CN113254988B - High-dimensional sensitive data privacy classified protection publishing method, system, medium and equipment - Google Patents

Abstract

The invention belongs to the technical field of information security data release, and discloses a method, a system, a medium and equipment for classified protection and release of high-dimensional sensitive data privacy, wherein the method for classified protection and release of the high-dimensional sensitive data privacy comprises the following steps: the issuing party carries out data acquisition and selects a total privacy protection parameter epsilon; after the data acquisition of the publisher is finished, carrying out data preprocessing; the data publisher evaluates the privacy protection level of the data attribute; the issuing party carries out data privacy noise disturbance; obtaining a k-dimensional low-dimensional matrix by using a sparse matrix transformation method; restoring the low-dimensional matrix to further restore the matrix before equalization; and forming m rows and n columns of matrixes and data tables for privacy protection and noise addition, and forming data for externally releasing a privacy version. According to the method, the operation efficiency of privacy protection release of the mass high-dimensional sensitive data is greatly improved under the condition that privacy protection degree requirements of all attributes of the mass high-dimensional sensitive data set are different; under the same privacy protection degree, the usability of the data is improved.

Description

Method, system, medium and device for hierarchical protection and release of high-dimensional sensitive data privacy

technical field

The invention belongs to the technical field of information security data release, and in particular relates to a method, system, medium and device for releasing high-dimensional sensitive data privacy graded protection.

Background technique

At present, with the advent of the era of big data cloud computing, the information age is filled with all kinds of high-dimensional data at any time. Authoritative departments related to medical care, people's livelihood, finance, public security and other fields often need to release some data to provide third parties for analysis and statistics. However, the information released by the above-mentioned authorities often contains extremely sensitive data. If the original information is directly released, the sensitive data of individual samples is likely to be used by third parties. When attackers have strong background knowledge and illegitimate intentions, the leakage of this sensitive information will have unpredictable consequences.

Data release is an important way of information disclosure, so ensuring the information security of sensitive data release is an important measure. In traditional sensitive data release, the classical Laplacian algorithm can be used to directly add noise disturbance, so that the information of individual samples cannot be leaked under the condition that the statistical results are relatively similar. However, the data that needs to be released in the fields of medical care, people's livelihood, finance, and public security are often high-dimensional and have many samples. If the traditional Laplacian mechanism plus noise perturbation method is used, the excessive amount of data will lead to too much noise, resulting in a large degree of distortion of the data. Such publishing will lead to poor data availability and reduce data publishing by authoritative departments. 's credibility. At the same time, in some huge data sets that need to be published, there are often a lot of data with different sensitivity levels. Some of the attributes require strong confidentiality, even at the expense of a certain degree of availability; but there are still some attributes that do not require strict confidentiality. , this part of the properties should still be based on the principle of ensuring the availability and reducing the error rate as much as possible. Therefore, it is of practical significance to formulate different privacy protection schemes according to the privacy protection requirements of different attributes.

Through the above analysis, the existing problems and defects in the prior art are:

(1) The traditional privacy protection mechanism is usually the Laplacian scrambling mechanism. When faced with massive high-dimensional data, this mechanism will introduce excessive noise. The problem of how to control the degree of data distortion needs to be solved urgently.

(2) Traditional privacy protection algorithms for massive high-dimensional data are often slow in computing and processing speed in existing hardware resources. How to improve the computing and processing efficiency of privacy-preserving algorithms for massive high-dimensional data under limited hardware resources also needs to be solve.

(3) The traditional privacy protection methods for datasets do not take into account the different requirements for the degree of privacy protection required by different attributes, so there is still some room for improvement in taking into account the differences in the degree of privacy protection between different attributes.

(4) At present, the existing privacy protection methods for massive high-dimensional data often first convert the original data matrix into a low-dimensional projection matrix, and then add disturbance noise to the projection matrix to restore the privacy of the same scale as the original data matrix. Protected version of the data matrix. However, most of this method belongs to adding noise perturbation after calculating the projection matrix, and there is still a large room for improvement in the balance between privacy protection and usability.

The difficulty of solving the above problems and defects is: designing a scheme needs to take into account the balance between privacy and availability of different attributes in the data set, and improve the operation efficiency of the overall privacy protection release of massive high-dimensional sensitive data sets.

The significance of solving the above problems and defects is that the present invention greatly improves the operation efficiency of the privacy protection release of the massive high-dimensional sensitive data while taking into account the different privacy protection requirements of each attribute of the massive high-dimensional sensitive data set. At the same time, compared with the traditional privacy protection processing method, the availability of data is improved under the same degree of privacy protection.

SUMMARY OF THE INVENTION

In view of the problems existing in the prior art, the present invention provides a method, system, medium and device for publishing high-dimensional sensitive data privacy hierarchical protection, in particular to a high-dimensional sensitive data privacy hierarchical protection publishing method based on block sparse matrix transformation , systems, media and equipment.

The present invention is implemented in this way, a method for publishing high-dimensional sensitive data privacy hierarchical protection, the high-dimensional sensitive data privacy hierarchical protection publishing method includes:

Accepts a total of m data samples of the input original data, the dimension of each sample is n, and the privacy protection parameter that needs to be set is ε; the output data of the privacy protection version is m data samples, and the dimension of each sample is n-dimensional, the perturbed data of the privacy-preserving version is regarded as the final public release data.

Wherein, the high-dimensional sensitive data privacy protection release method includes seven stages: data collection stage, data preprocessing stage, data privacy protection level evaluation stage, data disturbance stage, data transformation stage, data restoration stage and data release stage.

Further, the method for publishing high-dimensional sensitive data privacy classification protection includes the following steps:

Step 1, data collection: The publisher collects data and selects the appropriate total privacy protection parameter ε; data collection is the basis for the subsequent steps. Also select the corresponding parameters.

