A utility-enhanced conditional feature selection differential privacy data publishing method

Abstract

The utility model discloses a kind of utility enhanced conditional characteristic selection differential privacy data publishing method, the method is first to the original data set is preprocessed, then utilize conditional mutual information to the data set after pre-processing is associated with feature selection;Again normalization data, the data value corresponding to the feature that meets threshold condition is micro-aggregated and handled, obtain several cluster of the size k of scale, and using profile coefficient calculation obtains the local optimum k value;Then, according to the redefined feature dependence sensitivity, add the noise that meets condition to each cluster, reassign privacy budget to realize differential privacy;Finally, for the data after disturbance is published, the data published can be counted and classified etc. Task analysis is analyzed. The utility model can resist individual data privacy attack of enemy with strong background knowledge, improve the availability of data publishing under the premise that individual sensitive data is guaranteed privacy.

Description

Technical Field

[0001] This invention relates to the field of information security and privacy technology, and in particular to a utility-enhanced conditional feature selection differential privacy data publishing method. Background Technology

[0002] The rapid development of information sharing and knowledge exchange has led to an explosive growth in data, which (personal salaries, medical records, consumption habits, preferences, etc.) often contains a large amount of sensitive information. Service providers, in order to provide more accurate services, are increasingly eager to collect and analyze individual data, thus creating privacy threats to individuals or organizations. In reality, the privacy protection of published data lies in the correspondence between sensitive user data and individual identity. Tasks such as querying and analyzing published data only require blocking the identifier between sensitive individual information and identity.

[0003] To protect user privacy, k Traditional privacy models, such as anonymity and its extensions, have been studied and used extensively. However, due to the inability to ascertain the attacker's background knowledge, the emergence of new attack methods has exposed vulnerabilities in traditional privacy protection models, consistently demonstrating their limitations in protecting sensitive information. Furthermore, various correlations always exist between data or features, and these correlations can lead to serious privacy leaks through re-identification techniques such as multi-table joins. In response to the ever-evolving attack methods and the shortcomings of existing privacy protection mechanisms, the Dwork team at Microsoft Research proposed the differential privacy model. Differential privacy is a privacy protection model built on a solid mathematical foundation, rigorously defining privacy protection effectiveness and providing a method for quantitative evaluation. By adding perturbations to the data, it protects potentially sensitive user information in the published data. Even if an attacker possesses information about all records in the dataset except for a specific one, the attacker still cannot deduce the original data.

[0004] The mathematical definition of differential privacy is as follows: A random algorithm A satisfies e - Differential privacy, if and only if for all neighboring databases D 1 and D In section 2, all possible outputs of algorithm A S Range(A) satisfies the inequality Pr(A( D 1)∈S)≤exp( e )×Pr(A( D 2)). Among them, D 1 and D The difference between the two is only one record. e A value ≥ 0 represents a privacy budget, which can measure the degree of privacy protection for data. eThe closer the value is to 0, the higher the level of privacy protection. However, a higher degree of data perturbation also leads to greater errors. Furthermore, the level of data privacy protection is also related to the sensitivity of the query or classification algorithm; higher sensitivity requires a higher noise scale for differential privacy, resulting in greater errors. In practical applications, achieving strict differential privacy often results in high data distortion and low usability. Therefore, balancing data privacy and usability is a significant challenge in differential privacy data publishing and data mining scenarios.

[0005] To address the challenges of strong privacy protection and low data utility in differential privacy data publishing, a new model is combined with existing machine learning models to leverage the advantages of both and work synergistically. On one hand, this model can reduce the query sensitivity of privacy algorithms by utilizing the characteristics of machine learning models, thereby improving data usability. On the other hand, the algorithm can realize the basic definition of differential privacy, blocking the identification of sensitive individual information and identity. Applying this model to real-world datasets generates perturbed data for publication, which improves data utility while protecting sensitive individual information. Summary of the Invention

[0006] The purpose of this invention is to provide a utility-enhanced conditional feature selection differential privacy data publishing method, which on the one hand ensures strong privacy protection for user data; on the other hand, it can improve data availability by reducing the sensitivity of privacy algorithms, thereby improving the availability of tasks such as statistical queries and classification analysis.

