Data privacy protection method and system based on data annotation

Abstract

The invention provides a data privacy protection method and system based on data annotation, and the method comprises the steps: carrying out the data form reconstruction operation of a talent data set, obtaining a semantic data unit set, carrying out the feature cluster division operation of the semantic data unit set, generating a feature cluster set, and carrying out the data privacy protection based on the semantic density value and field association strength value of each feature cluster. Calculating a noise type and a noise parameter corresponding to each feature cluster by combining sensitive attribute tags of the feature clusters to obtain a feature cluster noise parameter comparison table, and mapping the semantic data unit set into the feature cluster noise parameter comparison table according to a feature cluster attribution relation; performing corresponding noise generation and injection on feature cluster data fields of different sensitive attribute tags to generate a disturbance data set, performing correlation degree evaluation on the disturbance data set, and judging whether a preset correlation maintaining threshold value is met or not by calculating a correlation degree change rate between fields in a semantic data unit before and after disturbance; and if yes, outputting a privacy protection data set for labeling the task. According to the method, the privacy protection accuracy and the data availability are improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a data privacy protection method and system based on data annotation. Background Technology

[0002] With the development of artificial intelligence technology, data annotation technology has been widely used in model training. This technology generates training samples with labeled information by annotating raw data, providing data support for model learning. However, how to protect the privacy of data containing sensitive information and avoid information leakage during the annotation process has become a key issue in the field of data processing. Common data privacy protection techniques based on data annotation usually isolate sensitive fields in the raw data, such as perturbing individual sensitive fields by replacement, generalization, or adding noise to reduce the risk of information leakage. However, this approach can easily overlook the inherent relationships and semantic context information between data fields, which may cause the processed data to lose its original semantic coherence, affecting the accuracy of the data annotation task. At the same time, due to the lack of full consideration of the overall semantic structure of the data, it may be impossible to fully identify potential combinations of sensitive information, resulting in incomplete privacy protection. Summary of the Invention

[0003] This invention provides a data privacy protection method and system based on data annotation.

[0004] In a first aspect, embodiments of the present invention provide a data privacy protection method based on data annotation. The method includes: performing data morphology reconstruction on a talent dataset; constructing a semantic association graph by parsing the semantic description information of each data field; identifying core data fields and auxiliary data fields based on the connection strength of the fields in the association graph; segmenting the original data records based on the semantic coverage of the core data fields to obtain a semantic data unit set, where each unit in the semantic data unit set contains multiple data fields with semantic associations; performing feature clustering on the semantic data unit set; constructing a semantic association matrix by calculating the semantic similarity between units; performing clustering on the semantic association matrix to generate an initial feature cluster set; and performing sensitive attribute matching on each initial feature cluster based on preset sensitive attribute identification rules to generate a set containing sensitive attributes. A set of feature clusters with sensitive attribute labels is generated. Based on the semantic density value and field association strength value of each feature cluster in the set of feature clusters containing sensitive attribute labels, the noise type and noise parameters corresponding to each feature cluster are calculated in combination with the sensitive attribute labels of the feature clusters, resulting in a feature cluster noise parameter lookup table. The set of semantic data units is mapped to the feature cluster noise parameter lookup table according to the feature cluster affiliation relationship. Corresponding noise generation and injection operations are performed on the data fields of feature clusters with different sensitive attribute labels. The noise distribution characteristics between data fields in the same cluster are maintained by controlling the random seed synchronization mechanism in the noise generation process, generating a perturbed data set. The association degree of the perturbed data set is evaluated. The change rate of association degree between fields in the semantic data unit before and after perturbation is calculated to determine whether the preset association preservation threshold is met. When the threshold requirement is met, a privacy-preserving dataset that can be used for annotation tasks is output.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory for storing computer-executable instructions or computer programs; and a processor for implementing the data privacy protection method based on data annotation as described above when executing the computer-executable instructions or computer programs stored in the memory.

[0006] The embodiments of this application have the following beneficial effects: This invention reconstructs the data form of talent datasets, builds a semantic association graph of fields, identifies core and auxiliary data fields, and divides them into semantic data unit sets. This transforms the data from a physical record form to a semantically structured form, avoiding over- or under-protection caused by traditional isolated field processing. By calculating the semantic similarity between units to construct a semantic association matrix and performing clustering, and combining it with sensitive attribute identification rules to generate a feature cluster set containing sensitive attribute labels, this invention upgrades from isolated field-level identification to semantic cluster-level association identification. It can capture implicit sensitive semantics formed by field combinations and improve the comprehensiveness of sensitive attribute identification. Noise types and parameters are calculated based on the semantic density value of feature clusters, field association strength value, and sensitive attribute labels to achieve differentiated noise customization. This ensures that noise parameters are accurately matched with the semantic characteristics of clusters, balancing privacy protection and data usability. A random seed synchronization mechanism maintains the noise distribution characteristics among data fields in the same cluster, preserving the semantic association structure within data units during noise injection and preventing independent noise from disrupting the intrinsic associations between fields. By calculating the rate of change in association degree between fields before and after perturbation and verifying whether it meets a preset threshold, the output privacy-preserving dataset can be directly used for annotation tasks, improving the usability of the data in annotation scenarios. This method effectively improves the accuracy of privacy protection and data usability in data annotation scenarios through the synergistic effects of semantic reconstruction, association cluster partitioning, dynamic noise customization, collaborative noise injection, and association degree verification. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the data privacy protection method based on data annotation provided in this application embodiment. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0009] The execution entity of this invention is a computer system. The computer system can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The computer system 200 can also be a laptop, desktop computer, or other device with data processing capabilities, but is not limited to these.

[0010] The following describes a computer system that implements the data privacy protection method based on data annotation provided in the embodiments of this application. See also Figure 1 This is a schematic diagram of the structure of the computer system provided in the embodiments of this application. Figure 1 The computer system 200 shown includes at least one processor 210, a memory 250, at least one network interface 220, and an external interface 230. The various components in the computer system 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 240.

[0011] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0012] External interface 230 may include, for example, one or more speakers and / or one or more visual displays. External interface 230 may also include one or more input devices 432, such as a keyboard, mouse, microphone, touch screen display, camera, etc.

[0013] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0014] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0015] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0016] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 252 is used to reach a device (such as a database server) that needs to communicate via one or more (wired or wireless) network interfaces 220. Exemplary network interfaces 220 include Bluetooth, WiFi, and Universal Serial Bus (USB), etc. Presentation module 253 is configured to enable the display of information (e.g., external interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., display screen, speaker, etc.) associated with external interface 230; The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0017] See Figure 2 This is a flowchart illustrating the data privacy protection method based on data annotation provided in this application embodiment.

[0018] Step S100: Perform data form reconstruction on the talent dataset. Construct a semantic association graph of fields by parsing the semantic description information of each data field. Identify core data fields and auxiliary data fields based on the connection strength of fields in the association graph. Divide the original data records into blocks based on the semantic coverage of the core data fields to obtain a semantic data unit set. Each unit in the semantic data unit set contains multiple data fields with semantic association.

[0019] The following examples all use talent big data as an example. In this application scenario, the talent dataset covers various aspects of data, such as personal basic information, educational background, work experience, and skills. Data format reconstruction aims to reorganize and structure the original, disorganized data, and uncover the semantic relationships behind it. Semantic description information provides a more detailed explanation and description of each data field, including its meaning, purpose, and contextual information. A field semantic association graph is a graph structure where nodes represent data fields, directed edges represent the direction of semantic association between fields, and edge weights reflect the strength of the association.

[0020] Core data fields are those that play a crucial role in the dataset and are indispensable for understanding and analyzing the overall business. For example, in a talent recruitment scenario, the "job title" field is a core data field, closely related to other fields such as "skill requirements" and "salary and benefits." Auxiliary data fields, on the other hand, supplement and support the core data fields. For instance, the "work location" field can help understand information related to the "job title."

[0021] Semantic coverage refers to the scope of semantic information involved in a core data field. Taking the "job title" field as an example, its semantic coverage might include information such as the job responsibilities, required skills, and career development path. Dividing the original data record into blocks based on the semantic coverage of the core data fields involves dividing the original data according to the semantic integrity of the core data fields, forming semantically defined data units with complete semantic contexts. For example, a talent data record containing information on multiple job titles can be divided into multiple semantically defined data units based on the semantic coverage of each job title. Each unit contains complete information related to that job title, such as job name, responsibilities, skill requirements, and salary.

[0022] In one implementation, step S100 may specifically include the following steps S110-S160: Step S110: Extract semantic metadata for each data field in the talent dataset, parse the word structure of the field name, the description text of the data type, and the explanatory paragraph of the business scenario, and generate a set of semantic description information for the field. The set of semantic description information for the field includes the terminology, contextual dependencies, and business affiliation features of each data field.

[0023] Semantic metadata extraction is the process of mining the inherent semantic information from data fields. Analyzing the word structure of field names helps in understanding their core meaning. For example, the "Work Experience" field, with its combination of "work" and "experience," clearly indicates that it relates to past work-related experiences. Descriptive text about data types helps understand how the field is stored and represented; for instance, the "Age" field's data type description is an integer, indicating that it stores an integer value. The business scenario description paragraph explains the data field's purpose and significance in actual business operations. For example, the "Project Outcomes" field, in a company's talent performance evaluation scenario, is used to measure an employee's contribution and ability in a project.

[0024] Terminology composition refers to the key terms contained in a data field, which are important elements for understanding the field's semantics. For example, the terminology composition of the "Education Background" field includes "education" and "background." Contextual dependency reflects the relationship between this data field and other data fields; for example, the "Salary" field may depend on fields such as "Position" and "Work Experience." Business affiliation dimensions indicate the business scope to which the data field belongs; for example, the "Training Records" field belongs to the talent development business dimension.

[0025] By parsing and extracting the word structure, data type description, and business scenario description of each data field, a detailed set of semantic description information for each field can be generated. For example, for the "skill certificate" field, analyzing its word structure reveals a connection to skills and certificates; the data type is likely text, used to record the certificate name; and the business scenario description indicates that it is used in talent skills assessment to demonstrate the professional skills possessed by talent. The resulting semantic description information includes terminology ("skills," "certificates"), contextual dependencies (depending on the "skill requirements" field), and business attribution dimensions (talent skills information dimension).

[0026] Step S120: Perform semantic disambiguation on the terms in the field semantic description information set. By comparing the explanations of polysemous terms in the domain terminology library and combining the business scenario description of the data field, determine the unique semantic reference and generate the disambiguated field semantic feature vector.

[0027] A domain terminology repository is a knowledge base that stores terms within a specific domain and their precise explanations. In the field of talent big data, the domain terminology repository contains detailed explanations of terms such as "recruitment process," "performance evaluation," and "career development planning." Polysemous terms are terms that have multiple meanings in different contexts; for example, "experience" has different meanings in "work experience" and "life experience."

[0028] The purpose of semantic disambiguation is to eliminate the ambiguity of these polysemous terms and determine their accurate semantics in a specific data field. For example, in the "Project Experience" field of talent data, the domain terminology library shows that "experience" here should be understood as the skills and knowledge accumulated during project implementation. Combined with the business scenario description (used to assess the talent's capabilities in the project), its unique semantic meaning can be clearly defined.

[0029] Generating a disambiguated semantic feature vector for a field involves quantifying the processed semantic information. This can be achieved using a vector space model, mapping each term to a point in a high-dimensional vector space, where each dimension represents a different semantic feature. For example, for the field "project experience," after semantic disambiguation, the elements in the generated semantic feature vector might represent semantic information related to project management skills, technical application skills, teamwork skills, etc.

[0030] Step S130: Based on the disambiguated field semantic feature vectors, identify the co-occurrence locations and contextual overlap areas of different data fields in the business description text, count the co-occurrence frequency and calculate the proportion of contextual overlap length, and generate semantic correlation values ​​between fields.

[0031] Business description text is text that provides a detailed description of business processes, rules, etc. In the field of talent big data, this could be a company's talent management manual, recruitment documents, etc. Co-occurrence locations refer to the positions where different data fields appear simultaneously in the business description text. For example, the fields "job title" and "skill requirements" may frequently appear together in a recruitment document. Contextual overlap areas refer to the overlapping portions of the contextual text containing these co-occurring fields. For instance, in the text describing "job title" and "skill requirements," the descriptions of the job duties and required skills constitute the contextual overlap area.