Step 2: Data preprocessing: After the data collection of the publisher is completed, data preprocessing is performed; each sample is scanned separately. If any of the n attribute values of the sample has a null value, it will be filled with 0 to ensure n dimension. Each attribute has a value; all data are integrated and arranged into a matrix with m rows and n columns, that is, the number of samples of the matrix is m and the dimension is n; the original data in step 1 is processed, form a data matrix.

Step 3: Evaluation of data privacy protection level: The data publisher evaluates the privacy protection level of the data attributes, and then rearranges the attributes; divides the data set into relatively low-dimensional block matrices, and after block, the dimension of each block matrix is p, and label the sensitivity level of the block matrix as a whole. The data matrix is divided into blocks first, in order to avoid the long processing time of directly calculating the original covariance matrix in step 4. At the same time, the sensitivity level of the block matrix is marked to pave the way for the subsequent privacy budget allocation.

Step 4: Data perturbation: The publisher adds Wishart privacy noise perturbation to the covariance matrix of the block matrix; performs data privacy protection noise perturbation to protect the original information from being leaked.

Step 5, data transformation: use the sparse matrix transformation method to obtain the eigenvector matrix and the corresponding eigenvalue diagonal matrix, and take the eigenvectors corresponding to the first k maximum values of the eigenvalue diagonal matrix; transpose the eigenvector matrix and combine it. Multiply with the averaged matrix to obtain a k-dimensional low-dimensional matrix, k≤p; using the sparse matrix transformation method can further greatly improve the processing efficiency of matrix dimensionality reduction.

Step 6, data restoration: restore the low-dimensional matrix, and then restore the matrix before averaging; restore the initial form of the matrix.

Step 7: Data release: splicing the m-row and p-column block matrices formed by restoring the low-dimensional matrix perturbed by all privacy noises to form an m-row and n-column matrix with privacy protection and noise, and add the corresponding data to the header of the data table. Attribute name; adjust and restore according to the attribute arrangement order of the original data table, form a complete data table, and form a private version of the data to be released to the outside world. On the basis of step 6, perform corresponding data processing, including restoring the data matrix before averaging, and removing the columns filled with 0 before in the last block. Finally, add the header name to form a complete data release table.

Further, the data collection stage includes a data collection stage and a parameter selection stage.

In the data collection stage, the data of each individual sample is collected, the specific values of n attributes in each sample are determined, and the value m of the total number of individual samples in the calculation data set is updated in real time;

In the parameter selection stage, the privacy protection parameter ε represents the budget of privacy allocation. If ε is larger, the degree of privacy protection is smaller, and the availability of data is stronger; if ε is smaller, the degree of privacy protection is larger and the availability of data is stronger. Difference. Therefore, the selection of the ε value needs to be continuously adjusted according to the actual privacy protection requirements, and finally the appropriate total privacy protection parameter ε is determined according to the actual privacy protection degree requirements.

Further, the data privacy protection level evaluation stage includes an attribute sensitivity level evaluation stage, an attribute rearrangement stage, a data set division stage, and a block sensitivity level labeling stage.

In the attribute sensitivity level evaluation stage, the sensitivity level of each column of attributes is evaluated, and each column level is marked according to the relative high, medium and low;

In the attribute rearrangement stage, after the sensitivity level is marked, the matrix is arranged according to the sensitivity level of each attribute. After arranging all attributes according to the sensitivity level, a new matrix with m rows and n columns is formed again;

In the data set division stage, all attributes are arranged according to the sensitivity level to form a new matrix with m rows and n columns, and then block according to the dimension threshold p; for the block matrix whose last block dimension is less than p, use the numerical value 0 is filled to form a p-dimensional filled matrix, p≤10;

In the block sensitivity level labeling stage, the level labeling is performed on the p-dimensional block matrix with m rows and p columns, because in the previous two stages, the attribute rearrangement stage and the data set division stage have already sorted the sensitivity levels of attributes according to high, high and low. Therefore, the overall privacy protection degree of the divided block matrix is divided into three levels: high, medium and low.

So far, a complete m-row and n-column block matrix X has been divided into several new m-row and p-column block matrices, and the block matrix is divided into three levels: high, middle and low according to the needs of different privacy protection degrees.

Further, the data perturbation module includes: a data averaging stage, a covariance matrix calculation stage, a privacy budget parameter allocation stage, a privacy noise extraction stage, and a data privacy noise addition stage.

In the data averaging stage, the mean value of the column is subtracted from each value to form an averaged matrix X;

In the stage of calculating the covariance matrix, X calculates its covariance matrix A after averaging each p-dimensional block matrix,

In the privacy budget parameter allocation stage, different privacy protection requirements are allocated according to the privacy protection level strength of different sensitive attributes; for differential privacy protection technology, the smaller the privacy protection parameter, the higher the corresponding privacy protection level and the lower the data availability; the The privacy budget is matched with the privacy budget of each attribute according to the high-sensitive attribute column, the medium-sensitive attribute column, and the low-sensitive attribute column according to the ratio of 1:9:90; Summation to get the privacy budget ε _i of the block matrix;

从Wishart分布W(m+1,C)中提取m行p列的噪声样本矩阵W；In the privacy noise extraction stage, for a block matrix with m rows and p columns, a positive definite matrix C with m rows and m columns is generated; the positive definite matrix C satisfies that the m eigenvalues are all equal, and all eigenvalues are set as

Extract the noise sample matrix W with m rows and p columns from the Wishart distribution W(m+1,C);

In the data privacy noise adding stage, the noise sample matrix W is added to the covariance matrix A to form a covariance matrix A' after scrambling noise, that is, A'=A+W.

Further, the data restoration stage includes a low-dimensional matrix restoration stage and an averaging restoration stage.

In the low-dimensional matrix restoration stage, the eigenvector matrix is multiplied by the k-dimensional low-dimensional matrix to obtain the restoration matrix, but the restoration matrix at this time has not been de-averaged;

In the mean restoration stage, add each element of the restoration matrix to the mean value of the column corresponding to the original matrix to obtain the mean restoration matrix, and remove the attribute column previously filled with 0 in the last block.