[0007] Technical solution: A utility-enhanced conditional feature selection differential privacy data publishing method, comprising the following steps:

[0008] S1. Preprocess the original dataset, including removing outliers and adding missing values;

[0009] S2. Separately process numerical data and categorical data. Discretize continuous numerical data and map and encode categorical data based on word vectors. Extract the feature domain of categorical data from WordNet 2.1.

[0010] S3. Divide the dataset into a sensitive feature set S and a quasi-identifier feature set Q based on the feature sensitivity of the dataset; calculate the conditional mutual information between the quasi-identifier feature set Q and the sensitive feature set S, and calculate the mutual information threshold. ;

[0011] S4. Select those that meet the threshold. The required feature set is used to normalize the data values, and micro-aggregation is performed on the data corresponding to the feature set. Different distance metrics are selected when micro-aggregating different types of data.

[0012] S5. Micro-aggregation yields several clusters of size... k cluster { C 1,…, C n / k The contour coefficient is used to measure local optima. k value;

[0013] S6, to n / k By adding perturbations to each cluster and reallocating the privacy budget, differential privacy is achieved, resulting in a dataset ready for publication that can be used for querying and classification tasks.

[0014] Based on the above scheme, further supplementary explanations are provided below.

[0015] In step S1, the original data undergoes preprocessing for missing and outlier values, and is divided into sensitive features and quasi-identifier features. Conditional mutual information selection is performed based on these features. Features with strong correlation that meet threshold requirements are selected, and micro-aggregation operations are performed on the corresponding categorical and numerical data. The unaggregated data is perturbed, and differential privacy Laplace and exponential mechanisms are implemented according to different data types to obtain data ready for publication that can be used for statistical queries or classification analysis. The original dataset is preprocessed by deleting or adding certain missing and outlier values.

[0016] Step S2: Process numerical data and categorical data separately. Discretize continuous numerical data and map and encode categorical data. Divide the dataset into a sensitive feature set S and a quasi-identifier feature set Q according to the feature sensitivity of the dataset.

[0017] Step S3: Calculate the conditional mutual information between the quasi-identifier feature set Q and the sensitive feature set S, and set the mutual information threshold. The formulas for calculating conditional mutual information and mutual information threshold are as follows:

[0018] Suppose that in a given dataset D Chinese characteristics When the value is [value], it is determined by other quasi-identifier characteristics. Knowledge about sensitive features The reduction in uncertainty. Then any feature and Given features The conditional mutual information at that time is as follows:

[0019] .

[0020] After obtaining the conditional mutual information between any quasi-identifier feature and the sensitive feature, all feature pairs need to undergo feature selection. The main process involves comparing the conditional mutual information of the feature pairs with a threshold. The formula for calculating the threshold is:

[0021] ,

[0022] in (0≤ ≤1) is a flexible parameter that determines the expected relevance; setting it... =0.25; Characteristics of quasi-identifiers domain, Indicates sensitive features The domain.

[0023] For step S4, select the feature set that meets the threshold requirement, and then calculate the conditional mutual information threshold based on the threshold obtained in each step. The greater the conditional mutual information, the stronger the dependency between features, and the more features are selected that exceed the threshold. All feature sets Normalization processes data values, standardizing them to [0,1]. The normalization formula is as follows:

[0024] ;

[0025] in, Let represent the i-th value of a certain feature, and let the domain range of the feature be []. ], This represents the normalized value data, where all feature values ​​are standardized to [0,1] using a standardization formula.

[0026] Step S5: Perform micro-aggregation on the data corresponding to the selected feature set. Micro-aggregation yields several clusters of size [missing information]. k cluster { C 1,…, C n / k Using the silhouette coefficient to measure local optima k value.

