Privacy-enhanced relational table data simulation generation method

Abstract

The invention provides a privacy-enhanced relational table data simulation generation method, which comprises the following steps of: mining highly related attributes in a linked table L and single tables U and V with foreign key association, and combining the attributes with the linked table L to obtain attributes in the U and V tables corresponding to k attribute pairs; according to the foreign key corresponding relation of the link table L, the obtained attributes and the link table L are spliced, a merge table # imgabs0 # is obtained, according to the attributes of the link table L, the merge table # imgabs1 # is segmented according to columns, and a link table # imgabs2 # generated through simulation is obtained; according to the synthesis result of the link table # imgabs3 #, using a Markov random field model to simulate and generate a table # imgabs4 #; and according to the synthesis result of the link table # imgabs5 #, simulating and generating a table # imgabs6 # by using a Markov random field model. When the simulation data is generated, the effectiveness of the simulation data is improved, and it is ensured that the authenticity and effectiveness of the simulation data can still be kept while privacy of the simulation data is protected.

Description

Technical Field

[0001] The present invention belongs to the field of table data processing, and in particular relates to a privacy-enhanced relational table data simulation generation method. Background Art

[0002] In various artificial intelligence applications, it is usually necessary to process and integrate data from multiple tables. Unlike traditional single-table data scenarios, multi-table data often has foreign key associations. This requires that the simulation data can not only reflect the statistical characteristics of the original single-table data, but also maintain the consistency and relevance of the data in a complex multi-table relationship network. Existing methods connect the parent table and the child table into a single table through foreign keys, and then use the single-table algorithm to synthesize simulation data. However, the connection of unrelated attributes not only reduces the operating efficiency of the simulation algorithm, but also can only maintain the association relationship between "U table and L table" or "V table and L table", which has an adverse effect on the utility of the simulation data. Not only is it not applicable to tabular data with "many-to-many" foreign key associations, but privacy enhancement technology is not applied to provide necessary privacy protection for the model and generated data. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a privacy-enhanced relational table data simulation generation method.

[0004] The present invention is implemented as follows: a privacy-enhanced relational table data simulation generation method, the method comprising the following steps: Step S1: mining highly correlated attributes in the link table L and the single tables U and V with foreign key associations, and merging the attributes with the link table L to obtain the attributes in the U and V tables corresponding to k attribute pairs; Step S2: According to the foreign key correspondence of the link table L, the attributes obtained in step S1 are concatenated with the link table L to obtain a merged table. , according to the attributes of the link table L, the merged table Split by column to get the link table generated by simulation ; Step S3: According to the link table The synthetic results are generated by simulating the Markov random field model. ; Step S4: According to the link table The synthetic results are generated by simulating the Markov random field model. .

[0005] A further technical solution of the present invention is: the step S1 comprises the following steps: Step S11: According to the foreign key matching relationship of the link table L, all attributes of the table U and the table V are merged; Step S12: Traverse the non-primary key attributes of table U , with each non-primary key attribute of table V Combination, using the correlation evaluation function To measure each attribute pair relevance; Step S13: Applying the index mechanism: using the correlation evaluation function As a utility function, For attribute pairs The output probability of k attribute pairs is randomly output from the attribute pair set in step S12. ; Step S14: Traverse the output results k times and collect the corresponding attributes of table U and table V respectively {U m |1≤m≤k}, {V m |1≤m≤k}.

[0006] A further technical solution of the present invention is: the step S2 comprises the following steps: Step S21: According to the foreign key correspondence of the link table L, the attribute {U m |1≤m≤k}, {V m |1≤m≤k} is concatenated with table L to obtain the merged table ; Step S22: Use the Markov random field model to model the original table U, and add Laplace noise in the modeling process to satisfy ε 21 -Differential privacy obtains graph structure; Step S23: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S24: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 23 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field to perform parameter learning in step S23; Step S25: Assume the privacy budget of step S2 is , ε 21 is the privacy budget of step S21, ε 22 is the privacy budget of step S22, ε 23 is the privacy budget of step S23. According to the serial mechanism of differential privacy, step S2 satisfies Differential privacy; Step S26: According to the attributes of the original link table L, the merged table Split by column to get the link table generated by simulation .

