Privacy measurement algorithm for structured data in data publishing environment

Abstract

The present invention belongs to the field of information security technology, and provides a privacy measurement algorithm for structured data in a data publishing environment. First, the data source collects user information and generates a user information summary, and filters the external data set based on the summary information and calculates the attribute similarity to distinguish private attributes from non-private attributes. Then, the external and internal average information volume and the distribution information volume of the private attributes are calculated, and the private attributes are priced based on demand and time loss to identify sensitive attributes, and the maximum entropy principle is applied to extract quasi-identifiers and identify risk factors in the data. Finally, the data privacy is measured according to the privacy model, and a privacy-oriented anonymity technology is given, which can prove that the privacy and availability under this orientation have achieved the best balance.

Description

Technical Field

[0001] The invention relates to a privacy measurement algorithm for structured data in a data publishing environment, and belongs to the technical field of information security. Background Art

[0002] The rapid development of big data-related technologies has promoted the importance of data in all walks of life. Various organizations can use large amounts of data through big data technology to optimize and upgrade their main business in order to gain a stronger competitive advantage in the market. At the same time, in order to prevent a few Internet giants who have a deeper grasp of big data technology from using their own advantages in big data technology to monopolize and affect market stability; the call for data sharing is growing. However, the privacy issues that come with data sharing have become another bottleneck affecting data sharing. If the privacy issues cannot be properly resolved, large-scale data sharing will be difficult to achieve, and it will become inevitable for giants to use their advantages to monopolize. As the main means of data sharing, the privacy protection issues in data publishing have received widespread attention at home and abroad, and various privacy protection algorithms and models have been generated.

[0003] Privacy protection technology in the publishing environment mainly studies the conversion of original data to anonymous data, and ensures that when anonymous data meets a certain distribution, it can effectively prevent the leakage of data privacy, and then publish the anonymous data. This type of technology is also called anonymous technology. However, no matter what kind of anonymous technology is used, when transforming the original data, the authenticity of a part of the data is always exchanged for the overall privacy enhancement of the data. Although the loss of too much data authenticity ensures privacy, it runs counter to the purpose of data sharing. If the shared data no longer has the value of the original data, data sharing loses its meaning.

[0004] Therefore, the research on anonymity technology has always been centered around the contradiction between privacy and availability, hoping to find an anonymity technology that can achieve the best balance between privacy and availability. However, existing anonymity technologies lack privacy orientation, that is, they do not protect data according to its characteristics before performing anonymization operations on the original data, but instead adopt a global unified operation, which causes some data that originally had a certain degree of privacy to lose this part of the privacy, and also loses the availability of this part of the data, which not only causes a waste of computing overhead, but also makes it always one step away from the goal of safe and efficient data sharing.

[0005] Privacy-oriented anonymization technology will definitely become the focus of future research on data anonymization technology. The privacy-oriented analysis of data mainly studies the following issues: for data that needs to be anonymous, how to analyze its privacy level so that the anonymization operations performed on it are as few as possible and as effective as possible. In the context of data release, data may already have certain privacy characteristics. Extracting the privacy characteristics of the data itself, connecting with the existing data anonymization technology, and achieving a smooth transition from secure data sharing to secure and efficient data sharing will become the focus of future big data research and attract worldwide attention. Summary of the invention

[0006] In order to effectively explore the privacy characteristics in structured data and provide accurate privacy guidance for anonymous technology, this paper proposes a privacy measurement algorithm for structured data. In this scheme, the subject of privacy measurement is the data source with data to be released. First, the data source collects published data sets from the outside world, uses attribute set pairing, classifies the number of paired entries, and separates the private attributes of the data source; then, information entropy pricing is applied to private attributes and sensitive attributes in the data are classified and located, and the maximum entropy principle is applied to non-private attributes to extract quasi-identifiers in the data; finally, a privacy model is defined, starting from the quasi-identifier, data privacy is measured, and the optimal trade-off to achieve the specified privacy level is given.

