Privacy information identification, classification and desensitization protection method based on large model

Abstract

The invention relates to the technical field of data processing and data security, in particular to a large model-based privacy information identification, classification and desensitization protection method, which comprises the following steps of: constructing a privacy template; on the basis of the privacy template, carrying out scientific classification and grading on the privacy information to form a classification and grading rule base; systematically storing the privacy template and the classification and grading rule base, and constructing a privacy knowledge base; selecting a suitable pre-training large model according to a business scene, constructing a large-scale training set based on the privacy knowledge base, performing fine tuning on the pre-training large model, and performing deployment; performing privacy identification on the to-be-identified data by using the deployed large model, classifying and grading the identified privacy information, and endowing a corresponding sensitive level; and according to the identified privacy information and the sensitivity level thereof, corresponding privacy desensitization strategies are generated for different types of privacy entities and are applied, so that efficient, refined and intelligent whole-process privacy protection is realized.

Description

Technical Field

[0001] The embodiments of this application relate to the fields of data processing and data security technology, and in particular to a method for privacy information identification, classification and grading, and de-identification protection based on a large model. Background Technology

[0002] With the rapid development of the digital economy, industries such as healthcare, finance, government affairs, e-commerce, education, and the internet are increasingly reliant on massive amounts of data for analysis, decision-making, and business innovation. Especially in scenarios such as electronic medical records, insurance claims documents, bank statements, credit materials, user behavior logs, customer service conversations, and knowledge document management systems, various types of data exhibit characteristics such as diverse sources, complex formats, highly irregular semantics, and rapid growth. This data often contains a large amount of privacy information directly or indirectly related to individuals. This privacy information includes direct identifiers such as ID card numbers, contact information, and addresses; quasi-identifiers such as age, occupation, region, and behavioral patterns; and highly sensitive attribute information such as medical diagnoses, medical history, income, asset status, family relationships, and psychological state.

[0003] However, existing data security systems largely focus on traditional security measures such as transmission encryption, access control, database permissions, and log auditing. They lack a deep understanding of the privacy implications, semantic structure, and sensitivity of data itself. This results in enterprises still heavily relying on manual review, expert experience, or coarse-grained security strategies for privacy protection during data collection, storage, analysis, and sharing. With data volumes growing exponentially, this reliance on manual methods is not only inefficient and costly, but also prone to overlooking privacy elements, making it difficult to meet the mandatory requirements of laws such as the Personal Information Protection Law, the Data Security Law, and the General Data Protection Regulation (GDPR) for "automatically identifying categories of personal information, determining sensitivity levels, and taking differentiated protection measures."

[0004] Meanwhile, the diversification of data types further increases the difficulty of privacy protection. While structured data (such as tables and database records) has relatively fixed fields, semantic ambiguity in privacy fields may still exist. Unstructured data (such as PDF documents, medical reports, communication records, customer service conversations, and mixed text and image content) is even more difficult to process automatically through rules or traditional feature engineering. Especially in scenarios such as medical diagnosis, financial behavior, and public opinion commentary, privacy risks are often hidden in the contextual semantics. For example, semantic privacy such as "depressive tendencies," "recent large debts," and "suspected tumor" cannot be analyzed through simple string retrieval or traditional NER (Named Entity Recognition) and must rely on models with deep semantic understanding capabilities.

[0005] Current industry-standard privacy protection technologies, including regular expressions, dictionary matching, shallow feature-based named entity recognition models, fixed-rule generalization strategies, and simple masking methods, are all inadequate for handling the complexity of modern data environments. They can only identify a limited number of entity types, cannot automatically determine the sensitivity level of different information in different business scenarios, and are even less capable of recognizing semantic or inferential privacy information such as medical diagnoses, financial status, and emotional states. Furthermore, existing masking methods typically employ deletion, masking, or fixed replacements, which easily disrupt the original semantic structure. This results in data unavailability in modeling, analysis, and sharing scenarios, and can even impact critical business decisions.

[0006] With the widespread application of LLMs (Large Language Models) in natural language understanding, semantic inference, and generative tasks, building a new technological system that combines deep semantic understanding capabilities to achieve automatic identification, classification, grading, and controllable de-identification protection of privacy information has become an urgent need in the industry. This new technological system not only needs to be adaptable to various industry scenarios such as healthcare, finance, and government affairs, but also must possess scalability, high recognition accuracy, strong semantic preservation capabilities, and anti-repeating recognition capabilities to simultaneously meet security, usability, and compliance requirements.

[0007] To achieve the identification, classification, grading, and anonymization of personal privacy in data, research teams have attempted to address the issue from different angles. However, in general, these solutions typically only cover one link in the privacy protection chain and cannot build a complete, integrated privacy protection system. In summary, current solutions suffer from the following four problems.

[0008] First, the ability to identify privacy information is limited. Existing privacy identification technologies mainly rely on named entity recognition models based on regular expressions, keyword matching, or shallow features. These models can only cover a portion of direct identifiers and cannot effectively identify semantic, inferential, or context-sensitive information (such as medical diagnoses, financial status, behavioral patterns, and psychological states). This makes it difficult to comprehensively grasp data risk points and fails to meet the needs of enterprises for full-scale privacy identification in actual business operations, resulting in vulnerabilities in privacy compliance and security management.

[0009] Second, the classification and grading system lacks intelligence and context awareness. Existing technologies largely rely on fixed rules or manually set labels in the classification and grading process, lacking the ability to automatically determine the sensitivity of information in different scenarios. The system cannot dynamically adjust sensitivity levels, nor can it comprehensively assess risks by associating multiple fields; for example, the same field may have different sensitivity levels in different business environments. Traditional methods are coarse-grained in classification, lack portability, and struggle to support differentiated de-identification strategies and fine-grained access control.

[0010] Third, desensitization methods are crude and struggle to balance semantic preservation with risk control. Currently used desensitization methods (such as deletion, masking, and fixed-rule generalization) easily disrupt the original semantic structure, reducing data usability in analysis, modeling, and sharing scenarios. While some generative replacement methods can preserve some semantics, they lack risk perception capabilities based on classification and grading, failing to generate desensitization results with matching strengths for different sensitivity levels, and are also ineffective at preventing linkability attacks.

[0011] Fourth, existing methods mostly rely on manual labor and high-cost annotation, making them difficult to adapt to large-scale data scenarios. When processing unstructured, multi-source, and multi-industry data, existing technologies rely on manual review, rule maintenance, or expert experience, resulting in low processing efficiency, large fluctuations in accuracy, and an inability to cope with the needs of massive data and rapidly iterating business operations. They also cannot continuously generate traceable, de-identified records, increasing enterprise compliance risks and operating costs. Summary of the Invention

[0012] In view of this, embodiments of this application propose a privacy information identification, classification, and de-identification protection method based on a large model. This method incorporates the semantic understanding and generative modeling capabilities of a large language model, enabling efficient and accurate automated privacy identification, intelligent classification, strategic de-identification protection, and traceable record generation for multi-source, multi-format, and multi-industry data without extensive manual intervention. Furthermore, through comprehensive analysis of sensitivity levels and contextual semantics, it effectively blocks potential re-identification risks, balancing data security, usability, and compliance. This provides a secure and reliable basic data environment for data governance, storage, analysis, sharing, and model training, thereby achieving efficient, refined, and intelligent end-to-end privacy protection.