[0032] Co-occurrence frequency records the number of times different data fields appear together in the business description text. A higher co-occurrence frequency indicates a potentially stronger semantic relationship between the two fields. Context overlap length ratio refers to the ratio of the length of the context overlap area to the total length of the context for both fields. A larger ratio indicates a stronger semantic relationship between the fields. Generating a semantic relationship value between fields can comprehensively consider both co-occurrence frequency and context overlap length ratio. For example, a weighted sum can be calculated for co-occurrence frequency and context overlap length ratio, with the weights determined based on specific business needs and data characteristics.

[0033] In practical applications, in order to overcome the problem of measuring "co-occurrence frequency" and "context overlap length ratio" with different dimensions and scales, those skilled in the art can, based on their basic knowledge, separately count the co-occurrence frequency and calculate the context overlap length ratio, normalize the co-occurrence frequency to obtain a co-occurrence intensity score, and then normalize the context overlap length ratio to obtain an overlap intensity score. The co-occurrence intensity score and the overlap intensity score are then weighted and summed to generate a semantic correlation value between fields, thereby overcoming this problem.

[0034] Step S140: Construct a directed weighted semantic association graph based on the semantic association degree values ​​between fields. The graph nodes are data fields, the directed edges represent the semantic association direction, and the edge weights are the semantic association degree values. Perform weight decay processing on the cyclic association paths in the graph.

[0035] A directed weighted semantic relationship graph is a graph structure used to represent the semantic relationships between data fields. Each data field is a node, and each node represents a data field. Directed edges indicate the direction of the semantic relationship between fields; for example, a directed edge from "educational background" to "job title" indicates that educational background affects job attainment. The weight of each edge represents the semantic relationship degree, reflecting the tightness of the semantic relationship between the two fields. A cyclic relationship path refers to a closed-loop relationship path in the graph, such as "A->B->C->A". Because cyclic relationship paths can lead to information duplication and redundancy, affecting the clarity of the graph and the accuracy of the analysis, weight attenuation is necessary. By reducing the weight of edges in cyclic paths, more meaningful relationships can be highlighted, allowing the graph to more accurately reflect the true semantic relationships between data fields.

[0036] In one implementation, step S140 may specifically include the following steps S141-S146: Step S141: Normalize the semantic correlation values ​​between fields, mapping the numerical range to a preset interval. The normalization process preserves the relative size relationship of the original correlation values.

[0037] Normalization is used to unify semantic relevance values ​​from different ranges into a preset interval for easier comparison and calculation. The preset interval can be set according to specific needs, such as [0,1]. During normalization, it's crucial to maintain the relative magnitudes of the original relevance values; that is, field pairs with high original relevance values ​​should retain relatively high values ​​after normalization. For example, a linear transformation method can be used for normalization. First, determine the minimum and maximum values ​​of the original semantic relevance values. Then, based on the preset interval, map each original relevance value to that interval through certain calculations. For instance, for the original relevance value of a field pair, convert it to its corresponding value within the preset interval based on its relative position within the original value range.

[0038] Step S142: Initialize the graph structure with data fields as nodes, add field identifiers, data types and business affiliation tags to each node, and establish a node attribute table to store basic node information.

[0039] Initializing the graph structure using data fields as nodes means treating each data field in the talent dataset as a node in the graph. A field identifier is a unique number or name used to identify each data field; for example, "F001" can represent the "Name" field. Data type refers to the data storage type of the field, such as "string," "integer," or "date." Business affiliation tag indicates the business category to which the field belongs, such as "Personal Information," "Occupational Information," or "Educational Information."

[0040] A node attribute table is a data table used to store basic information about each node. It can adopt a relational database table structure, containing columns such as field identifiers, data types, and business affiliation tags. By creating a node attribute table, the basic information of each node can be easily managed and queried. For example, for the "Job Start Date" field, the corresponding record in the node attribute table might have: field identifier "F003", data type "Date Type", and business affiliation tag "Occupation Information".

[0041] Step S143: Construct a set of directed edges based on the semantic association direction. Each directed edge contains a start node identifier, an end node identifier, and a normalized semantic association degree value. Generate an edge relationship table to record all association information.

[0042] Semantic association direction refers to the direction of the semantic association between two data fields, reflecting the flow and dependency of data at the semantic level. Constructing a set of directed edges based on semantic association direction involves connecting semantically related field pairs and determining the direction of the association.

[0043] Each directed edge includes a start node identifier, an end node identifier, and a normalized semantic correlation value. The start and end node identifiers identify the beginning and end points of the directed edge, respectively, while the normalized semantic correlation value reflects the strength of the semantic relationship between the two fields. The edge relationship table is a data table used to record information about all directed edges. It can adopt a relational database table structure and includes columns such as start node identifier, end node identifier, and normalized semantic correlation value. For example, a directed edge from "Skill Certificate" to "Job" has a start node identifier of "F004" (representing the "Skill Certificate" field), an end node identifier of "F005" (representing the "Job" field), and a relatively high normalized semantic correlation value. The edge relationship table will record relevant information about this edge.

[0044] Step S144: Verify the direction of the directed edges in the edge relationship table. By comparing the data flow direction in the business process, correct the association direction that does not match the actual business logic, so that the direction of the directed edges is consistent with the actual order of data use.

[0045] In a business process, the direction of data flow refers to the direction in which data moves from one data field to another during actual business operations. For example, in a talent recruitment process, the job position is usually determined first, and then candidates are screened based on their skills certificates, educational background, and other information according to the job requirements. Therefore, the direction of data flow is from "job position" to fields such as "skills certificates" and "educational background".

[0046] Verifying the direction of directed edges in the edge relationship table involves checking whether the direction of each directed edge aligns with the actual business logic. If the direction of a directed edge is found to be inconsistent with the actual data usage order, it needs to be corrected. For example, the edge relationship table might initially record a directed edge from "Skill Certificate" to "Job Title," but a comparison with the business process reveals that the actual data flow is from "Job Title" to "Skill Certificate," in which case the direction of this directed edge needs to be adjusted. This method ensures that the direction of directed edges accurately reflects the actual data usage in the business, improving the accuracy and usability of the graph.

[0047] Step S145: Detect cyclic association paths in the graph, identify cyclic paths with a length greater than a preset threshold by depth-first search, and perform attenuation processing on the weight of each directed edge in the cyclic path. The attenuation coefficient is directly proportional to the length of the cyclic path.

[0048] Depth-first search (DFS) is an algorithm for traversing or searching graphs. Starting from a node, it explores as deep as possible along a path until it cannot continue, then backtracks to the previous node and continues exploring other paths. When detecting cyclic paths in a graph, DFS is used to traverse each node, recording the nodes and edges visited. When a path returns to a previously visited node, a cyclic path has been found.

[0049] The preset threshold is a pre-defined limit for the length of a loop path. When the length of an identified loop path exceeds the preset threshold, it indicates that the loop path may contain a lot of redundant information and needs to be processed. The weight of each directed edge in the loop path is attenuated; the attenuation coefficient is directly proportional to the loop path length—the longer the loop path, the larger the attenuation coefficient, and the more the edge weight is attenuated. In this way, the interference of loop paths on the representation of relationships in the graph can be reduced, making the graph clearer and more accurate.

[0050] Step S146: Extract the largest connected subgraph from the graph, retain the field associations related to the core business process, delete isolated nodes and directed edges with weights lower than the association threshold, and generate a simplified field semantic association graph.

[0051] A maximum connected subgraph is a connected subgraph within a graph such that every two nodes in the subgraph are connected by a path, and no additional nodes or edges can be added to maintain this connectivity. In semantic association graphs, extracting the maximum connected subgraph allows us to focus on the relationships between closely connected fields, reducing interference from irrelevant information.

[0052] Core business processes refer to the business processes within an enterprise or organization that play a crucial role in achieving its main business objectives. In the field of talent big data, core business processes may include talent recruitment, talent development, and talent performance evaluation. Preserving the relationships between fields related to core business processes involves identifying the connections between fields that are closely related to data interaction and processing within those core business processes.

[0053] An isolated node is a node in the graph that is not connected to other nodes by directed edges. These nodes may be redundant information unrelated to the core business process. A directed edge with a weight below the association threshold is a directed edge whose weight is less than a pre-set association threshold. The fields represented by these edges may have weak associations and have little impact on the core business process.

[0054] Step S150: Traverse each node of the semantic association graph of the field, accumulate the sum of all incoming edge weights and outgoing edge weights of the node, compare the sum with the preset association strength benchmark value, mark the node that exceeds the benchmark value as the core data field, and mark the node that does not exceed the benchmark value as the auxiliary data field.

[0055] Traversing each node in the semantic association graph involves sequentially visiting each data field node in the graph. Incoming edge weights are the weights of directed edges pointing to that node, while outgoing edge weights are the weights of directed edges originating from that node. The sum of all incoming and outgoing edge weights for a node yields its association strength value.

[0056] The preset correlation strength benchmark is a pre-defined boundary value used to distinguish between core data fields and auxiliary data fields. When the correlation strength value of a node exceeds this benchmark, it indicates that the node is closely related to other nodes and holds an important position in the dataset, thus being marked as a core data field. Conversely, when the correlation strength value does not exceed the benchmark, it is marked as an auxiliary data field. For example, in a talent dataset, the "Position" field may have strong correlations with multiple other fields, and its correlation strength value exceeds the benchmark, so it is marked as a core data field; while the "Hobbies" field has weaker correlations with other fields, and its correlation strength value does not exceed the benchmark, so it is marked as an auxiliary data field.

[0057] Step S160: Extract the semantic coverage description of the core data field, combine it with the semantic dependency relationship of the auxiliary data field, and dynamically divide the original data record into windows. The starting position of the window is the first appearance of the core data field, and the ending position of the window is the semantic coverage boundary of the core data field, generating a set of semantic data units containing a complete semantic context.

[0058] The semantic coverage description of a core data field is a detailed explanation of the scope of semantic information involved in the core data field. For example, the semantic coverage of the "job title" field may include information related to the job responsibilities, skill requirements, work environment, and career development path. The semantic dependency of auxiliary data fields refers to the degree of semantic association and dependence between auxiliary data fields and core data fields.

[0059] Dynamic windowing is a method of dividing raw data records based on the semantic coverage of core data fields. The window starts at the first occurrence of the core data field, ensuring that each semantic data unit begins with the core data field. The window ends at the boundary of the semantic coverage of the core data field, i.e., the position containing all relevant semantic information related to that core data field.

[0060] In one implementation, step S160 may specifically include the following steps S161-S166: Step S161: Parse the semantic description information of the core data fields, extract the domain terms and business entities, and determine the starting and ending indexes of the semantic coverage by the distribution of the terms in the data records.

[0061] The semantic description information of core data fields includes the detailed meaning, purpose, and contextual information of the field. Domain terms are terms with specific meanings within a particular domain; for example, in the field of talent big data, "recruitment needs" and "performance evaluation" are domain terms. Business entities refer to objects that actually exist in the business, such as "enterprise" and "talent."

[0062] The semantic description information of core data fields is parsed to extract domain terms and business entities. Then, by locating these terms and entities within the data records, the starting and ending indices of the semantic coverage are determined. For example, for the core data field "Job Title," its semantic description information includes domain terms such as "Job Name," "Responsibilities," and "Skill Requirements." By locating these terms within the data records, the starting index (i.e., the first occurrence of "Job Name") and the ending index (i.e., the end of the "Skill Requirements" related information) of the semantic coverage of the "Job Title" field can be determined.

[0063] Step S162: Identify the auxiliary data fields corresponding to the core data fields, query the auxiliary data fields directly associated with the core data fields through the field semantic association graph, and record the semantic dependency direction and dependency strength of the auxiliary data fields.

[0064] The field semantic association graph records the semantic relationships between data fields. By querying the graph for nodes directly associated with core data fields, the corresponding auxiliary data fields can be identified. For example, for the core data field "job title," the field semantic association graph can be used to query auxiliary data fields directly associated with it, such as "skill requirements" and "salary and benefits."

[0065] Semantic dependency direction indicates the semantic association direction between auxiliary data fields and core data fields. For example, the dependency direction from "skill requirements" to "position" indicates that skill requirements are necessary to meet the needs of the position. Dependency strength indicates the tightness of this association, which can be reflected in the field semantic association graph through the edge weights.

[0066] Step S163: Adjust the semantic coverage boundary of the core data field according to the semantic dependency direction of the auxiliary data field. When the auxiliary data field is located after the core data field and the dependency strength exceeds the dependency threshold, extend the termination index to the semantic coverage boundary of the auxiliary data field.

[0067] The semantic dependency direction and strength of auxiliary data fields affect the semantic coverage of core data fields. When an auxiliary data field is located after a core data field and its dependency strength exceeds a dependency threshold, it indicates that the auxiliary data field is closely related to the core data field and needs to be included within the semantic coverage of the core data field.

