Sensitive data identification and desensitization method and system based on AI large model

Abstract

The invention relates to a sensitive data identification and desensitization method and system based on an AI large model, and belongs to the technical field of data security processing and artificial intelligence. The method comprises the following steps: receiving a multi-source data processing request, establishing a sensitive data identification framework, analyzing elements of sensitive data, and extracting a sensitive data feature vector through a natural language processing algorithm; classifying the multi-source data, dynamically adjusting an identification threshold value of the sensitive data identification framework through a sensitive data feature vector, and generating a self-adaptive sensitive data identification strategy; establishing a terminal desensitization verification model, performing desensitization verification on the identified sensitive data, performing mask replacement or generalization processing according to privacy requirements, and outputting a desensitization retry processing scheme; the sensitive data is subjected to block desensitization, a large data set is divided into small data sets for independent desensitization, replaceable entities in information are removed according to data attributes, and the desensitized data is further subjected to secondary desensitization.

Description

Technical Field

[0001] This invention belongs to the field of data security processing and artificial intelligence technology, specifically relating to a method and system for sensitive data identification and desensitization based on a large AI model. Background Technology

[0002] With the widespread application of cloud computing, big data, and artificial intelligence technologies, enterprises and institutions generate and process large amounts of data containing personal privacy, trade secrets, and sensitive information in multi-source, heterogeneous environments. Traditional methods for sensitive data identification and de-identification mainly rely on rule matching or keyword filtering, which are difficult to adapt to scenarios with complex semantic structures, cross-domain data fusion, and unstructured text, easily leading to low identification accuracy and high false positive rates. Meanwhile, existing de-identification technologies mostly use static rules or fixed thresholds, unable to dynamically adjust de-identification strategies based on data characteristics, making it difficult to balance data usability and privacy protection needs. Especially in artificial intelligence model training and large-scale data analysis, the risk of sensitive information leakage increases exponentially, urgently requiring a technical solution with intelligent identification, adaptive adjustment, and multi-layer de-identification capabilities. In recent years, semantic understanding and feature extraction technologies based on large AI models have demonstrated strong generalization capabilities in natural language processing, privacy computing, and data security, providing new solutions for the accurate identification and dynamic de-identification of sensitive data. Therefore, how to utilize large AI models to extract deep features from data, establish a self-learning identification threshold adjustment mechanism, and achieve segmented desensitization and differential consistency verification in a high-performance parallel environment has become an important technical direction for improving data security processing capabilities. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, this invention provides a sensitive data identification and de-identification method based on a large AI model. The objective of this invention can be achieved through the following technical solutions: S1: Receive a multi-source data processing request, establish a sensitive data identification framework based on the attributes of the multi-source data, parse the elements of the sensitive data according to the identified data attributes, and analyze the elements of the sensitive data through a natural language processing algorithm to extract the sensitive data feature vector; S2: Classify the multi-source data according to the sensitive data identification framework, and combine the classification status after classification, dynamically adjust the identification threshold of the sensitive data identification framework through the sensitive data feature vector to generate an adaptive sensitive data identification strategy; S3: Based on the adaptive sensitive data identification scheme, establish a terminal de-identification verification model. The terminal de-identification verification model performs de-identification verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a de-identification retry processing scheme. S4: Execute the desensitization and retry processing scheme, desensitize the sensitive data in blocks, divide the large dataset into small datasets for independent desensitization, remove replaceable entities in the information according to the data attributes, monitor the desensitization effect in real time, and if the privacy risk of the desensitized data still exceeds the threshold, further desensitize the desensitized data a second time.

[0004] Specifically, the attributes of the multi-source data serve as the identification method conditions of the sensitive data identification framework. By analyzing the sensitive data, the identification pattern in the sensitive data identification framework is matched according to the processing request.

[0005] Specifically, the method for constructing the sensitive data identification framework is as follows: the sensitive data is decomposed into data formats and corresponding data features are extracted; the data privacy support capabilities of the receiver and the capability profile of the sender are obtained through the natural language processing protocol to record the current identification status; and a mapping relationship is established between the data privacy format of the receiver and the capability profile of the sender.