[0027] Micro-clustering is a clustering-like approach that iteratively creates at least... k Clusters of elements C i ( i =2,…, n / kThe elements within each cluster should be as similar as possible, while the elements between different clusters should be as different as possible. Each cluster can select a representative record to represent its centroid, and use the centroid to replace other values ​​within the cluster. The specific implementation steps of micro-aggregation are as follows:

[0028] 1) Determine the dataset D Is the size greater than 3? k Calculate the dataset D Average record ;

[0029] 2) Select distance Record the furthest data record ;

[0030] 3) Select distance record The furthest data record ;

[0031] 4) with With the center of mass as the distance Recent k - 1 record constitutes a cluster ;

[0032] 5) with With the center of mass as the distance Recent k - 1 record constitutes a cluster

[0033] 6) From D Cross out the well-aggregated clusters:

[0034] 7) Determine the remaining dataset D Is it greater than 2? k Calculate the average number of records in the remaining dataset.

[0035] 8) Iterative calculation steps 1)-6)

[0036] 9) D The last remaining k (or insufficient) k The values ​​are divided into a cluster.

[0037] Among them, micro-aggregates yield several clusters of size... k cluster { C 1,…, C n / k The contour coefficient is used to measure local optima. k The value includes the following implementation process: given a value with n A dataset of records D Micro-aggregates n / kThere are clusters, and for each instance in a cluster... The contour coefficients are all:

[0038]

[0039] in, A ( a i )express The average distance to instances within the cluster. B ( )express To not include The minimum distance to other intra-cluster instances, S ( )∈[-1, 1], B ( )>> A ( )hour, S ( The closer the profile coefficient is to +1, the higher the intra-cluster cohesion and the lower the inter-cluster coupling. The selection of the profile coefficient should minimize the inclination of the cluster. S ( The largest k value.

[0040] Step S6, according to the different data types, n / k For each cluster, perturbations are added to achieve differential privacy: for numerical data, Euclidean distance is used for distance measurement during micro-clustering, and the Laplacian mechanism is used to perturb the data to achieve differential privacy; for categorical data, the feature values ​​are mapped to ontology knowledge (WordNet) to obtain a minimal hierarchical structure. When performing micro-aggregation of data values, a semantic distance metric is used, and an exponential mechanism is employed to perturb the data to achieve differential privacy. The resulting dataset, suitable for querying and classification, comprises the following steps:

[0041] The detailed process of implementing the differential privacy Laplace mechanism using the numerical data is as follows:

[0042] 1) Global sensitivity: , These are two adjacent datasets that differ by only one record.

[0043] 2) Laplace mechanism definition: For any query , Indicates privacy budget;

[0044] 3) Calculation of the sensitivity required for each cluster after micro-aggregation: ,in n For dataset DNumber of data records k The number of records for each cluster.

[0045] The detailed process of implementing the differential privacy index mechanism using the semantic distance definition and category data is as follows:

[0046] 1) Definition of semantic distance: The domain of feature A is For the two class values ​​of feature A and semantic distance Logarithmic function of non-common ancestor:

[0047] ,

[0048] In order to be in The eigenvalues ​​of the middle are The ancestor set of the classification has semantic distance that conforms to the mathematical distance measurement and satisfies the properties of nonnegativity, symmetry, reflexivity and triangle inequality.

[0049] 2) Definition of the index mechanism: Provide a scoring function. Global sensitivity is , For two adjacent datasets that differ by only one record, The output is an element from a set of discrete data. .

[0050] Specifically, a new adjustment has been made to the redistribution of the privacy budget, changing the traditional method of evenly distributing the total privacy budget to multidimensional features to a weighted allocation of the privacy budget. Let the original dataset be... D have d Dimensional features, total privacy budget is e After the differential privacy algorithm of conditional feature selection micro-aggregation, the weight of each feature is calculated. By redistributing the weighted privacy budget to the selected features based on their weight values, privacy data can be protected more effectively.

[0051] Beneficial effects: Compared with existing technologies, this invention employs a conditional feature-selective micro-aggregation differential privacy data publishing method, which can resist individual data privacy attacks by adversaries with strong background knowledge. This invention can effectively reduce the sensitivity of privacy algorithms while ensuring the privacy of individual sensitive data, thereby improving the usability of data publishing. Furthermore, this method can improve query accuracy and classification precision in tasks such as counting queries and data classification. Attached Figure Description

[0052] Figure 1This is a schematic diagram illustrating the interaction steps of the user data and privacy model provided by the present invention.