[0007] A further technical solution of the present invention is: after obtaining the generative model in step S2, sampling on the generative model to obtain a merged table The algorithm has two termination conditions: one is to maintain the same number of simulated data rows as the original table, and the algorithm terminates when it generates simulated data with the same number of rows as the original link table L; the other is to maintain the consistency of foreign keys, and when the algorithm generates simulated data with the same number of rows as the original link table L, it checks whether the foreign keys of the original link table L exist in the simulated data table. If all of them exist, the algorithm terminates. If there are any missing, the generation model sampling continues until all the foreign keys of the original link table L exist.

[0008] A further technical solution of the present invention is: if table U, table V and link table L are in a one-to-many foreign key relationship, it is necessary to perform aggregation operations on the attributes related to the two tables respectively. If the attribute is a categorical attribute, the categorical data is generated by sampling based on the proportion of each category in the group after grouping by the foreign key ID; if the attribute is a numerical attribute, the simulated data is replaced with its corresponding average based on the foreign key ID. Since the sampling proportion and the average are both calculated based on the simulation data, according to the nature of differential privacy post-processing, the privacy budget is not consumed.

[0009] A further technical solution of the present invention is: the step S3 comprises the following steps: Step S31: Use the Markov random field model to model the original table U, and add Laplace noise in the modeling process to make it satisfy ε 31 -Differential privacy obtains graph structure; Step S32: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S33: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 33 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; Step S34: To maintain the simulation table and simulation link table First, the U obtained in step S2 ID And the related attributes of table U are filled in and remain unchanged during the sampling process; Step S35: Given the Markov random field model and the simulation link table Under the filled attributes, sample and generate the remaining U table attributes; Step S36: Assume the privacy budget of step S3 is , ε 31 is the privacy budget of step S31, ε 32 is the privacy budget of step S32, ε 33 is the privacy budget of step S33. According to the serial mechanism of differential privacy, step S3 satisfies Differential privacy.

[0010] A further technical solution of the present invention is: the step S4 comprises the following steps: Step S41: Use the Markov random field model to model the original table V, and add Laplace noise in the modeling process to make it satisfy ε 41 -Differential privacy obtains graph structure; Step S42: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S43: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 43 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; Step S44: To maintain the simulation table and simulation link table First, the V obtained in step S2 ID And the related attributes of table V are filled in and remain unchanged during the sampling process; Step S45: Given the Markov random field model and the simulation link table Under the filled attributes, sample and generate the remaining V-table attributes; Step S46: Assume the privacy budget of step S4 is , ε 41 is the privacy budget of step S41, ε 42 is the privacy budget of step S42, ε 43 is the privacy budget of step S43. According to the serial mechanism of differential privacy, step S4 satisfies Differential privacy.

[0011] A further technical solution of the present invention is: it also includes step S5: based on the algorithm flow of steps S1 to S4, according to the serial mechanism of differential privacy, the overall algorithm satisfies Differential privacy.

[0012] The beneficial effects of the present invention are: by analyzing the associated attributes between multiple tables with foreign key relationships, attributes with significant associated relationships are selected so as to better maintain the association between child tables and multiple parent tables when generating simulation data, thereby improving the utility of simulation data, and integrating differential privacy mechanisms, including exponential mechanisms, Laplace mechanisms and Gaussian mechanisms. In the process of data simulation generation, privacy protection is provided for the generated model and simulation data to prevent information leakage, thereby ensuring that the simulation data can maintain its authenticity and validity while protecting privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is an example of relational table data with many-to-many foreign keys of the present invention; Figure 3 It is a multi-table data simulation generation process for many-to-many foreign keys of the present invention; Figure 4 The present invention simulates and generates a link table The process; Figure 5 It is the process of simulating and generating table U of the present invention; Figure 6 This is the process of simulating and generating Table V according to the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0015] The present invention provides a privacy-enhanced relational table data simulation generation method, which is oriented to relational table data with many-to-many foreign keys. When generating simulation data, the correlation between attributes of multiple tables under the foreign keys can be retained, overcoming the limitation that traditional methods can only be applied to "one-to-one" and "one-to-many".

[0016] Terminology Note: Differential privacy: Differential privacy is a privacy enhancement technology with a mathematical basis that can add or delete any data record without significantly affecting the query results. Therefore, even if an attacker knows all records except one in the data set, it will not lead to the leakage of the record. The relevant definitions of this patent are as follows: Definition 1, ε-differential privacy: For any two data sets D1 and D2 with the same data structure, there is only one record difference between D1 and D2 (adjacent data sets). Suppose there is a random algorithm M, the set of all possible output results of the algorithm is O, and any output result of the data sets D1 and D2 is , if the random algorithm M satisfies:

[0017] Where Pr[*] represents the probability of an event occurring, and the parameter ε represents the privacy protection budget, then the algorithm M satisfies ε-differential privacy protection.