[0006] Specifically, the sensitive data feature vector includes textual semantic features, structured field features, and contextual features. The textual semantic features are used to describe the sensitive words, entity names, and their semantic similarity contained in the data, and are obtained by vectorizing the word embedding results through a natural language processing model. The structured field features are used to reflect the field types, positional relationships, and format patterns of the data in database tables, log records, or message streams. The contextual features are used to measure the dependency and co-occurrence probability of sensitive data in the same semantic fragment or across document scenarios.

[0007] Specifically, the method for adjusting the identification threshold of the sensitive data identification framework is as follows: Based on the weights of privacy protection and data integrity, the adaptive sensitive data identification strategy analyzes the information types contained in the sensitive data, classifies the content of each single type in the sensitive data, and determines the identification threshold of the text content according to the ratio of the original number of tokens to the target number of tokens based on the data attribute format. The identification thresholds of the total dimension and the dimension after identification are dynamically adjusted in combination with the attention mechanism. The attention mechanism filters optional terms in the text and replaceable features in the image, and optimizes the weights of privacy protection and data integrity during the recognition process based on sensitive data characteristics and the recipient's capability profile.

[0008] Specifically, the method for establishing the terminal de-identification verification model is as follows: Based on the aforementioned sensitive data adaptive identification strategy, the format and desensitization method of the receiver's data content are obtained through a secure transmission protocol, and the obtained identification information and desensitization conversion strategy are stored in the sender's capability file. Based on the data characteristics in the sender's capability profile, the logical structure of the terminal desensitization verification model is established, and by detecting whether the identified content meets the recipient's privacy requirements and format support, the simulated desensitization results of the recipient for different sensitive data are generated. By analyzing the simulated desensitization results of different sensitive data, the desensitization processing strategy is run in different system states and operating systems, and the judgment rules and processing strategies of the model are dynamically updated according to the desensitization processing strategy to generate an adaptive and adjustable terminal desensitization verification system.

[0009] Specifically, the adaptive sensitive data identification strategy differentiates the identification thresholds for different categories of data based on the distribution density and feature similarity calculation results of the sensitive data feature vectors; and makes a sensitive judgment on the confidence level of the identification results to identify sensitive data in different situations.

[0010] Specifically, the process of de-identifying and verifying the identified sensitive data is as follows: The identified sensitive data is input into the terminal desensitization verification model; the terminal desensitization verification model analyzes the sensitive data feature information; matches the content features with the recipient's capability profile, and determines whether the content is within the recipient's privacy scope; For content deemed non-compliant, a de-identification feedback algorithm is used to convert the content into a format supported by the receiving end; if the conversion still fails to meet compliance requirements, the structured content is downgraded to anonymous content based on the recipient's capability profile. The processed content is sent to the receiver via a secure transmission protocol. The received content is verified by returning transmission and parsing feedback. The success of the de-identification verification is determined based on the feedback information of the received content. If the verification fails, the identification strategy or de-identification conversion strategy is dynamically adjusted based on the feedback, and a de-identification retry scheme is generated.

[0011] Specifically, the desensitization retry scheme is based on the desensitization retry mechanism, which analyzes the anomalies that occur in the parsing, storage or transmission of sensitive data. For content that does not comply with privacy regulations, the desensitization conversion is performed through the terminal desensitization verification model to generate a new format that the receiving end can support. During the retry process, the desensitization retry scheme is adaptively adjusted in combination with the information format of the receiving end.

[0012] Specifically, the segmented desensitization uses a data segmentation engine to logically divide the dataset to be desensitized according to data type, field correlation, and sensitivity level, generating multiple independent sub-data blocks. These sub-data blocks are then used to perform local desensitization operations in independent desensitization task threads. After desensitization is completed, each sub-data block is validated and reassembled.

[0013] Specifically, the sensitive data identification framework initiates privacy type detection to the recipient through sequence labeling, identifies the elements of the sensitive data, and writes the recipient's privacy compliance format into the sender's capability profile after the response, and parses the feature vectors of the sensitive data elements.