Another object of the present invention is to provide a high-dimensional sensitive data privacy hierarchical protection publishing system applying the high-dimensional sensitive data privacy hierarchical protection publishing method, and the high-dimensional sensitive data privacy hierarchical protection publishing system includes:

The data collection module is used for the publisher to collect data and select the appropriate total privacy protection parameter ε;

The data preprocessing module is used for data preprocessing after the publisher's data collection is completed; each sample is scanned separately. If a certain attribute of the n attribute values of the sample is empty, it will be filled with 0 to ensure n dimension. Each attribute has a value; all data are integrated and arranged into a matrix with m rows and n columns as a whole, that is, the number of samples of the matrix is m and the dimension is n;

The data privacy protection level evaluation module is used for the data publisher to evaluate the privacy protection level of the data attributes, and then rearrange the attributes; divide the data set into relatively low-dimensional block matrices, and after block, the dimension of each block matrix is p, and mark the overall sensitivity level of the block matrix;

The data perturbation module is used for the publisher to add Wishart privacy noise perturbation to the covariance matrix of the block matrix;

The data transformation module is used to obtain the eigenvector matrix and the corresponding eigenvalue diagonal matrix by using the sparse matrix transformation method, and obtain the eigenvectors corresponding to the first k maximum values of the eigenvalue diagonal matrix; transpose the eigenvector matrix and combine Multiply with the averaged matrix to obtain a k-dimensional low-dimensional matrix, k≤p;

The data restoration module is used to restore the low-dimensional matrix, and then restore the matrix before averaging;

The data release module is used to splicing all m-row and p-column block matrices formed by restoring the low-dimensional matrix perturbed by privacy noise to form a privacy-preserving noise-added m-row and n-column matrix, and add the corresponding data in the header of the data table. The attribute name of the original data table is adjusted and restored according to the attribute arrangement order of the original data table to form a complete data table, and form the data of the private version released to the outside world.

Another object of the present invention is to provide a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the following step:

Data collection, the publisher conducts data collection and selects the appropriate total privacy protection parameter ε;

Data preprocessing. After the data collection of the publisher is completed, data preprocessing is performed; each sample is scanned separately. If a certain attribute is empty in the n attribute values of the sample, it will be filled with 0 to ensure that each attribute in the n dimension is All data are integrated and arranged into a matrix with m rows and n columns as a whole, that is, the number of samples of the matrix is m and the dimension is n;

Data privacy protection level evaluation, the data publisher evaluates the privacy protection level of the data attributes, and then rearranges the attributes; the data set is divided into relatively low-dimensional block matrices. After block, the dimension of each block matrix is p, And mark the overall sensitivity level of the block matrix;

Data perturbation, which is used by the publisher to add Wishart privacy noise perturbation to the covariance matrix of the block matrix;

Data transformation, use the sparse matrix transformation method to obtain the eigenvector matrix and the corresponding eigenvalue diagonal matrix, and take the eigenvectors corresponding to the first k maximum values of the eigenvalue diagonal matrix; transpose the eigenvector matrix and average it. The resulting matrix is multiplied to obtain a k-dimensional low-dimensional matrix, k≤p;

Data restoration, restore the low-dimensional matrix, and then restore the matrix before averaging;

Data release, splicing all m-row and p-column block matrices scrambled and restored by transformed privacy-preserving noise to form a privacy-preserving noise-enhanced m-row n-column matrix, and add the corresponding attribute name to the header of the data table ; Adjust and restore according to the attribute arrangement order of the original data table, form a complete data table, and form a private version of the data released to the outside world.

Another object of the present invention is to provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, causes the processor to perform the following steps:

Data collection, the publisher conducts data collection and selects the appropriate total privacy protection parameter ε;

Data preprocessing. After the data collection of the publisher is completed, data preprocessing is performed; each sample is scanned separately. If a certain attribute is empty in the n attribute values of the sample, it will be filled with 0 to ensure that each attribute in the n dimension is All data are integrated and arranged into a matrix with m rows and n columns as a whole, that is, the number of samples of the matrix is m and the dimension is n;

Data privacy protection level evaluation, the data publisher evaluates the privacy protection level of the data attributes, and then rearranges the attributes; the data set is divided into relatively low-dimensional block matrices. After block, the dimension of each block matrix is p, And mark the overall sensitivity level of the block matrix;

Data perturbation, which is used by the publisher to add Wishart privacy noise perturbation to the covariance matrix of the block matrix;

Data transformation, use the sparse matrix transformation method to obtain the eigenvector matrix and the corresponding eigenvalue diagonal matrix, and take the eigenvectors corresponding to the first k maximum values of the eigenvalue diagonal matrix; transpose the eigenvector matrix and average it. The resulting matrix is multiplied to obtain a k-dimensional low-dimensional matrix, k≤p;

Data restoration, restore the low-dimensional matrix, and then restore the matrix before averaging;

Data release, splicing all m-row and p-column block matrices formed by restoring the low-dimensional matrix perturbed by privacy noise to form an m-row and n-column matrix with privacy protection and noise, and add the corresponding attribute name to the header of the data table ; Adjust and restore according to the attribute arrangement order of the original data table, form a complete data table, and form a private version of the data released to the outside world.

Another object of the present invention is to provide an information data processing terminal, the information data processing terminal is used to implement the high-dimensional sensitive data privacy classification protection publishing system.

Combined with all the above technical solutions, the advantages and positive effects of the present invention are as follows: the hierarchical protection and release method for privacy protection of high-dimensional sensitive data based on the block sparse matrix transformation provided by the present invention can take into account the various attributes of massive high-dimensional sensitive data sets. In the case of different privacy protection requirements, the operation efficiency of the privacy protection release of massive high-dimensional sensitive data is greatly improved. At the same time, compared with the traditional privacy protection processing method, the availability of data is improved under the same degree of privacy protection.

