A multi-field data desensitization method, device, equipment and medium

Abstract

The application discloses a multi-field data desensitization method and device, equipment and medium, and relates to the technical field of data security. The method comprises the following steps: performing semantic feature extraction on the field name and field value of a target field by using a target model to obtain field name semantic features and field value semantic features, and fusing the field name semantic features and the field value semantic features to obtain a target semantic vector; determining the target association strength between the target semantic vectors corresponding to any two target fields, and constructing a cross-field semantic association matrix based on the target association strength; determining a target field combination meeting a preset high semantic correlation standard from a database based on the cross-field semantic association matrix, performing sensitivity scoring on the target field combination to obtain a sensitivity score, determining the data type corresponding to the target field combination by using a preset classification model, and performing desensitization processing on the target field combination according to the sensitivity score and the data type. The application improves the privacy protection level while ensuring data availability.

Description

Technical Field

[0001] This application relates to the field of data security technology, and in particular to a method, apparatus, device and medium for desensitizing multi-field data. Background Technology

[0002] Currently, database anonymization has become a necessary means of privacy protection. However, existing technologies have the following key shortcomings: 1. Blind spots in identifying combined sensitive data: Existing technologies mostly rely on static rules or single-field machine learning models, which cannot capture the sensitive semantics formed by field combinations, leading to privacy leaks. 2. Lack of semantic relationships between fields: Traditional methods treat fields as independent entities, ignoring the emergent risks generated by field combinations. For example, in medical data, a diagnosis code plus a birth date can pinpoint a specific patient's condition, while the accuracy of cross-field identification is low. 3. Imbalance between data availability and privacy protection: To simplify processing, existing anonymization strategies, such as full-field hashing and fixed masks, destroy the original data format and statistical distribution characteristics, which may directly cause the data to lose its business value and restrict the rational utilization of data assets. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, apparatus, device, and medium for de-identifying multi-field data, which achieves high-precision identification and differentiated de-identification of combined sensitive data, significantly ensuring data availability while improving privacy protection. The specific solution is as follows:

[0004] Firstly, this application provides a method for desensitizing multi-field data, including:

[0005] The semantic features of the field name and field value of the target field stored in the database are extracted using a pre-trained target model to obtain the corresponding field name semantic features and field value semantic features. The field name semantic features and field value semantic features are then fused to obtain the target semantic vector corresponding to the target field.

[0006] Determine the target association strength between the target semantic vectors corresponding to any two target fields, and construct a cross-field semantic association matrix based on the target association strength between the target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level;

[0007] Based on the cross-field semantic association matrix, target field combinations that meet the preset high semantic relevance standard are determined from the database. A preset sensitivity scoring model is used to score the sensitivity of the target field combinations to obtain the corresponding sensitivity scores. A preset classification model is used to determine the data type corresponding to the target field combinations.

[0008] The target field combination is desensitized based on the sensitivity score and the data type.

[0009] Optionally, the target model is a dual-channel model;

[0010] Accordingly, the step of using a pre-trained target model to extract semantic features from the field names and values ​​of the target fields stored in the database, to obtain corresponding semantic features of field names and semantic features of field values, includes:

[0011] The target model uses a preset language model to extract semantic features from the field names of the target fields stored in the database through the first channel, thereby obtaining semantic features of the field names that represent the meaning of the field names.

[0012] The second channel of the target model uses a preset sequence labeling model to extract semantic features from the field values ​​of the target field stored in the database, thereby obtaining semantic features of the field values ​​that represent the meaning of the field values.

[0013] Optionally, determining the target association strength between the target semantic vectors corresponding to any two target fields includes:

[0014] Determine the cosine similarity between the target semantic vectors corresponding to any two target fields, and multiply the cosine similarity by a preset weight to obtain the target association strength between the target semantic vectors corresponding to the two target fields.