[0007] The technical solution of the present invention:

[0008] A privacy measurement algorithm for structured data in a data publishing environment, the steps are as follows:

[0009] (1) First, the data source generates a summary of the users involved in the local data, obtains published datasets with similar summaries, and pairs them with the attribute sets of these datasets. Then, the number of pairings is classified to separate the private attributes of the data source.

[0010] The specific process of generating spatiotemporal attribute summaries of datasets, pairing sets, and separating private attributes is as follows:

[0011] (1.1) When a data source collects relevant information about a specific user group, it needs to generate a user summary in a unified format. The information in the summary only describes the region where the data set was collected by users. Because it is time-sensitive, it is called a spatiotemporal summary. First, the composition of the data set is defined as follows:

[0012] D S ={I1, I2, ..., I i ..., I n}

[0013] D s is the database to be released collected by data source S, I iis the record of user i collected in the database, n is D s The number of records in , each record consists of the following:

[0014] I i ={Attr i1 , Attr i2 , ..., Attr ij ..., Attr im}

[0015] Among them, Attr ij is the value of the jth attribute of user i, and m=|I i | is the number of attributes of user i. For structured data sets, the number of attributes of any record is always the same, that is, m = |I1| = |I2| = ... = |I i |=…=|I n |, similarly, put Attr j Defined as the jth attribute in the data set, with n = |Attr1| = |Attr2| = ... = |Attr j |=…=|Attr m |. Among them, the attribute types of the same index are consistent, that is, there is a semantic value set S j , any element x,y∈S j , x and y belong to the same category semantically, if Attr ij ∈D s , there must be Attr ij ∈S j .

[0016] For the dataset D s , whose spatiotemporal summary is defined as follows:

[0017]

[0018] Abstract(I i ) is the administrative region where user i is located, Abstract(D s ) is a set of administrative districts, which covers the dataset D s The administrative regions of all users in .

[0019] (1.2) After generating the summary of the dataset, the data source searches for the published dataset. If the summary of the published dataset overlaps with the summary of the data source by more than a threshold, it will be used to perform set matching with the dataset to be published. The process of calculating summary overlap and set matching is as follows:

[0020] For a series of published datasets D1, D2, ..., D l, and the summaries of these datasets are Abstarct(D1), Abstract(D2,)..., Abstract(D l ), calculate the following value:

[0021]

[0022] Compared with calculating the similarity of the entire dataset, calculating only the summary similarity can help quickly filter out datasets that are completely different from the dataset records to be released, thereby greatly reducing overhead. i The dataset D to be released s Summary and published datasets D i The abstract similarity of , its inherent meaning is the proportion of administrative district duplication, when the proportion α i When the data is smaller than the preset threshold, the possibility of such data sets being used as background knowledge to launch privacy attacks and cause substantial privacy leakage is low, and they will be screened out. After screening out this part of the data, the remaining published data sets are merged, and the private attributes and non-private attributes of the data sets to be published are separated based on this data set. First, the merged data set is defined as follows:

[0023]

[0024] δ is the preset threshold. All the unscreened data sets are vertically merged to form the merged data set D Union , in order to make D Union In line with the definition of the data set, first let the attribute names of different records be compared, adjust the data of the same type to the same attribute, and re-index. In this process, some records may have attributes that other records do not have. Here we only need to add the relevant attributes to the records that do not have the attribute, but fill in the attribute value as empty, so as to ensure D Union It already meets the definition of a structured dataset. Union And the dataset D to be released s , we calculate each attribute Attr j The similarities are as follows:

[0025]

[0026] Then, we calculate the β j Classify, and according to the classification results, D s The attributes in are divided into private attributes and non-private attributes. We use the least squares method for classification. Since we are classifying one-dimensional data, we first set the fitting point as p and define the error sum of squares as follows:

[0027]

[0028] By solving the minimum value of the error function, we can obtain the similarity β j The minimum point of the fitting is used, and this point is used as the threshold. The relevant data greater than the threshold are taken as non-private attributes, and the relevant data less than the threshold are taken as private attributes. The method to solve the minimum value is to find the stationary point of the error function at point p. Since the function is a quadratic function with an opening upward, the stationary point is the minimum point. The solution is:

[0029]

[0030] So we can transform D s The attributes in are divided into two categories as follows:

[0031]

[0032] For attributes whose values ​​in the dataset to be released are mostly repeated in the published dataset, we believe that a large amount of knowledge has been obtained, which is enough to serve as background knowledge for launching privacy attacks. This is reflected in the fact that β j ≥b, and for j For the attribute of <b, its values ​​overlap less with those in the published datasets. It can be considered that other published datasets lack the relevant knowledge of this part. It is classified as a private attribute and should be protected.