The invention makes full use of the high-speed dimensionality reduction characteristics of the block sparse matrix transformation, and proposes a high-speed sparse matrix transformation based on the security risks that direct data release may suffer and the different privacy protection levels required by different types of attributes. Dimensional data privacy hierarchical protection solution. At the same time, the differential privacy technology is introduced into the privacy protection release method of high-dimensional sensitive data, which ensures the hierarchical and secure release of large-scale and massive high-dimensional sensitive data, and can resist the strong background knowledge attack of attackers. At the same time, the block sparse matrix transformation method proposed in the present invention can greatly improve the computing speed of privacy protection of massive data, so that the distortion of the data can be reduced while providing the same degree of privacy protection as the traditional method. The invention can be applied to solve the data release requirements of privacy classification protection of massive high-dimensional sensitive information data in the fields of medical treatment, people's livelihood, finance, public security and the like.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following will briefly introduce the accompanying drawings that need to be used in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative effort.

FIG. 1 is a flowchart of a method for publishing high-dimensional sensitive data privacy protection by hierarchical protection provided by an embodiment of the present invention.

2 is a structural block diagram of a high-dimensional sensitive data privacy hierarchical protection publishing system provided by an embodiment of the present invention;

In the figure: 1. Data acquisition module; 2. Data preprocessing module; 3. Data privacy protection level evaluation module; 4. Data perturbation module; 5. Data transformation module; 6. Data restoration module; 7. Data publishing module.

FIG. 3 is a schematic diagram of an application scenario of the method for publishing high-dimensional sensitive data privacy protection by classification according to an embodiment of the present invention.

FIG. 4 is an example data processing flow chart of the method for publishing high-dimensional sensitive data privacy protection by classification according to an embodiment of the present invention.

FIG. 5 is a schematic diagram comparing the traditional differential privacy Laplacian noise mechanism algorithm and the high-dimensional sensitive data privacy hierarchical protection and release method of the present invention provided by an embodiment of the present invention.

FIG. 6 is a schematic diagram of a comparative experiment provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Aiming at the problems existing in the prior art, the present invention provides a method, system, medium and device for publishing high-dimensional sensitive data privacy protection by grading. The present invention is described in detail below with reference to the accompanying drawings.

As shown in FIG. 1 , the method for publishing high-dimensional sensitive data privacy classification protection provided by an embodiment of the present invention includes the following steps:

S101, data collection: the publisher collects data and selects appropriate total privacy protection parameters;

S102, data preprocessing: after the data collection of the publisher is completed, data preprocessing is performed;

S103, data privacy protection level evaluation: the data publisher evaluates the privacy protection level of the data attribute;

S104, data disturbance: the publisher conducts data privacy noise disturbance;

S105, data transformation: use a sparse matrix transformation method to obtain a k-dimensional low-dimensional matrix;

S106, data restoration: restore the low-dimensional matrix, and then restore the matrix before averaging;

S107, data release: form an m-row and n-column matrix and a data table with added noise for privacy protection, and form data for releasing a private version to the outside world.

As shown in FIG. 2 , the high-dimensional sensitive data privacy classification protection publishing system provided by the embodiment of the present invention includes:

Data collection module 1, which is used by the publisher to collect data and select an appropriate total privacy protection parameter ε;

The data preprocessing module 2 is used for data preprocessing after the publisher's data collection is completed; each sample is scanned separately. If there is a null value for an attribute in the n attribute values of the sample, it will be filled with 0 to ensure n Each attribute of the dimension has a value; all data are integrated and arranged into a matrix with m rows and n columns as a whole, that is, the number of samples of the matrix is m, and the number of dimensions is n;

The data privacy protection level evaluation module 3 is used for the data publisher to evaluate the privacy protection level of the data attributes, and then rearrange the attributes; divide the data set into relatively low-dimensional block matrices, and after block, the dimension of each block matrix The number is p, and the sensitivity level of the whole block matrix is marked;

Data perturbation module 4, for the publisher to add Wishart privacy noise perturbation to the covariance matrix of the block matrix;

The data transformation module 5 is used to obtain the eigenvector matrix and the corresponding eigenvalue diagonal matrix by using the sparse matrix transformation method, and obtain the eigenvectors corresponding to the first k maximum values of the eigenvalue diagonal matrix; transpose the eigenvector matrix And multiplied with the averaged matrix to obtain a k-dimensional low-dimensional matrix, k≤p;

The data restoration module 6 is used to restore the low-dimensional matrix, and then restore the matrix before averaging;

The data release module 7 is used to splicing the m-row and p-column block matrix formed by restoring the low-dimensional matrix perturbed by all privacy noises to form an m-row and n-column matrix of privacy protection and noise, and add the corresponding data in the header of the data table. The attribute name of the original data table is adjusted and restored according to the attribute arrangement order of the original data table to form a complete data table, and form the data of the private version released to the outside world.

FIG. 3 is a schematic diagram of an application scenario of the method for publishing high-dimensional sensitive data privacy protection by classification provided by the embodiment of the present invention.

The technical solutions of the present invention will be further described below in conjunction with the embodiments.

The purpose of the present invention is to solve the problems of unsatisfactory privacy protection effect and long operation time of existing mass high-dimensional sensitive data publishing, and proposes a high-dimensional sensitive data privacy protection publishing method based on block sparse matrix transformation, so as to realize the large-scale high-dimensional sensitive data privacy protection publishing method. Efficiency of computational processing when high-dimensional sensitive data is released, and the balance between privacy protection and data errors.