[0015] Optionally, the step of using a preset sensitivity scoring model to perform sensitivity scoring on the target field combination to obtain the corresponding sensitivity score includes:

[0016] Determine the individual field sensitivity score corresponding to each field in the target field combination, and determine the field combination association strength corresponding to the target field combination based on the cross-field semantic association matrix;

[0017] The target sensitivity score corresponding to the target field combination is determined by a preset sensitivity scoring model based on the single-field sensitivity score, the pre-set single-field sensitivity score weights corresponding to each field in the target field combination, and the pre-set association strength amplification coefficient.

[0018] Wherein, if the target field combination contains more than two fields, determining the field combination association strength corresponding to the target field combination includes:

[0019] Determine the association strength between any two fields in the target field combination, and add the association strengths between any two fields in the target field combination to obtain the field combination association strength corresponding to the target field combination.

[0020] Optionally, determining the target sensitivity score corresponding to the target field combination using a preset sensitivity scoring model based on the single-field sensitivity score, the pre-set single-field sensitivity score weights corresponding to each field in the target field combination, and a pre-set association strength amplification coefficient includes:

[0021] The individual field sensitivity score of each field in the target field combination is multiplied by its corresponding individual field sensitivity score weight, and the multiplication results of each field are added together to obtain the first target result;

[0022] The second target result is obtained by multiplying the correlation strength of the field combination corresponding to the target field combination by the correlation strength amplification factor;

[0023] The first target result and the second target result are added together to obtain the target sensitivity score corresponding to the target field combination.

[0024] Optionally, determining the data type corresponding to the target field combination using a preset classification model includes:

[0025] A multilayer perceptron classifier is used to identify each field in the target field combination to determine the data type corresponding to each field in the target field combination; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include identity information, health information and financial information.

[0026] Optionally, the step of desensitizing the target field combination based on the sensitivity score and the data type includes:

[0027] For any field in the target field combination, a corresponding target desensitization algorithm is selected from the predefined desensitization strategies based on the sensitivity score and data type of the field, and the field is desensitized based on the target desensitization algorithm; the target desensitization algorithm includes format-preserving encryption, replacement algorithm and masking algorithm.

[0028] Secondly, this application provides a multi-field data desensitization device, comprising:

[0029] The semantic vector determination module is used to extract semantic features from the field names and field values ​​of the target field stored in the database using a pre-trained target model, to obtain the corresponding field name semantic features and field value semantic features, and to fuse the field name semantic features and field value semantic features to obtain the target semantic vector corresponding to the target field.

[0030] The matrix construction module is used to determine the target association strength between target semantic vectors corresponding to any two target fields, and to construct a cross-field semantic association matrix based on the target association strength between target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level;

[0031] The scoring and classification module is used to determine target field combinations that meet preset high semantic relevance standards from the database based on the cross-field semantic association matrix, to score the sensitivity of the target field combinations using a preset sensitivity scoring model to obtain corresponding sensitivity scores, and to determine the data type corresponding to the target field combinations using a preset classification model; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include personal identity and health information;

[0032] The desensitization module is used to perform desensitization processing on the target field combination based on the sensitivity score and the data type.

[0033] Thirdly, this application provides an electronic device, comprising:

[0034] Memory, used to store computer programs;

[0035] A processor is used to execute the computer program to implement the aforementioned multi-field data desensitization method.

[0036] Fourthly, this application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned multi-field data desensitization method.