[0033] (2) Apply information entropy to price private attributes, divide private attributes into sensitive attributes and non-sensitive attributes, and apply the maximum entropy principle to extract key quasi-identifiers from the remaining non-sensitive attributes and non-private attributes; the specific process is as follows:

[0034] (2.1) The pricing of private attributes is mainly considered from three aspects: pricing for internal data, pricing for external data, and distribution pricing. To facilitate the application of information entropy and help express it, we first define the internal probability of private attributes:

[0035]

[0036] Among them, |Attr ij | represents the data set D s The value is Attr ij The actual probability is difficult for people to know, so we will s Attr ij The frequency of is used as an approximate estimate of the probability, when D s As the number of data increases, the result approaches the true probability according to the law of large numbers.

[0037] By introducing this internal probability, we calculate the information entropy of a private attribute as follows:

[0038]

[0039] If the distribution of external data is consistent with that of the data set, we can use the entropy value to directly price the private attribute to represent the average amount of information obtained from a piece of data from the data set. However, in the publishing environment, the distribution of external data is generally not consistent with that of the local data set, because each data source can only collect a small sample of the overall data, and various errors will occur in the process. Even if the overall data is uniformly sampled, it is difficult for the sample to show all the characteristics of the overall data, resulting in the characteristics that cannot be shown in some samples being the key features in other samples. This also reflects the importance of data sharing and publishing. In the scenario we are considering, the distribution of external data is inconsistent with that of the local data set, so we also need to consider the external data scenario for private attributes to clarify external pricing. Similarly, we first define the probability of external distribution of data:

[0040]

[0041] Among them, N represents the total number of records of the dataset to be released and the released data. The average amount of information is calculated similarly as follows:

[0042]

[0043] Finally, we define the distribution probability of this part of the data as follows:

[0044]

[0045] Where C(n,|Attr ij |) is the number of combinations. It represents the probability of a private attribute Attrj appearing in the data set when combined with the external data distribution. Similarly, we calculate its information content as follows:

[0046] I d (Attr j )=-log(Probability d (Attr j ))

[0047] Based on the amount of information or entropy calculated by the above three processes, we price private attributes as follows:

[0048]

[0049] Among them, a is the attenuation factor, c is a constant related to demand, γ is the constant of the user-adjusted distribution pricing and external pricing ratio, and e is the base of the natural logarithm; at the same time, we introduce t to represent time, and the pricing of private attributes will change over time. By setting the time span between [0, 1], representing the validity period of the published data, and sampling the data pricing within the validity period, a multi-dimensional pricing group for each private attribute is formed, so that private attributes can be classified according to it; when the number of sampling is k, the private attribute Attr j The multidimensional pricing group is represented as follows:

[0050] Price sequence (Attr j )=[ξ j1 , ..., ξ ji ,...ξ jk ]st1≤i≤k

[0051] At the same time, take the first k-1 items and record them as Y j =[ξ j1 , ..., ξ ji ,...ξ j(k-1) ].