[0013] To achieve the above objectives, embodiments of this application provide a method for privacy information identification, classification, grading, and desensitization protection based on a large model. The method includes: conducting research and summarizing privacy information from different business systems, data types, and all potentially involved individuals or organizations; flexibly defining privacy entities, concepts, and their attributes to construct a privacy template; scientifically classifying and grading the privacy information based on the privacy template, and forming a classification and grading rule base based on the classification and grading results; systematically storing the privacy template and the classification and grading rule base to construct a privacy knowledge base that can be used for large model identification, classification, and desensitization; selecting a suitable pre-trained large model according to the business scenario; constructing a large-scale training set for fine-tuning based on the privacy knowledge base; fine-tuning the pre-trained large model based on the large-scale training set; deploying the fine-tuned large model; using the deployed large model to perform privacy identification on the data to be identified; classifying and grading the identified privacy information and assigning corresponding sensitivity levels; generating corresponding privacy desensitization strategies for different types of privacy entities based on the identified privacy information and their sensitivity levels; and applying the generated privacy desensitization strategies.

[0014] To achieve the above objectives, embodiments of this application also propose an electronic device, including a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement a privacy information identification, classification and grading and desensitization protection method based on a large model as described above.

[0015] To achieve the above objectives, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the aforementioned method for privacy information identification, classification, grading, and desensitization protection based on a large model.

[0016] The privacy information identification, classification, grading, and desensitization protection method proposed in this application has at least the following advantages compared with traditional methods.

[0017] First, this application dynamically constructs privacy templates associated with specific scenarios, enabling flexible definition of privacy entities, attributes, and their relationships based on different business and industry scenarios. It also embeds laws, regulations, and industry standards, avoiding inconsistencies and omissions caused by subjective human judgment. At the same time, the dynamic template update mechanism ensures real-time matching between the template and business scenarios and regulatory requirements, providing a reliable foundation for subsequent privacy identification and desensitization.

[0018] Second, this application scientifically and reasonably classifies privacy attributes and classifies them by combining direct and indirect identifiers. It also introduces scenario-aware sensitivity level mapping to achieve refined privacy classification and grading for different business scenarios, thus solving the problem that traditional fixed rules are difficult to adapt to changing scenarios.

[0019] Third, this application constructs a privacy knowledge base through multi-source data collection and virtual synthesis technology, uniformly processes structured and unstructured data, and embeds compliance rules to form standardized and computable representations of privacy entities, attributes, and sensitivity levels. The privacy knowledge base supports rapid querying, expansion, and updating, providing high-quality training data and rule support for large model fine-tuning and privacy identification, and significantly improving the accuracy and coverage of privacy identification.

[0020] Fourth, this application proposes a method for fine-tuning large models and improving privacy recognition capabilities. By combining semantic perception and attribute-subject linkage analysis, it can achieve comprehensive recognition and understanding of complex privacy relationships. Through automated recognition processes, it significantly reduces the cost of manual recognition and improves the system's privacy protection capabilities across multiple scenarios and data types.

[0021] Fifth, in the privacy desensitization stage, this application comprehensively considers semantic relevance and attribute-subject linkage to intelligently desensitize sensitive data, and optimizes the desensitization strategy in a closed loop through downstream task performance, semantic integrity and privacy leakage quantitative indicators to achieve efficient privacy protection while ensuring business availability.

[0022] Sixth, this application continuously optimizes the rules for associating privacy entities, attributes, and scenarios by using a dynamic update and system optimization mechanism for the privacy knowledge base, combined with semantic awareness, link management, and feedback on desensitization effects. This ensures the accuracy and completeness of the privacy knowledge base in the face of business changes and regulatory updates, while forming a closed-loop management system that achieves a dynamic balance between privacy protection, business availability, and compliance, significantly improving privacy management efficiency and system adaptability.

[0023] In summary, the overall solution of this application realizes a complete closed-loop process from privacy template construction, classification and grading, knowledge base construction, large model fine-tuning, privacy identification, intelligent desensitization to dynamic optimization. It solves the problems of imprecise privacy identification, single desensitization strategy, lagging knowledge base updates and poor cross-scenario applicability in traditional methods, and has wide applicability and scalability. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a method for privacy information identification, classification, grading, and desensitization protection based on a large model, provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the construction of a privacy template provided in one embodiment of this application; Figure 3 This is a schematic diagram illustrating the classification of privacy attributes provided in one embodiment of this application; Figure 4 This is a schematic diagram illustrating the construction of a privacy knowledge base provided in one embodiment of this application; Figure 5 This is a schematic diagram illustrating the enhancement of large model fine-tuning and privacy recognition capabilities provided in one embodiment of this application; Figure 6 This is a schematic diagram illustrating privacy identification, classification, and grading provided in one embodiment of this application; Figure 7 This is a schematic diagram illustrating the generation and application of a privacy de-identification strategy provided in one embodiment of this application; Figure 8 This is a schematic diagram illustrating the privacy knowledge base update and system dynamic optimization provided in one embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0027] In order to identify, classify, classify, and de-identify personal privacy in data, some technologies or tools at home and abroad have attempted to solve related problems from different angles. However, in general, these solutions usually only cover one link in the privacy protection chain and cannot build a complete integrated privacy protection system.

[0028] The closest existing implementation solutions are summarized below: 1. A privacy identification scheme based on rule and dictionary matching.

[0029] These solutions primarily rely on regular expressions, keyword libraries, or sensitive field dictionaries to automatically identify structured privacy information such as ID card numbers, bank card numbers, email addresses, and IP addresses. Their advantages lie in low implementation cost, clear rules, and the ability to quickly detect data with fixed formats. However, they heavily depend on manually maintaining the rule base, making them unable to identify semantic privacy (such as disease descriptions, financial difficulties, and behavioral patterns) or determine sensitivity levels based on context. The rule system lacks scalability and industry adaptability, making it difficult to handle unstructured text and massive amounts of diverse data, thus failing to meet the demand for accurate privacy identification in large-scale scenarios.

[0030] 2. Named entity recognition scheme based on deep learning.

[0031] In recent years, some organizations have chosen to use models such as BERT, RoBERTa, and BiLSTM-CRF to identify privacy-sensitive entities, which can identify basic privacy fields such as names, addresses, and contact information. However, these models are limited by predefined labeling systems, resulting in a limited range of recognition types and difficulty in covering sensitive entities in fields such as healthcare, finance, and government. Furthermore, they can only output entity categories, unable to determine sensitivity levels, and struggle to handle changes in the sensitivity of semantic privacy or quasi-identifiers across different contexts. In addition, these models lack industry adaptability, requiring a large number of labeled samples for deployment in new scenarios, leading to high costs and unstable performance.

[0032] 3. Privacy classification and grading scheme based on existing standards.

[0033] Currently, some privacy classification and grading standards exist in the industry, such as the national standard GB / T 35273 and industry standards (such as the "Basic Specifications for Electronic Medical Records" in the medical field and the "Guidelines for Financial Data Classification"). Some enterprises and institutions are attempting to develop automated classification and grading tools based on these standards. These solutions typically rely on fixed rules, such as field names, business tags, or simple characteristics, to classify data into risk categories, aiming to achieve preliminary data classification management. However, these solutions still have limitations in practical applications: their classification granularity is relatively coarse, mostly remaining at broad categories such as personal information and sensitive personal information, lacking detailed processing of specific business scenarios and semantic content; their contextual understanding capabilities are limited, making it difficult to determine the sensitivity level of the same field in different environments; and they lack modeling for privacy inference risks formed by combinations of multiple fields. Furthermore, these systems rely on manual maintenance of standards and rules, resulting in delayed updates and limited cross-industry migration capabilities.

