Intelligent privacy protection method and device for database and medium

Abstract

The embodiment of the invention discloses an intelligent privacy protection method and device for a database and a medium, belongs to the technical field of data privacy protection, and solves the problems that an existing privacy protection method is inaccurate in sensitive information identification and classification and poor in privacy data protection effect. Comprising the following steps: analyzing to-be-identified data through a preset sensitive information identification model to obtain reference sensitive information; performing secondary verification on the reference sensitive information, and feeding back a verification result to a preset sensitive information identification model so as to optimize the sensitive information identification model to obtain a target sensitive information identification model; based on a target sensitive information identification model, identifying data in the production environment to obtain target sensitive information; carrying out desensitization processing on the target sensitive information, and carrying out statistical analysis on the data after desensitization processing; the statistical analysis result is verified, and data access record analysis is carried out under the condition that verification is passed, so that database privacy protection is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of data privacy protection, and in particular to a method, device and medium for intelligent privacy protection of a database. Background Art

[0002] In today's data-driven era, databases, as the cornerstone of the information society, carry unprecedented amounts of data, including highly sensitive personal identity information, precise financial data, and vital health records. With the surge in data volume and the acceleration of data circulation, data security management faces unprecedented challenges, especially the protection of sensitive information, which has become the focus of attention from all walks of life.

[0003] Traditional privacy protection methods have gradually revealed their limitations in response to the complex and ever-changing threat environment of the modern world. On the one hand, although traditional data encryption and access control technologies can prevent unauthorized access to a certain extent, they often have difficulty accurately identifying the subtle boundary between legitimate needs and potential risks, resulting in "one-size-fits-all" protection measures that both limit the effective use of data and fail to completely block the channels for data leakage. On the other hand, with the continuous evolution of attack methods, such as advanced persistent threats, internal threats, and supply chain attacks, traditional defense systems seem to be unable to cope with the situation, and the protection effect is greatly reduced, resulting in existing privacy protection methods. Problems such as inaccurate identification and classification of sensitive information and poor privacy data protection. Summary of the invention

[0004] The embodiments of the present application provide a method, device and medium for intelligent privacy protection of a database, which are used to solve the following technical problems: the existing privacy protection methods have problems such as inaccurate identification and classification of sensitive information and poor privacy data protection effect.

[0005] The present application embodiment adopts the following technical solutions:

[0006] The embodiment of the present application provides an intelligent privacy protection method for a database. It includes: analyzing the data to be identified through a preset sensitive information identification model to obtain reference sensitive information; performing secondary verification on the reference sensitive information, and feeding back the verification result to the preset sensitive information identification model to optimize the sensitive information identification model to obtain a target sensitive information identification model; based on the target sensitive information identification model, identifying the data in the production environment to obtain the target sensitive information; desensitizing the target sensitive information, and performing statistical analysis on the desensitized data; verifying the statistical analysis results, and if the verification passes, performing data access record analysis to achieve database privacy protection.

[0007] The embodiment of the present application analyzes the data to be identified by presetting a sensitive information identification model to obtain reference sensitive information. The large model can understand complex language structures and contexts, thereby more accurately identifying private data. And after a large amount of diverse data training, it can identify various types of private data. Secondly, the embodiment of the present application desensitizes the target sensitive information and performs statistical analysis on the desensitized data to ensure that sensitive information is protected during the data processing process, while also obtaining valuable calculation results. This effectively prevents unauthorized access and data leakage, and significantly improves data security and the overall reliability of the system.

[0008] In one implementation of the present application, before analyzing the data to be identified by a preset sensitive information identification model to obtain reference sensitive information, the method also includes: preprocessing the database data; wherein the preprocessing includes at least data cleaning and data labeling; using the LLM large language model as a preset sensitive information identification model, and converting the preprocessed database data into an input format corresponding to the preset sensitive information identification model; based on the labeled database data and the supervised learning method, fine-tuning the preset sensitive information identification model so that the fine-tuned preset sensitive information identification model extracts features of sensitive information.