[0052] (2.2) After accurately pricing the private attributes, we use the least squares method to classify the priced private attributes into high-priced and low-priced categories, and regard the high-priced category as a sensitive attribute and the low-priced category as a non-sensitive attribute. Then, we apply the maximum entropy principle to extract quasi-identifiers to identify risk factors in the data. For the multidimensional pricing group of each private attribute, we set the fitted hyperplane as:

[0053] ζ k =ζ1x1+…+ζ i x i +…+ζ k-1 x k-1 =XZ st1≤i≤k

[0054] The error sum of squares function is defined as follows:

[0055]

[0056] To find the maximum value of the error square sum function, we can solve:

[0057]

[0058] From this, private attributes can be categorized as follows:

[0059]

[0060] (2.3) Then, we apply the maximum entropy principle to extract quasi-identifiers from the data. In the data set, quasi-identifiers play a role in distinguishing records, and attackers often use quasi-identifiers as background knowledge to launch privacy attacks. Therefore, accurate extraction of quasi-identifiers means accurate estimation of the attacker's background knowledge, which is of great significance for subsequent privacy protection work. The entropy of the remaining attributes is calculated as follows:

[0061]

[0062] Then, the calculated attributes are sorted according to the entropy value, with the attribute Attr with the largest entropy value. j As the basis, record it as Bse, and calculate its joint entropy with other remaining attributes as follows:

[0063]

[0064] Among them, Probability(Attr ij , Attr ik ) is the corresponding value Attr ij and Attr ik The joint probability of . Select and attribute Attr j The attribute with the largest joint entropy Attr k , forming a new basis <Attr j , Attr k > Update Base and calculate the joint entropy of base and other remaining attributes as follows:

[0065]

[0066] Among them, Base i is the value combination of the Base attribute corresponding to user i, and Probability (Base i , Attr il) Base i With Attr il Repeat the above steps and update the basis until the entropy of the basis is the maximum entropy of the internal data. Then, we can find a shortest quasi-identifier combination as the basis, whose value can uniquely distinguish any record in the data set. Obviously, when the attacker's background knowledge reaches the length of the basis, all information in the data will be leaked. Therefore, for privacy measurement, we can consider the maximum background knowledge length of the attacker to be the length of the basis.

[0067] (3) We construct a privacy model to describe the impact of data distribution on data privacy and the ability to resist privacy attacks, and provide a privacy security standard to determine which part of the data in the dataset should be protected.

[0068] (3.1) When discussing data privacy, the following consideration is often made: under inference-based attacks, the lower the probability of leaking exact information, the better the data privacy. Therefore, we define the privacy model as follows:

[0069] p(unknow|know)≤η

[0070] This formula defines the upper bound of the probability of using known information as background knowledge to infer unknown information in the data. If the probability of using known information to infer accurate unknown information from the outside world in the data set is less than or equal to η, then the data is said to satisfy the η privacy model; if there is an algorithm that operates on the data set so that it satisfies the η privacy model, then the algorithm is said to satisfy η inference privacy.

[0071] As mentioned earlier, we consider the most common situation in data release scenarios, that is, quasi-identifiers serve as background knowledge in the data set to infer other information. Any quasi-identifier may serve as background knowledge. Therefore, in this privacy model, known information is the quasi-identifier, and correspondingly, unknown information is a sensitive attribute. For non-sensitive attributes, we do not regard them as unknown information that needs to be inferred, because the content they contain does not have a certain amount of information and cannot be called privacy data.

[0072] (3.2) Next, we look for quasi-identifiers or combinations of quasi-identifiers that make the data not satisfy η privacy. We add these quasi-identifiers to the threat set. At the same time, we add the inferred information to the risk set. The length of the background knowledge in the search process is less than or equal to |Base|. Where |Base| is the length of Base.

[0073] The inference of known information generally adopts a linking operation, that is, linking data sets with a certain quasi-identifier combination to obtain matching and unmatched records, and then removing the known information from these records. The remaining information is the linked information, and then the privacy model is applied to determine whether this part meets η privacy.

[0074] In the process of extracting key quasi-identifiers, we found a Base that can uniquely distinguish a record in the data set. Unless η is greater than or equal to 1, there will be no algorithm that satisfies η inference privacy when this set of key quasi-identifiers is used as background knowledge. Therefore, as we continue to increase background knowledge to |Base|, there will inevitably be some combinations of quasi-identifiers that make the data not satisfy η privacy, that is, the above process converges.