[0068] For example, for the core data field "Job Title," the auxiliary data field "Salary and Benefits" is located after the "Job Title" field, and the dependency strength exceeds the dependency threshold. In this case, the terminating index of the semantic coverage of the core data field "Job Title" needs to be extended to the semantic coverage boundary of the "Salary and Benefits" field to ensure that the semantic data unit contains complete semantic information.

[0069] Step S164: Initialize the dynamic window parameters. The window size is the sum of the length of the semantic coverage of the core data field and the coverage length of the associated auxiliary data field. The window sliding step is the average length of the semantic coverage of the core data field.

[0070] The initialization of dynamic window parameters is to determine the specific method of partitioning. The window size is jointly determined by the length of the semantic coverage of the core data fields and the coverage length of the associated auxiliary data fields. Adding the two together ensures that the window can contain the core data fields and their related auxiliary data fields, forming a complete semantic data unit.

[0071] The window sliding step size is the average length of the semantic coverage of the core data fields. This setting ensures that the window slides across the data records with a certain regularity, avoiding data omissions or duplications. By setting a reasonable window size and sliding step size, the original data records can be effectively divided into blocks.

[0072] Step S165: Starting from the beginning of the original data record, data is extracted according to the dynamic window parameters. When the window contains at least one core data field and all associated auxiliary data fields, it is marked as a valid window, and the data in the window is extracted as candidate semantic data units.

[0073] Starting from the beginning of the original data record, data is extracted according to pre-defined dynamic window parameters. During extraction, each window is checked to see if it contains at least one core data field and all associated auxiliary data fields. If this condition is met, the window is marked as a valid window. For example, for a data record containing multiple job postings, if a window contains the core data field "Job Title" and associated auxiliary data fields such as "Skill Requirements" and "Salary," it is marked as a valid window. Then, the data within that window is extracted as candidate semantic data units. These units possess complete semantic information, providing a foundation for subsequent processing.

[0074] Step S166: Perform boundary optimization on the candidate semantic data units. By comparing the semantic overlap areas of adjacent units, merge units whose semantic overlap exceeds the overlap threshold, split units containing multiple independent core data fields, and generate the final set of semantic data units.

[0075] After initial truncation, candidate semantic data units may have some boundary issues that require optimization. Comparing the semantic overlap areas of adjacent units, the semantic overlap degree refers to the degree of repetition of semantic information between adjacent units. When the semantic overlap degree exceeds the overlap threshold, it indicates that the semantic information of these two units has a large overlap, and they can be merged into one unit to reduce data redundancy.

[0076] Simultaneously, units containing multiple independent core data fields need to be split. For example, a candidate semantic data unit might contain information about two different positions, each of which can be considered an independent core data field. In this case, the unit is split into two independent semantic data units, so that each unit contains only one core data field and its related auxiliary data fields. Through boundary optimization, a final set of semantic data units is generated, which have clearer semantic boundaries and complete semantic information.

[0077] Step S200: Perform feature clustering on the semantic data unit set, construct a semantic association matrix by calculating the semantic similarity between units, perform clustering on the semantic association matrix to generate an initial feature cluster set, and perform sensitive attribute matching on each initial feature cluster based on the preset sensitive attribute identification rules to generate a feature cluster set containing sensitive attribute labels.

[0078] In one implementation, step S200 may specifically include the following steps S210-S260: Step S210: Perform multimodal feature extraction on each unit in the semantic data unit set, extracting text description features, numerical distribution features, and business association features of the data fields within the unit, and concatenate the three types of features into a unit comprehensive feature vector, with the vector dimension adjusted according to the number of features.

[0079] Multimodal feature extraction involves extracting features from semantic data units from different perspectives. Textual description features refer to the features embodied in the text content of a data field; for example, the text content of the "Job Description" field can reflect the responsibilities and requirements of the position. Numerical distribution features refer to the distribution of numerical data in a data field; for example, the numerical distribution of the "Salary" field can reflect the salary range for the position. Business association features refer to the relationships between a data field and business processes and other data fields; for example, the relationship between the "Project Deliverables" field and fields such as "Project Name" and "Project Time".

[0080] The extracted textual description features, numerical distribution features, and business-related features are concatenated into a unit-level comprehensive feature vector. The vector's dimension is adjusted based on the number of extracted features to ensure a comprehensive representation of the semantic data unit's feature information. For example, for a semantic data unit containing multiple data fields, different features from each field are extracted and combined into a high-dimensional unit-level comprehensive feature vector.

[0081] Step S220: Calculate the similarity between the comprehensive feature vectors of any two semantic data units. By comparing the cosine value of word vectors of text description features, the cumulative distribution function difference of numerical distribution features, and the path matching degree of business association features, a weighted comprehensive semantic similarity value between units is generated.

[0082] Calculating the similarity between the combined feature vectors of any two semantic data units requires consideration from different feature dimensions. For textual descriptive features, the cosine similarity of word vectors is used for comparison. Word vectors represent words in text as vectors; by calculating the cosine of two word vectors, their semantic similarity can be measured.

[0083] For numerical distribution characteristics, the degree of difference in the cumulative distribution function is calculated. The cumulative distribution function describes the distribution of numerical data. By comparing the cumulative distribution functions of two units' numerical distribution characteristics, the degree of difference between them in terms of numerical distribution can be obtained.

[0084] For business-related features, path matching degree is calculated. Business-related features reflect the association paths between data fields. By comparing the degree of matching of the business-related feature paths of two units, their similarity in terms of business association can be measured. Finally, the similarity values ​​of these three dimensions are weighted and summed to obtain the comprehensive semantic similarity value between units. The weight allocation can be adjusted according to specific business needs and data characteristics to highlight the importance of different feature dimensions in the comprehensive similarity calculation.

[0085] Step S230: Construct a semantic association matrix based on the comprehensive semantic similarity value between units. The rows and columns of the matrix correspond to semantic data units, and the matrix elements are the comprehensive semantic similarity values ​​of the corresponding unit pairs. The matrix is ​​symmetric to ensure the consistency of similarity calculation.

[0086] A semantic association matrix is ​​constructed based on the comprehensive semantic similarity value between units, with rows and columns corresponding to semantic data units. Matrix elements represent the comprehensive semantic similarity value of corresponding unit pairs. For example, the element in the i-th row and j-th column of the matrix represents the comprehensive semantic similarity value between the i-th and j-th semantic data units. The matrix is ​​symmetric because semantic similarity has symmetry; that is, the similarity between unit A and unit B should be equal to the similarity between unit B and unit A. Symmetry ensures the symmetry of the elements in the matrix, guaranteeing consistency in similarity calculation. Symmetry can be achieved by copying the upper triangular elements of the matrix to the lower triangular elements.

[0087] Step S240: Perform clustering on the semantic association matrix. Iteratively merge the pairs of units with the highest similarity values. After merging, recalculate the comprehensive semantic similarity value of the new unit with other units until the similarity value of all units is lower than the merging threshold, and obtain the initial feature cluster set.

[0088] The semantic association matrix is ​​clustered to group semantically similar data units into the same cluster. The system iteratively merges the pairs of units with the highest similarity values, forming a new unit after each merge. Then, the combined semantic similarity value of this new unit with other unmerged units is recalculated. A pre-defined similarity threshold is used as the merging threshold. When the similarity values ​​between all units fall below the merging threshold, it indicates that no more sufficiently similar unit pairs can be found for merging, and the merging process stops, resulting in the initial set of feature clusters. Units within each cluster have high semantic similarity, while units between different clusters have low similarity.

[0089] In one implementation, step S240 may specifically include the following steps S241-S245: Step S241: Preprocess the semantic association matrix by replacing the zero values ​​in the matrix with the preset proportion of the minimum similarity value, and sorting the similarity values ​​in each row in descending order to accelerate the search for the maximum similarity.

[0090] Preprocessing the semantic association matrix improves the efficiency and accuracy of subsequent processing. Zero values ​​in the matrix may indicate a lack of significant similarity between two elements; however, in practice, to avoid zero values ​​affecting calculations, they are replaced with a preset proportion of the minimum similarity value. The minimum similarity value can be set according to specific circumstances, and the preset proportion can be adjusted based on experience or requirements.

[0091] Sort the similarity values ​​of each row in descending order. This allows for faster location of the maximum similarity value within each row. Sort in descending order reduces the time complexity of finding the maximum similarity value, thus improving processing efficiency.

[0092] Step S242: Initialize the cluster set. Each semantic data unit is added to the cluster set as an independent cluster. Record the list of member units and the cluster center feature vector of each cluster. The cluster center feature vector is the average value of the comprehensive feature vectors of the member units.

[0093] Initialize the cluster set by adding each semantic data unit as an independent cluster. Each cluster has its own list of member units, recording the semantic data units it contains. The cluster center feature vector is the average of the combined feature vectors of all member units within that cluster. The cluster center feature vector can be obtained by summing and averaging vectors. For each cluster, sum the corresponding elements of the combined feature vectors of its member units, then divide by the number of member units to obtain the cluster center feature vector. The cluster center feature vector represents the overall characteristics of the cluster and is used for subsequent similarity calculations and cluster merging operations.

[0094] Step S243: Find the unit pair corresponding to the maximum similarity value in the semantic association matrix, extract the two clusters to which the unit pair belongs, calculate the similarity value between the cluster center feature vectors of the two clusters, and if the value exceeds the cluster merging threshold, perform a merging operation; when merging two clusters, merge the member unit list into the member list of the new cluster, recalculate the cluster center feature vector of the new cluster, update all similarity values ​​related to the new cluster in the semantic association matrix, and delete the relevant records of the original two clusters.

[0095] Find the pair of units with the highest similarity value in the semantic association matrix, and then locate the clusters to which these two units belong. Calculate the similarity value between the cluster center feature vectors of these two clusters, and determine if this value exceeds the cluster merging threshold. If it exceeds the threshold, it indicates that the two clusters have high similarity, and a merging operation can be performed.

[0096] When merging two clusters, their member lists are combined into the member list of the new cluster. The cluster center feature vector of the new cluster is recalculated, which is the average of the combined feature vectors of all member units within the new cluster. Simultaneously, all similarity values ​​related to the new cluster in the semantic association matrix are updated, as the characteristics of the new cluster have changed, and its similarity to other units will change accordingly. Finally, related records of the original two clusters in the semantic association matrix and cluster set are deleted to maintain data consistency.

[0097] Step S244: Repeat the maximum similarity search and cluster merging operation. After each merge, update the cluster set and semantic association matrix until all similarity values ​​in the matrix are lower than the cluster merging threshold, and then stop the merging process.

[0098] Repeat the maximum similarity search and cluster merging operation. After each merge, the cluster set and semantic association matrix will change. Update the cluster set by adding the new merged cluster to the set and deleting the two original clusters. Update the semantic association matrix by recalculating the similarity values ​​of the new cluster with other units and updating the corresponding elements in the matrix.

[0099] This process is repeated until all similarity values ​​in the matrix are below the cluster merging threshold. At this point, it indicates that no more clusters with sufficiently high similarity can be found for merging, and the merging process is stopped, resulting in the final initial set of feature clusters.

[0100] Step S245: Perform cluster size verification on the merged cluster set, split the clusters with more than the preset upper limit into multiple sub-clusters, and split them based on the distribution of the comprehensive semantic similarity value of the units within the cluster, so that the similarity of the units within the sub-clusters is higher than the similarity between the sub-clusters, and generate the final initial feature cluster set.

[0101] The merged cluster set is then subjected to cluster size verification to check whether the number of member units in each cluster exceeds a preset upper limit. This preset upper limit can be set according to specific needs and data characteristics. If the number of member units in a cluster exceeds the preset upper limit, it indicates that the cluster is too large, hindering subsequent processing and analysis, and needs to be split.

[0102] The splitting is based on the distribution of comprehensive semantic similarity values ​​among units within a cluster. By analyzing the distribution of similarity values ​​among units within a cluster, units with high similarity are grouped into the same sub-cluster, ensuring that the similarity among units within a sub-cluster is higher than the similarity between sub-clusters. Clustering algorithms, such as K-means clustering, can be used to split large clusters into multiple sub-clusters. Through cluster size verification and splitting operations, the final initial feature cluster set is generated.

[0103] Step S250: Traverse each cluster in the initial feature cluster set, extract the description text and sample data of all data fields in the cluster, and perform field-by-field matching with the preset sensitive attribute identification rules. The rules include keyword matching patterns, semantic patterns and data distribution patterns.