[0014] Specifically, a sensitive data identification and de-identification system based on an AI large model is characterized by comprising: Sensitive data acquisition module: Receives multi-source data processing requests, establishes a sensitive data identification framework based on the attributes of the multi-source data, parses the elements of the sensitive data according to the identified data attributes, and analyzes the elements of the sensitive data through natural language processing algorithms to extract sensitive data feature vectors; Data feature recognition module: Classifies the multi-source data according to the sensitive data recognition framework, and dynamically adjusts the recognition threshold of the sensitive data recognition framework by combining the classification status after classification, and generates an adaptive sensitive data recognition strategy by using sensitive data feature vectors; Desensitization verification processing module: Based on the adaptive sensitive data identification scheme, a terminal desensitization verification model is established. The terminal desensitization verification model performs desensitization verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a desensitization retry processing scheme. Batch privacy enhancement module: Executes the desensitization and retry processing scheme, desensitizes the sensitive data in blocks, divides the large dataset into small datasets for independent desensitization, removes replaceable entities in the information according to data attributes, monitors the desensitization effect in real time, and further desensitizes the desensitized data a second time when the privacy risk of the desensitized data still exceeds the threshold.

[0015] The beneficial effects of this invention are as follows: Compared with existing traditional methods based on rule matching or static de-identification strategies, the sensitive data identification and de-identification method based on AI large-scale models proposed in this invention has significant technological advancements and practical value. First, by introducing the semantic understanding and feature extraction capabilities of AI large-scale models, deep semantic recognition of sensitive elements in multi-source heterogeneous data can be achieved, significantly improving the accuracy and recall rate of sensitive data identification. Second, by constructing sensitive data feature vectors and introducing adaptive identification strategies, this invention achieves dynamic optimization of the identification threshold, enabling the system to automatically adjust the identification intensity according to data type and privacy level, improving the algorithm's adaptability in different scenarios. Third, the terminal de-identification verification model uses a differential consistency verification algorithm to verify the consistency of the de-identification results, maintaining the effectiveness of data analysis while ensuring privacy security. Finally, through block-based de-identification and secondary de-identification mechanisms, this invention can achieve high-parallelism de-identification processing in large-scale dataset environments and perform real-time monitoring and reinforcement correction of residual privacy risks, thereby significantly improving de-identification performance and data security. In summary, this invention can achieve refined and intelligent privacy protection while ensuring data usability, and has good prospects for widespread application. Attached Figure Description

[0016] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of the structure of a sensitive data identification and desensitization method and system based on an AI large model according to the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the overall technical flow of a sensitive data identification and desensitization method and system based on an AI large model according to the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0020] Please see Figure 1 A sensitive data identification and de-identification method based on AI large-scale models: S1: Receive a multi-source data processing request, establish a sensitive data identification framework based on the attributes of the multi-source data, parse the elements of the sensitive data according to the identified data attributes, and analyze the elements of the sensitive data through a natural language processing algorithm to extract the sensitive data feature vector; S2: Classify the multi-source data according to the sensitive data identification framework, and combine the classification status after classification, dynamically adjust the identification threshold of the sensitive data identification framework through the sensitive data feature vector to generate an adaptive sensitive data identification strategy; S3: Based on the adaptive sensitive data identification scheme, establish a terminal de-identification verification model. The terminal de-identification verification model performs de-identification verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a de-identification retry processing scheme. S4: Execute the desensitization and retry processing scheme, desensitize the sensitive data in blocks, divide the large dataset into small datasets for independent desensitization, remove replaceable entities in the information according to the data attributes, monitor the desensitization effect in real time, and if the privacy risk of the desensitized data still exceeds the threshold, further desensitize the desensitized data a second time.

[0021] Specifically, the attributes of the multi-source data serve as the identification method conditions of the sensitive data identification framework. By analyzing the sensitive data, the identification pattern in the sensitive data identification framework is matched according to the processing request.

