity automatic classification system, characterized in that, The system includes: The feature extraction module is used to acquire the data to be classified and extract data attribute features from the data to be classified. The semantic understanding and recognition module is used to enhance the large model with pre-trained compliance knowledge to perform sensitive semantic understanding and recognition on the data to be classified, and obtain a structured sensitive recognition result, which includes: sensitive content features, initial sensitivity level and confidence level; The rule matching module is used to match the data attribute features and the sensitive content features with the rules in the compliance rule base. If a mandatory classification rule is matched, the sensitivity level of the data to be classified is output; otherwise, the compliance constraints are output. The prediction module is used to input the data attribute features, the compliance constraints, and the structured sensitivity identification results into a pre-trained hierarchical decision model to obtain prediction results; the prediction results include each sensitivity level and its prediction probability; The inspection module is used to check whether the highest predicted sensitivity level is inconsistent with the initial sensitivity level, and whether the difference between the highest predicted probability and the second highest predicted probability is less than a set threshold. If so, the compliance knowledge enhancement model is triggered to perform secondary semantic analysis on the data to be classified and output calibration suggestions. Based on the compliance constraints, the prediction results and the calibration suggestions, the final sensitivity level and its confidence level are output.

10. The multi-dimensional data sensitivity automatic classification system according to claim 9, characterized in that, The system also includes: The compliance knowledge enhancement model construction module is used to train the base model to obtain the compliance knowledge enhancement model through incremental fine-tuning. The compliance knowledge-enhanced big model construction module includes: The dataset generation unit is used to construct a multi-industry compliance-sensitive dataset. Each sample in the dataset contains the original text, sensitive semantic annotation labels, and compliance knowledge annotation labels. The instance conversion unit is used to convert each sample in the dataset into a structured training instance based on a standardized prompt word template for a sensitive identification and classification task. A matrix injection unit is used to freeze a certain number of network parameters in the base model and inject LoRA matrices only in the self-attention layer of the base model. The loss function determination unit is used to determine the multi-task joint loss function, which includes: autoregressive cross-entropy loss function, sensitivity level classification cross-entropy loss function, and contrastive learning loss function; The training unit is used to update the parameters of the LoRA matrix of the self-attention layer only, with the goal of minimizing the multi-task joint loss function, while keeping the frozen parameters in the base large model unchanged, and iteratively training until convergence to obtain the compliance knowledge augmentation large model.

11. The multi-dimensional data sensitivity automatic classification system according to claim 9 or 10, characterized in that, The system also includes: The report generation module is used to generate standardized hierarchical decision-making reports, which include one or more of the following information: basic data information, a list and location of sensitive content, multi-dimensional feature weight analysis, compliance rule matching, corresponding legal and regulatory provisions, the hierarchical decision-making process and the basis for final classification; and / or The annotation module is used to highlight sensitive content in the data to be classified.

12. The multi-dimensional data sensitivity automatic classification system according to claim 11, characterized in that, The system also includes: The evidence storage module is used to store evidence for all operations, input data, model outputs, rule execution status, and classification results throughout the entire hierarchical process, generating a unique audit traceability ID.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the automatic classification method for multi-dimensional data sensitivity according to any one of claims 1 to 8.

Images