[0104] Iterate through each cluster in the initial feature cluster set and process all data fields within each cluster. Extract the descriptive text and sample data for each data field. The descriptive text contains the meaning and purpose of the field, while the sample data is the actual data stored for that field.

[0105] The pre-defined sensitive attribute identification rules include keyword matching, semantic patterns, and data distribution patterns. Keyword matching determines whether a data field is sensitive by searching for specific keywords in its description text or sample data, such as "ID number" or "salary." Semantic patterns determine sensitivity based on the semantic information of the field; for example, fields with semantics closely related to personal privacy may be considered sensitive. Data distribution patterns determine sensitivity by analyzing the numerical distribution of data fields; for example, data with distributions exhibiting clear personal characteristics may be considered sensitive. The extracted description text and sample data are then matched field-by-field against the pre-defined sensitive attribute identification rules to check if each data field conforms to the rules.

[0106] Step S260: When the number of successfully matched sensitive attribute identification rules within a cluster exceeds the rule threshold, add a corresponding sensitive attribute label to the cluster. If multiple sensitive attribute labels are matched, determine the main label based on the sum of rule matching degrees, and generate a feature cluster set containing a cluster identifier, a list of member units, and sensitive attribute labels.

[0107] The rule threshold is a pre-defined limit on the number of matching rules. When the number of successfully matched sensitive attribute identification rules within a cluster exceeds the rule threshold, it indicates that there is a large amount of sensitive attribute data in that cluster, and a corresponding sensitive attribute label is added to that cluster.

[0108] If a cluster matches multiple sensitive attribute labels, the primary label needs to be determined based on the sum of rule matching scores. The sum of rule matching scores refers to the total degree of matching between the cluster and each sensitive attribute identification rule. The label with the highest sum of matching scores is selected as the primary label to highlight the main sensitive attribute features of the cluster.

[0109] The final result is a set of feature clusters containing cluster identifiers, a list of member units, and sensitive attribute labels. The cluster identifier uniquely identifies each cluster, the list of member units records the semantic data units contained in the cluster, and the sensitive attribute labels specify the sensitive attributes of the cluster.

[0110] Step S300: Based on the semantic density value and field association strength value of each feature cluster in the feature cluster set containing sensitive attribute labels, calculate the noise type and noise parameters corresponding to each feature cluster in combination with the sensitive attribute labels of the feature clusters, and obtain the feature cluster noise parameter reference table.

[0111] Different sensitive attribute labels correspond to different levels of privacy protection requirements. Combining the semantic density value of the feature cluster, the field association strength value, and the sensitive attribute labels, the noise type and noise parameters corresponding to each feature cluster are calculated. Noise types include substitution, additive, perturbation, and combined types. Different noise types have different impacts on privacy protection and data availability. Noise parameters are specific parameters set for different noise types, such as the substitution probability and size of the substitution value set for substitution noise, and the noise mean and variance for additive noise.

[0112] In one implementation, step S300 may specifically include the following steps S310-S360: Step S310: For each cluster in the feature cluster set containing sensitive attribute labels, count the number of terms in the semantic description text of the data fields within the cluster, calculate the ratio of the number of terms to the total number of fields, and generate the semantic density value of the cluster by combining the importance weight of the terms in the domain terminology library.

[0113] The ratio of the number of terms to the total number of fields is calculated; this ratio can initially reflect the semantic richness of the data fields within the cluster. Then, the importance weight of each term in the domain terminology corpus is considered. Each term in the domain terminology corpus has a corresponding importance weight, the magnitude of which reflects the term's generality and professional relevance within the domain.

[0114] In one implementation, step S310 may specifically include the following steps S311-S316: Step S311: Extract the semantic description text of all data fields within the feature cluster, perform word segmentation, remove stop words and punctuation marks, retain nouns, verbs and adjectives with actual semantic meaning, and generate a set of field terms.

[0115] Step S312: Match the field term set with the terms in the domain terminology library, identify words that belong to the core terms of the domain, and record the importance weight of each matched term in the domain terminology library. The weight value reflects the degree of generality and professional relevance of the term in the domain.

[0116] The process involves matching the set of field terms with terms in the domain terminology database to find words in the field terminology set that belong to the core domain terms. Core domain terms are those that hold an important position and are widely used within a specific domain. For example, in the field of talent big data, "recruitment," "training," and "performance evaluation" are all core domain terms.

[0117] The importance weight of each matching term in the domain terminology database is recorded. The weight value reflects the term's generality and professional relevance within the domain. Terms with high generality and strong professional relevance have relatively higher importance weights. For example, the term "recruitment" may have a high importance weight in the field of talent big data because it is one of the core business processes in this field.

[0118] Step S313: Count the number of core domain terms in the field term set, calculate the ratio of this number to the total number of data fields in the cluster, and obtain the basic term density value. When the total number of fields is zero, the basic term density value is set to zero.

[0119] The term density is calculated by determining the number of core domain terms in the terminology set of the statistical field, i.e., the number of matched core domain terms. This number is then divided by the total number of data fields within the cluster to obtain the basic term density value. This basic term density value provides a preliminary indication of the richness of core domain terms within the data fields of that cluster. When the total number of data fields within a cluster is zero, it indicates that the cluster has no data fields; in this case, the basic term density value is set to zero to avoid division by zero errors.

[0120] Step S314: Weight the basic term density value by summing the importance weights of the core terms in each domain and multiplying the sum by the basic term density value to obtain the weighted density value. The higher the weight of a term, the greater its contribution to the weighted density value.

[0121] The base term density values ​​are weighted by summing the importance weights of core terms in each domain. Then, the sum is multiplied by the base term density values ​​to obtain the weighted density value. Terms with higher weights contribute more to the weighted density value because they are more important within the domain.

[0122] Step S315: Detect the uniformity of term distribution in the term set of the field, calculate the standard deviation of the distribution of terms in the semantic description text of different data fields. The smaller the standard deviation, the more uniform the distribution. Correct the uniformity of the weighted density value. The more uniform the distribution, the larger the correction coefficient.

[0123] The uniformity of term distribution within a term set is measured by calculating the standard deviation of the term distribution across the semantic descriptions of different data fields. A smaller standard deviation indicates a more uniform distribution of terms across the data fields. The weighted density value is then adjusted for uniformity, with the adjustment coefficient determined based on the uniformity of the term distribution. A more uniform distribution results in a larger adjustment coefficient. For example, if terms are evenly distributed across the semantic descriptions of each data field, it indicates that the semantics of the data fields within that cluster are more coherent and consistent; in this case, a larger adjustment coefficient can be applied, increasing the weighted density value. This uniformity adjustment allows for a more comprehensive consideration of the impact of term distribution on the semantic density value.

[0124] In practical applications, to clarify quantification relationships and ensure dimensional consistency, those skilled in the art can calculate the distribution uniformity coefficient of terms in semantic description texts across different data fields based on their common knowledge. This coefficient is positively correlated with the concentration of term distribution. The weighted density value is then corrected using this distribution uniformity coefficient to generate a corrected weighted density value. Step S316: Map the corrected weighted density values ​​to a preset semantic density value range to generate the final cluster semantic density values. The range ensures that the semantic density values ​​of different feature clusters are comparable, and the mapping process preserves the relative relationship of the original weighted density values.

[0125] The corrected weighted density values ​​are mapped to a preset semantic density value range. This range can be set according to specific needs, such as the [0,1] range. The mapping process needs to preserve the relative relationships of the original weighted density values; that is, clusters with large original weighted density values ​​should retain relatively large values ​​after mapping. A linear mapping method can be used to convert the corrected weighted density values ​​to their corresponding values ​​within the preset range, based on their relative positions within the original value range. Through this mapping operation, the final cluster semantic density values ​​are generated. The semantic density values ​​of different feature clusters are within the same range, ensuring comparability.

[0126] Step S320: Extract the subgraph structure of each feature cluster in the field semantic association graph, accumulate the weight values ​​of all directed edges in the subgraph, and use the ratio of the accumulated result to the number of nodes in the subgraph as the field association strength value of the cluster, which represents the overall degree of association between fields within the cluster.

[0127] Extract the subgraph structure of each feature cluster in the field semantic association graph. A subgraph is a graph composed of nodes and edges related to that feature cluster selected from the field semantic association graph. In the subgraph, nodes represent data fields, directed edges represent semantic associations between fields, and edge weights indicate the strength of the association.

[0128] The weights of all directed edges in the subgraph are summed to obtain a total value. This total value is then divided by the number of nodes in the subgraph to obtain the cluster's field association strength value. The field association strength value characterizes the overall degree of association between data fields within the cluster. For example, if the directed edge weights between nodes in the subgraph of a feature cluster are relatively large, and the number of nodes is moderate, then the cluster's field association strength value will be high, indicating a strong association between fields within the cluster.

[0129] Step S330: Analyze the privacy protection requirement level corresponding to the sensitive attribute labels of the feature clusters. Different labels correspond to different privacy protection strength requirements. The level classification is based on the degree of leakage risk of the sensitive information identified by the label.

[0130] This analysis examines the privacy protection requirements corresponding to the sensitive attribute tags of feature clusters. Different sensitive attribute tags represent different types of sensitive information, and their levels of leakage risk vary. For example, the "ID number" tag indicates a higher risk of information leakage, requiring a higher level of privacy protection; while the "hobby" tag indicates a relatively lower risk of information leakage, requiring a lower level of privacy protection.

[0131] Privacy protection requirements are categorized into different levels—high, medium, and low—based on the degree of risk of leakage of sensitive information identified by the tags. Different levels correspond to different levels of privacy protection intensity, providing a basis for subsequent selection of noise types and parameters.

[0132] Step S340: Based on the privacy protection requirement level, semantic density value, and field association strength value, determine the noise type corresponding to the feature cluster. The noise type includes replacement type, addition type, perturbation type, and combination type. The selection criterion is the comprehensive impact of noise on privacy protection effect and data availability.

[0133] Privacy protection requirement levels are categorized based on the degree of risk of leakage of sensitive information identified by the sensitive attribute tags of feature clusters. When the leakage of sensitive information would pose a serious security threat, economic loss, or reputational damage to an individual or organization, the corresponding privacy protection requirement level is determined to be high. For example, feature clusters containing sensitive information such as ID card numbers and bank card numbers would have a high privacy protection requirement level. If the leakage would only cause minor impacts, such as general hobbies or interests, the privacy protection requirement level would be relatively low.

[0134] Semantic density reflects the semantic richness of data fields within a cluster. A high semantic density value is determined when the semantic description text of data fields within a feature cluster contains a large number of domain-specific terms crucial for business understanding, and these terms are relatively evenly distributed and involve multiple business dimensions. For example, in talent data, if a feature cluster covers detailed job descriptions, professional skill requirements, complex career development paths, and other information, its semantic density value will be high; conversely, if it only contains simple, broad descriptions, the semantic density value will be low.

[0135] The field association strength value reflects the degree of correlation between fields within a cluster. If, in the semantic association graph of a feature cluster, the directed edge weights between nodes in the corresponding subgraph are generally large, and there are many direct and indirect association paths between nodes, it indicates a high degree of interdependence and mutual influence among fields within the cluster, and the field association strength value is considered high. For example, in a talent recruitment scenario, fields such as "job title," "skill requirements," and "salary" are closely related, and the field association strength value of their respective feature clusters is high; if the correlation between fields is loose, the association strength value is low.

[0136] For feature clusters with high privacy protection requirements, high semantic density, and strong field correlation, combined noise is the preferred choice. This is because high privacy protection demands robust privacy safeguards, and combined noise, by combining the characteristics of multiple noise types, can more comprehensively hide sensitive information. High semantic density and strong field correlation imply complex and information-rich data structures; combined noise can protect privacy while maximizing the maintenance of data structure and correlation, thus ensuring data availability.

[0137] Replacement noise is suitable when privacy protection requirements are high, but semantic density and field association strength are relatively low. Replacement noise can directly replace the original data values, effectively masking sensitive information, and because the data structure is relatively simple, it has little impact on the overall data structure.

[0138] If the level of privacy protection requirements is moderate, and the semantic density and field association strength values ​​are moderate, additive noise is a good choice. Additive noise protects privacy by adding noise values ​​to the original data. It can hide sensitive information to a certain extent without excessively destroying the original characteristics of the data, thus achieving a good balance between privacy protection and data usability.