[0022] Specifically, the method for constructing the sensitive data identification framework is as follows: the sensitive data is decomposed into data formats and corresponding data features are extracted; the data privacy support capabilities of the receiver and the capability profile of the sender are obtained through the natural language processing protocol to record the current identification status; and a mapping relationship is established between the data privacy format of the receiver and the capability profile of the sender.

[0023] Specifically, the sensitive data feature vector includes textual semantic features, structured field features, and contextual features. The textual semantic features are used to describe the sensitive words, entity names, and their semantic similarity contained in the data, and are obtained by vectorizing the word embedding results through a natural language processing model. The structured field features are used to reflect the field types, positional relationships, and format patterns of the data in database tables, log records, or message streams. The contextual features are used to measure the dependency and co-occurrence probability of sensitive data in the same semantic fragment or across document scenarios.

[0024] This embodiment provides a sensitive data identification and de-identification technology process based on a large AI model, such as... Figure 2 As shown, it includes the following steps: Feature vector construction: Receives a multi-source dataset D = {d1, d2, ..., d...} n The system performs attribute parsing on each data sample, extracting semantic, structural, and contextual features to construct a sensitive data feature vector. The feature vector is defined as follows: , Among them, E s E is a semantic embedding function. t For embedding structured fields, E c For contextual relevance features, ƒ NLP This is the feature fusion mapping function.

[0025] Adaptive recognition threshold update: After classifying different data categories, the system calculates the classification confidence P. c (d) i ), and dynamically adjust the recognition threshold based on the rate of change of the feature vector: , Among them, T old S(V) is the initial recognition threshold, α is the learning rate coefficient, and S(V) is the learning rate coefficient. i Let S(V) represent the feature similarity score of the current sample, and S(V) be the sample mean. This formula enables self-learning optimization of the recognition threshold, allowing the system to dynamically balance the recognition intensity of sensitive information under different data scenarios.

[0026] Desensitization verification model: The system establishes a desensitization verification model based on the differential consistency verification algorithm to verify the identified sensitive data.

[0027] Differential consistency is defined as: , Among them, V ori V is the original feature vector. mask Let be the feature vector after desensitization, and ϵ be a small constant to prevent division by zero. If Δ diff <δ th If the desensitization intensity is insufficient, the process will proceed to a retry procedure.

[0028] Chunking and Secondary Masking: Performing chunking masking on large datasets involves dividing the data into several subsets. Each D k Independent de-identification. The system monitors the privacy risk value R after de-identification in real time. k The calculation is as follows: , Among them, P link Indicates linkability risk, S sim This represents the risk of feature similarity, where β1 and β2 are weighting factors. When R... k >R th When the risk value is below the threshold, a secondary desensitization mechanism is triggered, which performs enhanced generalization or pseudo-random replacement operations on the subset until the risk value is below the threshold.

[0029] Through the above steps, the present invention realizes a complete closed-loop process from intelligent identification of sensitive information → threshold adaptive optimization → differential verification → block segmentation and secondary desensitization, which not only ensures privacy and security but also preserves the analytical usability of the data.

[0030] Specifically, the method for adjusting the identification threshold of the sensitive data identification framework is as follows: Based on the weights of privacy protection and data integrity, the adaptive sensitive data identification strategy analyzes the information types contained in the sensitive data, classifies the content of each single type in the sensitive data, and determines the identification threshold of the text content according to the ratio of the original number of tokens to the target number of tokens based on the data attribute format. The identification thresholds of the total dimension and the dimension after identification are dynamically adjusted in combination with the attention mechanism. The attention mechanism filters optional terms in the text and replaceable features in the image, and optimizes the weights of privacy protection and data integrity during the recognition process based on sensitive data characteristics and the recipient's capability profile.

[0031] Specifically, the method for establishing the terminal de-identification verification model is as follows: Based on the aforementioned sensitive data adaptive identification strategy, the format and desensitization method of the receiver's data content are obtained through a secure transmission protocol, and the obtained identification information and desensitization conversion strategy are stored in the sender's capability file. Based on the data characteristics in the sender's capability profile, the logical structure of the terminal desensitization verification model is established, and by detecting whether the identified content meets the recipient's privacy requirements and format support, the simulated desensitization results of the recipient for different sensitive data are generated. By analyzing the simulated desensitization results of different sensitive data, the desensitization processing strategy is run in different system states and operating systems, and the judgment rules and processing strategies of the model are dynamically updated according to the desensitization processing strategy to generate an adaptive and adjustable terminal desensitization verification system.