[0037] In this application, a pre-trained target model is used to extract semantic features from the field names and values ​​of target fields stored in the database, obtaining corresponding semantic features of field names and semantic features of field values. These semantic features are then fused to obtain the target semantic vector corresponding to the target field. The target association strength between the target semantic vectors corresponding to any two target fields is determined, and a cross-field semantic association matrix is ​​constructed based on this strength. This cross-field semantic association matrix reflects the semantic relevance strength of different target fields. Based on the cross-field semantic association matrix, a combination of target fields that meets a preset high semantic relevance standard is determined from the database. A preset sensitivity scoring model is used to score the sensitivity of the target field combination, obtaining a corresponding sensitivity score. A preset classification model is used to determine the data type corresponding to the target field combination. Finally, the target field combination is de-identified based on the sensitivity score and the data type. As shown above, this application generates target semantic vectors by fusing semantic features of field names and field values ​​through a pre-trained model. It determines the target association strength between any two field vectors and constructs a cross-field semantic association matrix. Finally, based on the matrix, it selects field combinations that meet the "high semantic relevance" requirement. This process breaks the limitations of traditional single-field independent judgment, proactively identifying sensitive associations in field combinations and avoiding the blind spot in sensitive data identification. By fusing dual-dimensional semantic features of field names and values, it ensures that the target semantic vectors fully carry the business semantics of the fields. The cross-field semantic association matrix intuitively presents the association strength between target semantic vectors, making the semantic association between fields measurable and traceable. This effectively compensates for the lack of semantic association in traditional methods and improves the accuracy of cross-field sensitive identification. A preset sensitivity scoring model quantifies the risk of target field combinations, a preset classification model determines the data type, and finally, a desensitization algorithm is matched based on "score + type," avoiding a "one-size-fits-all" desensitization approach. This balances privacy and security while maximizing the preservation of data's business value. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0039] Figure 1 This is a flowchart of a multi-field data anonymization method disclosed in this application;

[0040] Figure 2This is a schematic diagram of a specific system architecture disclosed in this application;

[0041] Figure 3 This is a schematic diagram of the structure of a multi-field data desensitization device disclosed in this application;

[0042] Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0044] Currently, database anonymization has become an essential means of privacy protection. However, existing technologies suffer from key drawbacks such as blind spots in the identification of combined sensitive data, lack of semantic relationships between fields, and an imbalance between data availability and privacy protection. To address these shortcomings, this application provides a multi-field data anonymization method that achieves high-precision identification and differentiated anonymization of combined sensitive data, significantly ensuring data availability while improving privacy protection.

[0045] See Figure 1 As shown in the figure, this application discloses a method for desensitizing multi-field data, including:

[0046] Step S11: Use the pre-trained target model to extract semantic features from the field names and field values ​​of the target fields stored in the database, obtain the corresponding field name semantic features and field value semantic features, and fuse the field name semantic features and field value semantic features to obtain the target semantic vector corresponding to the target field.

[0047] In this embodiment, the target model is a dual-channel model. Accordingly, the pre-trained target model is used to extract semantic features from the field names and values ​​of the target fields stored in the database, resulting in corresponding field name semantic features and field value semantic features. This can include: using the first channel of the target model to extract semantic features from the field names of the target fields stored in the database using a preset language model, obtaining field name semantic features representing the meaning of the field names; and using the second channel of the target model to extract semantic features from the field values ​​of the target fields stored in the database using a preset sequence labeling model, obtaining field value semantic features representing the meaning of the field values.

[0048] It should be noted that in order to improve the semantic understanding ability of the field content in the database, a pre-trained language model can be used to extract semantic features for each target field, and the semantic features of the field name and the semantic features of the field value extracted are fused to obtain a high-dimensional semantic vector. In this way, not only can the semantic features of the field itself be captured, but also the potential sensitive semantic expressions in the field, such as semantic representations of key information types like "name", "diagnosis code", "address", etc., can be effectively identified. Among them, the database in this embodiment can be a medical industry database, a financial industry database, etc.

[0049] The semantic features of the field name and the semantic features of the field value can be concatenated using the following formula to obtain the target semantic vector:

[0050] ;

[0051] where field_name represents the field name and data_value represents the field value.

[0052] For example, # Input example: Medical table field combination

[0053] input_data =

[0054] ("patient_name", "Zhang Ming"),

[0055] ("diagnosis_code", "F33.1"),

[0056] ("address", "Some city, some district")

[0058] Use a two-channel pre-trained model to perform field semantic vectorization on the above medical table field combination:

[0059] Channel 1: Use BERT (Bidirectional Encoder Representation from Transformers, a pre-trained language model) to process the field name, such as identifying "diagnosis_code" as "diagnosis code".

[0060] Channel 2: Use BiLSTM (Bidirectional Long Short-Term Memory) and CRF (Conditional Random Fields) to process the field value, such as identifying "F33.1" as the depression code.