[0009] In one implementation of the present application, the data to be identified is analyzed by a preset sensitive information identification model to obtain reference sensitive information, which specifically includes: inputting the data to be identified into a fine-tuned preset sensitive information identification model, extracting sensitive information features by the fine-tuned preset sensitive information identification model, and outputting the identified sensitive information in a preset format; wherein the sensitive information includes at least the type of sensitive information and the location of the sensitive information; and testing the fine-tuned preset sensitive information identification model through a verification set, and when the test passes, determining that the output sensitive information is correct.

[0010] In one implementation of the present application, a secondary check is performed on the reference sensitive information, and the check result is fed back to the preset sensitive information recognition model to optimize the sensitive information recognition model and obtain a target sensitive information recognition model, specifically including: a secondary check is performed on the reference sensitive information through a regular expression; wherein the secondary check includes at least one of a format check, a logic check, and a data consistency check; error information corresponding to the secondary check is determined, and the error information is added to a training set as new annotated data; the preset sensitive information recognition model is iteratively optimized by adding the training set with the new annotated data; until the output recognition result of the sensitive information meets the preset requirements, the parameter combination corresponding to the preset sensitive information recognition model at this time is determined to obtain the target sensitive information recognition model.

[0011] In one implementation of the present application, before performing statistical analysis on the desensitized data, the method also includes: in a test environment, inserting the desensitized data into a database for structural verification; wherein the structural verification includes at least one of field type verification, length consistency verification, field format verification, and database constraint verification; formatting the desensitized data into structured data, and processing the abnormal values ​​and missing values ​​detected by the verification.

[0012] In one implementation of the present application, statistical analysis is performed on the desensitized data, specifically including: determining a plurality of custom fusion algorithms in a database; performing corresponding operations on the desensitized data through a plurality of custom fusion algorithms to achieve statistical analysis; wherein the statistical analysis includes at least one of mean value processing, variance processing, correlation processing, trend processing, and outlier detection processing.

[0013] In one implementation of the present application, the statistical analysis results are verified, specifically including: inputting the statistical analysis results into a preset model; comparing the statistical analysis results with the original data for key statistical indicators through the preset model; wherein the key statistical indicators include at least one of the mean, variance, and distribution form; and comparing the classification accuracy corresponding to the desensitized data in the statistical analysis results with the classification accuracy corresponding to the original data; and detecting abnormal data in the statistical analysis results to verify the business logic corresponding to the desensitized data through the detected abnormal data.

[0014] In one implementation of the present application, when verification is passed, data access record analysis is performed, specifically including: defining different access permission levels according to user roles; recording access information of sensitive information and generating access logs; wherein the access information includes at least the accessing user, access time, and access operation content; and regularly performing anomaly detection on the access logs, and when it is determined that abnormal access behavior exists, adjusting the user's access permission level according to the detection results.

[0015] An embodiment of the present application provides an intelligent privacy protection device for a database, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: analyze the data to be identified by a preset sensitive information identification model to obtain reference sensitive information; perform a secondary verification on the reference sensitive information, and feed the verification result back to the preset sensitive information identification model to optimize the sensitive information identification model to obtain a target sensitive information identification model; identify data in a production environment based on the target sensitive information identification model to obtain target sensitive information; desensitize the target sensitive information, and perform statistical analysis on the desensitized data; verify the statistical analysis results, and if the verification passes, perform data access record analysis to achieve database privacy protection.

[0016] A non-volatile computer storage medium provided in an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: analyze the data to be identified through a preset sensitive information identification model to obtain reference sensitive information; perform a secondary verification on the reference sensitive information, and feed the verification result back to the preset sensitive information identification model to optimize the sensitive information identification model to obtain a target sensitive information identification model; identify data in a production environment based on the target sensitive information identification model to obtain target sensitive information; desensitize the target sensitive information, and perform statistical analysis on the desensitized data; verify the statistical analysis results, and if the verification passes, perform data access record analysis to achieve database privacy protection.