[0053] Figure 2 A schematic diagram illustrating the implementation process of the differential privacy data publishing model for enhancing utility through conditional feature selection provided by this invention. Detailed Implementation

[0054] To enable those skilled in the art to fully understand and implement the technical solutions provided by this invention, the invention will be further described below in conjunction with the accompanying drawings.

[0055] Background setting of the embodiments of the present invention: Trusted third-party servers collect user datasets, such as population census datasets and medical datasets collected by government departments. These datasets contain a large amount of sensitive information about individual users, and direct release would lead to the leakage of a large amount of sensitive information. At the same time, these data have great analytical value. For example, the group analysis of infectious diseases in medical data can have a positive effect on human health development, and it is necessary to maintain a high degree of data availability.

[0056] This invention provides a utility-enhanced conditional feature selection differential privacy data publishing method. First, the original data undergoes preprocessing for missing and outlier values, and is divided into sensitive features and quasi-identifier features. Conditional mutual information selection is then performed based on these sensitive and quasi-identifier features. Features with strong correlation that meet threshold requirements are selected, and micro-aggregation operations are performed on the corresponding categorical and numerical data. The unaggregated data is perturbed, and differential privacy Laplace and exponential mechanisms are implemented according to different data types, thereby obtaining data to be published that can be used for statistical queries or classification analysis.

[0057] See Figure 1 and Figure 2 The implementation process of the method described in this invention will be explained in detail below.

[0058] Figure 1 This is a schematic diagram of the user data and privacy model interaction process provided by the present invention; the specific steps include:

[0059] In step 1, a trusted third-party server categorizes and organizes the collected raw data. This data protects a large amount of user-submitted personal information, including both sensitive and non-sensitive data. Publishing this data directly without processing would harm users' privacy; therefore, it is necessary to protect user privacy data.

[0060] In practice, users submit their individual data to a trusted third-party server. The server then categorizes and organizes the collected raw user data to generate a true raw dataset. D。

[0061] In step 2, existing machine learning model algorithms and differential privacy models are integrated.

[0062] In practice, machine learning algorithms such as feature selection, generalization, supervised classification, and unsupervised clustering can effectively perform data analysis tasks, but these algorithms do not adequately address the issue of data privacy leakage. Differential privacy offers stronger privacy guarantees for data, and its mathematical definition is as follows:

[0063] A random algorithm A satisfies e - Differential privacy, if and only if for all neighboring databases D 1 and D In section 2, all possible outputs of algorithm A S Range(A) satisfies the inequality Pr(A( D 1)∈S)≤exp( e )×Pr(A( D 2)). Among them, D 1 and D The difference between the two is only one record. e A value ≥ 0 represents a privacy budget, which can measure the degree of privacy protection for data. e The closer a value is to 0, the higher the level of privacy protection; however, the higher the degree of data perturbation, the greater the error. When... e For a small value, differential privacy can easily lead to significant data distortion. Therefore, for specific tasks, integrating machine learning models and differential privacy to work synergistically can leverage the advantages of both models and address more privacy issues. In this invention, conditional feature selection and micro-aggregation are first performed on the dataset, and then different differential privacy mechanisms are implemented for different types of data to generate publishable data.

[0064] In step 3, the integrated privacy model is applied to the raw data to generate a dataset that can be published. D ′.

[0065] In practice, the dataset to be published is generated by the perturbation model. D ′, D A statistical distribution that approximates the original dataset as closely as possible can be used for querying or classifying statistical data. Wherein: the data queryer directs the data to the dataset... D Send some query requests, D It responds to query requests and returns the query results to the data queryer.

[0066] See Figure 2 This is a schematic diagram illustrating the implementation process of the utility-enhanced conditional feature selection differential privacy data publishing model provided by the present invention.

[0067] In this embodiment of the invention, the original dataset is first preprocessed, such as by deleting or adding certain missing values ​​or outliers; the data type of the original dataset is determined, and numerical data and categorical data are processed separately. Continuous numerical data is discretized, and categorical data is mapped and encoded. The dataset is then divided into a sensitive feature set S and a quasi-identifier feature set Q according to the feature sensitivity of the dataset.