[0018] Definition 2, Sensitivity: Given a query function , D is the input data set, R d is the output data set. On any pair of D1 and D2, the sensitivity of function f is:

[0019] in yes and The first-order distance between them. Sensitivity measures the impact of changes in the input data set on the corresponding output results. Based on sensitivity, we can give an implementation mechanism for differential privacy.

[0020] Definition 3, Laplace mechanism: Given a dataset D and a privacy budget ε, the sensitivity of the function f is , when the output of f satisfies:

[0021] Then algorithm A is said to satisfy ε-differential privacy, where is a random noise that satisfies the Laplace distribution.

[0022] Definition 4, Exponential Mechanism: Given a dataset D and a privacy budget ε, the output of the randomized algorithm M is an entity object r∈R, μ(D,R) is the utility function, is the sensitivity of the function μ(D,R), if it is proportional to The probability of selecting and outputting r from the input is , then the algorithm M satisfies ε-differential privacy.

[0023] Definition 5: Serial combination mechanism: For Differential Privacy Sub-Algorithms , the processes of any two sub-algorithms are independent of each other, then the overall output of the algorithm is satisfy Differential privacy.

[0024] Foreign key: A foreign key in a database is a field that references the primary key in another table and is used to establish and express the relationship between tables. There are mainly the following types of foreign key relationships: one-to-one, one-to-many, and many-to-many.

[0025] In order to facilitate the description of the algorithm involved in the present invention, for many-to-many foreign key relational table data, a triple B=(U,V,L) is used to represent the result of the model generation. Among them, the elements U and V represent the table data of two foreign key associations, and the element L, as the representation of the many-to-many foreign key relationship, not only has the association relationship between the attributes of the two single tables, but also has the attributes under the many-to-many foreign key association (such as Figure 2 as shown).

[0026] A many-to-many foreign key relationship table data such as Figure 2 As shown. Attribute User ID and movies ID The rating table is the link table L in the many-to-many foreign key relationship, where there are not only two foreign key user tables (U) and movie tables (V), but also two foreign key user tables (U and V). ID and movies ID It represents the foreign key association between the user table and the movie table, and contains other necessary attributes, such as rating values. The existing generation model first models a single table user table U or movie table V, and then merges the user table U or movie table V with the rating table L into a single table through a foreign key relationship, and then synthesizes the data based on the single table simulation algorithm. The existing method has to deal with a contradiction: when modeling the rating table L, the existing method cannot decide whether to base it on the data of the user table U or the data of the movie table V, and it is difficult to ensure that the generated rating table L conforms to the logical consistency of the user table U and the movie table V. In order to overcome this problem, the present invention adopts a new generation order: first generate the rating table L, and then generate the user table U and the movie table V respectively based on the already generated rating table L. The specific process is as follows. Figure 3 shown.

[0027] The following is a detailed description of each step: The scheme of step S1 is used to mine the associated attribute pairs between tables. The single tables U and V and the link table L with foreign key association are input, and the attributes in the U and V tables corresponding to k attribute pairs are output.

[0028] The purpose of mining associated attribute pairs is to find highly correlated attributes in tables U and V and merge them with the link table L, which not only avoids the excessively high data dimension caused by merging all attributes, but also prevents the negative impact of irrelevant attributes on the effectiveness of the simulation algorithm. Figure 3 As shown, the algorithm flow is as follows: according to the foreign key matching relationship of the link table L, all attributes of tables U and V are merged; the non-primary key attributes of table U are traversed. , and each non-primary key attribute of table V Combination, using the correlation evaluation function To measure each attribute pair The Laplace and Gaussian mechanisms are mainly applicable to the privacy protection of numerical data, while the exponential mechanism can be applied to selective problems, such as selecting the best option under privacy protection from a set of discrete outputs. Applying the exponential mechanism: Evaluating the function with correlation As a utility function (its sensitivity is 2), with For attribute pairs The output probability of k attribute pairs is randomly output from the attribute pair set. ; Traverse the output results k times and collect the corresponding attributes of table U and table V respectively {U m |1≤m≤k}, {V m |1≤m≤k}.