[0032] Specifically, the adaptive sensitive data identification strategy differentiates the identification thresholds for different categories of data based on the distribution density and feature similarity calculation results of the sensitive data feature vectors; and makes a sensitive judgment on the confidence level of the identification results to identify sensitive data in different situations.

[0033] Specifically, the process of de-identifying and verifying the identified sensitive data is as follows: The identified sensitive data is input into the terminal desensitization verification model; the terminal desensitization verification model analyzes the sensitive data feature information; matches the content features with the recipient's capability profile, and determines whether the content is within the recipient's privacy scope; For content deemed non-compliant, a de-identification feedback algorithm is used to convert the content into a format supported by the receiving end; if the conversion still fails to meet compliance requirements, the structured content is downgraded to anonymous content based on the recipient's capability profile. The processed content is sent to the receiver via a secure transmission protocol. The received content is verified by returning transmission and parsing feedback. The success of the de-identification verification is determined based on the feedback information of the received content. If the verification fails, the identification strategy or de-identification conversion strategy is dynamically adjusted based on the feedback, and a de-identification retry scheme is generated.

[0034] In this embodiment, the AI-based big data model-based sensitive data identification and de-identification system of this invention is deployed in the hospital data center. The system first receives multi-source data input from electronic medical records, laboratory information systems, and image archiving systems. The sensitive data identification module automatically parses potential sensitive elements in text, images, and structured fields, such as patient names, visit numbers, and medical image tags. The AI ​​big data model utilizes semantic understanding and contextual analysis capabilities to identify information that can be directly or indirectly used for identity inference and extracts corresponding feature vectors.

[0035] Subsequently, the adaptive recognition strategy module automatically adjusts the recognition threshold according to different data types (such as text, tables, and images), distinguishes between high-sensitivity and low-sensitivity data, and generates the optimal desensitization strategy. After performing differential consistency detection, the desensitization verification module performs mask replacement or semantic generalization on the identified sensitive fields, such as replacing "Zhang San" with "Patient A" and generalizing specific ward numbers to "general ward".

[0036] Finally, the system uses a block-based de-identification mechanism to process large-scale datasets in parallel on a distributed server, monitoring the de-identification effect in real time. When a batch of de-identified results is detected to still pose privacy risks, the system automatically triggers a secondary de-identification strategy to reprocess the high-risk data. The processed data not only meets the needs of scientific research statistics and model training but also ensures that the patient's true identity cannot be recovered through reverse inference, thereby achieving privacy protection and secure use in the process of medical data sharing.

[0037] Specifically, the desensitization retry scheme is based on the desensitization retry mechanism, which analyzes the anomalies that occur in the parsing, storage or transmission of sensitive data. For content that does not comply with privacy regulations, the desensitization conversion is performed through the terminal desensitization verification model to generate a new format that the receiving end can support. During the retry process, the desensitization retry scheme is adaptively adjusted in combination with the information format of the receiving end.

[0038] Specifically, the segmented desensitization uses a data segmentation engine to logically divide the dataset to be desensitized according to data type, field correlation, and sensitivity level, generating multiple independent sub-data blocks. These sub-data blocks are then used to perform local desensitization operations in independent desensitization task threads. After desensitization is completed, each sub-data block is validated and reassembled.

[0039] Specifically, the sensitive data identification framework initiates privacy type detection to the recipient through sequence labeling, identifies the elements of the sensitive data, and writes the recipient's privacy compliance format into the sender's capability profile after the response, and parses the feature vectors of the sensitive data elements.