[0034] 4. Industry-specific desensitization systems and rule-based protection tools.

[0035] Some industries (such as healthcare and finance) have developed data anonymization tools tailored to specific business needs, such as healthcare de-identification systems and financial account masking tools. These systems combine entity recognition and fixed-rule replacement strategies to mask or generalize some sensitive information. However, they typically cannot flexibly select anonymization strategies based on the level of sensitivity and are difficult to migrate across industries. For unstructured text, highly semantically sensitive content, or cross-field related privacy risks, these systems have limited processing capabilities; furthermore, fixed strategies may corrupt data semantics, reducing their usability in statistical analysis, risk assessment, and model training.

[0036] 5. Privacy handling schemes for generative replacement or language model rewriting.

[0037] With the development of large-scale language models, some exploratory solutions attempt to replace privacy fields with generative models to achieve desensitization while maintaining semantics. These methods have a certain semantic understanding capability and generate natural content, but they still have limitations: they lack the ability to perceive risks based on privacy levels, the generated content may be inconsistent or distorted, they do not systematically model linkability, and they lack a unified cross-industry classification system, making it difficult to form a complete privacy protection chain.

[0038] With the acceleration of digitalization, sectors such as healthcare, finance, government affairs, education, and internet companies generate and process massive amounts of data assets containing personal information and sensitive data every day. This data is characterized by its wide range of sources, complex structure, diverse semantics, and frequent updates, and is widely found in business logs, electronic medical records, financial statements, contracts, communication records, customer service conversations, spreadsheets, and semi-structured and unstructured data.

[0039] However, existing privacy protection technologies are insufficient to meet the actual needs of automated identification, accurate classification and grading, and controllable desensitization of personal information in large-scale data processing scenarios, resulting in the following prominent problems in data processing processes in real applications.

[0040] First, existing privacy identification technologies rely on traditional regular expression matching, keyword dictionaries, or named entity recognition models based on shallow features. These technologies can only cover a limited range of privacy types and can only identify some direct identifiers, such as names, ID numbers, or contact information. They cannot effectively identify semantic, inferential, or contextual sensitive information such as medical diagnoses, financial status, behavioral patterns, psychological states, and lifestyle habits. As a result, enterprises cannot fully grasp data risk points in actual business operations and find it difficult to meet privacy compliance requirements.

[0041] Secondly, existing technologies lack a unified and intelligent system for classifying and grading the sensitivity of personal information. Different systems use inconsistent methods for categorizing privacy information, making it difficult to support the differentiated security requirements of various industries such as education, healthcare, finance, and government. Furthermore, traditional methods cannot combine contextual semantics to identify the degree of privacy sensitivity. For example, traditional methods cannot distinguish the sensitivity level difference between "region" and "diagnosed with lung cancer," nor can they identify the different sensitivity levels of the same field in different scenarios. For instance, age may be considered low-sensitivity data in some statistical scenarios, but when combined with medical history, family members, or other attributes, it may become a quasi-identifier, thereby increasing the risk of re-identification. The lack of refined classification and grading capabilities makes it difficult for enterprises to implement differentiated storage, access control, permission management, and compliance assessments.

[0042] Furthermore, in the data anonymization stage, current mainstream methods such as deletion, masking, and fixed-rule generalization often damage the original semantic structure, leading to decreased data readability, distorted business analysis, or severe degradation of model training performance. While some generative replacement methods can preserve some semantics, they often fail to generate anonymized results with matching strengths based on different sensitivity levels due to the lack of risk awareness and grading mechanisms. In addition, existing methods generally lack the ability to "break linkability," failing to prevent attackers from using background knowledge or external data for re-identification, thus posing significant security risks in scenarios such as medical image analysis, risk control, and user behavior analysis.

[0043] Meanwhile, with the continuous strengthening of privacy regulations (such as the Personal Information Protection Law, the Data Security Law, GDPR, and HIPAA), enterprises need to continuously and automatically identify, classify, and grade massive amounts of data, and generate traceable de-identification records. Existing technologies generally rely on manual review, costly annotation, or industry expert experience, which cannot cope with the ever-increasing data scale, nor can they effectively handle the privacy risks brought about by the complex semantics in unstructured data, resulting in high labor costs, large fluctuations in accuracy, and unsustainable processing.

[0044] Therefore, there is an urgent need for an intelligent privacy protection technology system for complex data environments. This system should be able to automatically perform the entire process of privacy information identification, classification and grading, and de-identification protection for multi-source, multi-format, and multi-industry data without requiring extensive human intervention. This would enable comprehensive control and dynamic risk assessment of personal information and provide secure and compliant basic data for subsequent data governance, storage, analysis, sharing, and model training.

[0045] In view of this, one embodiment of this application proposes a method for privacy information identification, classification and grading, and de-identification protection based on a large model. The implementation details of the method for privacy information identification, classification and grading, and de-identification protection based on a large model proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0046] This embodiment proposes a privacy information identification, classification, grading, and de-identification protection scheme based on a large model. It requires the prior construction of a multi-source data privacy knowledge base, collecting existing privacy classification standards, industry norms, and de-identification strategies, extracting privacy types, sensitivity levels, and de-identification rules, and generating privacy processing templates to build a privacy information processing template library. In practical applications, the system first matches suitable templates to different data sets based on data sources, formats, and business scenarios, achieving preliminary classification, grading, and sensitivity assessment of multi-source, multi-format data, thereby establishing a unified structured description system for privacy risks. After template matching, the system uses the semantic understanding capabilities of the large model to perform full-scale privacy identification on the data, including direct identifiers, quasi-identifiers, and semantic and inferred sensitive information. Subsequently, based on contextual semantics and business scenarios, it automatically determines the sensitivity level of various types of privacy information, achieving intelligent, fine-grained classification and grading, and comprehensively analyzing privacy inference risks arising from multi-field combinations or contextual associations. In the privacy anonymization process, the system combines classification and grading results with a strategy-based generative anonymization method to apply matching strength anonymization operations to data of different sensitivity levels, while preserving the original semantic structure of the data to ensure its usability in analysis, modeling, and sharing scenarios. The system can automatically generate traceable anonymization processing records and has the ability to interrupt linkability attacks, thereby reducing the risk of re-identification. Furthermore, the system supports expansion to multiple industries and business scenarios and has adaptive capabilities. By collecting and analyzing data processing flows, business rules, and log information, the system can continuously optimize templates and classification strategies, achieving fully automated processing. The entire process of privacy information identification, classification, grading, and anonymization protection is completed automatically by computer, requiring minimal manual intervention, thus significantly improving processing efficiency, reducing operating costs, and ensuring data compliance and security.

[0047] The purpose of this embodiment is to provide a privacy information identification, classification, grading, and de-identification protection method based on a large model, aiming to solve problems such as incomplete privacy identification, unintelligent classification and grading, crude de-identification methods, and difficulty in preventing re-identification risks in existing technologies. By introducing the semantic understanding and generative modeling capabilities of a large model, this embodiment can achieve automated privacy identification, intelligent classification and grading, strategic de-identification protection, and traceable record generation for multi-source, multi-format, and multi-industry data without extensive manual intervention. Simultaneously, through comprehensive analysis of sensitivity levels and contextual semantics, this embodiment effectively blocks potential re-identification risks, balancing data security, usability, and compliance. It provides a secure and reliable basic data environment for data governance, storage, analysis, sharing, and model training, achieving efficient, refined, and intelligent end-to-end privacy protection.