[0061] Output: 128-dimensional semantic vector.

[0062] ​Step S12: Determine the target association strength between the target semantic vectors corresponding to any two target fields, and construct a cross-field semantic association matrix based on the target association strength between the target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level.

[0063] In this embodiment, the cosine similarity between the target semantic vectors corresponding to any two target fields can be determined. The cosine similarity is then multiplied by a preset weight to obtain the target association strength between the target semantic vectors corresponding to any two target fields. A cross-field semantic association matrix can then be constructed based on the target association strength between the target semantic vectors corresponding to any two target fields.

[0064] It should be noted that the above process further introduces an attention mechanism, constructing a semantic association matrix between fields. This allows for the discovery of data combinations that, when viewed individually, may not be sensitive, but when combined, pose a high privacy risk. For example, "name + date of birth + address" constitutes a complete identity identifier. This modeling method breaks through the limitations of traditional methods that only focus on independent judgment of single fields, achieving more granular and realistic identification of sensitive information.

[0065] The target association strength between any two target semantic vectors can be determined using the following formula:

[0066] ;

[0067] Preset weights.

[0068] Step S13: Based on the cross-field semantic association matrix, determine the target field combination that meets the preset high semantic relevance standard from the database, use the preset sensitivity scoring model to score the sensitivity of the target field combination, obtain the corresponding sensitivity score, and use the preset classification model to determine the data type corresponding to the target field combination.

[0069] In this embodiment, each element in the cross-field semantic association matrix corresponds to the association strength of a pair of fields. The higher the association strength value, the stronger the semantic relevance of the pair of fields, and the higher the risk of privacy leakage after combination. Therefore, the cross-field semantic association matrix can be used to determine the target field combination that meets the preset high semantic relevance standard from the database.

[0070] Furthermore, the individual field sensitivity scores for each field in the target field combination can be determined, and the field combination association strength corresponding to the target field combination can be determined based on the cross-field semantic association matrix. Then, the target sensitivity score corresponding to the target field combination is determined using a preset sensitivity scoring model based on the individual field sensitivity scores, pre-set weights for the individual field sensitivity scores corresponding to each field in the target field combination, and a pre-set association strength amplification coefficient. If the target field combination contains more than two fields, the above determination of the field combination association strength corresponding to the target field combination can include: determining the association strength between any two fields in the target field combination, and adding the association strengths between any two fields in the target field combination to obtain the field combination association strength corresponding to the target field combination.

[0071] The process involves determining the target sensitivity score for a target field combination using a pre-set sensitivity scoring model. This determination is based on the individual field sensitivity scores, pre-set individual field sensitivity score weights corresponding to each field in the target field combination, and a pre-set association strength amplification coefficient. The steps include: first, multiplying the individual field sensitivity score of each field in the target field combination by its corresponding individual field sensitivity score weight, and then summing the results to obtain a first target result. Next, multiplying the association strength of the target field combination by the association strength amplification coefficient to obtain a second target result. Finally, adding the first and second target results yields the target sensitivity score for the target field combination.

[0072] For example, the target sensitivity score corresponding to the combination of target fields can be determined using the following formula:

[0073] ;

[0074] in, The sensitivity score for a single field can be calculated using a CNN (Convolutional Neural Network) classifier, with a value ranging from 0 to 1. , Weighting for single-field sensitive scores; This is the correlation strength amplification factor, with a default value of 1.5; To determine the strength of the association between fields.

[0075] Furthermore, a multilayer perceptron (MLP) classifier can be used to identify each field in the target field combination to determine the data type corresponding to each field in the target field combination, thereby achieving accurate data classification. The data type is a category based on the privacy attributes of the data content, and the privacy attributes include, but are not limited to, identity information, health information, and financial information.

[0076] Step S14: Desensitize the target field combination according to the sensitivity score and the data type.

[0077] In this embodiment, for any field in the target field combination, a corresponding target desensitization algorithm can be selected from the predefined desensitization strategy according to the sensitivity score and data type of any field, and the target desensitization algorithm is used to desensitize any field; wherein, the target desensitization algorithm includes, but is not limited to, format preservation encryption, replacement algorithm and masking algorithm.