[0068] In this embodiment of the invention, the conditional mutual information between the quasi-identifier feature set Q and the sensitive feature set S is calculated, and a mutual information threshold is set. Calculate according to the following formula:

[0069] ,

[0070] in Given the features, find the features of other quasi-identifiers. With sensitive features The conditional mutual information. The threshold formula for calculating conditional mutual information is:

[0071] ,

[0072] in (0≤ ≤1) is a flexible parameter that determines the expected relevance; setting it... =0.25; Characteristics of quasi-identifiers domain, Indicates sensitive features The domain.

[0073] In this embodiment of the invention, the greater the conditional mutual information, the stronger the dependency relationship between features, and features with a mutual information greater than the threshold are selected. All feature sets Normalization processes data values, standardizing them to [0,1]. The normalization formula is as follows:

[0074] ,

[0075] in, Let i represent the i-th value of a certain feature, and let the domain range of the feature be [ ], This represents the normalized value data, where all feature values ​​are standardized to [0,1] using a standardization formula.

[0076] In this embodiment of the invention, the micro-aggregation step can iteratively create at least k Clusters of elements C i ( i=2, …, n / k The elements within each cluster should be as similar as possible, while the elements between different clusters should be as different as possible. Each cluster can select a representative record to represent its centroid, and use the centroid to replace other values ​​within the cluster. The specific implementation steps of micro-aggregation are as follows:

[0077] 1) Determine the dataset D Is the size greater than 3? k Calculate the dataset D Average record ;

[0078] 2) Select distance Record the furthest data record ;

[0079] 3) Select distance record The furthest data record ;

[0080] 4) with With the center of mass as the distance Recent k - 1 record constitutes a cluster ;

[0081] 5) with With the center of mass as the distance Recent k - 1 record constitutes a cluster

[0082] 6) From D Cross out the well-aggregated clusters: ;

[0083] 7) Determine the remaining dataset D Is it greater than 2? k Calculate the average number of records in the remaining dataset.

[0084] 8) Iterative calculation steps 1) - 6).

[0085] 9) D The last remaining k (or insufficient) k Each set of values ​​is divided into a cluster.

[0086] In this embodiment of the invention, the cluster size in the micro-aggregation step k The choice of value affects the overall result of the algorithm; therefore, it is necessary to select the optimal value. k Value. After obtaining several values ​​of size... k cluster { C 1,…, C n / kAfter that, the silhouette coefficient is used here to measure the local optimum. k Value. Given a value with n A dataset of records D Micro-aggregates n / k There are clusters, and for each instance in a cluster... The contour coefficients are all:

[0087] ,

[0088] in, A ( a i )express The average distance to instances within the cluster. B ( )express To not include The minimum distance to other intra-cluster instances, S ( )∈[-1, 1], B ( )>> A ( )hour, S ( The closer a value is to +1, the higher the intra-cluster cohesion and the lower the inter-cluster coupling. This is based on the silhouette coefficient. S ( The calculated value of ) is selected to make S ( The largest k value.

[0089] In this embodiment of the invention, the data types are adjusted accordingly. n / k Differential privacy perturbations are added to each cluster. For numerical data, Euclidean distance is used as the distance metric during micro-clustering, and a Laplacian perturbation mechanism is used to achieve differential privacy. For categorical data, feature values ​​are mapped to ontology knowledge (WordNet), resulting in a minimal hierarchical structure. When performing micro-aggregation of data values, semantic distance is used as a metric, and an exponential mechanism is used to perturb the data to achieve differential privacy.

[0090] In this embodiment of the invention, after obtaining the dataset to be published, it can be used for querying and classification. The specific perturbation process is as follows:

[0091] The detailed process of implementing the differential privacy Laplace mechanism using the numerical data is as follows:

[0092] 1) Global sensitivity: , These are two adjacent datasets that differ by only one record.

[0093] 2) Laplace mechanism definition: For any query , Indicates privacy budget;

[0094] 3) Calculation of the sensitivity required for each cluster after micro-aggregation: ,in n For dataset D Number of data records k The number of records for each cluster.