[0075] (3.3) Construct a background knowledge inference graph, map each element in the threat set into a point on the graph, and fill in an edge for the corresponding point of the quasi-identifier sequence that constitutes the parent-child sequence, and finally form a |Base| partial graph.

[0076] By adjusting the order of nodes in the |Base| graph, we can represent it as a hierarchical structure, with the parent sequence at the top and the child sequence at the bottom. Obviously, the sequence at the highest level is the key factor that causes the data to not meet the η inferred privacy. An attacker using any node in the graph as background knowledge can cause privacy leakage exceeding η. For this problem, we can consider the method of separating background knowledge to improve data privacy.

[0077] (3.4) The main purpose of background knowledge separation is to limit any background knowledge to the parent sequence of the highest-level node in the |Base| graph, so as to ensure that the data meets η privacy. In the process of background knowledge separation, we use the method of calculating node similarity to plan the path of background knowledge separation, so as to find a method that requires the least operation division and maximizes the availability of data.

[0078] Beneficial effects of the present invention: Data sharing plays a huge role in big data technology, and the publishing scenario, as the main scenario for data sharing, has received widespread attention. In the publishing scenario, the main research is on how to effectively publish data while ensuring the privacy and security of the data. However, current work directly uses various anonymous technologies to impose privacy protection on the data without measuring the privacy of the data. Such research work that ignores the privacy characteristics of the data itself maintains the balance between privacy and availability after the data is anonymized at a low level, reducing the efficiency of data publishing. To this end, we designed a privacy measurement method for structured data in a publishing environment, which can specifically explore the privacy characteristics of the data and perform anonymous protection based on the privacy characteristics of the data, thereby maintaining the balance between privacy and availability at a high level and improving the efficiency of privacy protection in the publishing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is an organizational chart of the privacy measurement algorithm described in the present invention.

[0080] Figure 2 This is a flow chart of separating private attributes according to the present invention.

[0081] Figure 3 A flow chart for identifying risk factors according to the present invention.

[0082] Figure 4 A flowchart of output privacy measurement and guidance according to the present invention. DETAILED DESCRIPTION

[0083] In order to express the purpose, technical solutions and advantages of the present invention more clearly, the present invention will be further described in detail below through embodiments and drawings.

[0084] A privacy measurement algorithm for structured data in a data publishing environment. The method includes the judgment of proprietary data, the identification of risk factors, the measurement of data privacy and the privacy orientation of anonymity technology.

[0085] Reference Figure 2 , the specific operation process of separating private attributes is as follows:

[0086] Step 1. Data source S collects target user information to form data set D s .

[0087] Step 2. Data source S is based on data set D s The user features of the dataset are used to generate the spatiotemporal summary. The formula is as follows:

[0088]

[0089] Step 3. Calculate the dataset D s The similarity between the spatiotemporal summary of the published dataset and the spatiotemporal summary of the published dataset is calculated as follows:

[0090]

[0091] Step 4. Merge the published data whose similarity is greater than the threshold. The formula is as follows:

[0092]

[0093] Step 5. For each attribute of the published dataset, calculate its similarity with the same attribute of the merged published dataset. The formula is as follows:

[0094]

[0095] Step 6. For the similarity calculated for each attribute, use the least squares method to classify, calculate the classification threshold, and define the error as follows:

[0096]

[0097] Step 7. Find the maximum value of the error term and solve for the threshold value:

[0098]

[0099] Step 8. According to the threshold, the published dataset D s The attributes are classified as follows:

[0100]

[0101] Reference Figure 3 The specific operation process of identifying risk factors is as follows:

[0102] Step 9. Calculate the internal probability for private attributes using the following formula:

[0103]

[0104] Step 10. Calculate the internal information entropy of private attributes as internal pricing, the formula is as follows:

[0105]

[0106] Step 11. Calculate the external probability for private attributes using the following formula:

[0107]

[0108] Step 12. Calculate the external average information content of private attributes as the external pricing, the formula is as follows:

[0109]

[0110] Step 13. Calculate the distribution probability for private attributes using the following formula:

[0111]