[0078] It should be noted that the de-identification strategy allows users to set different de-identification algorithms based on different types of data. When matching de-identification algorithms, the user-configured de-identification algorithm will be used first.

[0079] See Figure 2 As shown, the technical solution in this embodiment will be explained below using a specific system architecture as an example.

[0080] First, a communication connection with the database is established to provide a channel for subsequent data operations and ensure access to the database. Metadata is extracted from the database, containing structural information such as field names, data types, and table relationships, providing foundational information for subsequent operations such as field semantic analysis. Then, a pre-trained model is used to convert field content into semantic vectors, and a cross-field semantic association matrix is ​​constructed based on these vectors to achieve cross-field association modeling. Further, a collaborative sensitivity identification engine is used to perform sensitivity scoring and data type classification on target field combinations identified from the database that meet preset high semantic relevance criteria. This allows for desensitization of target field combinations based on sensitivity scores and data types via a management console. The management console includes:

[0081] Strategy configuration module: Used to configure de-identification strategies, such as de-identification rules for different types of data.

[0082] Dynamic data anonymization decision module: Based on sensitivity scores, data types, and strategy configurations, it dynamically determines which data anonymization method to use.

[0083] Key Management Module: Manages the keys used for encryption and other operations, ensuring key security to guarantee the security of encryption-related operations such as de-identification.

[0084] Monitoring and auditing module: Monitors and audits the entire de-identification process, records the operation process, facilitates subsequent viewing and traceability, and ensures the compliance of the process.

[0085] Based on the results of dynamic desensitization decision-making, a suitable strategy can be selected from a variety of desensitization methods:

[0086] Format-preserving encryption: When encrypting data, the original format of the data is preserved, so that the encrypted data can still be used in business scenarios that depend on the format.

[0087] Generalization: This involves summarizing and abstracting data to reduce its granularity and thus protect privacy, such as generalizing a specific age into an age range.

[0088] Suppression and Replacement: Directly suppress sensitive data, such as by not displaying it or replacing it with other data, to achieve the purpose of desensitization.

[0089] Finally, the anonymized data is output for subsequent business use.

[0090] As shown above, this embodiment improves the accuracy of sensitive information identification through semantic vectorization and cross-field association modeling. Through strategy configuration and dynamic desensitization decision-making, different desensitization strategies can be flexibly selected for different sensitivity scores and data types, maximizing data value while ensuring security. This achieves a multi-level desensitization strategy, balancing data usability and security. The use of a lightweight model based on an attention mechanism significantly improves the speed of operations such as semantic processing of database fields.

[0091] See Figure 3 As shown in the figure, this application also discloses a multi-field data desensitization device, including:

[0092] The semantic vector determination module 11 is used to extract semantic features from the field name and field value of the target field stored in the database using a pre-trained target model, to obtain the corresponding field name semantic features and field value semantic features, and to fuse the field name semantic features and the field value semantic features to obtain the target semantic vector corresponding to the target field.

[0093] The matrix construction module 12 is used to determine the target association strength between target semantic vectors corresponding to any two target fields, and to construct a cross-field semantic association matrix based on the target association strength between target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level.

[0094] The scoring and classification module 13 is used to determine the target field combination that meets the preset high semantic relevance standard from the database based on the cross-field semantic association matrix, to score the sensitivity of the target field combination using a preset sensitivity scoring model to obtain the corresponding sensitivity score, and to determine the data type corresponding to the target field combination using a preset classification model; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include personal identity and health information;

[0095] The desensitization module 14 is used to perform desensitization processing on the target field combination based on the sensitivity score and the data type.

[0096] In some specific implementations, the target model is a dual-channel model;

[0097] Accordingly, the semantic vector determination module 11 includes:

[0098] The first feature extraction unit is used to extract semantic features from the field names of the target fields stored in the database through the first channel of the target model using a preset language model, so as to obtain the semantic features of the field names that represent the meaning of the field names.