[0017] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects: the embodiments of the present application analyze the data to be identified by presetting the sensitive information identification model to obtain reference sensitive information, and can understand complex language structures and contexts through the large model, so as to more accurately identify private data. And after a large amount of diversified data training, it can identify various types of private data. Secondly, the embodiments of the present application desensitize the target sensitive information and perform statistical analysis on the desensitized data to ensure that sensitive information is protected during the data processing process, while also obtaining valuable calculation results. Thereby effectively preventing unauthorized access and data leakage, and significantly improving data security and the overall reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0019] Figure 1 A flowchart of an intelligent privacy protection method for a database provided in an embodiment of the present application;

[0020] Figure 2 A schematic diagram of the structure of an intelligent privacy protection device for a database provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The embodiments of the present application provide a method, device and medium for intelligent privacy protection of a database.

[0022] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0023] The technical solution proposed in the embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0024] Figure 1 A flowchart of an intelligent privacy protection method for a database provided in an embodiment of the present application is as follows: Figure 1 As shown, the intelligent privacy protection method of the database includes the following steps:

[0025] S101. Analyze the data to be identified by using a preset sensitive information identification model to obtain reference sensitive information.

[0026] In one embodiment of the present application, the database data is preprocessed; wherein the preprocessing includes at least data cleaning and data labeling. The LLM large language model is used as a preset sensitive information recognition model, and the preprocessed database data is converted into an input format corresponding to the preset sensitive information recognition model. Based on the labeled database data and the supervised learning method, the preset sensitive information recognition model is fine-tuned so that the fine-tuned preset sensitive information recognition model extracts features of sensitive information.

[0027] Specifically, first clarify sensitive information, and according to the definition of sensitive information, set specific standards and thresholds for identifying sensitive information. Collect data from the database, clean, preprocess and annotate the collected data to ensure the quality and availability of the data. Use a large language model (LLM) as the basic model for sensitive information identification and classification, and represent the data containing sensitive information in the input format required by the model. Use annotated datasets to fine-tune the pre-trained model using supervised learning methods. According to specific tasks and data characteristics, adjust the model's parameters to optimize its performance in sensitive information identification tasks, so that it can learn the characteristics and patterns of sensitive information. Input the sensitive data to be identified into the fine-tuned LLM large language model.

[0028] Furthermore, sensitive information usually includes personal identity information, such as name, ID number, phone number, financial information, health data, business secrets, etc. If this information is improperly used or leaked, it may cause losses to individuals or organizations. According to business needs and data protection regulations, set specific sensitive information identification standards. For example, ID numbers must meet specific formats, and phone numbers must meet specific country / region code plus number combinations.

[0029] Furthermore, data sets that may contain sensitive information are extracted from the database to ensure data quality by removing duplicates, processing missing values, and correcting erroneous data. Text data is converted to a unified format, such as converting to lowercase, removing punctuation, and performing word segmentation, and sensitive information in the data is annotated to create an annotated data set.

[0030] Furthermore, a pre-trained LLM is selected as the base model, and the cleaned and pre-processed data is converted into an input format acceptable to the model. The pre-trained model is fine-tuned for supervised learning using a labeled dataset. By adjusting the model's parameters, such as learning rate, batch size, number of iterations, etc., the model can be made to more accurately identify sensitive information. For example, the model's ability to recognize digital sequences and specific formats can be optimized for the identification of ID card numbers. New, unlabeled data is input into the fine-tuned LLM model, and the model outputs the type of sensitive information contained in each piece of data and its location.

[0031] In one embodiment of the present application, the data to be identified is input into a fine-tuned preset sensitive information identification model, sensitive information features are extracted by the fine-tuned preset sensitive information identification model, and the identified sensitive information is output in a preset format; wherein the sensitive information at least includes the type of sensitive information and the location of the sensitive information. The fine-tuned preset sensitive information identification model is tested by a verification set, and if the test passes, it is determined that the output sensitive information is correct.