[0029] Application example of step S1: As shown in Table 1, U and table V have non-primary key attributes U1, U2 and V1, V2 respectively, and the combination can obtain 4 different attribute pairs. i ,V j ), respectively calculate the association of different attribute pairs under foreign key association. According to the index mechanism, each attribute (U i ,V j ) is proportional to The probability of random output.

[0030] Table 1:

[0031] Assume that the number of outputs k=2, and the results of the two random outputs according to probability are (U1, V2) and (U2, V1), then the corresponding attributes of the collected tables U and V are {U1, U2} and {V1}.

[0032] Step S2 is used to simulate and generate a link table , input the link table L and the attributes in the corresponding U and V tables of k attribute pairs, and output the link table generated by the simulation .

[0033] According to the foreign key correspondence of the link table L, the attribute {U m |1≤m≤k}, {V m |1≤m≤k} is concatenated with table L to obtain the merged table .

[0034] like Figure 4 As shown in Figure 1, under the restriction of differential privacy, the Markov random field model is used to simulate and generate the merged table U. The specific steps are as follows: the original table U is modeled using the Markov random field model, and Laplace noise is added during the modeling process to satisfy ε 21-Differential privacy obtains the graph structure; Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy: In each iteration, in order to maximize the accuracy of the Markov random field model, the present invention calculates the difference between the marginal distribution of the model and the marginal distribution in the original data. , and apply the Laplace mechanism to add noise to satisfy ε 23 -Differential privacy. Differential privacy from noisy versions Select the marginal distribution corresponding to the maximum value and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field. Suppose the privacy budget of step S2 is , according to the serial mechanism of differential privacy, step S2 satisfies Differential privacy.

[0035] After obtaining the generative model, sample the generative model to obtain the merged table According to specific business requirements, the algorithm has two termination conditions: if the number of rows in the simulated data is to be maintained the same as that in the original table, the algorithm terminates when it generates the same number of rows in the simulated data as the original link table L; if the foreign key is to be maintained consistent, when the algorithm generates the same number of rows in the simulated data as the original link table L, check whether the foreign key of the original link table L exists in the simulated data table. If all of them exist, the algorithm terminates. If there are any missing, the generation model sampling continues until all the foreign keys of the original link table L exist.

[0036] In addition, since the table U, table V and the link table L may have a one-to-many foreign key relationship, it is necessary to perform aggregation operations on the attributes related to the two tables respectively: if the attribute is a categorical attribute, such as the user's gender, then the categorical data is generated by sampling according to the proportion of each category in the group after grouping by the foreign key ID. If the attribute is a numerical attribute, such as the user's age, then the simulation data is replaced with its corresponding average according to the foreign key ID. Since the sampling ratio and average are both calculated based on the simulation data, according to the nature of differential privacy post-processing, this step does not consume the privacy budget.

[0037] Finally, according to the attributes of the original link table L, the merged table Split by column to get the link table generated by simulation .

[0038] Step S3 generates a single table based on the synthesis result of the link table L , input the original data table U and the simulation link table The attribute U' corresponding to table U is output as simulation generated table .

[0039] Assume that the privacy budget of step S3 is ε3. Under the restriction of differential privacy, the Markov random field model is used to simulate and generate the merged table U. The specific steps are as follows: The original table U is modeled using the Markov random field model, and Laplace noise is added during the modeling process to satisfy ε 31 -Differential privacy obtains graph structure; According to the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning parameters to satisfy Differential privacy; In each iteration process, in order to improve the model accuracy of the Markov random field as much as possible, the present invention calculates the difference between the marginal distribution of the model and the marginal distribution in the original data. , and apply the Laplace mechanism to add noise to satisfy ε 33 -Differential privacy. Differential privacy from noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; like Figure 5 As shown, in order to maintain the simulation table and simulation link table First, the U obtained in step S2 ID And the related attributes of table U are filled in and remain unchanged during the sampling process.

[0040] Given a Markov random field model and a simulation link table Under the filled attributes, sample to generate the remaining U table attributes.

[0041] Suppose the privacy budget of step S3 is , according to the serial mechanism of differential privacy, step S3 satisfies Differential privacy.

[0042] Step S4 generates a single table V based on the synthesis result of the link table L, input: original data table V and simulation link table The attribute V' corresponding to table V, output: simulation generates table .