As shown in FIG. 4 , the disclosure method for privacy protection of high-dimensional sensitive data based on block sparse matrix transformation of the present invention. Accepts a total of m data samples of the input original data, the dimension of each sample is n, and the privacy protection parameter that needs to be set is ε; the output data of the privacy protection version is m data samples, and the dimension of each sample is n-dimensional, the perturbed data of the privacy-preserving version is regarded as the final public release data. The whole process includes seven stages: data collection stage, data preprocessing stage, data privacy protection level assessment stage, data disturbance stage, data transformation stage, data restoration stage, and data release stage. The specific stages are as follows:

(1) Data collection stage, including two sub-stages: data collection stage and parameter selection stage.

Data collection stage: collect the data of each individual sample, determine the specific values of n attributes in each sample, and update the value m of the total number of individual samples in the data set in real time.

Parameter selection stage: The privacy protection parameter ε represents the budget of privacy allocation. If ε is larger, the degree of privacy protection is smaller and the availability of data is stronger; if ε is smaller, the degree of privacy protection is larger and the availability of data is stronger. Difference. Therefore, the selection of the ε value needs to be continuously adjusted according to the actual privacy protection requirements, and finally the appropriate total privacy protection parameter ε is determined according to the actual privacy protection degree requirements.

(2) Data preprocessing stage: Scan each sample separately. If there is a null value for an attribute in the n attribute values of the sample, it will be filled with 0 to ensure that each attribute in the n dimension has a value. Arrange all the data into a matrix with m rows and n columns as a whole, that is, the number of samples of this matrix is m and the dimension is n.

(3) Data privacy protection level evaluation stage, including four sub-stages: attribute sensitivity level evaluation stage, attribute rearrangement stage, dataset division stage, and block sensitivity level labeling stage.

Attribute sensitivity level evaluation stage: evaluate the sensitivity level of each column of attributes, and each column level is marked according to the relative high, medium and low.

Attribute rearrangement stage: After the sensitivity level is marked, the matrix is arranged according to the sensitivity level of each attribute. If the sensitivity level of the attribute is higher, the arranged dimension will be higher (the lower the dimension). After arranging all attributes according to the sensitivity level, a new matrix with m rows and n columns is reconstructed.

Data set division stage: Because the matrix dimension value n is usually large after being arranged according to attributes and sensitivity levels, it greatly affects the operation efficiency of calculating the covariance matrix in the subsequent data processing stage. For example, the dimension in the current census data table is very large. It is huge, and there are usually hundreds of attribute categories. For the calculation of the covariance matrix, it will undoubtedly greatly affect the operation efficiency. Therefore, dividing the n-dimensional matrix arranged according to the sensitivity level into a plurality of relatively low-dimensional matrices is of great significance to improve the operation efficiency of subsequent covariance processing. The effect after division should control the dimension of each block matrix within 10. Therefore, a new matrix with m rows and n columns is formed after arranging all the attributes according to the sensitivity level, and then blocks according to the dimension threshold value p (p≤10). The last block matrix whose block dimension is less than p is filled with a value of 0 to form a p-dimensional filled matrix.

Block Sensitivity Level Labeling Stage: For the level labeling of the p-dimensional block matrix with m rows and p columns, since in the previous two stages, the attribute rearrangement stage and the data set division stage have already sorted the sensitivity levels of the attributes according to high, high and low. Divide. Therefore, the sensitivity level of the block matrix at this stage can also be easily divided into three levels: high, medium and low, that is, the overall privacy protection degree of the divided block matrix is divided into three levels: high, medium and low.

So far, a complete m-row and n-column block matrix X has been divided into several new m-row and p-column block matrices, and the block matrix is divided into three levels: high, middle and low according to the needs of different privacy protection degrees.

(4) Data perturbation stage, including five sub-stages: data averaging stage, covariance matrix calculation stage, privacy budget parameter allocation stage, privacy noise extraction stage, and data privacy noise addition stage.

Data averaging stage: In order to avoid the influence of dimension, each value is subtracted from the mean of the column, and then an averaged matrix X is formed.

In the stage of calculating the covariance matrix, X calculates its covariance matrix A after averaging each p-dimensional block matrix,

Privacy budget parameter allocation stage: allocate different privacy protection requirements according to the privacy protection level strength of different sensitive attributes. For differential privacy protection technology, the smaller the privacy protection parameter, the higher the corresponding privacy protection level and the lower the data availability. Therefore, the allocation method at this stage is to match the privacy budget of each attribute according to the high-sensitive attribute column, medium-sensitive attribute column, and low-sensitive attribute column in a ratio of 1:9:90. After that, sum the privacy budgets of all attribute columns in each block matrix to obtain the privacy budget ε _i of the block matrix.

进而从Wishart分布W(m+1,C)中提取一个m行p列的噪声样本矩阵W。这样保证了加入噪声扰动的整个分块矩阵满足(ε,0)-差分隐私，即对于相邻的分块矩阵(两个分块矩阵只有一个元素不同)，经过噪声扰动后，它们的输出几乎是完全一致的。这样即使攻击者具备了除了某个元素之外的所有信息，也无法获知该元素的具体值，这样就具备了强大的隐私保护特性。Privacy noise extraction stage: Since the covariance matrix has characteristics such as symmetry and positive semi-definite, the disturbance noise added to the covariance needs to satisfy these two properties. Specifically, for a block matrix with m rows and p columns, first generate a positive definite matrix C with m rows and m columns. This positive definite matrix C is required to satisfy m eigenvalues that are all equal, and all eigenvalues are set as

Then, a noise sample matrix W with m rows and p columns is extracted from the Wishart distribution W(m+1,C). This ensures that the entire block matrix added with noise perturbation satisfies (ε, 0)-differential privacy, that is, for adjacent block matrices (two block matrices differ only by one element), after noise perturbation, their outputs are almost is exactly the same. In this way, even if the attacker has all the information except a certain element, he cannot know the specific value of the element, which has a strong privacy protection feature.

Data privacy and noise addition stage: add the noise sample matrix W in this module to the covariance matrix A to form the covariance matrix A' after scrambled noise, that is, A'=A+W.