[0032] Specifically, the sensitive data to be identified is input into the fine-tuned LLM large language model. The LLM large language model performs feature extraction and other operations based on the input data to analyze and identify the sensitive information. The performance of the model is verified and monitored using the evaluation indicators on the validation set. The LLM large language model outputs the identified sensitive information in a specific format. The model outputs the sensitive information and its location range, marking the specific location or type of the sensitive information. The types of sensitive information include but are not limited to personal identity information, transaction information, financial information, contact information, etc.

[0033] S102: Perform a secondary check on the reference sensitive information, and feed the check result back to the preset sensitive information recognition model to optimize the sensitive information recognition model and obtain a target sensitive information recognition model.

[0034] In one embodiment of the present application, a second verification is performed on the reference sensitive information through a regular expression; wherein the second verification includes at least one of a format verification, a logic verification, and a data consistency verification. The error information corresponding to the second verification is determined, and the error information is added to the training set as new annotated data. The preset sensitive information recognition model is iteratively optimized by adding the training set with the new annotated data. Until the output recognition result of the sensitive information meets the preset requirements, the parameter combination corresponding to the preset sensitive information recognition model at this time is determined to obtain the target sensitive information recognition model.

[0035] Specifically, the sensitive information identified by the model is double-checked for format, logic or data consistency using regular expression matching to ensure the accuracy and reliability of the identification. The verification results are fed back to the model, and the model is iteratively trained and optimized to improve its accuracy and robustness in the sensitive information identification task. Specifically, the errors found after the regular expression verification are used as new labeled data and added to the training set or validation set. This process is equivalent to supplementing the model with "counterexamples" to help it better understand the error pattern. After adding the new labeled data, the model is retrained or fine-tuned. This training includes two parts: the original training data and the verification data generated based on feedback. The model will gradually correct previous errors in the new round of training and learn more accurate recognition patterns. Multiple iterations, the goal of each round of iteration is to reduce the false alarm rate and false negative rate of the model and improve the accuracy of sensitive information identification. Adjust the model's learning rate, batch size, regularization coefficient and other hyperparameters to find the best combination that can improve model performance. Expand the diversity of the training data set through data augmentation to make the model more robust and able to handle various complex text structures. Select some layers of the model to fine-tune, freeze the rest of the layers, and add a new linear layer to the last layer of the model to generate prediction results specific to the task of identifying sensitive information.

[0036] Specifically, define corresponding regular expression rules according to the type of sensitive information, such as ID number, phone number, email address, etc. Take the sensitive information identified by the model as input and use the defined regular expression for matching verification. If the match is successful, the sensitive information is considered to be in the correct format; if the match fails, it is considered to be an incorrect recognition. Record the verification results, especially those sensitive information that are incorrectly recognized by the model. The errors found after the regular expression verification are added as new annotated data to the training set or validation set. These new annotated data actually provide "counterexamples" for the model, which helps the model learn how to avoid similar mistakes.

[0037] Furthermore, the original training data is merged with the newly generated verification data to form a new training set. As needed, the model's hyperparameters such as learning rate, batch size, and regularization coefficient are adjusted to optimize the training process. Fine-tune the model using the new training set. You can choose to fine-tune some layers of the model (such as the last few layers) while freezing other layers to maintain the stability of the underlying features. Add a new linear layer to the last layer of the model to generate specific prediction results for the sensitive information identification task. During the fine-tuning process, the model learns new feature representations and tries to correct previous errors. Through multiple iterative training, the model will gradually improve its accuracy and robustness on the sensitive information identification task.

[0038] Furthermore, in order to further improve the generalization ability of the model, the diversity of the training data set can be expanded through data augmentation. Specifically, the training data can be randomly transformed, such as synonym replacement, sentence reorganization, noise addition, etc., to generate new training samples. According to the format and rules of sensitive information, synthetic data that meets the requirements can be automatically generated. For example, a program can be written to generate a large number of random ID numbers and phone numbers as training data.