[0112] Step 14. Calculate the distribution information of private attributes as the distribution pricing, the formula is as follows:

[0113] I d (Attr j )=-log(Probability d (Attr j ))

[0114] Step 15. Pricing private attributes, the formula is as follows:

[0115]

[0116] Step 16. Sample the attribute pricing within a given time range to obtain the specific pricing of the private attribute at certain time points. When the number of samplings is k, we set the private attribute Attr j The multidimensional pricing group is represented as follows:

[0117] Price sequence (Attr j )=[ξ j1 , ..., ξ ji ,...ξ jk ]st1≤i≤k

[0118] At the same time, take the first k-1 items and record them as Y j =[ξ j1, ..., ξ ji ,...ξ j(k-1) ].

[0119] Step 17. Apply the least squares method to classify private attributes, and the fitted hyperplane is:

[0120] ζ k =ζ1x1+…+ζ i x i +…+ζ k-1 x k-1 =XZ st1≤i≤k

[0121] in

[0122]

[0123] Step 18. According to the positional relationship with the fitted hyperplane, the private attributes are classified as follows:

[0124]

[0125] Step 19. Calculate the entropy of the remaining attributes using the following formula:

[0126]

[0127] Step 20. For the remaining attributes, take the attribute with the largest entropy as the basis and calculate the joint entropy of the basis and other attributes. The formula is as follows:

[0128]

[0129] Step 21. Take the attribute with the largest joint entropy and merge it with the original basis to form a new basis, and calculate the joint entropy of this basis and other attributes. The formula is as follows:

[0130]

[0131] Step 22. Repeat step 21 until the joint entropy reaches the maximum entropy of the internal data, at which point the basis obtained is the key quasi-identifier.

[0132] Reference Figure 4 ,The specific operation process of outputting privacy measurement and guidance is as follows:

[0133] Step 23. Starting from any quasi-identifier, perform a link operation on the dataset to be published, find the known information and location information sets that do not satisfy η privacy, and add them to the threat set and risk set respectively.

[0134] Step 24. Construct a background knowledge inference graph, map each element in the threat set into a point on the graph, and fill in an edge for the corresponding point for the quasi-identifier sequence that constitutes the parent-child sequence, and finally form a |Base| partial graph.

[0135] Step 25: Layer the graph and calculate the similarity of the highest-level nodes.

[0136] Step 26. Separate background knowledge based on node similarity and output privacy guidance.