(5) Data transformation stage:

A projection operator is constructed using a sparse matrix transformation algorithm based on eigenvector estimation. The orthogonal eigenmatrix of sparse matrix transformation can be expressed as a series of finite times of continuous Gives rotation multiplication. The specific implementation of sparse matrix transformation is as follows:

for h=1toH:

达到最大值①Adjust i,j to satisfy

Reaches the maximum value

②Calculate the angle

③ The sparse matrix G _h of each round is to perform Givens rotation transformation on the coordinates (i _h , j _h ), G _h =I+Θ(i _h , j _h , θ _h ), where I is the identity matrix.

④Calculation

⑤Calculation

并计算

得到特征值，之后将特征向量按照其所对应的特征值从大到小排列。并计算特征值求和

的累计贡献率，将特征向量的累计贡献率和大于95％的特征值矩阵中的前k个最大值组成Λ＝{λ₁,λ₂,......,λ_k}，并得到特征值所对应的特征向量矩阵E_k，通常主成分越靠前则包含的信息越充足。进而将特征向量转置并与均值化后的矩阵相乘，得到k维(k≤p)的低维矩阵。Among them, the iterative method is used to obtain G _h of each step, and the number of iterations H can be obtained by cross-validation, and then the sparse matrix transformation method is used to obtain the eigenvector matrix

and calculate

The eigenvalues are obtained, and then the eigenvectors are arranged in descending order according to their corresponding eigenvalues. and calculate the sum of the eigenvalues

The cumulative contribution rate of , the cumulative contribution rate of the eigenvectors and the top k maximum values in the eigenvalue matrix greater than 95% are composed of Λ={λ ₁ ,λ ₂ ,...,λ _k }, and get The eigenvector matrix E _k corresponding to the eigenvalues usually contains more sufficient information as the principal components are in the front. Then, the eigenvectors are transposed and multiplied by the averaged matrix to obtain a k-dimensional (k≤p) low-dimensional matrix.

(6) Data restoration stage: It includes two sub-stages, a low-dimensional matrix restoration stage and an average restoration stage.

Low-dimensional matrix restoration stage: The restoration matrix is obtained by multiplying the eigenvector matrix with the k-dimensional low-dimensional matrix. However, the restoration matrix at this time has not been de-averaged.

The mean value restoration stage: add each element of the restoration matrix to the mean value of the column corresponding to the original matrix to obtain the mean value restoration matrix.

(7) Data release stage: splicing all m-row and p-column block matrices formed by restoring the low-dimensional matrix perturbed by privacy noise to form a privacy-preserving noise-added m-row and n-column matrix, and add it to the header of the data table. The corresponding attribute name is adjusted and restored according to the attribute arrangement order of the original data table to form a complete data table. So far, the private version of the data can be formed.

Fig. 5 is a comparison between the traditional differential privacy Laplacian noise mechanism algorithm and the high-dimensional sensitive data privacy protection and publishing method based on block sparse matrix transformation provided by the present invention. When controlling the same attributes in the above two methods On the premise that the privacy protection budget is equal, the approximate error rate is compared.

FIG. 6 is a comparative experiment provided by an embodiment of the present invention, the purpose of which is to examine the impact on the processing speed of the privacy operation after the block is divided. The specific comparison is to control other dependent variables unchanged, and only change the variable "whether to adopt the block scheme" to compare the operating efficiency between the two. While the running time of the traditional unblocked matrix is 1893×10 ^-4 seconds, it can be concluded that the block scheme greatly improves the processing efficiency of privacy-preserving operations.

Table 1 Approximate error rate comparison table

privacy protection budget Traditional Laplace method error rate Error rate of the present invention 0.1 1670% 152% 0.5 580% 98% 1 180% 46% 10 35% 5% 100 12% 0.90%

Table 2 Sub-library matrix running time comparison

number of blocks The running time of the block method of the present invention/10^-4 seconds 5 26 6 20 7 18 8 16 9 twenty two 10 26

While the running time of the traditional unblocked matrix is 1893×10 ^-4 seconds, it can be concluded that the block scheme greatly improves the processing efficiency of privacy-preserving operations.

In the above-mentioned embodiments, it may be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be downloaded from a website site, computer, server or data center Transmission to another website site, computer, server, or data center by wire (eg, coaxial cable, fiber optic, digital subscriber line (DSL), or wireless (eg, infrared, wireless, microwave, etc.)). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, a data center, or the like that includes an integration of one or more available media. The usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVD), or semiconductor media (eg, Solid State Disk (SSD)), among others.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited to this, any person skilled in the art is within the technical scope disclosed by the present invention, and all within the spirit and principle of the present invention Any modifications, equivalent replacements and improvements made within the scope of the present invention should be included within the protection scope of the present invention.

Claims (9)

1. A classified protection and release method for privacy of high-dimensional sensitive data is characterized by comprising the following steps:

receiving m data samples in total of input original data, wherein the dimension of each sample is n-dimension, and the privacy protection parameter needing to be set is epsilon; outputting data of the privacy protection version as m data samples, wherein the dimension of each sample is n-dimension, and taking the disturbance data of the privacy protection version as the data which is finally published to the outside;

the classified protection and release method for the privacy of the high-dimensional sensitive data comprises seven major stages: the method comprises a data acquisition stage, a data preprocessing stage, a data privacy protection level evaluation stage, a data disturbance stage, a data transformation stage, a data restoration stage and a data release stage;

the privacy classified protection publishing method of the high-dimensional sensitive data comprises the following steps:

step one, data acquisition: the issuing party performs data acquisition and selects a proper total privacy protection parameter epsilon;