[0043] Assume that the privacy budget of step S4 is ε4. Under the restriction of differential privacy, the Markov random field model is used to simulate and generate the merged table V. The specific steps are as follows: The original table V is modeled using the Markov random field model, and Laplace noise is added during the modeling process to satisfy ε 41 -Differential privacy obtains graph structure; According to the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning parameters to satisfy Differential privacy; In each iteration process, in order to improve the model accuracy of the Markov random field as much as possible, the present invention calculates the difference between the marginal distribution of the model and the marginal distribution in the original data. , and apply the Laplace mechanism to add noise to satisfy ε 43 -Differential privacy. Differential privacy from noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; like Figure 6 As shown, in order to maintain the simulation table and simulation link table First, the V obtained in step S2 ID And the related attributes of table V are filled in and remain unchanged during the sampling process; Given a Markov random field model and a simulation link table Under the filled attributes, sample and generate the remaining V-table attributes.

[0044] Suppose the privacy budget of step S4 is , according to the serial mechanism of differential privacy, step S4 satisfies Differential privacy.

[0045] Based on the algorithm flow of steps S1 to S4, according to the serial mechanism of differential privacy, the overall algorithm satisfies Differential privacy.

[0046] The existing method connects the parent table and the child table into a single table through foreign keys, and then uses a single table algorithm to synthesize simulation data. However, the connection of unrelated attributes not only reduces the operating efficiency of the simulation algorithm, but also can only maintain the association relationship between "U table and L table" or "V table and L table", which has an adverse effect on the utility of the simulation data. The present invention proposes a privacy-enhanced multi-table data generation method, which mines the association relationship between attributes between tables with foreign key relationships under differential privacy, and retains the attributes with significant association relationships and the link table for simulation data generation, while maintaining the association relationship between the link table L and the two parent tables U and V; the existing relational table data simulation generation algorithm is not only unable to apply to the table data with "many-to-many" foreign key associations, but also does not apply privacy enhancement technology to the necessary privacy protection of the model and generated data. On the basis of the entire algorithm flow, the present invention applies differential privacy mechanisms such as exponential mechanism, Laplace mechanism and Gaussian mechanism to protect the privacy security of simulation data and generated models based on the theory of differential privacy.

[0047] The present invention aims at the problem that the traditional method merges multiple tables into a single table, resulting in low association efficiency and poor data utility, and proposes a method for generating multi-table simulation data that can effectively retain the association relationship. By analyzing the association attributes between multiple tables with foreign key relationships, attributes with significant association relationships are selected so as to better maintain the association between the child table and multiple parent tables (i.e., L table and U table, V table) when generating simulation data, thereby improving the utility of simulation data. In order to solve the risk of privacy leakage that may occur in the existing method under the "many-to-many" foreign key relationship, the present invention integrates differential privacy mechanisms in the algorithm flow, including exponential mechanism (association attribute pair output), Laplace mechanism (construction of graph model structure, calculation of edge distribution gap) and Gaussian mechanism (mirror gradient descent). These mechanisms provide privacy protection for the generated model and simulation data during the data simulation generation process to prevent information leakage and ensure that the simulation data can maintain its authenticity and effectiveness while protecting privacy.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A privacy-enhanced relational table data simulation generation method, characterized in that: The method comprises the following steps: Step S1: mining highly correlated attributes in the link table L and the single tables U and V with foreign key associations, and merging the attributes with the link table L to obtain the attributes in the U and V tables corresponding to k attribute pairs; Step S2: According to the foreign key correspondence of the link table L, the attributes obtained in step S1 are concatenated with the link table L to obtain a merged table. , according to the attributes of the link table L, the merged table Split by column to get the link table generated by simulation ; Step S3: According to the link table The synthetic results are generated by simulating the Markov random field model. ; Step S4: According to the link table The synthetic results are generated by simulating the Markov random field model. .

2. The privacy-enhanced relational table data simulation generation method according to claim 1, characterized in that: The step S1 comprises the following steps: Step S11: According to the foreign key matching relationship of the link table L, all attributes of the table U and the table V are merged; Step S12: Traverse the non-primary key attributes of table U , with each non-primary key attribute of table V Combination, using the correlation evaluation function To measure each attribute pair relevance; Step S13: Applying the index mechanism: using the correlation evaluation function As a utility function, For attribute pairs The output probability of k attribute pairs is randomly output from the attribute pair set in step S12. ; Step S14: Traverse the output results k times and collect the corresponding attributes of table U and table V respectively {U m |1≤m≤k}, {V m |1≤m≤k}.