Claims

1. A privacy measurement algorithm for structured data in a data publishing environment, characterized by: Here are the steps: (1) The data source generates summaries of users involved in local data, obtains published datasets with similar summaries, pairs them with the datasets’ attribute sets, classifies the number of pairs, and separates private attributes of the data source; The specific process of generating spatiotemporal attribute summaries of datasets, pairing sets, and separating private attributes is as follows: (1.1) When a data source collects relevant information about a specific user group, the data source needs to generate a user summary in a unified format and define the composition of the data set as follows: D S ={I1,I2,…,I i …,I n } Among them, D s is the database to be released collected by data source S, I i is the record of user i collected in the database, n is D s The number of records in , each record consists of the following: I i ={Attr i1 ,Attr i2 ,…,Attr ij …,Attr im } Among them, Attr ij is the value of the jth attribute of user i, m=|I i | is the number of attributes of user i. For structured data sets, the number of attributes of any record is always the same, m = |I1| = |I2| = ... = |I i |=…=|I n |, put Attr j Defined as the jth attribute in the dataset, n = |Attr1| = |Attr2| = ... = |Attr j |=…=|Attr m |; Among them, the attribute types of the same index are consistent, and there is a semantic value set S j , any element x,y∈S j , x and y belong to the same semantic category, if Attr ij ∈D s , then Attr ij ∈S j ; For the dataset D s , whose spatiotemporal summary is defined as follows: Abstract(I i ) is the administrative region where user i is located, Abstract(D s ) is a set of administrative districts, which covers the dataset D s Administrative regions of all users in; (1.2) After the summary of the dataset is generated, the data source searches for published datasets. If the summary of the published dataset overlaps with the summary of the data source by more than a threshold, it is used to perform set matching with the dataset to be published. The process of calculating summary overlap and set matching is as follows: For a series of published datasets D1, D2, ..., D l , and the summaries of these datasets are Abstarct(D1),Abstract(D2,)...,Abstract(D l ), calculate the following value: Where 1≤i≤l Among them, α i The dataset D to be released s Summary and published datasets D i The summary similarity of i When it is less than the preset threshold, the corresponding data set is screened out; the remaining published data sets are merged, and the merged data set is used to separate the private attributes and non-private attributes of the data set to be published; the merged data set is as follows: Among them, δ is the preset threshold, and all the unscreened data sets are vertically merged to form a merged data set D Union By comparing the attribute names of different records, the same type of data is adjusted to the same attribute, the index is rearranged, the attributes between records complement each other, and the attribute values ​​are filled in as blanks to ensure D Union It meets the definition of structured data set; for data set D Union And the dataset D to be released s , calculate each attribute Attr j The similarities are as follows: Through the β j Classification, according to the classification results, D s The attributes in are divided into private attributes and non-private attributes, and the classification method adopts the least squares method; let the fitting point be p, and define the error sum of squares as follows: By solving the minimum value of the error function, we can obtain the similarity β j The minimum point of the error function is fitted, and the minimum point is used as the threshold. The relevant data greater than the threshold are non-private attributes, and the relevant data less than the threshold are private attributes. The method for solving the minimum value is to find the stationary point of the error function at point p, and the solution is: D s The attributes in are divided into two categories as follows: Where Attr j ∈D s For the attributes with values in the dataset to be released that are repeated in the released dataset, they are defined as attributes satisfying β j ≥ p, and for the attributes satisfying β j < p, they are classified as private attributes; (2) Apply information entropy to price private attributes, divide private attributes into sensitive attributes and non-sensitive attributes, and apply the maximum entropy principle to extract quasi-identifiers from the remaining non-sensitive attributes and non-private attributes; the specific process is as follows: (2.1) Pricing of private attributes: Pricing of internal data, pricing of external data, and distribution pricing; First, define the internal probability of private attributes: Among them, |Attr ij | represents the data set D s The value is Attr ij The number of D s Attr ij The frequency of is used as an approximate estimate of probability; By introducing this internal probability, the information entropy of a private attribute is calculated as follows: If the distribution of external data is consistent with that of the data set, the entropy value is used to price the private attribute, which indicates the average amount of information obtained from a piece of data in the data set. If the distribution of external data is inconsistent with that of the local data set, the external data scenario is considered for the private attribute to clarify the external pricing. The probability of external distribution of data is as follows: Among them, N represents the total number of records of the data set to be released and the released data. The average amount of information is calculated as follows: The distribution probability of this part of data is as follows: Among them, C(n,|Attr ij |) is the number of combinations, representing a private attribute Attr j In combination with the external data distribution, the probability of its appearance in the data set is calculated as follows: I d (Attr j )=-log(Probability d (Attr j )) According to the amount of information or entropy calculated by the above three processes, the