[0095] The detailed process of implementing the differential privacy index mechanism using the semantic distance definition and category data is as follows:

[0096] 1) Definition of semantic distance: The domain of feature A is For the two class values ​​of feature A and semantic distance Logarithmic function of non-common ancestor:

[0097] ,

[0098] In order to be in The eigenvalues ​​of the middle are The ancestor set of the classification has semantic distance that conforms to the mathematical distance measurement and satisfies the properties of nonnegativity, symmetry, reflexivity and triangle inequality.

[0099] 2) Definition of the index mechanism: Provide a scoring function. Global sensitivity is , For two adjacent datasets that differ by only one record, The output is an element from a set of discrete data. .

[0100] In this embodiment of the invention, a new adjustment is made to the problem of privacy budget redistribution, changing the traditional equal allocation of the total privacy budget to multidimensional features to a weighted privacy budget allocation. The specific steps are as follows: Let the original dataset D have... d Dimensional features, total privacy budget is e After the differential privacy algorithm of conditional feature selection micro-aggregation, the weight of each feature is calculated. By redistributing the weighted privacy budget to the selected features based on their weight values, privacy data can be protected more effectively.

[0101] The following are experimental results of the utility-enhanced conditional feature selection differential privacy data publishing method.

[0102] The dataset used in the experiment was the Adult dataset, a dataset from the UCI Machine Learning Library. After processing outliers and missing values ​​in the original dataset, the dataset contained 30,162 data records and 15 features. Ten privacy levels and privacy budgets were implemented in the experiment. e The values ​​were set as follows: 0.2, 0.4, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0. Furthermore, the method described in this invention requires calculating the cluster size after micro-aggregation, which has been experimentally verified on the Adlut dataset. k =59 is a local optimum.

[0103] The experimental results of the utility-enhanced conditional feature selection differential privacy data publishing method on the Adult dataset are shown in Tables 1, 2, and 3. The main metrics are the application of the method described in this invention on real-world datasets.

[0104] 1. Under different privacy budgets e In this case, the mean absolute error (MAE) of the datasets before and after the privacy perturbation is compared, as shown in Table 1. The formula for calculating MAE is as follows:

[0105]

[0106] in, The query result is the actual data for the i-th query. The query results are after differential privacy perturbation.

[0107] Different privacy budgets e In this embodiment, 5000 balanced sample data records were selected from the original dataset. SVM and Logistic Regression, two machine learning classification methods, were used to measure the classification accuracy of the proposed method, as shown in Tables 2 and 3. The term "no differential privacy" in the tables indicates a method that only performs micro-aggregation on the original data without adding differential privacy perturbations; therefore, this method is not subject to privacy budgets. e The influence of this remains constant. As can be seen from the table, the method of this invention maintains high accuracy for machine learning classification problems and significantly improves data utility.

[0108] Table 1. MAE Measurement Results under Different Privacy Budgets in the Adult Dataset

[0109]

[0110] Table 2. SVM classification results under different privacy budgets in the Adult dataset.

[0111]

[0112] Table 3. Logistic Regression Results for Different Privacy Budgets in the Adult Dataset

[0113]

[0114] Overall, the utility-enhanced conditional feature selection differential privacy data publishing method described in this invention combines relevant models from machine learning with differential privacy to work synergistically. It first performs conditional feature selection and micro-aggregation operations on the dataset, and then implements different differential privacy mechanisms for different types of data, thereby generating publishable data. Thus, on the one hand, user data can be strongly protected in terms of privacy; on the other hand, data usability can be improved by reducing the sensitivity of privacy algorithms, thereby enhancing the usability of tasks such as statistical queries and classification analysis.

[0115] In the process of platform construction and implementation, the method described in this invention can also adopt the following system structure.

[0116] It includes: a data processing unit, a model perturbation unit, and a statistical analysis unit, wherein:

[0117] The data processing unit is used by the central server to integrate, classify, and preprocess user data, classifying the data according to features into identifier features, quasi-identifier features, and sensitive features;

[0118] The model perturbation unit is used to integrate relevant models in machine learning with differential privacy to work together. It performs conditional feature selection on the input data, selects feature data that meets the threshold requirements for micro-aggregation operation, and processes different types of data separately during micro-aggregation and perturbs different data separately to realize the differential privacy Laplace mechanism or exponential mechanism.