[0040] Specifically, a sensitive data identification and de-identification system based on an AI large model is characterized by comprising: Sensitive data acquisition module: Receives multi-source data processing requests, establishes a sensitive data identification framework based on the attributes of the multi-source data, parses the elements of the sensitive data according to the identified data attributes, and analyzes the elements of the sensitive data through natural language processing algorithms to extract sensitive data feature vectors; Data feature recognition module: Classifies the multi-source data according to the sensitive data recognition framework, and dynamically adjusts the recognition threshold of the sensitive data recognition framework by combining the classification status after classification, and generates an adaptive sensitive data recognition strategy by using sensitive data feature vectors; Desensitization verification processing module: Based on the adaptive sensitive data identification scheme, a terminal desensitization verification model is established. The terminal desensitization verification model performs desensitization verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a desensitization retry processing scheme. Batch privacy enhancement module: Executes the desensitization and retry processing scheme, desensitizes the sensitive data in blocks, divides the large dataset into small datasets for independent desensitization, removes replaceable entities in the information according to data attributes, monitors the desensitization effect in real time, and further desensitizes the desensitized data a second time when the privacy risk of the desensitized data still exceeds the threshold.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for sensitive data identification and de-identification based on a large AI model, characterized in that, include: S1: Receive a multi-source data processing request, establish a sensitive data identification framework based on the attributes of the multi-source data, parse the elements of the sensitive data according to the identified data attributes, and analyze the elements of the sensitive data through a natural language processing algorithm to extract the sensitive data feature vector; S2: Classify the multi-source data according to the sensitive data identification framework, and combine the classification status after classification, dynamically adjust the identification threshold of the sensitive data identification framework through the sensitive data feature vector to generate an adaptive sensitive data identification strategy; S3: Based on the adaptive sensitive data identification scheme, establish a terminal de-identification verification model. The terminal de-identification verification model performs de-identification verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a de-identification retry processing scheme. S4: Execute the desensitization and retry processing scheme, desensitize the sensitive data in blocks, divide the large dataset into small datasets for independent desensitization, remove replaceable entities in the information according to the data attributes, monitor the desensitization effect in real time, and if the privacy risk of the desensitized data still exceeds the threshold, further desensitize the desensitized data a second time.

2. The method according to claim 1, characterized in that, The attributes of the multi-source data serve as the identification method conditions of the sensitive data identification framework. By analyzing the sensitive data, the identification pattern in the sensitive data identification framework is matched according to the processing request.

3. The method according to claim 1, characterized in that, The method for constructing the sensitive data identification framework is as follows: decompose the sensitive data into data formats and extract the corresponding data features; obtain the data privacy support capabilities of the receiver and the capability profile of the sender through the natural language processing protocol to record the current identification status; and establish a mapping relationship between the data privacy format of the receiver and the capability profile of the sender.

4. The method according to claim 1, characterized in that, The sensitive data feature vector includes textual semantic features, structured field features, and contextual features. The textual semantic features are used to describe the sensitive words, entity names, and their semantic similarity contained in the data, and are obtained by vectorizing the word embedding results through a natural language processing model. The structured field features are used to reflect the field types, positional relationships, and format patterns of the data in database tables, log records, or message streams. The contextual features are used to measure the dependency and co-occurrence probability of sensitive data in the same semantic fragment or cross-document scenarios.

5. The method according to claim 2, characterized in that, The method for adjusting the identification threshold of the sensitive data identification framework is as follows: Based on the weights of privacy protection and data integrity, the adaptive sensitive data identification strategy analyzes the information types contained in the sensitive data, classifies the content of each single type in the sensitive data, and determines the identification threshold of the text content according to the ratio of the original number of tokens to the target number of tokens based on the data attribute format. The identification thresholds of the total dimension and the dimension after identification are dynamically adjusted in combination with the attention mechanism. The attention mechanism filters optional terms in the text and replaceable features in the image, and optimizes the weights of privacy protection and data integrity during the recognition process based on sensitive data characteristics and the recipient's capability profile.