step two, data preprocessing: after the data acquisition of the publisher is finished, carrying out data preprocessing; scanning each sample respectively, and if n attribute values of the samples have a certain attribute as a null value, filling the n attribute values with 0 to ensure that each attribute of n dimensions has a numerical value; integrating and arranging all data into a matrix which is m rows and n columns as a whole, namely the number of samples of the matrix is m, and the dimension is n;

step three, evaluating the data privacy protection level: the data publisher evaluates the privacy protection level of the data attributes, and then rearranges the attributes; dividing the data set into relatively low-dimensional block matrixes, wherein after the data set is divided into blocks, the dimension of each block matrix is p, and the sensitivity level of the whole block matrix is marked;

step four, data disturbance: adding Wishart privacy noise disturbance to a covariance matrix of a block matrix by a publisher;

step five, data transformation: obtaining a characteristic vector matrix and a corresponding characteristic value diagonal matrix by using a sparse matrix transformation method, and taking characteristic vectors corresponding to the first k maximum values of the characteristic value diagonal matrix; transposing the characteristic vector matrix and multiplying the transposed characteristic vector matrix and the averaged matrix to obtain a k-dimensional low-dimensional matrix, wherein k is less than or equal to p;

step six, data restoration: restoring the low-dimensional matrix to further restore the matrix before equalization;

step seven, data release: splicing m rows and p columns of block matrixes formed by restoring all privacy noise disturbed low-dimensional matrixes to form m rows and n columns of matrixes with privacy protection and noise, and adding corresponding attribute names to the head of a data table; and adjusting and restoring the attribute arrangement sequence of the original data table to form a complete data table and form data of the external release privacy version.

2. The privacy-based hierarchical protection and release method for high-dimensional sensitive data according to claim 1, wherein the data acquisition phase comprises a data collection phase and a parameter selection phase;

a data collection stage, namely collecting the data of each individual sample, determining specific values of n attributes in each sample, and updating and calculating the value m of the total number of the individual samples in the data set in real time;

in the parameter selection stage, a privacy protection parameter epsilon represents the budget of privacy allocation, and if epsilon is larger, the smaller the privacy protection degree is, and the stronger the data availability is; if epsilon is smaller, the greater the degree of privacy protection, the poorer the usability of data; therefore, the value of epsilon needs to be continuously adjusted according to the actual privacy protection requirement, and finally, a proper total privacy protection parameter epsilon is determined according to the actual privacy protection requirement.

3. The method for privacy-preserving and-protecting issuance of high-dimensional sensitive data according to claim 1, wherein the data privacy-preserving level evaluation phase comprises an attribute sensitivity level evaluation phase, an attribute rearrangement phase, a data set division phase, and a blocking sensitivity level labeling phase;

an attribute sensitivity level evaluation stage, wherein the sensitivity level of each row of attributes is evaluated, and each row of levels are labeled according to relative high, medium and low;

in the attribute rearrangement stage, after the sensitivity levels are marked, the matrix is arranged according to the sensitivity level of each attribute, and if the sensitivity level of the attribute is higher, the arranged dimensionality is closer to the front, and the dimensionality is lower; after all attributes are arranged according to the sensitivity levels, a new matrix with m rows and n columns is formed again;

in the data set division stage, all attributes are arranged according to the sensitivity level to form a new matrix with m rows and n columns, and then blocking is carried out according to the dimension threshold value as p; filling the block matrix with the last block dimension less than p by using a numerical value 0 to form a p-dimension filling matrix, wherein p is less than or equal to 10;

in the blocking sensitivity level marking stage, level marking is carried out on a p-dimensional blocking matrix of m rows and p columns, and since the sensitivity levels of the attributes are sorted and divided according to high, medium and low in the attribute rearranging stage and the data set dividing stage in the last two stages, the integral privacy protection degree of the divided blocking matrix is divided into three levels of high, medium and low;

at this point, a complete m rows and n columns are divided to form a new block matrix X with a plurality of m rows and p columns, and the block matrix is divided into three levels of high, medium and low according to the requirements of different privacy protection degrees.

4. The privacy-preserving publication method for high-dimensional sensitive data according to claim 1, wherein the data perturbation module comprises: a data averaging stage, a covariance matrix calculation stage, a privacy budget parameter allocation stage, a privacy noise extraction stage and a data privacy noise adding stage;

in the data equalization stage, subtracting the average value of the columns of the block matrix X where the element is positioned from each numerical value to form an equalized matrix X;

a stage of calculating covariance matrix, namely calculating covariance matrix A of X after averaging each p-dimensional block matrix,

in the privacy budget parameter distribution stage, different privacy protection budget parameters are distributed according to the privacy protection level strengths of different sensitive attributes; for the differential privacy protection technology, the smaller the privacy protection parameter is, the higher the corresponding privacy protection level is, and the lower the data availability is; the privacy budget is calculated according to the proportion of a high sensitive attribute column, a medium sensitive attribute column and a low sensitive attribute column of 1:9: a ratio of 90 matches the privacy budget for each attribute; summing the privacy budgets of all attribute columns in each block matrix to obtain the privacy budget epsilon of the block matrix _i ；

In the privacy noise extraction stage, for a block matrix with m rows and p columns, a positive definite matrix C with m rows and m columns is generated; the positive definite matrix C satisfies that m eigenvalues are all equal, and all eigenvalues are set as

Extracting a noise sample matrix W with m rows and p columns from the Wishart distribution W (m +1,C);

and a data privacy and noise adding stage, namely adding the noise sample matrix W to a covariance matrix A to form a covariance matrix A 'after scrambling noise, namely A' = A + W.

5. The privacy-preserving and publishing method for high-dimensional sensitive data according to claim 1, wherein the data recovery phase comprises a low-dimensional matrix recovery phase and an averaging recovery phase;

in the low-dimensional matrix restoration stage, a feature vector matrix is multiplied by a k-dimensional low-dimensional matrix to obtain a restoration matrix, but the restoration matrix is not subjected to averaging;

and in the equalization restoration stage, adding the average value of the column where the corresponding original matrix is positioned to each element of the restoration matrix to obtain an equalized and restored matrix, and removing the column filled with 0 before in the last block.