[0139] For feature clusters with low privacy protection requirements and low semantic density and field association strength, perturbation noise is sufficient. Perturbation noise moderately agitates the original data, providing basic privacy protection while minimizing its impact on the data and maximizing its usability. By comprehensively considering these three factors, the noise type can be precisely selected according to different situations to achieve the optimal balance between privacy protection and data usability.

[0140] Step S350: For a given noise type, calculate the corresponding noise parameters. Replacement noise parameters include replacement probability and size of the replacement value set. Addition noise parameters include noise mean and variance. Disturbance noise parameters include disturbance amplitude and disturbance correlation. The parameter values ​​are directly proportional to the semantic density value and the field association strength value.

[0141] In one implementation, step S350 may specifically include the following steps S351-S356: Step S351: For substitution noise, multiply the semantic density value by the preset substitution probability benchmark value to obtain the initial substitution probability, multiply the field association strength value by the substitution value set size benchmark value to obtain the initial substitution value set size, and truncate the initial value to ensure it is within the effective range.

[0142] For substitution-type noise, the substitution probability and the size of the substitution value set are calculated. The semantic density value is multiplied by a preset substitution probability baseline value to obtain the initial substitution probability. The preset substitution probability baseline value is a pre-defined base value used to calculate the substitution probability.

[0143] The initial size of the replacement value set is obtained by multiplying the field association strength value by the baseline value for the replacement value set size. The baseline value for the replacement value set size is a pre-defined base value used to calculate the size of the replacement value set.

[0144] The initial replacement probabilities and the size of the replacement value set are truncated to ensure they are within a valid range. For example, the valid range of the replacement probability might be [0,1]. If the initial replacement probability exceeds 1, it is truncated to 1; if it is less than 0, it is truncated to 0. This truncation ensures that the calculated noise parameters are reasonable and valid.

[0145] Step S352: For added noise, the semantic density value is mapped to the noise variance. The higher the semantic density value, the larger the variance. The field association strength value is mapped to the noise mean. The higher the association strength value, the closer the mean is to zero, so as to avoid the data distribution shift caused by the mean being too large.

[0146] For additive noise, calculate the noise variance and noise mean. Map semantic density values ​​to noise variance; the higher the semantic density value, the larger the noise variance. This is because feature clusters with high semantic density values ​​may contain more sensitive information, requiring a larger noise variance to protect privacy. Map field association strength values ​​to noise mean; the higher the association strength value, the closer the noise mean is to zero. This is to avoid an excessively large mean causing data distribution shift and affecting data usability.

[0147] Step S353: For perturbation-type noise, the product of semantic density value and field association strength value is used as the basis for calculating perturbation amplitude. The larger the base value, the larger the perturbation amplitude. The field association strength value is used as the basis for calculating perturbation correlation. The higher the association strength value, the higher the perturbation correlation between different fields.

[0148] For perturbation-type noise, the perturbation amplitude and perturbation correlation are calculated. The product of the semantic density value and the field association strength value is used as the basis for calculating the perturbation amplitude; the larger the base value, the larger the perturbation amplitude. This is because feature clusters with both large semantic density values ​​and large field association strength values ​​require larger perturbation amplitudes to protect privacy. The field association strength value is used as the basis for calculating the perturbation correlation; the higher the association strength value, the higher the perturbation correlation between different fields.

[0149] Step S354: For combined noise, calculate the parameter values ​​of each basic noise type separately, determine the weight ratio of each basic noise type according to the sensitive attribute label of the feature cluster, and generate combined noise parameters by weighted merging of parameter values. The weight ratio represents the contribution of different noise types to privacy protection.

[0150] For combined noise, parameter values ​​for basic noise types such as substitution, addition, and perturbation are calculated separately. The weight percentage of each basic noise type is determined based on the sensitivity attribute labels of the feature cluster. This weight percentage reflects the contribution of different noise types to privacy protection. For example, for a feature cluster containing highly sensitive information, substitution noise may contribute more to privacy protection and is therefore given a higher weight percentage.

[0151] The parameter values ​​of each basic noise type are weighted and combined to generate the combined noise parameter. For example, for a certain parameter of the combined noise, the parameter value of the substitution noise is multiplied by its weight ratio, the parameter value of the additive noise is multiplied by its weight ratio, and the parameter value of the perturbation noise is multiplied by its weight ratio. The results are then added together to obtain the parameter value of the combined noise. In this way, the advantages of different noise types can be comprehensively utilized to achieve better privacy protection.

[0152] Step S355: Compare the calculated noise parameters with the preset parameter constraints. The constraints include the upper and lower limits of the impact of noise on privacy protection and data availability. If the parameters do not meet the constraints, adjust them proportionally until they do.

[0153] The calculated noise parameters are compared with preset parameter constraints, which include upper and lower limits on the impact of noise on privacy protection and data availability. For example, the privacy protection effect cannot be lower than a certain level, and the data availability cannot be lower than a certain threshold.

[0154] When the calculated noise parameters do not meet the constraints, the parameter values ​​are adjusted proportionally. If the impact of noise on privacy protection is below the lower limit, it indicates insufficient privacy protection, so the relevant noise parameters are increased; if data availability is below the lower limit, it indicates excessive interference from noise, so the relevant noise parameters are decreased. The adjustment process continues, constantly checking whether the parameters meet the constraints, until all parameters are within the specified upper and lower limits. For example, if the replacement probability of replacement noise is too high, leading to a decrease in data availability, the replacement probability is appropriately reduced; if the variance of additive noise is too small, resulting in poor privacy protection, the variance is increased.

[0155] Step S356: Verify the adjusted noise parameters. Calculate privacy protection indicators and data availability indicators by simulating injected noise. The indicator values ​​must meet the preset standards simultaneously. Parameters that do not meet the standards are recalculated and adjusted until all indicators meet the requirements.

[0156] Validating the adjusted noise parameters is a crucial step in ensuring a balance between privacy protection and data availability. Simulated noise injection involves processing sample data according to the adjusted noise parameters without affecting the original data. This process requires calculating two important metrics: privacy protection metrics and data availability metrics.

[0157] Privacy protection metrics measure the degree to which sensitive information is protected in data after noise injection. This can be assessed in various ways, such as by examining the extent to which sensitive information is obfuscated or hidden in the data, and the difficulty for external attackers to infer the original sensitive information from the processed data. Data usability metrics focus on whether the processed data can still be used for normal business analysis and annotation tasks, such as data integrity, consistency, and fit with business logic.

[0158] The preset standards are pre-defined thresholds used to determine whether privacy protection and data availability meet the requirements. When the calculated indicator values ​​after simulating noise injection simultaneously meet the preset standards, it indicates that the adjusted noise parameters are appropriate. If any indicator fails to meet the standards, the parameters need to be recalculated and adjusted. For example, if the privacy protection indicator fails to meet the standards, the noise parameters may need to be increased further; if the data availability indicator fails to meet the standards, the noise parameters need to be appropriately decreased. This process is repeated until all indicators meet the preset standards, thus determining the final usable noise parameters.

[0159] Step S360: Integrate the feature cluster identifier, sensitive attribute label, noise type and calculated noise parameters into a record to generate a feature cluster noise parameter lookup table. Each record in the table contains complete noise generation configuration information.

[0160] Integrating feature cluster identifiers, sensitive attribute labels, noise types, and calculated noise parameters into a single record facilitates accurate noise generation for each feature cluster. The feature cluster identifier uniquely identifies each cluster, enabling rapid location and differentiation of different clusters when processing large amounts of data. The sensitive attribute label clarifies the type of sensitive information contained within the feature cluster, providing targeted guidance for noise generation. The noise type determines the data processing method, while the noise parameters specify the operational details for that noise type. The feature cluster noise parameter lookup table is presented in tabular form, with each record corresponding to one feature cluster. Each column in the table stores the feature cluster identifier, sensitive attribute label, noise type, and various noise parameters.

[0161] Step S400: Map the semantic data unit set to the feature cluster noise parameter lookup table according to the feature cluster affiliation relationship, perform corresponding noise generation and injection operations on the feature cluster data fields with different sensitive attribute labels, maintain the noise distribution characteristics between data fields in the same cluster by controlling the random seed synchronization mechanism in the noise generation process, and generate a perturbed data set.

[0162] In one implementation, step S400 may specifically include the following steps S410-S460: Step S410: Traverse each unit in the semantic data unit set, extract the identifiers of all data fields within the unit, query the feature cluster identifiers to which the fields belong, generate a unit-field-cluster attribution table, and record the cluster attribution information of each data field.

[0163] Traversing each cell in the semantic data unit set is to gain a comprehensive understanding of the distribution and affiliation of data fields. For each cell, the identifiers of all data fields within it are extracted; these identifiers are codes or names that uniquely identify each data field. Then, based on the previously constructed feature cluster partitioning results, the feature cluster identifier to which each data field belongs is queried.

[0164] A unit-field-cluster attribution table is generated, which records detailed attribution information for each data field in tabular form. Each row of the table corresponds to a data field, and the columns are the unit identifier, data field identifier, and feature cluster identifier, respectively. For example, for a semantic data unit "U001" containing the data field "F001", a query reveals that this field belongs to the feature cluster "C002". Therefore, the attribution table will contain a record with the following information: Unit identifier "U001", Data field identifier "F001", and Feature cluster identifier "C002". This attribution table clearly shows the correspondence between each data field and the feature cluster, providing an accurate basis for subsequent noise configuration and injection.

[0165] Step S420: Based on the unit-field-cluster attribution table and the feature cluster noise parameter comparison table, match the corresponding noise type and noise parameter for each data field to generate a field noise configuration table. The configuration table includes field identifier, noise type, noise parameter and sensitive attribute label.

[0166] Based on the unit-field-cluster attribution table and the feature cluster noise parameter lookup table, the corresponding noise type and noise parameters for each data field are found. The attribution table clarifies the feature cluster to which the data field belongs, while the feature cluster noise parameter lookup table records the noise generation configuration information for each feature cluster.

[0167] A field noise configuration table is generated, where each row corresponds to a data field, and the columns are the field identifier, noise type, noise parameters, and sensitivity attribute label. For example, for the data field "F001", the attribution table shows that it belongs to feature cluster "C002". The noise type corresponding to "C002" is found to be "perturbation type" from the feature cluster noise parameter lookup table. The noise parameters include the specific perturbation amplitude and perturbation correlation. The sensitivity attribute label for this feature cluster is "skill certificate". This information will then be recorded in the field noise configuration table. The field noise configuration table provides detailed noise generation guidelines for each data field, ensuring that subsequent noise injection operations are performed accurately according to the configuration.

[0168] Step S430: Initialize the random number generator and assign a unique basic random seed to each feature cluster. The basic random seed is generated by combining the hash value of the feature cluster identifier with the current timestamp to ensure that the seeds of different clusters are unique.

[0169] Initializing the random number generator prepares for the subsequent generation of random noise. A random number generator is an algorithm or device capable of generating sequences of random numbers. Assigning a unique base random seed to each feature cluster is crucial to ensuring consistent noise distribution characteristics within the same cluster and that noise is independent between different clusters.

[0170] The base random seed is generated by combining the hash value of the feature cluster identifier and the current timestamp. The hash value is a fixed-length numerical value obtained by transforming the feature cluster identifier using a hash function; it is unique and deterministic, meaning that the same feature cluster identifier will produce the same hash value. The current timestamp is the system's current time information and is constantly changing. Combining these two values ​​ensures that the base random seed for each feature cluster is unique.

[0171] In one implementation, step S430 may specifically include the following steps S431-S436: Step S431: Extract the data type description and business scenario description from the semantic description text of all data fields within the feature cluster, identify the description fragments with domain-specific meanings, count the frequency and distribution of each description fragment in the field descriptions within the cluster, and generate a feature cluster semantic fingerprint containing a frequency sequence and a position offset sequence of description fragments.

[0172] Extract data type descriptions and business scenario descriptions from the semantic description text of all data fields within the feature cluster. This information contains the key features and usage context of the data fields. Within these descriptions, identify descriptive fragments with domain-specific meanings—words or phrases that have unique significance within a specific domain. For example, in the field of talent big data, terms like "recruitment process optimization" and "performance evaluation system" are descriptive fragments with domain-specific meanings.

[0173] The frequency and distribution of each descriptive fragment within the cluster's field descriptions are statistically analyzed. Frequency reflects the importance of the descriptive fragment within the cluster, while distribution reflects its contextual relationship in the semantic description. The frequency occurrences are compiled into a descriptive fragment frequency sequence, and the distribution positions are compiled into a position offset sequence; these two sequences together constitute the semantic fingerprint of the feature cluster. The semantic fingerprint of a feature cluster can be seen as a unique identifier for that feature cluster, reflecting the semantic characteristics and distribution patterns of the data fields within the feature cluster.