[0119] The statistical analysis unit generates publishable data based on the model perturbation unit, which can be used for statistical counting queries and classification analysis.

Claims

1. A utility-enhanced conditional feature selection differential privacy data publishing method, characterized in that: Includes the following steps: S1. Preprocess the original dataset, including removing outliers and adding missing values; S2. Separately process numerical data and categorical data. Discretize continuous numerical data and map and encode categorical data based on word vectors. Extract the feature domain of categorical data from WordNet2.

1. S3. Divide the dataset into a sensitive feature set S and a quasi-identifier feature set Q based on the feature sensitivity of the dataset; calculate the conditional mutual information between the quasi-identifier feature set Q and the sensitive feature set S, and calculate the mutual information threshold. ; S4. Select those that meet the threshold. The required feature set is used to normalize the data values, and micro-aggregation is performed on the data corresponding to the feature set. Different distance metrics are selected for micro-aggregation of different types of data. The specific steps include the following: S41. Based on the conditional mutual information threshold obtained in each calculation. The greater the conditional mutual information, the stronger the dependency between features; select all feature sets greater than the threshold. , , Dimensions representing features; S42. Normalize the data values ​​of each feature using the following formula: , in, Indicates the first feature Individual data points, the domain range of the features is [ ], This represents the normalized value data, where all feature values ​​are standardized to [0,1] using a standardization formula; S43, Micro-aggregates are created iteratively by creating at least Clusters of elements , Elements within each cluster are as similar as possible, elements between different clusters are as different as possible, and each cluster can select a representative record to represent the centroid and use the centroid to replace other values ​​within the cluster; S5. Micro-aggregation yields several clusters of size... clusters Using the silhouette coefficient to measure local optima The value includes the following steps: Given with A dataset of records Micro-aggregates There are clusters, and for each instance in a cluster... The contour coefficients are all: express The average distance to instances within the cluster. B ( )express To not include The minimum distance to other intra-cluster instances, , B ( )>> A ( )hour, S ( The closer the profile coefficient is to +1, the higher the intra-cluster cohesion and the lower the inter-cluster coupling. Therefore, choosing a profile coefficient that minimizes the inward cohesion of the cluster and minimizes the outward coupling of the cluster is beneficial. S ( The largest k value; S6, to By adding perturbations to each cluster and reallocating the privacy budget, differential privacy is achieved, resulting in a dataset ready for publication that can be used for querying and classification tasks.

2. The utility-enhanced conditional feature selection differential privacy data publishing method according to claim 1, characterized in that: In step S2, after extracting the feature domain of the categorical data, the feature values ​​in the features are mapped to the ontology knowledge WordNet2.1, resulting in a minimum hierarchical structure. ; For feature domain extraction, for each classification feature, the feature domain is captured and modeled by ontology to measure the semantic similarity between two concepts; the ontology for each feature domain includes extraction from existing knowledge sources, as well as creation of the ontology by generalizing and classifying the concepts in the feature domain.

3. The utility-enhanced conditional feature selection differential privacy data publishing method according to claim 1, characterized in that: In step S6, for n / k Differential privacy is achieved by adding perturbations to each cluster and reallocating the privacy budget, resulting in a dataset ready for publication that can be used for querying and classification tasks. The process includes the following steps: S61. For numerical data, Euclidean distance is used for distance measurement during micro-aggregation, and Laplace mechanism is used to perturb the data to achieve differential privacy. S62. For categorical data, map the feature values ​​in the features to WordNet 2.1 ontology knowledge to obtain a minimum hierarchical structure. When performing micro-aggregation of data values, a semantic distance metric is used, and an exponential mechanism is used to perturb the data to achieve differential privacy. S63. A new adjustment has been made to the redistribution of the privacy budget, changing the traditional equal allocation of the total privacy budget to multidimensional features to a weighted allocation of the privacy budget. The specific steps are as follows: Let the original dataset be... D have d Dimensional features, total privacy budget is ε After the differential privacy algorithm of conditional feature selection micro-aggregation, the weight of each feature is calculated. The weighted privacy budget is redistributed to the selected features based on their weight values, thereby protecting privacy data.

Images