[0048] Before introducing the specific process of the privacy information identification, classification and grading and desensitization protection method based on a large model proposed in this embodiment, the meaning of the technical terms involved in this embodiment will be explained first.

[0049] Large models (large language models) are machine learning models trained on large-scale data that have the ability to understand and process textual information, and are used for privacy information identification and processing.

[0050] Privacy information refers to identifiable information relating to an individual or organization, including identity, financial, health, location, and behavioral information, which is data that needs to be protected.

[0051] Privacy templates are used to structure and describe privacy information and processing rules, including field types, sensitivity levels, and processing requirements.

[0052] Privacy classification and grading (sensitivity level) categorizes privacy information according to type, sensitivity, and risk level to guide identification, processing, and de-identification operations.

[0053] A privacy knowledge base is a database that stores privacy templates, classification and grading rules, and historical processing records to support automated identification and de-identification.

[0054] Fine-tuning involves retraining the pre-trained large model using domain-specific data to improve its privacy detection and classification capabilities.

[0055] Privacy identification involves automatically identifying potential privacy information from raw data, including both structured and unstructured data, using models or algorithms.

[0056] Privacy desensitization involves processing identified privacy information to make it unidentifiable during use or sharing. This includes methods such as desensitization, encryption, and obfuscation.

[0057] The specific process of the privacy information identification, classification, grading, and de-identification protection method based on a large model proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Conduct research and summarize privacy information for different business systems, data types, and all potentially involved individuals or organizations, and flexibly define privacy entities, concepts, and their attributes to construct a privacy template.

[0058] In practice, before formally building a privacy knowledge base, it is necessary to first conduct research and summarize the privacy information of different business systems, data types, and all possible individuals or organizations involved, and flexibly define privacy entities, concepts and their attributes in order to build a privacy template.

[0059] In one example, the process of building a privacy template can be as follows: Figure 2 As shown.

[0060] First, we conducted research and compilation on different business systems, data types, and all potentially involved personal or organizational privacy information, including identity information, financial information, health information, location information, and behavioral information. We also referenced relevant industry laws, regulations, and standards, such as the Personal Information Protection Law, Medical Data Protection Standards, and Financial Data Security Guidelines, to ensure that the constructed privacy template complies with compliance requirements. Next, based on the survey and summary results, and under the constraints of laws, regulations and standards, a privacy template will be defined, including the categories of privacy entities, field attributes, sensitivity levels, possible sources of risk and handling requirements, to provide a standardized basis for subsequent automated identification and classification.

[0061] Then, by combining domain expert knowledge, we supplement the privacy entities that cannot be automatically obtained from existing data, ensuring that the privacy template covers all privacy information that may be involved in the business process and meets the compliance requirements of different industries.

[0062] Finally, the defined privacy templates are validated and optimized, including deduplication, attribute checking, and normalization, to ensure that the privacy template structure is complete and consistent, and can be flexibly extended to adapt to different business scenarios and industry standards.

[0063] Step 102: Based on the privacy template, scientifically classify and grade the privacy information, and form a classification and grading rule base based on the classification and grading results.

[0064] In practical implementation, after the privacy template is constructed, it is necessary to scientifically classify and grade the privacy information based on the privacy template, and form a classification and grading rule base based on the classification and grading results, so as to facilitate subsequent automated identification, risk assessment and de-identification processing.

[0065] In one example, such as Figure 3 As shown, privacy information is classified according to privacy entities and their attributes, and divided into seven major categories and thirty-six subcategories, as detailed below.

[0066] Biometric features: fingerprints, facial recognition.

[0067] Identification documents include: gender, facial features, ID number, date of birth, passport number, name, nationality, and address.

[0068] Medical and health category: medications, diagnosis results, hospital information, examination details, doctor's name, date of visit, and medical images.

[0069] Financial accounts: spending records, bank card numbers, invoice records.

[0070] Location tracking data includes: flight information, delivery addresses, activity tracks, landmarks, geographical locations, and road signs.

[0071] Property and identity information: license plate number, vehicle brand and model, personal property.

[0072] Social attributes include: company information, hobbies, social relationships, religious beliefs, occupation, phone number, and email address.

[0073] In addition to classification, privacy information also needs to be categorized into direct identifiers and indirect identifiers based on its sensitivity and identification risks. Direct identifiers can identify an individual's privacy information independently in most scenarios, such as ID card numbers, passport numbers, names, mobile phone numbers, bank card numbers, and medical images. The sensitivity level of direct identifiers is typically defined as high-risk. Indirect identifiers require the combination of other information to identify an individual's privacy information, such as addresses, dates of birth, activity patterns, and interests. The sensitivity level of indirect identifiers is typically defined as medium-risk or low-risk.

[0074] Sensitivity levels need to be adjusted based on scenario association. By constructing privacy templates, each type of privacy information is mapped to specific business and industry scenarios. The sensitivity level classification is dynamically adjusted based on scenario sensitivity, legal requirements, industry standards, and the complexity of data combinations. For example, in a medical scenario, date of birth might be classified as high-risk, while in a social media platform scenario, hobbies and interests might be classified as medium-risk.

[0075] After classification and grading are completed, a classification and grading rule base can be formed based on the results. This base includes classification labels, sensitivity levels, applicable scenarios, processing requirements, and compliance basis for privacy information, providing a standardized reference for subsequent large-scale model recognition, automated classification, and de-identification strategy generation.

[0076] After establishing the classification and grading rule base, it is also necessary to verify the classification and grading rule base by combining actual business data and domain expert review to ensure that it covers all kinds of privacy information in the platform. Based on the verification results, iterative optimization should be carried out to solve problems such as ambiguous classification, sensitivity level deviation or mismatch of business scenarios, so as to ensure that the privacy classification and grading is both scientific and reasonable and meets the requirements of the scenario and compliance requirements.

[0077] Step 103: Systematically store the privacy templates and classification and grading rule base to build a privacy knowledge base that can be used for large model recognition, classification and de-identification.

[0078] In practical implementation, after completing the construction of privacy templates and the definition of classification and grading, the privacy templates and classification and grading rule base can be systematically stored to build a privacy knowledge base that can be used for large model recognition, classification and de-identification.

[0079] In one example, the process of building a privacy knowledge base can be as follows: Figure 4 As shown.

[0080] The first step involves collecting and virtually synthesizing privacy information. This includes gathering data from business systems, databases, log files, documents, third-party data, and industry standards. Simultaneously, virtual synthesis technology is used to generate privacy samples with insufficient coverage or sensitivity. The collected and generated data then undergoes cleaning and structuring to create structured privacy entities and attribute information. This cleaning and structuring process includes deduplication, error correction, standardization of field names, and uniform unit formats. For unstructured data, such as text, images, and PDF files, information extraction is performed to create structured privacy entities and attribute information, laying the foundation for knowledge base construction.

[0081] The next step is to perform hierarchical information mapping, apply the classification and hierarchical rule base to all privacy entities, and combine different business scenarios to associate the sensitivity levels of direct and indirect identifiers with actual application scenarios to form scenario-aware privacy information records, so that the privacy knowledge base can accurately reflect the privacy risk level in different scenarios.