[0099] The second feature extraction unit is used to extract semantic features from the field values ​​of the target field stored in the database through the second channel of the target model using a preset sequence labeling model, so as to obtain the semantic features of the field values ​​that represent the meaning of the field values.

[0100] In some specific embodiments, the matrix construction module 12 includes:

[0101] The first association strength determination unit is used to determine the cosine similarity between the target semantic vectors corresponding to any two target fields, and multiply the cosine similarity by a preset weight to obtain the target association strength between the target semantic vectors corresponding to any two target fields.

[0102] In some specific embodiments, the scoring and classification module 13 includes:

[0103] The association strength determination submodule is used to determine the single-field sensitivity score corresponding to each field in the target field combination, and to determine the field combination association strength corresponding to the target field combination based on the cross-field semantic association matrix.

[0104] The sensitivity scoring submodule is used to determine the target sensitivity score corresponding to the target field combination based on the single-field sensitivity score, the pre-set single-field sensitivity score weights corresponding to each field in the target field combination, and the pre-set association strength amplification coefficient through a preset sensitivity scoring model.

[0105] Wherein, if the target field combination contains more than two fields, the second association strength determination submodule includes:

[0106] The second association strength determination unit is used to determine the association strength between any two fields in the target field combination, and to add the association strengths between any two fields in the target field combination to obtain the field combination association strength corresponding to the target field combination.

[0107] In some specific embodiments, the sensitivity scoring submodule includes:

[0108] The first target result determination unit is used to multiply the single-field sensitivity score of each field in the target field combination by the corresponding single-field sensitivity score weight, and add the multiplication results corresponding to each field to obtain the first target result;

[0109] The second target result determination unit is used to multiply the correlation strength of the field combination corresponding to the target field combination with the correlation strength amplification factor to obtain the second target result;

[0110] The sensitivity score determination unit is used to add the first target result and the second target result to obtain the target sensitivity score corresponding to the target field combination.

[0111] In some specific embodiments, the scoring and classification module 13 includes:

[0112] The type identification unit is used to identify each field in the target field combination using a multilayer perceptron classifier to determine the data type corresponding to each field in the target field combination; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include identity information, health information and financial information.

[0113] In some specific embodiments, the desensitization processing module 14 includes:

[0114] The data desensitization unit is used to select a corresponding target desensitization algorithm from a predefined desensitization strategy based on the sensitivity score and data type of any field in the target field combination, and to perform desensitization processing on any field based on the target desensitization algorithm; the target desensitization algorithm includes format-preserving encryption, replacement algorithm and masking algorithm.

[0115] Furthermore, embodiments of this application also disclose an electronic device, Figure 4 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0116] Figure 4This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the multi-field data desensitization method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0117] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0118] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0119] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the multi-field data desensitization method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0120] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned multi-field data desensitization method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0121] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0122] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0124] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0125] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for desensitizing multi-field data, characterized in that, include: The semantic features of the field name and field value of the target field stored in the database are extracted using a pre-trained target model to obtain the corresponding field name semantic features and field value semantic features. The field name semantic features and field value semantic features are then fused to obtain the target semantic vector corresponding to the target field. Determine the target association strength between the target semantic vectors corresponding to any two target fields, and construct a cross-field semantic association matrix based on the target association strength between the target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level; Based on the cross-field semantic association matrix, target field combinations that meet the preset high semantic relevance standard are determined from the database. A preset sensitivity scoring model is used to score the sensitivity of the target field combinations to obtain the corresponding sensitivity scores. A preset classification model is used to determine the data type corresponding to the target field combinations. The target field combination is desensitized based on the sensitivity score and the data type.

2. The multi-field data anonymization method according to claim 1, characterized in that, The target model is a dual-channel model; Accordingly, the step of using a pre-trained target model to extract semantic features from the field names and values ​​of the target fields stored in the database, to obtain corresponding semantic features of field names and semantic features of field values, includes: The target model uses a preset language model to extract semantic features from the field names of the target fields stored in the database through the first channel, thereby obtaining semantic features of the field names that represent the meaning of the field names. The second channel of the target model uses a preset sequence labeling model to extract semantic features from the field values ​​of the target field stored in the database, thereby obtaining semantic features of the field values ​​that represent the meaning of the field values.