[0174] Step S432: Traverse all direct and indirect association paths of the feature cluster in the field semantic association graph, record the node field identifiers and the number of edges traversed by each path, calculate the cumulative sum of the weights of all edges in each path, and take the cumulative value corresponding to the path with the largest cumulative sum as the path feature value.

[0175] The system iterates through all direct and indirect association paths of the feature clusters in the field semantic association graph. A direct association path is a path that directly connects two nodes, while an indirect association path is a path connected through other nodes. During the traversal, the system records the node field identifiers and the number of edges traversed by each path.

[0176] Calculate the cumulative sum of all edge weights in each path, where edge weights represent the semantic association strength between nodes. The larger the cumulative sum of edge weights in a path, the more important the path is in terms of semantic association. The cumulative value corresponding to the path with the largest cumulative sum is taken as the path feature value. The path feature value reflects the core association strength of the feature cluster in the field semantic association graph, and together with the feature cluster semantic fingerprint, provides more comprehensive feature information for generating the basic random seed.

[0177] Step S433: Element-wise fusion of the frequency sequence of the description fragments in the semantic fingerprint of the feature cluster with the path feature values. The fusion method is to add the frequency values ​​and path feature values ​​according to the sequence position to generate a fusion sequence. The length of the fusion sequence is consistent with the length of the frequency sequence in the semantic fingerprint.

[0178] The semantic fingerprint of the feature cluster is fused with the description fragment frequency sequence and path feature values ​​at the element-wise level. Element-wise fusion means that each frequency value in the description fragment frequency sequence is added to the path feature value according to the position of the sequence. For example, if the description fragment frequency sequence is [a1,a2,a3,...], the path feature value will be added to these frequency values ​​in sequence during fusion to obtain the fused sequence [b1,b2,b3,...], where bi = ai + path feature value (corresponding position).

[0179] The length of the generated fused sequence is consistent with the length of the frequency sequence in the semantic fingerprint. This ensures that the fused sequence still accurately reflects the semantic and association features of the feature clusters. The fused sequence integrates the semantic and association strength information of the feature clusters, providing richer and more representative material for the subsequent generation of basic random seeds.

[0180] It is understandable that, in practical applications, in order to overcome the problems of dimensionality and scale consistency, those skilled in the art can normalize the frequency sequence of description fragments in the semantic fingerprint of the feature cluster to obtain a normalized frequency sequence, and then scale the path feature values ​​to match their magnitude with the normalized frequency sequence. After that, the scaled path feature values ​​are weighted and combined with each element in the normalized frequency sequence to generate a fusion sequence.

[0181] Step S434: Perform iterative transformation on the fused sequence. In each transformation operation, the difference between two adjacent elements in the sequence is used as the corresponding element of the new sequence. Repeat the transformation operation until the sequence length is reduced to the same as the preset seed length to generate the seed material sequence.

[0182] Iterative transformations are performed on the fused sequence to further extract key features and compress them to a suitable length. Each transformation operation uses the difference between two adjacent elements in the original sequence as the corresponding element in the new sequence. For example, for the fused sequence [a1, a2, a3, ...], after one transformation, the new sequence [b1, b2, b3, ...] is obtained, where b1 = a2 - a1, b2 = a3 - a2, and so on.

[0183] This transformation operation is repeated until the sequence length is reduced to match the preset seed length. The preset seed length is pre-set according to the requirements of the random number generator. Through this iterative transformation, redundant information in the sequence is removed, the most critical features are extracted, and a seed material sequence is generated. The seed material sequence contains the core feature information of the feature clusters and is an important source for generating the basic random seed.

[0184] Step S435: Convert the seed material sequence into numerical form and compare it one by one with the seed values ​​of other feature clusters stored in the seed management database. If there are duplicate values, adjust the weight allocation of the description fragment in the semantic fingerprint and regenerate the fusion sequence until a unique value that does not exist in the database is obtained.

[0185] Converting the seed material sequence into numerical form facilitates comparison with existing seed values ​​in the seed management database. The seed management database stores the base random seed values ​​previously generated for other feature clusters, making it crucial to ensure that the seed for each feature cluster is unique.

[0186] The converted values ​​are compared one by one with other seed values ​​in the database. If duplicate values ​​are found, it indicates that the currently generated seed is not unique. In this case, the weight allocation of the description fragments in the semantic fingerprint needs to be adjusted. Adjusting the weight allocation changes the frequency sequence of the description fragments, thus affecting the generation of the fusion sequence. After regenerating the fusion sequence, iterative transformation is performed again to obtain a new seed material sequence, which is then converted into numerical form for comparison. This process is repeated until a unique value not found in the database is obtained, ensuring that the basic random seed for each feature cluster is unique.

[0187] It is understandable that if numerical duplicates exist, a random perturbation based on the current system time can be introduced into the current seed material sequence to generate a new seed material sequence and reconvert it into numerical form for comparison until a unique value that does not exist in the database is obtained.

[0188] Step S436: Load the unique value as the base random seed into the random number generator, configure the generator's internal state parameters so that the random number sequence output by the generator satisfies the distribution characteristics of the corresponding noise type, and record the seed value, semantic fingerprint and path feature value used in the generation process into the seed audit log.

[0189] The obtained unique value is used as the base random seed and loaded into the random number generator. After loading the base random seed, the random number generator generates a sequence of random numbers based on that seed. The generator's internal state parameters are configured; these parameters affect the method and characteristics of random number generation. Depending on the type of noise, the internal state parameters are adjusted so that the generator's output random number sequence conforms to the distribution characteristics of the corresponding noise type. For example, for replacement noise, it may be necessary to generate random numbers that conform to a set probability distribution to determine the replacement value; for additive noise, it is necessary to generate random numbers that conform to a specified mean and variance.

[0190] Simultaneously, the seed value, semantic fingerprint, and path feature values ​​used in the generation process are recorded in the seed audit log. The seed audit log records the generation process and related information of the basic random seed, which is crucial for subsequent auditing and traceability. By reviewing the seed audit log, one can understand how the basic random seed for each feature cluster is generated and what key information is used, ensuring the traceability and security of the entire random seed generation process.

[0191] Step S440: Perform the corresponding noise generation operation according to the noise type of the data field. Replacement noise generates random replacement values ​​that conform to the distribution of the replacement value set. Addition noise generates random noise values ​​that conform to the mean and variance requirements. Disturbance noise generates disturbance vectors that conform to the disturbance amplitude and correlation.

[0192] Noise generation is performed according to the noise type of the data field. For replacement noise, replacement values ​​are randomly generated based on the previously determined distribution of the replacement value set. The distribution of the replacement value set specifies the probability of each possible replacement value, and the randomly generated replacement values ​​must conform to this distribution.

[0193] For additive noise, random noise values ​​that meet the requirements are generated based on the previously calculated noise mean and variance. The noise mean determines the average level of the random noise values, while the variance controls the dispersion of the noise values. A random number generator is used to generate random noise values ​​that meet specific mean and variance, so that they can be subsequently added to the original data field values.

[0194] For perturbation-type noise, a perturbation vector is generated that conforms to the perturbation amplitude and correlation. The perturbation amplitude specifies the degree to which the original data values ​​are perturbed, while the perturbation correlation reflects the relationship between the perturbations in different data fields. Based on these parameters, a corresponding perturbation vector is generated to perform vector operations on the original data field values, thereby changing their numerical values.

[0195] Step S450: During the noise generation process, the data fields in the same cluster use a derived seed from the basic random seed. The derived seed is generated by performing an XOR operation between the basic random seed and the hash value of the field identifier, so that the noise distribution of the fields in the same cluster has a preset correlation.

[0196] Using derived seeds to generate noise ensures that the noise distribution of data fields within the same cluster has a certain correlation. Because clustered data fields share the same basic random seed, and derived seeds are generated by combining the basic random seed with field identifiers, their noise distributions will exhibit similar patterns, but will differ due to the different field identifiers. This ensures the correlation between clustered data fields while avoiding complete noise uniformity, allowing the data to retain certain characteristics and structure while protecting privacy.

[0197] In one implementation, step S450 may specifically include the following steps S451-S456: Step S451: Parse the semantic dependency tree of the data fields in the semantic data unit, determine the dependency level of the field relative to the core data field, and increase the level value as the length of the dependency chain increases to generate the field dependency level value.

[0198] This section analyzes the semantic dependency tree of clustered data fields within semantic data units. The semantic dependency tree describes the semantic associations and dependencies between data fields, with the core data field at the core of the tree, and other data fields connected to the core data field through different dependency chains.

[0199] Determine the dependency level of a field relative to the core data field. The dependency level reflects the distance and degree of dependence between the field and the core data field. The dependency level value increases with the length of the dependency chain; that is, the farther away from the core data field, the larger the dependency level value. For example, a field directly connected to the core data field has a dependency level value of 1, a field connected to the core data field through an intermediate field has a dependency level value of 2, and so on.

[0200] Step S452: Extract the logical position information of the field within the feature cluster, including the field's arrangement order number in the data unit and the number of interval fields between it and adjacent fields. Add the order number and the number of interval fields to obtain the position feature value.

[0201] Extract the logical position information of a field within a feature cluster. This logical position information reflects the field's relative position and contextual relationship within the data unit. It includes the field's order number within the data unit and the number of other fields that separate it from its adjacent fields. The order number refers to the field's sequential number within the data unit, and the number of other fields that separate it from its adjacent fields.

[0202] The positional feature value is obtained by adding the sequential number to the interval number. This value comprehensively considers both the field's order and its distance from adjacent fields, reflecting the field's specific location and context within the feature cluster. The positional feature value, along with the field dependency hierarchy value, is used in a weighted combination to generate a comprehensive field identifier.

[0203] Step S453: Combine the field dependency level value and the location feature value with a weighted sum. The weight allocation ratio is determined based on the connection strength value of the field in the field semantic association graph. The higher the connection strength value, the greater the weight of the dependency level value. Generate a comprehensive field identifier.

[0204] The field dependency hierarchy value and location feature value are weighted and combined. The weighting ratio is determined based on the field's connection strength value in the field semantic association graph. The connection strength value reflects the degree of semantic association between the field and other fields; the higher the connection strength value, the more important the field is in semantic association.

[0205] When the connection strength value is high, the weight of the dependency level value is increased, because the dependency level value better reflects the core position of the field in the semantic dependency relationship. Through weighted combination, the field dependency level value and positional feature value are merged into a single value to generate a comprehensive field identifier. This comprehensive field identifier integrates the semantic dependency and logical positional information of the field, providing more comprehensive feature information for the generation of derivation seeds.

[0206] It is understandable that, in practical applications, those skilled in the art can normalize the field dependency hierarchy value and position feature value based on basic common sense to obtain the normalized hierarchy value and position value. Then, the weight allocation ratio is determined based on the connection strength value of the field in the field semantic association graph. The higher the connection strength value, the greater the weight of the normalized hierarchy value. The weighted normalized hierarchy value and the normalized position value are added together to generate the comprehensive field identifier.

[0207] Step S454: Perform data fusion processing on the basic random seed of the feature cluster and the comprehensive field identifier. During the fusion process, the numerical characteristics of the basic random seed and the uniqueness of the comprehensive field identifier are preserved to generate a preliminary derived seed.

[0208] The basic random seed of the feature cluster and the comprehensive field identifier are fused together. Data fusion combines two different values ​​while preserving their individual characteristics. During the fusion process, it is necessary to retain the numerical characteristics of the basic random seed to ensure the correlation of noise distribution among data fields in the same cluster, while also preserving the uniqueness of the comprehensive field identifier to ensure that the noise distribution of each field is different.

[0209] A fusion algorithm can be used to combine the basic random seed and the field comprehensive identifier to generate a preliminary derived seed. The preliminary derived seed combines the overall features of the feature cluster and the individual features of the field, providing a foundation for subsequent generation of random numbers that satisfy a specific noise distribution.

[0210] Step S455: Adjust the numerical distribution characteristics of the initial derived seed based on the number of successful rule matches in the sensitive attribute matching process. The more successful matches, the higher the concentration of the seed numerical distribution.

[0211] The numerical distribution characteristics of the initial derived seed are adjusted based on the number of successful rule matches for a field during the sensitive attribute matching process. The number of successful rule matches reflects the sensitivity of the field; the more successful matches, the more sensitive information the field contains, requiring stronger privacy protection.