3. The privacy-enhanced relational table data simulation generation method according to claim 1, characterized in that: The step S2 comprises the following steps: Step S21: According to the foreign key correspondence of the link table L, the attribute {U m |1≤m≤k}, {V m |1≤m≤k} is concatenated with table L to obtain the merged table ; Step S22: Use the Markov random field model to model the original table U, and add Laplace noise in the modeling process to satisfy ε 21 -Differential privacy obtains graph structure; Step S23: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S24: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 23 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field to perform parameter learning in step S23; Step S25: Assume the privacy budget of step S2 is , ε 21 is the privacy budget of step S21, ε 22 is the privacy budget of step S22, ε 23 is the privacy budget of step S23. According to the serial mechanism of differential privacy, step S2 satisfies Differential privacy; Step S26: According to the attributes of the original link table L, the merged table Split by column to get the link table generated by simulation .

4. The privacy-enhanced relational table data simulation generation method according to claim 3, characterized in that: After the generative model is obtained in step S2, the merged table is obtained by sampling on the generative model The algorithm has two termination conditions: one is to maintain the same number of simulated data rows as the original table, and the algorithm terminates when it generates simulated data with the same number of rows as the original link table L; the other is to maintain the consistency of foreign keys, and when the algorithm generates simulated data with the same number of rows as the original link table L, it checks whether the foreign keys of the original link table L exist in the simulated data table. If all of them exist, the algorithm terminates. If there are any missing, the generation model sampling continues until all the foreign keys of the original link table L exist.

5. The privacy-enhanced relational table data simulation generation method according to claim 3, characterized in that: If table U, table V and link table L are in a one-to-many foreign key relationship, it is necessary to perform aggregation operations on the attributes related to the two tables respectively. If the attribute is a categorical attribute, the categorical data is generated by sampling based on the proportion of each category in the group after grouping by the foreign key ID; if the attribute is a numerical attribute, the simulated data is replaced with its corresponding average based on the foreign key ID. Since the sampling proportion and average are both calculated based on the simulated data, according to the nature of differential privacy post-processing, the privacy budget is not consumed.

6. The privacy-enhanced relational table data simulation generation method according to claim 1, characterized in that: The step S3 comprises the following steps: Step S31: Use the Markov random field model to model the original table U, and add Laplace noise in the modeling process to satisfy ε 31 -Differential privacy obtains graph structure; Step S32: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S33: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 33 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; Step S34: To maintain the simulation table and simulation link table First, the U obtained in step S2 ID And the related attributes of table U are filled in and remain unchanged during the sampling process; Step S35: Given the Markov random field model and the simulation link table Under the filled attributes, sample and generate the remaining U table attributes; Step S36: Assume the privacy budget of step S3 is , ε 31 is the privacy budget of step S31, ε 32 is the privacy budget of step S32, ε 33 is the privacy budget of step S33. According to the serial mechanism of differential privacy, step S3 satisfies Differential privacy.

7. The privacy-enhanced relational table data simulation generation method according to claim 1, characterized in that: The step S4 comprises the following steps: Step S41: Use the Markov random field model to model the original table V, and add Laplace noise in the modeling process to make it satisfy ε 41 -Differential privacy obtains graph structure; Step S42: Based on the existing graph structure, the parameters of the Markov random field model are iteratively learned using mirror gradient descent. The Gaussian mechanism is applied in the process of learning the parameters to satisfy Differential privacy; Step S43: Calculate the difference between the marginal distribution of the model and the marginal distribution of the original data , and apply the Laplace mechanism to add noise to satisfy ε 43 -Differential privacy, from the difference between noisy versions Select the marginal distribution corresponding to the maximum value, and add the edge corresponding to the marginal distribution to the graph structure of the Markov random field; Step S44: To maintain the simulation table and simulation link table First, the V obtained in step S2 ID And the related attributes of table V are filled in and remain unchanged during the sampling process; Step S45: Given the Markov random field model and the simulation link table Under the filled attributes, sample and generate the remaining V-table attributes; Step S46: Assume the privacy budget of step S4 is , ε 41 is the privacy budget of step S41, ε 42 is the privacy budget of step S42, ε 43 is the privacy budget of step S43. According to the serial mechanism of differential privacy, step S4 satisfies Differential privacy.

8. The privacy-enhanced relational table data simulation generation method according to claim 1, characterized in that: It also includes step S5: Based on the algorithm flow of steps S1 to S4, according to the serial mechanism of differential privacy, the overall algorithm satisfies Differential privacy.

Images