3. The multi-field data anonymization method according to claim 1, characterized in that, Determining the target association strength between target semantic vectors corresponding to any two target fields includes: Determine the cosine similarity between the target semantic vectors corresponding to any two target fields, and multiply the cosine similarity by a preset weight to obtain the target association strength between the target semantic vectors corresponding to the two target fields.

4. The multi-field data anonymization method according to claim 1, characterized in that, The step of using a preset sensitivity scoring model to perform sensitivity scoring on the target field combination to obtain the corresponding sensitivity score includes: Determine the individual field sensitivity score corresponding to each field in the target field combination, and determine the field combination association strength corresponding to the target field combination based on the cross-field semantic association matrix; The target sensitivity score corresponding to the target field combination is determined by a preset sensitivity scoring model based on the single-field sensitivity score, the pre-set single-field sensitivity score weights corresponding to each field in the target field combination, and the pre-set association strength amplification coefficient. Wherein, if the target field combination contains more than two fields, determining the field combination association strength corresponding to the target field combination includes: Determine the association strength between any two fields in the target field combination, and add the association strengths between any two fields in the target field combination to obtain the field combination association strength corresponding to the target field combination.

5. The multi-field data desensitization method according to claim 4, characterized in that, The step of determining the target sensitivity score corresponding to the target field combination by means of a preset sensitivity scoring model based on the single-field sensitivity score, the pre-set single-field sensitivity score weights corresponding to each field in the target field combination, and a pre-set association strength amplification coefficient includes: The individual field sensitivity score of each field in the target field combination is multiplied by its corresponding individual field sensitivity score weight, and the multiplication results of each field are added together to obtain the first target result; The second target result is obtained by multiplying the correlation strength of the field combination corresponding to the target field combination by the correlation strength amplification factor; The first target result and the second target result are added together to obtain the target sensitivity score corresponding to the target field combination.

6. The multi-field data desensitization method according to claim 1, characterized in that, The step of determining the data type corresponding to the target field combination using a preset classification model includes: A multilayer perceptron classifier is used to identify each field in the target field combination to determine the data type corresponding to each field in the target field combination; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include identity information, health information and financial information.

7. The multi-field data desensitization method according to any one of claims 1 to 6, characterized in that, The process of desensitizing the target field combination based on the sensitivity score and the data type includes: For any field in the target field combination, a corresponding target desensitization algorithm is selected from the predefined desensitization strategies based on the sensitivity score and data type of the field, and the field is desensitized based on the target desensitization algorithm; the target desensitization algorithm includes format-preserving encryption, replacement algorithm and masking algorithm.

8. A multi-field data desensitization device, characterized in that, include: The semantic vector determination module is used to extract semantic features from the field names and field values ​​of the target field stored in the database using a pre-trained target model, to obtain the corresponding field name semantic features and field value semantic features, and to fuse the field name semantic features and field value semantic features to obtain the target semantic vector corresponding to the target field. The matrix construction module is used to determine the target association strength between target semantic vectors corresponding to any two target fields, and to construct a cross-field semantic association matrix based on the target association strength between target semantic vectors corresponding to any two target fields; the cross-field semantic association matrix is ​​used to reflect the correlation strength of different target fields at the semantic level; The scoring and classification module is used to determine target field combinations that meet preset high semantic relevance standards from the database based on the cross-field semantic association matrix, to score the sensitivity of the target field combinations using a preset sensitivity scoring model to obtain corresponding sensitivity scores, and to determine the data type corresponding to the target field combinations using a preset classification model; the data type is a category based on the privacy attributes of the data content, and the privacy attributes include personal identity and health information; The desensitization module is used to perform desensitization processing on the target field combination based on the sensitivity score and the data type.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the multi-field data desensitization method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs, which, when executed by a processor, implement the multi-field data desensitization method as described in any one of claims 1 to 7.

Images