[0039] S103. Based on the target sensitive information identification model, the data in the production environment is identified to obtain the target sensitive information.

[0040] In one embodiment of the present application, the trained model is deployed in the production environment to realize the real-time sensitive information identification and classification function of the data in the production environment. The embodiment of the present application designs appropriate interfaces and services for different application scenarios and requirements to support the user's identification and classification requirements for sensitive information. According to the identification results, the sensitive information is stored by category.

[0041] S104: Desensitize the target sensitive information and perform statistical analysis on the desensitized data.

[0042] In one embodiment of the present application, in a test environment, the desensitized data is inserted into a database for structural verification, wherein the structural verification includes at least one of field type verification, length consistency verification, field format verification, and database constraint verification. The desensitized data is formatted into structured data, and the abnormal values ​​and missing values ​​detected by the verification are processed.

[0043] Specifically, desensitize the identified sensitive information, replace sensitive information with non-real information or delete certain sensitive information fields. Ensure that the desensitized data still meets the structural requirements of the original data, keep the field type and length consistent, and follow the field format and database constraints. Write an automated script to verify the desensitized data to ensure that sensitive information has been effectively protected. Insert the desensitized data into the database in a test environment to observe whether there are any errors and ensure that it does not trigger constraint errors or format inconsistencies. Format the desensitized data into structured data to ensure data consistency and integrity, and handle missing values ​​and outliers.

[0044] Further, identify the fields that need to be desensitized, such as name, ID number, address, phone number, bank account, etc., and determine which data environments need to be desensitized. Replace sensitive information with fixed values, placeholders, or randomly generated but formatted data, retain some information, but hide key parts, such as using asterisks to mask the middle digits of a phone number, or directly delete fields or records containing sensitive information. Depending on the desensitization method selected, write SQL scripts, data migration tool scripts, or use specialized desensitization tools to process the data. Write automated scripts to check the desensitized data to ensure that sensitive information has been correctly replaced or deleted, verify that the data format and type are consistent with the original data, and ensure that database constraints are not violated. Format the desensitized data as structured data to ensure data consistency and integrity, and handle missing values ​​and outliers.

[0045] In one embodiment of the present application, a plurality of custom fusion algorithms are determined in the database. Through the plurality of custom fusion algorithms, the desensitized data is subjected to corresponding operation processing to achieve statistical analysis. Among them, the statistical analysis includes at least one of average value processing, variance processing, correlation processing, trend processing and outlier detection processing.

[0046] Specifically, the programming language supported by the database is used to integrate a variety of custom algorithms to perform specific operations on the desensitized data, so that it can automatically adjust parameters to perform a variety of statistical analysis tasks such as calculating mean values, variances, correlations, trends, and outlier detection.

[0047] Specifically, by integrating multiple custom algorithms, the integrated algorithms are used to operate on the desensitized data. For example, a custom linear regression algorithm can use the desensitized age and years of work to predict income. Although the income data is replaced with approximate values, the model can still identify trends and relationships between variables. Alternatively, through a custom unsupervised learning algorithm, assuming that the transaction data is desensitized, the algorithm is used to discover abnormal transaction behaviors of users, such as sudden large transactions or abnormally frequent operations.

[0048] S105. Verify the statistical analysis results, and if the verification is passed, perform data access record analysis to achieve database privacy protection.

[0049] In one embodiment of the present application, the statistical analysis results are input into a preset model. The statistical analysis results are compared with the original data for key statistical indicators through the preset model; wherein the key statistical indicators include at least one of the mean value, variance, and distribution form. And the classification accuracy corresponding to the desensitized data in the statistical analysis results is compared with the classification accuracy corresponding to the original data. And the abnormal data in the statistical analysis results is detected to verify the business logic corresponding to the desensitized data through the detected abnormal data.