6. The method according to claim 5, characterized in that, The method for establishing the terminal de-identification verification model is as follows: Based on the aforementioned sensitive data adaptive identification strategy, the format and desensitization method of the receiver's data content are obtained through a secure transmission protocol, and the obtained identification information and desensitization conversion strategy are stored in the sender's capability file. Based on the data characteristics in the sender's capability profile, the logical structure of the terminal desensitization verification model is established, and by detecting whether the identified content meets the recipient's privacy requirements and format support, the simulated desensitization results of the recipient for different sensitive data are generated. By analyzing the simulated desensitization results of different sensitive data, the desensitization processing strategy is run in different system states and operating systems, and the judgment rules and processing strategies of the model are dynamically updated according to the desensitization processing strategy to generate an adaptive and adjustable terminal desensitization verification system.

7. The method according to claim 4, characterized in that, The adaptive sensitive data identification strategy adjusts the identification thresholds for different categories of data based on the distribution density and feature similarity calculation results of the sensitive data feature vectors; and makes a sensitive judgment on the confidence level of the identification results to identify sensitive data in different situations.

8. The method according to claim 2, characterized in that, The process of de-identifying and verifying the identified sensitive data is as follows: The identified sensitive data is input into the terminal desensitization verification model; the terminal desensitization verification model analyzes the sensitive data feature information; matches the content features with the recipient's capability profile, and determines whether the content is within the recipient's privacy scope; For content deemed non-compliant, a de-identification feedback algorithm is used to convert the content into a format supported by the receiving end; if the conversion still fails to meet compliance requirements, the structured content is downgraded to anonymous content based on the recipient's capability profile. The processed content is sent to the receiver via a secure transmission protocol. The received content is verified by returning transmission and parsing feedback. The success of the de-identification verification is determined based on the feedback information of the received content. If the verification fails, the identification strategy or de-identification conversion strategy is dynamically adjusted based on the feedback, and a de-identification retry scheme is generated.

9. The method according to claim 4, characterized in that, The desensitization retry scheme is based on the desensitization retry mechanism. It analyzes the anomalies that occur in the parsing, storage or transmission of sensitive data. For content that does not comply with privacy regulations, it performs desensitization conversion through the terminal desensitization verification model to generate a new format that the receiving end can support. During the retry process, it adaptively adjusts the desensitization retry scheme based on the information format of the receiving end.

10. The method according to claim 4, characterized in that, The segmented desensitization uses a data segmentation engine to logically divide the dataset to be desensitized according to data type, field correlation, and sensitivity level, generating multiple independent sub-data blocks. These sub-data blocks are then used to perform local desensitization operations in independent desensitization task threads. After desensitization is completed, each sub-data block is validated and reassembled.

11. The method according to claim 7, characterized in that, The sensitive data identification framework initiates privacy type detection to the recipient through sequence labeling, identifies the elements of the sensitive data, and writes the recipient's privacy compliance format into the sender's capability profile after the response, and parses the feature vector of the sensitive data elements.

12. A sensitive data identification and de-identification system based on an AI large model, characterized in that, include: Sensitive data acquisition module: Receives multi-source data processing requests, establishes a sensitive data identification framework based on the attributes of the multi-source data, parses the elements of the sensitive data according to the identified data attributes, and analyzes the elements of the sensitive data through natural language processing algorithms to extract sensitive data feature vectors; Data feature recognition module: Classifies the multi-source data according to the sensitive data recognition framework, and dynamically adjusts the recognition threshold of the sensitive data recognition framework by combining the classification status after classification, and generates an adaptive sensitive data recognition strategy by using sensitive data feature vectors; Desensitization verification processing module: Based on the adaptive sensitive data identification scheme, a terminal desensitization verification model is established. The terminal desensitization verification model performs desensitization verification on the identified sensitive data through the differential consistency verification algorithm, and performs mask replacement or generalization processing according to privacy requirements, and outputs a desensitization retry processing scheme. Batch privacy enhancement module: Executes the desensitization and retry processing scheme, desensitizes the sensitive data in blocks, divides the large dataset into small datasets for independent desensitization, removes replaceable entities in the information according to data attributes, monitors the desensitization effect in real time, and further desensitizes the desensitized data a second time when the privacy risk of the desensitized data still exceeds the threshold.

Images