[0082] Next, we will design the structure of the privacy knowledge base. The privacy knowledge base uses triples or plurals to represent privacy entities, attributes, their levels, and contextual information, such as (Entity: Name, Attribute: ID Number, Level: High Risk, Context: Medical System) and (Entity: Activity Tracking, Attribute: Address, Level: Medium Risk, Context: E-commerce Platform). The privacy knowledge base structure supports rapid querying, expansion, and updating, while also meeting compliance auditing and tracking requirements.

[0083] Next, compliance rules and standards are embedded, including industry laws and regulations such as the Personal Information Protection Law, medical privacy regulations, and financial data compliance requirements, as well as internal security policies, into a knowledge base for subsequent automated privacy processing and risk assessment. Clear processing requirements are defined for each privacy entity record, including whether storage is permitted, whether de-identification is required, and the priority of de-identification strategies.

[0084] In addition, the privacy knowledge base needs to be verified and optimized. Verification should be carried out using real business data and virtual synthetic data to ensure the integrity, accuracy and scenario adaptability of privacy entities. Combined with expert review and model recognition feedback, the privacy knowledge base should be continuously updated and optimized to ensure coverage of emerging privacy attributes and changing scenarios.

[0085] Finally, there is the interface design of the privacy knowledge base, which provides standardized interfaces for large models to call the privacy knowledge base, supports privacy information query, sensitivity level acquisition, and scenario-adaptive rule calling. The privacy knowledge base also supports the automated process of large model fine-tuning and privacy identification, forming a closed-loop privacy management system.

[0086] Step 104: Select a suitable pre-trained large model based on the business scenario, construct a large-scale training set for fine-tuning based on the privacy knowledge base, fine-tune the pre-trained large model based on the large-scale training set, and deploy the fine-tuned large model.

[0087] In practice, after establishing a privacy knowledge base, a suitable pre-trained large model can be selected based on the business scenario. A large-scale training set for fine-tuning is constructed based on the privacy knowledge base. The pre-trained large model is then fine-tuned based on the large-scale training set, and the fine-tuned large model is deployed.

[0088] In one example, the process of fine-tuning large models and improving privacy awareness can be as follows: Figure 5 As shown.

[0089] The first step is to select and initially adapt a pre-trained large-scale model. Based on the business scenario, a suitable pre-trained large-scale model is selected, and its basic capabilities are evaluated to ensure that its ability to understand text, images, or structured data meets the basic requirements for privacy recognition. Initial adaptation of the pre-trained large-scale model is then performed, including adjusting the input format, output labeling system, and embedding privacy attributes and their sensitivity levels (seven major categories and thirty-six subcategories) from a privacy knowledge base, enabling the pre-trained large-scale model to identify and distinguish different types of privacy information.

[0090] Next, fine-tuning training data preparation is performed. Data in the privacy knowledge base undergoes preprocessing, sampling, and augmentation to generate a large-scale training set for fine-tuning. This large-scale training set includes both real-world business data and virtual synthetic data to cover privacy attributes and sensitivity levels across different scenarios. Data preprocessing includes deduplication, standardization, structured mapping, and sensitivity level labeling to ensure that the fine-tuning process effectively improves the model's accuracy in identifying privacy information.

[0091] Next, large-scale model fine-tuning and evaluation are performed. Using supervised or semi-supervised learning, the pre-trained large model is fine-tuned based on a large training set to ensure it can accurately identify privacy entities, attributes, and sensitivity levels in different business scenarios. During fine-tuning, a scenario-aware strategy is employed, mapping the sensitivity levels of different scenarios into the training objective. After fine-tuning, an independent test set is used for evaluation to verify the accuracy, recall, and F1 score of the fine-tuned large model in identifying different types of privacy information. The fine-tuned large model is then adjusted and optimized based on expert review results.

[0092] Next, continuous feedback and incremental optimization will be implemented. A continuous learning mechanism will be established to feed back new privacy information, misidentification cases, and scenario changes identified during the operation of the business system into the fine-tuning process, enabling incremental training and model updates. Through closed-loop feedback, the model's ability to identify complex and diverse privacy information will be continuously improved, ensuring the model's adaptability and robustness in practical applications.

[0093] Finally, model interface and deployment are implemented. Standardized interfaces are provided to seamlessly integrate the finely tuned large model with the privacy knowledge base and business systems, supporting real-time identification, classification, and subsequent anonymization of privacy information. The interface design balances performance, scalability, and security, ensuring efficient model operation in production environments while meeting compliance requirements and audit trail needs.

[0094] Step 105: Using the deployed large model, perform privacy identification on the data to be identified, classify and grade the identified privacy information, and assign corresponding sensitivity levels.

[0095] In practice, after the large model is deployed, it can be used to perform privacy identification on the data to be identified, classify and classify the identified privacy information, and assign corresponding sensitivity levels.

[0096] In one example, a large deployed model is used for privacy identification and classification processes, such as... Figure 6 As shown.

[0097] First, leveraging a large-scale deployed model and combining it with domain expert knowledge, the system automatically scans and analyzes business systems, databases, log files, documents, and unstructured data to identify privacy entities and their attribute information. During this process, scenario templates and seven major categories and thirty-six subcategories of privacy attributes from a knowledge base are used to ensure that different types of privacy information can be accurately located and labeled.

[0098] The identified privacy information will then be processed according to a classification and grading rule base. Direct and indirect identifiers will be assigned corresponding sensitivity levels, and these levels will be dynamically adjusted based on different business scenarios and industry standards. For example, in a medical setting, an ID card number might be considered high-risk, while on an e-commerce platform, some behavioral patterns might only be considered medium-risk. The classification and grading results will be recorded in a privacy knowledge base, forming a scenario-aware privacy information labeling system.

[0099] Finally, the identification and classification results are verified both automatically and manually to ensure the accuracy and completeness of the labeling of privacy entities, attributes, and sensitivity levels. By comparing the results with historical data, virtual synthetic data, and expert review results, the identification model and classification rules are optimized to form a continuously updated privacy information identification system. Simultaneously, the identification results are fed back into the large model fine-tuning process to achieve closed-loop optimization, improving the model's adaptability and robustness in different scenarios.

[0100] Step 106: Based on the identified privacy information and its sensitivity level, generate corresponding privacy desensitization strategies for different types of privacy entities, and apply the generated privacy desensitization strategies.

[0101] In practical implementation, after identifying the privacy information and its sensitivity level, it is necessary to generate corresponding privacy desensitization strategies for different types of privacy entities based on the identified privacy information and its sensitivity level, and then apply the generated privacy desensitization strategies.

[0102] In one example, the process of generating and applying privacy de-identification strategies can be as follows: Figure 7 As shown.

[0103] First, based on the identified privacy information and its sensitivity level, corresponding privacy desensitization strategies are generated for different types of privacy entities. Strategy types include data masking, hashing, generalization, random perturbation, and encryption. Strict desensitization methods are adopted for high-risk and highly sensitive direct identifiers, while lightweight processing is used for medium-risk, low-risk, or indirect identifiers. Semantic relevance is considered as an important factor in strategy design, and highly semantically related privacy information is jointly desensitized to prevent the reconstruction of sensitive data through semantic reasoning. The privacy desensitization strategy focuses on breaking the direct or indirect links between privacy attributes and the subject, ensuring that the subject's identity cannot be recovered through association analysis or reasoning. The priority and processing order of the privacy desensitization strategy are adjusted according to business scenarios and compliance requirements to ensure the security of data during use, storage, and transmission.