[0212] The numerical distribution characteristics of the initial derived seeds are adjusted so that the concentration of the seed numerical distribution is correlated with the number of successful rule matches. The more successful matches, the higher the concentration of the seed numerical distribution. This means that when generating noise, for highly sensitive fields, the noise distribution will be more concentrated, thus better protecting sensitive information. This adjustment mechanism ensures that fields with different levels of sensitivity receive appropriate privacy protection during noise generation.

[0213] Step S456: Load the adjusted derived seed into the temporary buffer of the random number generator. After generating the noise sequence required for the current data field, immediately clear the seed information in the buffer and reset the internal state of the generator to prevent cross-interference of seed information of different fields.

[0214] The adjusted derived seed is loaded into the random number generator's temporary buffer. The random number generator generates the noise sequence required for the current data field based on the derived seed in the temporary buffer. After generating the noise sequence, the seed information in the buffer is immediately cleared, and the generator's internal state is reset.

[0215] Clearing seed information and resetting internal states prevents cross-interference between seed information from different fields. If not cleared and reset in a timely manner, subsequent noise generation for other data fields may be affected by previous seed information, leading to chaotic noise distribution. This operation ensures that noise generation for each data field is independent and accurate, guaranteeing the pre-defined correlation of noise distribution within the same data cluster and the independence between different fields.

[0216] Step S460: Inject the generated noise values ​​into the original data field values. Replacement noise directly replaces the original values, addition noise is added to the original values, and perturbation noise is vector-operated with the original values. After injection, the data fields are reorganized into semantic data units to generate a perturbation data set containing noise injection records.

[0217] The generated noise values ​​are injected into the original data field values, with different injection methods used depending on the type of noise. For replacement noise, the generated random replacement value directly replaces the original data field value. For example, if the original data field value is "value A" and the generated replacement value is "value B", then the field value is replaced with "value B".

[0218] For additive noise, the generated random noise value is added to the original data field value. For example, if the original data field value is 10 and the generated random noise value is 2, then the field value becomes 12 after the noise is injected.

[0219] For perturbation-type noise, the generated perturbation vector is used to perform vector operations with the original data field values. Vector operations can be defined according to specific business needs and data characteristics, such as vector addition and vector multiplication. Through vector operations, the numerical values ​​of the original data fields are changed.

[0220] After injection, the data fields are reorganized into semantic data units. A semantic data unit is a set of data fields with a complete semantic context. The injected data fields are combined according to their original semantic relationships to restore their semantic structure. Simultaneously, relevant information about the noise injection is recorded, including noise type, noise parameters, and injected fields, generating a perturbed data set containing the noise injection records. The data in the perturbed data set protects sensitive information to a certain extent while preserving data usability.

[0221] Step S500: Perform an association evaluation operation on the perturbed dataset. Determine whether the preset association preservation threshold is met by calculating the rate of change in association between fields within semantic data units before and after the perturbation. When the threshold requirement is met, output a privacy-preserving dataset that can be used for annotation tasks.

[0222] In one implementation, step S500 may specifically include the following steps S510-S570: Step S510: Extract the original semantic data unit set corresponding to the perturbation data set from the original talent data set, so that the number, field composition and arrangement order of the two sets of data units are completely consistent, and establish a one-to-one correspondence at the unit level.

[0223] Extract the original semantic data unit set corresponding to the perturbation data set from the original talent dataset. To accurately assess the impact of noise injection on data correlation, it is necessary to ensure that the two sets of data units are completely identical in number, field composition, and arrangement order. This ensures that the correlation relationships between the same fields before and after the perturbation can be accurately compared in subsequent correlation calculations.

[0224] A one-to-one correspondence is established at the unit level, meaning that each semantic data unit in the perturbed dataset corresponds to a unit in the original semantic data unit set. This correspondence facilitates unit-by-unit comparison and analysis of the data before and after the perturbation, providing a foundation for calculating the rate of change in correlation.

[0225] Step S520: For each group of corresponding data units, select all field pairs and calculate the original correlation value of the original data unit field pairs and the perturbation correlation value of the perturbed data unit field pairs. The correlation calculation adopts the mutual information method to reflect the statistical dependency between fields.

[0226] For each corresponding data unit, select all possible field pair combinations. For example, for a data unit containing three data fields A, B, and C, the field pair combinations are (A,B), (A,C), and (B,C).

[0227] Calculate the original correlation degree value for field pairs of original data units and the perturbed correlation degree value for field pairs of perturbed data units. The correlation degree calculation uses the mutual information method, which is an indicator that measures the statistical dependency between two random variables. By calculating the mutual information between field pairs, the strength of their correlation can be understood. For original data units, the original correlation degree value is obtained by calculating the mutual information of their field pairs; for perturbed data units, the perturbed correlation degree value is obtained by similarly calculating the mutual information of their field pairs.

[0228] Step S530: Calculate the rate of change of the correlation degree for each field pair. The rate of change is the ratio of (perturbed correlation degree value - original correlation degree value) to the original correlation degree value. When the original correlation degree value is zero, the rate of change is set to zero to avoid division by zero error.

[0229] Calculate the rate of change of relevance for each field pair. This rate of change reflects the degree of change in the relevance of the field pair after noise injection. The rate of change is obtained by calculating the ratio of (perturbed relevance value - original relevance value) to the original relevance value. When the original relevance value is zero, the rate of change is set to zero to avoid division by zero errors.

[0230] The rate of change in correlation degree can intuitively show the impact of noise injection on the correlation relationship of each field. If the rate of change in correlation degree is large, it means that the noise injection has had a significant impact on the correlation relationship of that field pair; if the rate of change in correlation degree is small, it means that the noise injection has had a small impact on the correlation relationship of that field pair. By calculating the rate of change in correlation degree, a comprehensive understanding of the overall impact of noise injection on the data correlation degree can be obtained.

[0231] Step S540: Calculate the absolute value of the rate of change of the correlation degree of all field pairs in each semantic data unit, and calculate the average of the absolute values ​​as the unit correlation change index, which reflects the degree of data correlation retention of the unit as a whole.

[0232] Calculate the absolute value of the rate of change of association degree for all field pairs within each semantic data unit. The absolute value of the rate of change of association degree eliminates the influence of the positive or negative sign, focusing only on the magnitude of the change. Calculate the average of these absolute values ​​to obtain the unit association change index.

[0233] The unit association change index reflects the degree to which the overall data association of a unit is maintained. A small unit association change index indicates that the associations between fields in the semantic data unit remain largely unchanged after noise injection, and the data associations are well preserved. Conversely, a large unit association change index indicates that noise injection has significantly damaged the data associations of the unit, affecting its data associations. The unit association change index allows for the evaluation of the association maintenance of each semantic data unit.

[0234] Step S550: Compare the unit association change index with the preset association maintenance threshold. When the index is less than or equal to the threshold, the unit is determined to pass the evaluation. When the index is greater than the threshold, it is marked as failing the evaluation and needs to be optimized.

[0235] The unit association change index is compared with a preset association retention threshold. The preset association retention threshold is a limit value pre-set based on business needs and data characteristics. When the unit association change index is less than or equal to the threshold, it indicates that the change in data association of the semantic data unit after noise injection is within an acceptable range, and the unit is deemed to have passed the evaluation.

[0236] When the unit correlation change index exceeds a threshold, it indicates that the data correlation of that unit has changed too much, which may affect subsequent data annotation and analysis. This unit is marked as failing the evaluation and requires optimization. This comparison and judgment mechanism can filter out semantic data units with good data correlation and identify units that require further processing.

[0237] Step S560: For units that fail the evaluation, analyze the field pairs with the largest correlation change rate among the field pairs they contain, query the noise parameter of the feature cluster to which the corresponding field belongs, reduce the noise intensity parameter of the cluster and re-perform the noise injection operation, and recalculate the unit correlation change index until the threshold requirement is met.

[0238] For units that failed the evaluation, analyze the field pairs with the largest rate of change in correlation among their constituent field pairs. The field pair with the largest rate of change in correlation indicates that the correlation of that field pair was most affected after noise injection, possibly due to excessive noise intensity.

[0239] Query the noise parameters of the feature cluster to which the corresponding field belongs, and find the noise type and noise parameters corresponding to that feature cluster. Reduce the noise intensity parameters of that cluster. For example, for replacement noise, reduce the replacement probability; for additive noise, reduce the noise mean and variance; for perturbation noise, reduce the perturbation amplitude.

[0240] The noise injection operation is repeated, and the units that failed the evaluation are processed according to the adjusted noise parameters. The unit correlation change index is recalculated to check whether it meets the preset correlation preservation threshold. If it still does not meet the threshold, the noise intensity parameters are adjusted again, and the noise injection and correlation calculation are repeated until the unit correlation change index meets the threshold requirement. Through this optimization mechanism, it is ensured that each semantic data unit can maintain good data correlation while protecting privacy.

[0241] Step S570: Collect all perturbation data units that have passed the evaluation, reorganize them in the order of the original data records, and generate a privacy-preserving dataset that includes the cell correlation change index, noise injection records, and privacy protection level. The dataset can be directly used for subsequent annotation tasks.

[0242] Collect all perturbation data units that passed the evaluation. These units, after noise injection, both preserved sensitive information and maintained good data correlation. Reassemble these perturbation data units according to the original data recording order to restore the original data structure and order.

[0243] Generate a privacy-preserving dataset that includes a cell association change index, noise injection records, and a privacy protection level. The cell association change index records the changes in the data association of each semantic data unit after noise injection; the noise injection records contain detailed information such as noise type, noise parameters, and injected fields, facilitating subsequent data traceability and analysis; the privacy protection level is determined based on the sensitive attribute labels of the feature clusters and the privacy protection requirement level, reflecting the degree of privacy protection in different parts of the dataset.

[0244] This privacy-preserving dataset can be directly used for subsequent data annotation tasks. Because the data in the dataset retains a certain degree of usability and relevance while protecting privacy, it can provide valuable information for annotation tasks, while ensuring that sensitive information is effectively protected. Through the above series of steps, data privacy protection based on data annotation is achieved, meeting the needs of data annotation and analysis while ensuring data privacy.

[0245] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A data privacy protection method based on data annotation, characterized in that, The method includes: The talent dataset is reconstructed by parsing the semantic description information of each data field to construct a semantic association graph. The core data fields and auxiliary data fields are identified based on the connection strength of the fields in the association graph. The original data records are divided into blocks based on the semantic coverage of the core data fields to obtain a semantic data unit set. Each unit in the semantic data unit set contains multiple data fields with semantic association. The semantic data unit set is subjected to feature clustering. A semantic association matrix is ​​constructed by calculating the semantic similarity between units. The semantic association matrix is ​​then subjected to clustering to generate an initial feature cluster set. Based on the preset sensitive attribute identification rules, a sensitive attribute matching operation is performed on each initial feature cluster to generate a feature cluster set containing sensitive attribute labels. Based on the semantic density value and field association strength value of each feature cluster in the feature cluster set containing sensitive attribute labels, the noise type and noise parameters corresponding to each feature cluster are calculated in combination with the sensitive attribute labels of the feature clusters, and a feature cluster noise parameter reference table is obtained. The semantic data unit set is mapped to the feature cluster noise parameter lookup table according to the feature cluster affiliation relationship. Corresponding noise generation and injection operations are performed on the feature cluster data fields with different sensitive attribute labels. The noise distribution characteristics between data fields in the same cluster are maintained by controlling the random seed synchronization mechanism in the noise generation process, and a perturbation data set is generated. The perturbation dataset is evaluated for correlation. The change rate of correlation between fields in semantic data units before and after perturbation is calculated to determine whether the preset correlation preservation threshold is met. When the threshold requirement is met, a privacy-preserving dataset that can be used for annotation tasks is output.

2. The method according to claim 1, characterized in that, The data format reconstruction operation of the talent dataset involves constructing a semantic association graph by parsing the semantic description information of each data field, identifying core data fields and auxiliary data fields based on the connection strength of the fields in the association graph, and segmenting the original data records based on the semantic coverage of the core data fields to obtain a set of semantic data units, including: Semantic metadata is extracted from each data field in the talent data set. The word structure of the field name, the description text of the data type, and the explanatory paragraph of the business scenario are parsed to generate a set of field semantic description information. The set of field semantic description information includes the terminology composition, contextual dependencies, and business affiliation dimensions of each data field. The terms in the semantic description information set of the field are semantically disambiguated. By comparing the explanations of polysemous terms in the domain terminology library and combining the business scenario description of the data field, a unique semantic orientation is determined, and a disambiguated field semantic feature vector is generated. Based on the disambiguated field semantic feature vectors, the co-occurrence locations and contextual overlap areas of different data fields in the business description text are identified, the co-occurrence frequency is counted and the proportion of contextual overlap length is calculated, and the semantic correlation value between fields is generated. A directed weighted semantic association graph is constructed based on the semantic association degree values ​​between the fields. The graph nodes are data fields, the directed edges represent the semantic association direction, and the edge weights are the semantic association degree values. The weights of the cyclic association paths in the graph are reduced. Traverse each node of the semantic association graph of the field, accumulate the sum of all incoming and outgoing edge weights of the node, compare the sum with the preset association strength benchmark value, mark the node that exceeds the benchmark value as the core data field, and mark the node that does not exceed the benchmark value as the auxiliary data field; Extract the semantic coverage description of the core data field, combine it with the semantic dependency relationship of the auxiliary data field, and dynamically divide the original data record into windows. The starting position of the window is the first appearance of the core data field, and the ending position of the window is the semantic coverage boundary of the core data field, generating a set of semantic data units containing a complete semantic context.