[0050] Specifically, the calculation results are input into the big model for preliminary verification to ensure the accuracy and rationality of the processing. Ensure that the desensitized data is consistent with the original data in statistical characteristics so that the data is still practical during the query and analysis process.

[0051] Furthermore, statistical characteristics comparison: by calculating and comparing the key statistical indicators (such as mean, variance, distribution form, etc.) of the desensitized data and the original data, the rationality of the data processing can be verified. For example, if the income field of the desensitized data is close to the mean value of the original data, and the distribution curve of the data remains similar, it means that the desensitization processing is relatively reasonable.

[0052] Model performance comparison: Use the model's evaluation indicators to verify. For example, if the accuracy of the classification model on the desensitized data is not much different from that on the original data, it means that the desensitized data still retains enough information.

[0053] The big model can use its natural language understanding and reasoning capabilities, combined with the contextual information of the data, to verify the business logic. For example, if the data involves financial transaction records, the big model can infer whether certain transaction behaviors are reasonable and give feedback based on the business logic. Alternatively, if the big model detects that a user has an abnormally high transaction frequency in the desensitized data, it will prompt that the data may not conform to the expected business logic.

[0054] Furthermore, statistical characteristics refer to the basic statistical properties and laws of data. For example, the overall distribution of data should remain unchanged. Whether it is normal distribution, exponential distribution, uniform distribution, etc., the desensitized data should retain the distribution characteristics of the original data as much as possible. Or, when processing salary data, although sensitive individual salary information is desensitized, the average salary level of the entire data set should remain unchanged.

[0055] In one embodiment of the present application, different access rights levels are defined according to user roles. Access information of sensitive information is recorded to generate access logs; wherein the access information at least includes the access user, access time, and access operation content. The access logs are regularly checked for abnormalities, and when abnormal access behavior is determined, the user's access rights level is adjusted according to the detection results.

[0056] Specifically, define different levels of access rights based on user roles. Record the user, time, and operation content of each access to sensitive information for auditing and tracking purposes. Regularly analyze access logs to identify abnormal access behaviors. Manually review abnormal access behaviors to confirm whether there are security risks. Based on the review results, adjust the user's access rights to prevent further leakage of sensitive information.

[0057] Figure 2 This is a schematic diagram of the structure of an intelligent privacy protection device for a database provided in an embodiment of the present application. Figure 2 As shown, the intelligent privacy protection device for a database includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: analyze the data to be identified through a preset sensitive information identification model to obtain reference sensitive information; perform secondary verification on the reference sensitive information, and feed the verification result back to the preset sensitive information identification model to optimize the sensitive information identification model to obtain a target sensitive information identification model; identify data in a production environment based on the target sensitive information identification model to obtain target sensitive information; desensitize the target sensitive information, and perform statistical analysis on the desensitized data; verify the statistical analysis results, and if the verification passes, perform data access record analysis to achieve database privacy protection.

[0058] A non-volatile computer storage medium provided in an embodiment of the present application stores computer executable instructions, and the computer executable instructions are configured to: analyze the data to be identified through a preset sensitive information identification model to obtain reference sensitive information; perform a secondary verification on the reference sensitive information, and feed the verification result back to the preset sensitive information identification model to optimize the sensitive information identification model to obtain a target sensitive information identification model; identify data in a production environment based on the target sensitive information identification model to obtain target sensitive information; desensitize the target sensitive information, and perform statistical analysis on the desensitized data; verify the statistical analysis results, and if the verification passes, perform data access record analysis to achieve database privacy protection.

[0059] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0060] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various changes and variations. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A method for intelligent privacy protection of a database, characterized in that: The method comprises: The data to be identified is analyzed by using a preset sensitive information identification model to obtain reference sensitive information; Performing a secondary check on the reference sensitive information, and feeding back the check result to the preset sensitive information recognition model to optimize the sensitive information recognition model to obtain a target sensitive information recognition model; Based on the target sensitive information identification model, the data in the production environment is identified to obtain the target sensitive information; Desensitizing the target sensitive information and performing statistical analysis on the desensitized data; The statistical analysis results are verified, and if the verification is passed, data access record analysis is performed to achieve database privacy protection.