[0104] Subsequently, the generated privacy de-identification strategy is applied to business systems, databases, log files, and unstructured data, automatically processing the identified privacy information and its semantic related information. It supports batch processing and real-time processing, allowing for flexible selection based on business needs, and seamlessly integrates with existing data processing workflows. During the de-identification process, the processing procedure and results are recorded simultaneously to ensure traceability and compliance. Special attention is paid to whether the link between the de-identified privacy attributes and the subject is effectively severed.

[0105] Finally, the anonymized data is verified to ensure that sensitive information is effectively protected without affecting the availability and analytical value of business data. Verification methods include statistical analysis, re-identification risk assessment, semantic relevance checks, and expert review. Based on the verification results, the privacy anonymization strategy and application process are continuously optimized to form a closed-loop management system. This ensures that the privacy anonymization strategy can effectively destroy the association between sensitive information and the subject in different scenarios, different data types, and information environments with highly semantic relevance, while maintaining the business availability of the data.

[0106] In one example, after applying the generated privacy de-identification policy, further steps are required, such as... Figure 8 The privacy knowledge base updates and system dynamic optimizations shown include the following:

[0107] The privacy knowledge base is dynamically updated. New privacy data and related information are collected periodically or triggered by business needs. The identified privacy entities, attributes and sensitivity levels are updated to the privacy knowledge base. During the update process, virtual synthetic data, unstructured data and third-party data are also structured to maintain the integrity and accuracy of the privacy knowledge base.

[0108] Semantic-aware management performs semantic analysis on newly added or updated privacy information, identifies highly semantically related privacy entities and their attributes, and ensures that their semantic relationships are marked in the privacy knowledge base. Through semantic awareness, relevant information can be jointly processed in subsequent desensitization, classification, and risk assessment stages, preventing the leakage of sensitive information through semantic reasoning.

[0109] Linkage management monitors the direct or indirect links between privacy attributes and the subject in real time, ensuring that new information does not undermine the effectiveness of existing privacy desensitization strategies. By analyzing and maintaining the links, sensitive relationships are marked in the privacy knowledge base, and corresponding desensitization or risk control measures are automatically triggered to ensure that the data can still effectively protect the subject's privacy in a dynamically changing environment.

[0110] The de-identification effect is quantified and feedback is updated. The updated privacy knowledge base is tested on downstream tasks, including data analysis, recommendation, retrieval and business processes. The de-identification effect and privacy leakage situation are quantified. The execution effect of the privacy de-identification strategy is evaluated through comprehensive indicators (such as classification accuracy, semantic integrity, privacy leakage risk, etc.). The feedback results are used to optimize the privacy attribute classification, de-identification strategy and privacy knowledge base structure to achieve closed-loop update.

[0111] The optimization and strategy iteration of the privacy knowledge base, combined with the execution effect, identification accuracy, business usability and compliance audit results of the privacy de-identification strategy, optimizes the structure, index and query mechanism of the privacy knowledge base. At the same time, through expert review, model feedback and automated analysis, the annotation rules of privacy entities, attributes and semantic associations are continuously optimized so that the privacy knowledge base can adapt to emerging privacy attributes, changes in scenarios and updates of laws and regulations.

[0112] Closed-loop management and system self-adaptation form a closed-loop management mechanism for updating and optimizing the privacy knowledge base. This enables the entire privacy management system to adapt to business changes and data flow. Through semantic awareness and link-based management, the system continuously monitors the risks of sensitive information, achieving a balance between privacy protection, business availability, and compliance, thus forming a dynamic, secure, and intelligent privacy knowledge base management system.

[0113] In one example, besides dynamic template construction based on scenario association, a static template library combined with a rule engine can also be used. This involves selecting appropriate templates based on business type and industry standards, and then using automated rules to quickly map new scenarios, achieving similar privacy entity and attribute definition functionality. The advantage of this approach is its relative simplicity, ease of deployment, and ability to quickly cover common scenarios. However, static templates struggle to handle complex or emerging scenarios, lack flexibility, require manual maintenance for updates, and cannot automatically adapt to business changes like dynamic templates, potentially leading to incomplete privacy coverage or inaccurate sensitivity levels.

[0114] In one example, besides the fixed classification method of seven major categories and thirty-six subcategories, machine learning or knowledge graph reasoning can be used to automatically identify and classify privacy attributes. Sensitivity levels can be determined by combining contextual semantics and historical access patterns, achieving scenario-aware hierarchical management. The advantage of this approach is its ability to dynamically adapt to new data types and business scenarios, reducing manual intervention. However, its disadvantages include the need for large amounts of labeled data for training or rule design, weak model interpretability, and potential risks in compliance auditing and sensitivity level controllability.

[0115] In one example, different data representation methods can be used, such as graph databases or relational databases combined with embedded vector representations, to store privacy entities, attributes, their sensitivity levels, and contextual information; alternatively, structured or semi-structured data can be used, as long as it supports fast querying, updating, and compliance auditing. The advantage of this alternative is that it enables flexible and scalable storage and query mechanisms, but the disadvantage is that without semantic and link management, the knowledge base may not fully reflect the relationships between attributes, leading to inaccurate anonymization or risk assessment.

[0116] In one example, besides fine-tuning, zero-shot or few-shot learning, rule-enhanced models, or multi-model fusion can be used to improve privacy recognition capabilities. Alternatively, transfer learning can be used to train an existing model on similar scenarios and then directly apply it to achieve the same privacy recognition effect. These alternatives reduce training costs or data dependence, but their recognition accuracy may be limited by the coverage of the existing model, and their adaptability to emerging privacy attributes or special scenarios is not as good as targeted fine-tuning.

[0117] In one example, privacy identification and sensitivity level mapping can be achieved using pure rule-driven methods, pure statistical methods, or a hybrid model-rule approach. Alternatively, multimodal data analysis (such as joint identification of text, images, and log information) can be added to achieve the same identification goal. The advantage of these alternatives is the variety of implementation methods, allowing for flexible selection based on business needs. However, the disadvantage is that pure rule-driven or pure statistical methods may fail to handle complex semantics or cross-attribute relationships, easily leading to missed or false positives, thus affecting the accuracy of sensitivity level classification.

[0118] In one example, besides semantic and link-aware desensitization strategies, rule-based static desensitization, randomization-based desensitization, or encryption-based desensitization methods can also be used. Desensitization strategies can be used in combination or applied in stages to protect sensitive data while maintaining business availability. The advantage of alternatives is lower implementation cost and ease of standardization, but the disadvantage is that static or single strategies cannot be dynamically adjusted according to data semantics and subject relationships, which may lead to insufficient desensitization or damage to data availability, making it difficult to balance privacy protection and business analysis needs.

[0119] In one example, in addition to periodic updates, event-driven or incremental update mechanisms can be used to automatically trigger knowledge base updates by monitoring data changes or business events. Dynamic optimization can be combined with feedback mechanisms or reinforcement learning methods to achieve adaptive adjustments to the knowledge base and desensitization strategies. The advantage of this approach is that it can automatically manage changes in the knowledge base, but the disadvantage is that it may not be able to fully capture changes in the semantic relationships between attributes and the main links. If there is a lack of closed-loop verification and quantification of the effects of downstream tasks, it may lead to fluctuations in the privacy protection effect or the failure of the strategy.

[0120] The privacy information identification, classification, grading, and desensitization protection method based on a large model proposed in this embodiment has at least the following beneficial effects compared with traditional methods.