3. The method according to claim 2, characterized in that, The construction of a directed weighted semantic association graph based on the semantic association values ​​between the fields, where graph nodes are data fields, directed edges represent semantic association directions, and edge weights are semantic association values, includes weight decay processing for cyclic association paths in the graph, including: The semantic correlation values ​​between the fields are normalized to map the numerical range to a preset interval. Initialize the graph structure with data fields as nodes, add field identifiers, data types and business affiliation tags to each node, and establish a node attribute table to store basic node information; A set of directed edges is constructed based on the semantic association direction. Each directed edge contains a start node identifier, an end node identifier, and a normalized semantic association degree value. An edge relationship table is generated to record all association information. The directed edges in the edge relationship table are verified for direction. By comparing the data flow direction in the business process, the association direction that does not match the actual business logic is corrected so that the direction of the directed edges is consistent with the actual order of data use. Circular association paths in the graph are detected. Circular paths with a length greater than a preset threshold are identified by depth-first search. The weight of each directed edge in the circular path is attenuated. The attenuation coefficient is directly proportional to the length of the circular path. Extract the largest connected subgraph from the graph, retain the field associations related to the core business process, delete isolated nodes and directed edges with weights below the association threshold, and generate a simplified field semantic association graph.

4. The method according to claim 2, characterized in that, The process involves extracting the semantic coverage description of the core data field, combining it with the semantic dependencies of auxiliary data fields, and dynamically dividing the original data record into windows. The window starts at the first occurrence of the core data field and ends at the boundary of the semantic coverage of the core data field, generating a set of semantically encoded data units containing a complete semantic context, including: The semantic description information of the core data fields is parsed, and the domain terms and business entities are extracted. The starting and ending indices of the semantic coverage are determined by the distribution of the terms in the data records. Identify the auxiliary data fields corresponding to the core data fields, query the auxiliary data fields directly associated with the core data fields through the field semantic association graph, and record the semantic dependency direction and dependency strength of the auxiliary data fields; Based on the semantic dependency direction of the auxiliary data field, adjust the semantic coverage boundary of the core data field. When the auxiliary data field is located after the core data field and the dependency strength exceeds the dependency threshold, extend the termination index to the semantic coverage boundary of the auxiliary data field. Initialize the dynamic window parameters. The window size is the sum of the length of the semantic coverage of the core data field and the coverage length of the associated auxiliary data field. The window sliding step is the average length of the semantic coverage of the core data field. Starting from the beginning of the original data record, data is extracted according to the dynamic window parameters. When the window contains at least one core data field and all associated auxiliary data fields, it is marked as a valid window, and the data within the window is extracted as candidate semantic data units. Boundary optimization is performed on candidate semantic data units. By comparing the semantic overlap areas of adjacent units, units with semantic overlap exceeding the overlap threshold are merged, and units containing multiple independent core data fields are split to generate the final set of semantic data units.

5. The method according to claim 1, characterized in that, The step involves performing feature cluster partitioning on the semantic data unit set, constructing a semantic association matrix by calculating the semantic similarity between units, performing cluster partitioning on the semantic association matrix to generate an initial feature cluster set, and performing sensitive attribute matching on each initial feature cluster based on preset sensitive attribute identification rules to generate a feature cluster set containing sensitive attribute labels, including: Multimodal feature extraction is performed on each unit in the semantic data unit set. The text description features, numerical distribution features and business association features of the data fields within the unit are extracted. The three types of features are concatenated into a unit comprehensive feature vector, and the vector dimension is adjusted according to the number of features. Calculate the similarity between the comprehensive feature vectors of any two semantic data units. By comparing the cosine value of word vectors of text description features, the difference of cumulative distribution function of numerical distribution features, and the path matching degree of business-related features, a weighted comprehensive semantic similarity value between units is generated. A semantic association matrix is ​​constructed based on the comprehensive semantic similarity value between the units. The rows and columns of the matrix correspond to semantic data units, and the matrix elements are the comprehensive semantic similarity values ​​of the corresponding unit pairs. The matrix is ​​then symmetrically processed. The semantic association matrix is ​​subjected to cluster partitioning. The units with the highest similarity values ​​are iteratively merged. After merging, the comprehensive semantic similarity value of the new unit with other units is recalculated until the similarity value of all units is lower than the merging threshold, thus obtaining the initial feature cluster set. Traverse each cluster in the initial feature cluster set, extract the description text and sample data of all data fields in the cluster, and perform field-by-field matching with the preset sensitive attribute identification rules. The rules include keyword matching patterns, semantic patterns and data distribution patterns. When the number of successfully matched sensitive attribute identification rules within a cluster exceeds the rule threshold, a corresponding sensitive attribute label is added to the cluster. If multiple sensitive attribute labels are matched, the main label is determined based on the sum of the rule matching degrees, and a feature cluster set containing a cluster identifier, a list of member units, and sensitive attribute labels is generated.

6. The method according to claim 5, characterized in that, The semantic association matrix is ​​subjected to cluster partitioning, which involves iteratively merging the pairs of units with the highest similarity values. After merging, the combined semantic similarity value of the new unit with other units is recalculated until the similarity value of all units is lower than the merging threshold, resulting in an initial feature cluster set, including: The semantic association matrix is ​​preprocessed by replacing zero values ​​in the matrix with a preset proportion of minimum similarity values, and sorting the similarity values ​​in each row in descending order. Initialize the cluster set, add each semantic data unit as an independent cluster to the cluster set, record the list of member units of each cluster and the cluster center feature vector, and the cluster center feature vector is the average value of the comprehensive feature vectors of the member units; Find the pair of units with the maximum similarity value in the semantic association matrix, extract the two clusters to which the pair of units belongs, calculate the similarity value between the cluster center feature vectors of the two clusters, and if the value exceeds the cluster merging threshold, perform a merging operation; when merging two clusters, merge the member unit list into the member list of the new cluster, recalculate the cluster center feature vector of the new cluster, update all similarity values ​​related to the new cluster in the semantic association matrix, and delete the relevant records of the original two clusters; Repeat the maximum similarity search and cluster merging operation, updating the cluster set and semantic association matrix after each merge, until all similarity values ​​in the matrix are below the cluster merging threshold, at which point the merging process stops; The merged cluster set is subjected to cluster size verification. Clusters with more than the preset upper limit of member units are split into multiple sub-clusters. The splitting is based on the distribution of the comprehensive semantic similarity value of the units within the cluster, so that the similarity of units within the sub-cluster is higher than the similarity between sub-clusters, thus generating the final initial feature cluster set.

7. The method according to claim 1, characterized in that, The semantic density value and field association strength value of each feature cluster in the feature cluster set containing sensitive attribute labels are used to calculate the noise type and noise parameters corresponding to each feature cluster, resulting in a feature cluster noise parameter lookup table, including: For each cluster in the feature cluster set containing sensitive attribute tags, the number of terms in the semantic description text of the data fields within the cluster is counted, the ratio of the number of terms to the total number of fields is calculated, and the semantic density value of the cluster is generated by combining the importance weight of the terms in the domain terminology library. Extract the subgraph structure of each feature cluster in the field semantic association graph, accumulate the weight values ​​of all directed edges in the subgraph, and use the ratio of the accumulated result to the number of nodes in the subgraph as the field association strength value of the cluster, which represents the overall degree of association between fields within the cluster. The privacy protection requirement level corresponding to the sensitive attribute labels of the feature clusters is analyzed. Different labels correspond to different privacy protection strength requirements. The level classification is based on the degree of risk of leakage of the sensitive information identified by the label. Based on the privacy protection requirement level, semantic density value, and field association strength value, the noise type corresponding to the feature cluster is determined. The noise type includes replacement type, addition type, perturbation type, and combination type. The selection criterion is the comprehensive impact of noise on privacy protection effect and data availability. For a given noise type, calculate the corresponding noise parameters. Replacement noise parameters include replacement probability and size of the replacement value set; addition noise parameters include noise mean and variance; and disturbance noise parameters include disturbance amplitude and disturbance correlation. The parameter values ​​are directly proportional to the semantic density value and the field association strength value. The feature cluster identifier, sensitive attribute label, noise type, and calculated noise parameters are integrated into a record to generate a feature cluster noise parameter lookup table. Each record in the table contains complete noise generation configuration information.

8. The method according to claim 7, characterized in that, For each cluster in the feature cluster set containing sensitive attribute tags, the number of terms in the semantic description text of the data fields within the cluster is counted, the ratio of the number of terms to the total number of fields is calculated, and the semantic density value of the cluster is generated by combining the importance weight of the terms in the domain terminology library, including: Extract the semantic description text of all data fields within the feature cluster, perform word segmentation, and generate a set of field terms; The terminology set is matched with terms in the domain terminology library to identify words that belong to the core domain terms. The importance weight of each matched term in the domain terminology library is recorded. The weight value reflects the degree of generality and professional relevance of the term in the domain. The number of core domain terms in the field terminology set is counted, and the ratio of this number to the total number of data fields in the cluster is calculated to obtain the basic term density value. When the total number of fields is zero, the basic term density value is set to zero. The term density base value is weighted by summing the importance weights of the core terms in each domain and multiplying the sum by the term density base value to obtain the weighted density value. The higher the weight of the term, the greater its contribution to the weighted density value. The uniformity of term distribution in the term set of the field is detected, and the standard deviation of the term distribution in the semantic description text of different data fields is calculated. The smaller the standard deviation, the more uniform the distribution. The weighted density value is corrected for uniformity. The more uniform the distribution, the larger the correction coefficient. The corrected weighted density values ​​are mapped to a preset semantic density value range to generate the final cluster semantic density value. The range ensures that the semantic density values ​​of different feature clusters are comparable, and the mapping process preserves the relative relationship of the original weighted density values.

9. The method according to claim 1, characterized in that, The process involves mapping the semantic data unit set to the feature cluster noise parameter lookup table according to the feature cluster affiliation relationship, performing corresponding noise generation and injection operations on the feature cluster data fields with different sensitive attribute labels, maintaining the noise distribution characteristics among data fields in the same cluster by controlling the random seed synchronization mechanism during the noise generation process, and generating a perturbed data set, including: Traverse each unit in the semantic data unit set, extract the identifiers of all data fields within the unit, query the feature cluster identifiers to which the fields belong, generate a unit-field-cluster attribution table, and record the cluster attribution information of each data field; Based on the unit-field-cluster attribution table and the feature cluster noise parameter comparison table, match the corresponding noise type and noise parameter for each data field to generate a field noise configuration table. The configuration table includes field identifier, noise type, noise parameter and sensitive attribute label. Initialize the random number generator and assign a unique basic random seed to each feature cluster. The basic random seed is generated by combining the hash value of the feature cluster identifier with the current timestamp to ensure that the seeds of different clusters are unique. According to the noise type of the data field, the corresponding noise generation operation is performed. Replacement noise generates random replacement values ​​that conform to the distribution of the replacement value set. Addition noise generates random noise values ​​that conform to the mean and variance requirements. Disturbance noise generates disturbance vectors that conform to the disturbance amplitude and correlation. During the noise generation process, the data fields in the same cluster use a derived seed from the basic random seed. The derived seed is generated by performing an XOR operation between the basic random seed and the hash value of the field identifier, so that the noise distribution of the fields in the same cluster has a preset correlation. The generated noise values ​​are injected into the original data field values. Replacement noise directly replaces the original value, additive noise is added to the original value, and perturbation noise is vector-operated with the original value. After injection, the data fields are reorganized into semantic data units to generate a perturbation data set containing noise injection records.

10. A computer system, characterized in that, include: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the data privacy protection method based on data annotation as described in any one of claims 1 to 9.

Images