private attributes are priced as follows: Where a is the attenuation factor, c is a constant related to demand, γ is a constant for the ratio of user-adjusted distributed pricing to external pricing, and e is the base of the natural logarithm; t represents time, and the pricing of private attributes changes over time; the time span is between [0,1], representing the validity period of the published data. During the validity period, the data pricing is sampled to form a multidimensional pricing group for each private attribute, and the private attributes are classified according to it; when the number of sampling is k, the private attribute Attr j The multidimensional pricing group is represented as follows: Price sequence (Attr j )=[ξ j1 ,…,x ji ,…x jk ], among which 1≤i≤k Take the first k-1 items and record them as Y j =[ξ j1 ,…,ξ ji ,…ξ j(k-1) ]; (2.2) After pricing the private attributes, the least squares method is used to classify the priced private attributes into high-price and low-price categories, and the high-price category is regarded as a sensitive attribute and the low-price category as a non-sensitive attribute; for the multidimensional pricing group of each private attribute, the fitting hyperplane is set as: g k =ζ1x1+…+ζ i x i +…+g k-1 x k-1 , among which 1≤i≤k The error sum of squares function is defined as follows: Among them, Label(Attr j )=private Find the maximum value of the error square sum function and we get: Among them, Label(Attr j )=private,1≤i <k Therefore, private attributes are classified as follows: Among them, Label(Attr j )=private (2.3) Apply the maximum entropy principle to extract quasi-identifiers from the data. In the data set, quasi-identifiers play the role of distinguishing records; the entropy of the remaining attributes is calculated as follows: Sort the calculated attributes by entropy value, with the attribute Attr with the largest entropy value j As the basis, record it as Base, and calculate its joint entropy with other remaining attributes as follows: Among them, Probability (Attri j ,Attr ik ) is the corresponding value Attr ij and Attr ik The joint probability of; select and attribute Attr j The attribute with the largest joint entropy Attr k , forming a new set of bases <Attr j , Attr k > Update Base and calculate the joint entropy of base and other remaining attributes as follows: Among them, Base i is the value combination of the Base attribute corresponding to user i, Probability(Base i ,Attr il ) is Base i With Attr il Repeat the above steps and update the basis until the entropy of the basis is the maximum entropy of the internal data, and find a shortest quasi-identifier combination as the basis, whose value is used to distinguish any record in the data set; (3) Construct a privacy model to describe the impact of data distribution on data privacy and the ability to resist privacy attacks; provide privacy security standards to determine the data in the data set that needs to be protected; (3.1) When discussing data privacy, the possibility of exact information leakage is low and data privacy is good; the privacy model is defined as follows: p(unknow|know)≤η This formula defines the upper bound of the probability of using known information as background knowledge to infer unknown information in the data. For the external known and unknown information in the data set, if the probability of using known information to infer unknown information is less than or equal to η, the data is said to satisfy the η privacy model; if there is an algorithm that operates on the data set to satisfy the η privacy model, then the algorithm is said to satisfy η inference privacy; Quasi-identifiers act as background knowledge in the data set to infer other information. In this privacy model, known information is a quasi-identifier, and unknown information is a sensitive attribute. Non-sensitive attributes are not considered unknown information that needs to be inferred and are not called private data. (3.2) Find the quasi-identifier or combination of quasi-identifiers that causes the data to not satisfy η privacy, add the quasi-identifier to the threat set, and at the same time, add the information that can be inferred to the risk set. The length of the background knowledge in the search process is less than or equal to |Base|, where |Base| is the length of Base; For the inference of known information, a link operation is adopted. A quasi-identifier is used to combine and link the data set to obtain matching and unmatched records. The known information is removed from the records. The remaining information is the linked information. The privacy model is applied to determine whether this part meets η privacy. In the process of extracting quasi-identifiers, a Base is found to uniquely distinguish a record in the data set. When η is greater than or equal to 1, the group of quasi-identifiers is used as background knowledge to satisfy the algorithm of η inference privacy. When the background knowledge is continuously increased to |Base|, some combinations of quasi-identifiers appear, making the data not satisfy η privacy, and the above process converges. (3.3) Construct a background knowledge inference graph, map each element in the threat set into a point on the graph, and fill in an edge for the corresponding point of the quasi-identifier sequence that constitutes the parent-child sequence, and finally form a |Base| partial graph; By adjusting the order of nodes in the |Base| graph, it is represented as a hierarchical structure, with parent sequences on top and child sequences on the bottom. The highest-level sequence is the key factor that causes the data to not meet the η inference privacy. The attacker uses any node in the graph as background knowledge to cause privacy leakage exceeding η. The background knowledge separation method is used to improve the privacy of the data. (3.4) Background knowledge separation is to restrict any background knowledge to the parent sequence of the highest-level node in the |Base| graph to ensure that the data meets η privacy. In the process of background knowledge separation, the path of background knowledge separation is planned by calculating node similarity to ensure data availability.

Images