[0121] First, this embodiment dynamically constructs privacy templates associated with specific scenarios, enabling flexible definition of privacy entities, attributes, and their relationships based on different business and industry scenarios. It also embeds laws, regulations, and industry standards, avoiding inconsistencies and omissions caused by subjective human judgment. At the same time, the dynamic template update mechanism ensures real-time matching between the template and business scenarios and regulatory requirements, providing a reliable foundation for subsequent privacy identification and desensitization.

[0122] Second, this embodiment scientifically and reasonably classifies privacy attributes and classifies them by combining direct and indirect identifiers. It also introduces scene-aware sensitivity level mapping to achieve refined privacy classification and grading for different business scenarios, solving the problem that traditional fixed rules are difficult to adapt to changing scenarios.

[0123] Third, this embodiment constructs a privacy knowledge base through multi-source data acquisition and virtual synthesis technology, uniformly processes structured and unstructured data, and embeds compliance rules to form standardized and computable representations of privacy entities, attributes, and sensitivity levels. The privacy knowledge base supports rapid querying, expansion, and updating, providing high-quality training data and rule support for large model fine-tuning and privacy identification, significantly improving the accuracy and coverage of privacy identification.

[0124] Fourth, the method for fine-tuning the large model and improving privacy recognition capabilities proposed in this embodiment, combined with semantic perception and attribute-subject linkage analysis, can achieve the recognition and understanding of complex privacy relationships. Through automated recognition processes, it significantly reduces the cost of manual recognition and improves the system's privacy protection capabilities across multiple scenarios and data types.

[0125] Fifth, in the privacy desensitization stage, this embodiment comprehensively considers semantic relevance and attribute-subject linkage to intelligently desensitize sensitive data, and optimizes the desensitization strategy in a closed loop through downstream task performance, semantic integrity and privacy leakage quantitative indicators to achieve efficient privacy protection while ensuring business availability.

[0126] Sixth, this embodiment continuously optimizes the rules for associating privacy entities, attributes, and scenarios by using a dynamic update and system optimization mechanism for the privacy knowledge base, combined with semantic awareness, link management, and feedback on desensitization effects. This ensures the accuracy and completeness of the privacy knowledge base in the face of business changes and regulatory updates, while forming a closed-loop management system that achieves a dynamic balance between privacy protection, business availability, and compliance, significantly improving privacy management efficiency and system adaptability.

[0127] In summary, the overall solution of this embodiment realizes a complete closed-loop process from privacy template construction, classification and grading, knowledge base construction, large model fine-tuning, privacy identification, intelligent desensitization to dynamic optimization. It solves the problems of imprecise privacy identification, single desensitization strategy, lagging knowledge base updates and poor cross-scenario applicability in traditional methods, and has wide applicability and scalability.

[0128] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0129] Another embodiment of this application proposes an electronic device, such as Figure 9 As shown, it includes a processor 201 and a memory 202. The memory 202 stores instructions that the processor 201 can execute. When the processor 201 is configured to execute the instructions, the electronic device can realize a privacy information identification, classification and grading and desensitization protection method based on a large model as described in the above method embodiment.

[0130] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium.

[0131] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0132] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a privacy information identification, classification, grading, and desensitization protection method based on a large model as described in the above method embodiments.

[0133] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0134] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for privacy information identification, classification, grading, and de-identification protection based on a large model, characterized in that, The method includes: Research and summarize the privacy information of different business systems, data types, and all individuals or organizations that may be involved, and flexibly define privacy entities, concepts and their attributes to build privacy templates; Based on the privacy template, privacy information is scientifically classified and graded, and a classification and grading rule base is formed based on the classification and grading results; Systematically store privacy templates and classification and grading rule bases to build a privacy knowledge base that can be used for large model recognition, classification and de-identification; Select a suitable pre-trained large model based on the business scenario, build a large-scale training set for fine-tuning based on the privacy knowledge base, fine-tune the pre-trained large model based on the large-scale training set, and deploy the fine-tuned large model. Using the deployed large model, privacy identification is performed on the data to be identified, and the identified privacy information is classified and graded, and assigned a corresponding sensitivity level. Based on the identified privacy information and its sensitivity level, corresponding privacy desensitization strategies are generated for different types of privacy entities, and the generated privacy desensitization strategies are applied.

2. The method according to claim 1, characterized in that, For different business systems, data types, and all potentially involved personal or organizational privacy information, conduct research and summarize the findings, flexibly define privacy entities, concepts, and their attributes to construct a privacy template, including: We conduct research and compilation on different business systems, data types, and all possible privacy information of individuals or organizations involved, while referring to relevant industry laws, regulations, and standards to ensure that the privacy templates comply with requirements. Based on the results of the survey and summary, and under the constraints of laws, regulations and standards, a privacy template is defined, including the categories of privacy entities, field attributes, sensitivity levels, possible sources of risk and handling requirements. By combining domain expert knowledge, privacy entities that cannot be automatically obtained from existing data are supplemented, ensuring that privacy templates cover all privacy information that may be involved in business processes and meet the compliance requirements of different industries; The defined privacy templates are validated and optimized, including deduplication, attribute checking and normalization, to ensure that the privacy template structure is complete and consistent, and can be flexibly extended to adapt to different business scenarios and industry standards.

3. The method according to claim 1, characterized in that, Based on the privacy template, privacy information is scientifically classified and graded, and a classification and grading rule base is formed based on the classification and grading results, including: Privacy information is categorized based on privacy entities and their attributes. The categorization results include biometrics, identity documents, medical and health information, financial accounts, location tracking, property and identity information, and social attributes. Based on the sensitivity and identification risks of privacy information, privacy information is divided into direct identifiers and indirect identifiers. The sensitivity level of direct identifiers is defined as high risk level, and the sensitivity level of indirect identifiers is defined as medium risk level or low risk level. The sensitivity level classification results are dynamically adjusted according to the sensitivity of the scenario, legal requirements, industry standards and the complexity of data combination. A classification and grading rule base is formed based on the classification and grading results, including classification labels, sensitivity levels, applicable scenarios, processing requirements and compliance basis for privacy information; By combining actual business data and domain expert reviews, the classification and grading rule base is verified to ensure coverage of all types of privacy information on the platform, and iterative optimization is carried out based on the verification results.

4. The method according to claim 1, characterized in that, A systematic storage of privacy templates and classification / grading rule bases is implemented to construct a privacy knowledge base that can be used for large-scale model recognition, classification, and de-identification, including: Collect data from business systems, databases, log files, documents, third-party data, and industry standards. At the same time, introduce virtual synthesis technology to generate privacy samples that are insufficiently covered or sensitive. Clean and structure the collected and generated data to form structured privacy entities and attribute information. The classification and grading rule base is applied to all privacy entities. Combined with different business scenarios, the sensitivity levels of direct and indirect identifiers are associated with actual application scenarios to form scenario-aware privacy information records. Privacy entities, attributes, sensitivity levels, and scenario information are represented in the form of triples or plurals to build a privacy knowledge base that supports rapid querying, expansion, and updating, while also meeting compliance auditing and tracking needs. Industry laws and regulations and internal security policies are embedded in the privacy knowledge base, and clear processing requirements are recorded for all privacy entities, including whether storage is allowed, whether de-identification is required, and the priority of privacy de-identification policies. Verification is conducted using real business data and virtual synthetic data to ensure the integrity, accuracy, and scenario adaptability of privacy entities. Combined with expert review and model recognition feedback, the privacy knowledge base is continuously updated and optimized to ensure coverage of emerging privacy attributes and changing scenarios. It provides standardized interfaces for large models to call the privacy knowledge base, supports privacy information query, sensitivity level acquisition, and scenario-adaptive rule calling. The privacy knowledge base also supports automated processes for large model fine-tuning and privacy identification, forming a closed-loop privacy management system.