2. According to claim 1, the method for intelligent privacy protection of a database is characterized in that: Before analyzing the data to be identified by using the preset sensitive information identification model to obtain reference sensitive information, the method further includes: Preprocessing the database data; wherein the preprocessing at least includes data cleaning and data labeling; Using the LLM large language model as the preset sensitive information recognition model, and converting the preprocessed database data into an input format corresponding to the preset sensitive information recognition model; Based on the labeled database data and the supervised learning method, the preset sensitive information recognition model is fine-tuned so that the fine-tuned preset sensitive information recognition model extracts features of the sensitive information.

3. The intelligent privacy protection method for a database according to claim 1, characterized in that: The data to be identified is analyzed by using a preset sensitive information identification model to obtain reference sensitive information, specifically including: Inputting the data to be identified into the fine-tuned preset sensitive information identification model, extracting sensitive information features through the fine-tuned preset sensitive information identification model, and outputting the identified sensitive information in a preset format; wherein the sensitive information at least includes the type of sensitive information and the location of the sensitive information; The fine-tuned preset sensitive information recognition model is tested using a verification set, and if the test passes, it is determined that the output sensitive information is correct.

4. The intelligent privacy protection method for a database according to claim 1, characterized in that: The second verification of the reference sensitive information and feeding back the verification result to the preset sensitive information recognition model to optimize the sensitive information recognition model to obtain a target sensitive information recognition model specifically includes: Performing a secondary check on the reference sensitive information by using a regular expression; wherein the secondary check includes at least one of a format check, a logic check, and a data consistency check; Determine error information corresponding to the secondary verification, and add the error information as new annotation data to the training set; Iteratively optimizing the preset sensitive information recognition model by adding the training set after the new annotated data; Until the output sensitive information recognition result meets the preset requirements, the parameter combination corresponding to the preset sensitive information recognition model at this time is determined to obtain the target sensitive information recognition model.

5. The intelligent privacy protection method for a database according to claim 1, characterized in that: Before performing statistical analysis on the desensitized data, the method further comprises: In the test environment, the desensitized data is inserted into the database for structural verification; wherein the structural verification includes at least one of field type verification, length consistency verification, field format verification and database constraint verification; Format the desensitized data into structured data, and process the detected outliers and missing values.

6. The method for intelligent privacy protection of a database according to claim 1, characterized in that: The statistical analysis of the desensitized data specifically includes: Identify multiple custom fusion algorithms in the database; Through the multiple custom fusion algorithms, the desensitized data is processed accordingly to achieve the statistical analysis; The statistical analysis includes at least one of average value processing, variance processing, correlation processing, trend processing and outlier detection processing.

7. The method for intelligent privacy protection of a database according to claim 1, characterized in that: The verification of the statistical analysis results specifically includes: Input statistical analysis results into pre-set models; Comparing the statistical analysis results with the original data for key statistical indicators through the preset model; wherein the key statistical indicators include at least one of the mean value, variance and distribution form; and Comparing the classification accuracy rate corresponding to the desensitized data in the statistical analysis result with the classification accuracy rate corresponding to the original data; and The abnormal data in the statistical analysis results is detected to verify the business logic corresponding to the desensitized data through the detected abnormal data.

8. The method for intelligent privacy protection of a database according to claim 1, characterized in that: When the verification is passed, data access record analysis is performed, specifically including: Define different access rights levels based on user roles; Record access information of sensitive information and generate access logs; wherein the access information at least includes the access user, access time and access operation content; The access log is regularly tested for abnormalities, and when abnormal access behavior is determined, the user's access authority level is adjusted according to the test result.

9. An intelligent privacy protection device for a database, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 8.

Images