6. A privacy-preserving and-publishing system for high-dimensional sensitive data, which implements the privacy-preserving and-publishing method for high-dimensional sensitive data according to any one of claims 1 to 5, wherein the privacy-preserving and-publishing system for high-dimensional sensitive data comprises:

the data acquisition module is used for acquiring data by a publisher and selecting a proper total privacy protection parameter epsilon;

the data preprocessing module is used for preprocessing the data after the data acquisition of the publisher is finished; scanning each sample respectively, and if some attribute of n attribute values of the samples is a null value, filling the n attribute values with 0 to ensure that each attribute of n dimensions has a numerical value; integrating and arranging all data into a matrix which is m rows and n columns as a whole, namely the number of samples of the matrix is m, and the dimension is n;

the data privacy protection level evaluation module is used for evaluating the privacy protection level of the data attribute by the data publisher so as to rearrange the attribute; dividing the data set into relatively low-dimensional block matrixes, wherein after the data set is divided into blocks, the dimension of each block matrix is p, and the sensitivity level of the whole block matrix is marked;

the data disturbance module is used for adding Wishart privacy noise disturbance to the covariance matrix of the block matrix by the publisher;

the data transformation module is used for obtaining a characteristic vector matrix and a corresponding characteristic value diagonal matrix by using a sparse matrix transformation method, and taking characteristic vectors corresponding to the first k maximum values of the characteristic value diagonal matrix; transposing the characteristic vector matrix and multiplying the transposed characteristic vector matrix and the averaged matrix to obtain a k-dimensional low-dimensional matrix, wherein k is less than or equal to p;

the data recovery module is used for recovering the low-dimensional matrix so as to recover the matrix before equalization;

the data publishing module is used for splicing m rows and p columns of block matrixes formed by recovering all low-dimensional matrixes disturbed by privacy noise to form m rows and n columns of matrixes with privacy protection and noise, and adding corresponding attribute names to the head of a data table; and adjusting and restoring the attribute arrangement sequence of the original data table to form a complete data table and form data of the external release privacy version.

7. A computer device, characterized in that the computer device comprises a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to carry out the steps of:

data acquisition, namely acquiring data by a publisher, and selecting a proper total privacy protection parameter epsilon;

data preprocessing, namely preprocessing the data after the data acquisition of the issuing party is finished; scanning each sample respectively, and if some attribute of n attribute values of the samples is a null value, filling the n attribute values with 0 to ensure that each attribute of n dimensions has a numerical value; integrating and arranging all data into a matrix which is m rows and n columns as a whole, namely the number of samples of the matrix is m, and the dimension is n;

evaluating the privacy protection level of the data, and evaluating the privacy protection level of the data attribute by a data publisher so as to rearrange the attribute; dividing the data set into relatively low-dimensional block matrixes, wherein after the data set is divided into blocks, the dimension of each block matrix is p, and the sensitivity level of the whole block matrix is marked;

data disturbance: adding Wishart privacy noise disturbance to a covariance matrix of a block matrix by a publisher;

data transformation, namely obtaining a characteristic vector matrix and a corresponding characteristic value diagonal matrix by using a sparse matrix transformation method, and taking characteristic vectors corresponding to the first k maximum values of the characteristic value diagonal matrix; transposing the characteristic vector matrix and multiplying the transposed characteristic vector matrix and the averaged matrix to obtain a k-dimensional low-dimensional matrix, wherein k is less than or equal to p;

data recovery, recovering the low-dimensional matrix, and further recovering the matrix before equalization;

data publishing, namely splicing m rows and p columns of block matrixes formed by restoring all privacy noise disturbed low-dimensional matrixes to form m rows and n columns of matrixes for privacy protection and noise addition, and adding corresponding attribute names to a data table header; and adjusting and restoring the attribute arrangement sequence of the original data table to form a complete data table and form data of the external release privacy version.

8. A computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of:

data acquisition, namely acquiring data by a publisher, and selecting a proper total privacy protection parameter epsilon;

data preprocessing, namely preprocessing the data after the data acquisition of the issuing party is finished; scanning each sample respectively, and if some attribute of n attribute values of the samples is a null value, filling the n attribute values with 0 to ensure that each attribute of n dimensions has a numerical value; integrating and arranging all data into a matrix which is m rows and n columns as a whole, namely the number of samples of the matrix is m, and the dimension is n;

evaluating the privacy protection level of the data, and evaluating the privacy protection level of the data attribute by a data publisher so as to rearrange the attribute; dividing the data set into relatively low-dimensional block matrixes, wherein after the data set is divided into blocks, the dimension of each block matrix is p, and the sensitivity level of the whole block matrix is marked;

data disturbance: adding Wishart privacy noise disturbance to a covariance matrix of a block matrix by a publisher;

data transformation, namely obtaining a characteristic vector matrix and a corresponding characteristic value diagonal matrix by using a sparse matrix transformation method, and taking characteristic vectors corresponding to the first k maximum values of the characteristic value diagonal matrix; transposing the characteristic vector matrix and multiplying the transposed characteristic vector matrix and the averaged matrix to obtain a k-dimensional low-dimensional matrix, wherein k is less than or equal to p;

data recovery, namely recovering the low-dimensional matrix so as to recover the matrix before equalization;

data publishing, namely splicing m rows and p columns of block matrixes formed by recovering all privacy noise disturbed low-dimensional matrixes to form m rows and n columns of matrixes for privacy protection and noise, and adding corresponding attribute names to the head of a data table; and adjusting and restoring the attribute arrangement sequence of the original data table to form a complete data table and form data of the external release privacy version.

9. An information data processing terminal, characterized in that the information data processing terminal is used for implementing the privacy-preserving and publishing system of high-dimensional sensitive data according to claim 6.

Images