5. The method according to claim 1, characterized in that, Select a suitable pre-trained large model based on the business scenario, construct a large-scale training set for fine-tuning based on a privacy knowledge base, fine-tune the pre-trained large model based on the large-scale training set, and deploy the fine-tuned large model, including: Select a suitable pre-trained large model based on the business scenario, evaluate the basic capabilities of the pre-trained large model to ensure that its ability to understand text, images or structured data meets the basic requirements of privacy recognition, and perform preliminary adaptation of the pre-trained large model, including adjusting the input format, output label system, embedding privacy attributes and their sensitivity levels of the privacy knowledge base, so that the pre-trained large model can identify and distinguish different types of privacy information. Data preprocessing, sampling, and augmentation are performed on the data in the privacy knowledge base to generate a large-scale training set for fine-tuning. The training data in the large-scale training set includes real business data and virtual synthetic data to cover privacy attributes and sensitivity levels in different scenarios. By using supervised or semi-supervised learning, a pre-trained large model is fine-tuned based on a large-scale training set to enable it to accurately identify privacy entities, attributes, and sensitivity levels in different business scenarios. During the fine-tuning process, a scenario-aware strategy is adopted to map the sensitivity levels of different scenarios into the training objective. After fine-tuning, an independent test set is used for evaluation to verify the accuracy, recall, and F1 score of the fine-tuned large model in the identification of different types of privacy information. The fine-tuned large model is then adjusted and optimized based on the results of expert review. Establish a continuous learning mechanism to feed back new privacy information, misidentification cases, and scenario changes identified during the operation of the business system to the fine-tuning process, thereby enabling incremental training and model updates; It provides standardized interfaces to enable seamless integration of fine-tuned large models with privacy knowledge bases and business systems, supporting real-time identification, classification and grading of privacy information, and subsequent de-identification processing.

6. The method according to claim 5, characterized in that, By combining domain expert knowledge and utilizing a deployed large-scale model, privacy identification is performed on the data to be identified. The identified privacy information is then classified and graded, and assigned corresponding sensitivity levels, including: By combining domain expert knowledge and utilizing the deployed large model, the system can automatically scan and analyze business systems, databases, log files, documents, and unstructured data to identify privacy entities and their attribute information. The identified privacy information is processed according to the classification and grading rule base, and direct and indirect identifiers are assigned corresponding sensitivity levels. The sensitivity levels are then dynamically adjusted in combination with different business scenarios and industry standards. The identification and classification results are verified both automatically and manually to ensure the accuracy and completeness of the labeling of privacy entities, attributes and sensitivity levels. At the same time, the identification results are fed back to the large model fine-tuning process to achieve closed-loop optimization.

7. The method according to claim 1, characterized in that, Based on the identified privacy information and its sensitivity level, corresponding privacy de-identification strategies are generated for different types of privacy entities, and the generated privacy de-identification strategies are applied, including: Based on the identified privacy information and its sensitivity level, corresponding privacy desensitization strategies are generated for different types of privacy entities. The strategy types include data masking, hashing, generalization, random perturbation, and encryption. Strict desensitization methods are adopted for high-risk and highly sensitive direct identifiers, while lightweight processing is adopted for medium-risk, low-risk, or indirect identifiers. In the strategy design, semantic relevance is taken as an important consideration. Highly semantically related privacy information is jointly desensitized to prevent the reconstruction of sensitive data through semantic reasoning. The privacy desensitization strategy focuses on destroying the direct or indirect links between privacy attributes and subjects, ensuring that the data cannot recover the subject's identity through association analysis or reasoning. The priority and processing order of privacy desensitization strategies are adjusted according to business scenarios and compliance requirements to ensure the security of data during use, storage, and transmission. The generated privacy de-identification strategy is applied to business systems, databases, log files and unstructured data. It automatically processes the identified privacy information and its semantic related information, supports batch processing and real-time processing, can be flexibly selected according to business needs, and is seamlessly integrated with existing data processing processes. During the de-identification process, the processing process and results are recorded at the same time to ensure the traceability and compliance of the processing. Special attention is paid to whether the link between the de-identified privacy attributes and the subject is effectively severed. The anonymized data is verified to ensure that sensitive information is effectively protected without affecting the availability and analytical value of business data. Verification methods include statistical analysis, re-identification risk assessment, semantic relevance checks, and expert review. Based on the verification results, the privacy anonymization strategy and application process are continuously optimized to form a closed-loop management system. This ensures that the privacy anonymization strategy can effectively destroy the association between sensitive information and the subject in different scenarios, different data types, and information environments with highly semantic relevance, while maintaining the business availability of the data.

8. The method according to any one of claims 1 to 7, characterized in that, After applying the generated privacy de-identification policy, the method further includes: The privacy knowledge base is dynamically updated. New privacy data and related information are collected periodically or triggered by business needs. The identified privacy entities, attributes and sensitivity levels are updated to the privacy knowledge base. During the update process, virtual synthetic data, unstructured data and third-party data are also structured to maintain the integrity and accuracy of the privacy knowledge base. Semantic awareness management performs semantic analysis on newly added or updated privacy information, identifies highly semantically related privacy entities and their attributes, and ensures that their semantic relationships are marked in the privacy knowledge base; Linkage management monitors the direct or indirect links between privacy attributes and the subject in real time to ensure that new information does not undermine the effectiveness of existing privacy desensitization strategies. By analyzing and maintaining the links, sensitive relationships are marked in the privacy knowledge base, and corresponding desensitization or risk control measures are automatically triggered to ensure that the data can still effectively protect the subject's privacy in a dynamically changing environment. The de-identification effect is quantified and feedback is updated. The updated privacy knowledge base is tested on downstream tasks, including data analysis, recommendation, retrieval and business processes. The de-identification effect and privacy leakage are quantified. The execution effect of the privacy de-identification strategy is evaluated through comprehensive indicators. The feedback results are used to optimize the privacy attribute classification, de-identification strategy and privacy knowledge base structure to achieve closed-loop update. The optimization and strategy iteration of the privacy knowledge base, combined with the execution effect, identification accuracy, business usability and compliance audit results of the privacy desensitization strategy, optimizes the structure, index and query mechanism of the privacy knowledge base. At the same time, through expert review, model feedback and automated analysis, the rules for labeling privacy entities, attributes and semantic associations are continuously optimized so that the privacy knowledge base can adapt to emerging privacy attributes, changes in scenarios and updates in laws and regulations. Closed-loop management and system self-adaptation form a closed-loop management mechanism for updating and optimizing the privacy knowledge base. This enables the entire privacy management system to adapt to business changes and data flow. Through semantic awareness and link-based management, the system continuously monitors the risks of sensitive information, achieving a balance between privacy protection, business availability, and compliance, thus forming a dynamic, secure, and intelligent privacy knowledge base management system.

9. An electronic device, characterized in that, include: The processor and memory, wherein the memory stores instructions executable by the processor, and the processor is configured to, when executing the instructions, enable the electronic device to implement a privacy information identification, classification, grading, and desensitization protection method based on a large model as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement a method for privacy information identification, classification, grading, and desensitization protection based on a large model, as described in any one of claims 1 to 8.

Images