Social network-oriented privacy ontology construction method and system

Abstract

The invention discloses a social network-oriented privacy ontology construction method and system, and relates to the technical field of security information processing, and the method comprises the steps: obtaining a social network data set, and extracting candidate privacy attributes; calculating information entropies of the candidate privacy attributes and mutual information of the candidate privacy attributes and identity or target sensitive attributes, normalizing the information entropies, and weighting the information entropies and the identity or target sensitive attributes to obtain a comprehensive sensitivity score so as to screen a privacy attribute set; conditional mutual information of the privacy attribute pair is calculated under the constraint of conditional attributes to serve as dependency intensity, an inference direction is determined based on a conditional entropy reduction criterion and normalized to obtain a relation weight, and a privacy ontology containing attribute nodes and weighted dependency edges is constructed; and receiving data batch according to time slices, executing node addition detection, statistic increment updating and relation evidence accumulation, and incrementally updating an ontology structure or weight. Through the technical scheme of the invention, privacy attribute sensitivity quantification and inference relation modeling are realized, and privacy risk identification accuracy and interpretability are improved.

Description

Technical Field

[0001] This invention relates to the field of secure information processing technology, and in particular to a method for constructing a privacy ontology for social networks and a system for constructing a privacy ontology for social networks. Background Technology

[0002] Online social networking platforms accumulate vast amounts of user profiles and interaction data. User records typically contain multiple attribute variables and their values, which are correlated with user identity attributes or other sensitive attributes. In privacy protection and risk assessment scenarios, it is necessary to structure the semantic concepts of privacy attributes and their interdependencies to support privacy leakage path analysis, risk measurement inference, and governance strategy formulation.

[0003] Existing privacy ontology construction solutions mainly revolve around extracting privacy policies / compliance clauses, modeling predefined privacy concept systems, and organizing rules or inference mechanisms, or forming an ontology structure based on a pre-defined list of privacy attributes and relational rules for specific business scenarios. These solutions generally suffer from the following shortcomings in practical applications: First, the sensitivity of privacy attributes often relies on human experience or a single indicator for judgment, lacking a unified quantitative measurement framework for all user samples. This makes it difficult to compare the sensitivity of different attributes horizontally, and the screening criteria for privacy attributes are inconsistent.

[0004] Secondly, attribute dependencies are often characterized by correlation or rules, lacking a quantitative determination of the strength of dependency under conditional attribute constraints. This makes it difficult to distinguish between direct dependencies and indirect couplings, and easily introduces unstable relationships.

[0005] Third, the direction of attribute relationships usually depends on manual setting, lacking an inference direction determination mechanism based on uncertainty reduction, making it difficult to form a directed dependency structure that can be used for inference path analysis.

[0006] Fourth, social network data is continuously updated. Existing ontologies are mostly built offline in one go, lacking a mechanism for incremental updates of statistics on batch data at the time granularity and for unified determination and update of structure and relation weights at the time slice boundary. This makes it difficult for ontologies to maintain consistency and effectiveness as the data evolves.

[0007] Therefore, there is an urgent need for a privacy ontology construction method for social networks. This method should perform sensitive quantitative screening of candidate privacy attributes based on the statistics of all user samples, determine the dependence strength of privacy attribute pairs under conditional constraints, determine the inference direction and relation weight, and support incremental updates and unified judgment updates of data in batches by time slice, thereby forming a sustainable evolutionary privacy ontology structure. Summary of the Invention

[0008] To address the aforementioned issues, this invention provides a method and system for constructing a privacy ontology for social networks. It extracts candidate privacy attributes from a social network dataset and calculates the information entropy and mutual information between the entropy and user identity attribute variables or target sensitive attribute variables based on the probability distribution of all user samples. A comprehensive sensitivity score is constructed using normalization and weighting, achieving unified quantitative screening of privacy attributes. This avoids inconsistencies and insufficient comparability caused by relying solely on human experience or single indicators. Under the constraint of a preset set of conditional attributes, the conditional mutual information of privacy attribute pairs is calculated as the dependency strength. The inference direction is determined based on the conditional entropy reduction criterion, and relation weights are formed using normalized dependency strength. This yields attribute dependency edges with directional and strength characterization, reducing the risk of misjudging indirect coupling as direct dependency and improving the stability and interpretability of inference path analysis. By dividing time slices by time granularity, performing node addition detection on batch data, incrementally updating statistics, and accumulating observational evidence, and uniformly determining and incrementally updating the ontology structure or relation weights at the end of each time slice, the privacy ontology can maintain consistency and effectiveness as the data continues to evolve, improving the real-time performance and engineering implementation capabilities of privacy risk assessment and governance.

[0009] To achieve the above objectives, this invention provides a method for constructing a privacy ontology for social networks, comprising: Obtain a social network dataset, which includes multiple user records. Each user record contains multiple privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. Candidate privacy attributes are extracted from the social network dataset to form a candidate privacy attribute set; For each candidate privacy attribute in the candidate privacy attribute set, the information entropy is calculated based on the probability distribution obtained from the statistics of all user samples, and the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable is calculated. After normalizing the information entropy and the mutual information respectively, a comprehensive sensitivity score is constructed according to a preset weight, and a set of privacy attributes is obtained by filtering based on the comprehensive sensitivity score; For each privacy attribute pair in the privacy attribute set, conditional mutual information is calculated under the constraints of a preset conditional attribute set, and the conditional mutual information is used as the dependency strength between the privacy attribute pairs; when the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pairs is determined based on the conditional entropy reduction criterion, and the dependency strength is normalized to obtain the relation weight; A privacy ontology is constructed based on the set of privacy attributes. The privacy ontology includes attribute nodes and attribute dependency edges represented by the inference direction and the relation weight. The social network data is divided into multiple time slices according to a preset time granularity, and the corresponding data batch is received in each time slice; for the data batch, the addition detection and initialization of attribute nodes, incremental updates of node statistics, and accumulation of attribute relationship observation evidence are performed. At the end of the time slice, the candidate attribute dependencies accumulated within the time slice are uniformly determined. When the dependency strength corresponding to the candidate attribute dependency exceeds the preset threshold, the structure of the privacy ontology or the relation weight is incrementally updated.

[0010] In the above technical solution, preferably, the normalization processing of the information entropy and the mutual information adopts extreme value normalization based on the candidate privacy attribute set, so that the normalization result falls into the [0,1] interval, and a very small positive number is introduced to avoid the denominator being zero; the comprehensive sensitivity score is the weighted sum of the normalized information entropy and the normalized mutual information according to a preset weight, and the weight sum of the preset weights is 1.

[0011] In the above technical solution, preferably, the probability distribution used to calculate the information entropy and the mutual information is estimated from the statistical results of all user samples in the social network dataset, and the probability distribution includes at least the marginal probability distribution of the candidate privacy attribute and the joint probability distribution of the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable.

[0012] In the above technical solution, preferably, the preset condition attribute set is a set of privacy attribute variables other than the privacy attribute pairs to be analyzed, or a subset of condition attributes selected from the privacy attribute set based on correlation and comprehensive sensitivity scores.

[0013] In the above technical solution, preferably, when making a unified determination on the privacy attribute pair at the end of the time slice, a statistical significance test is performed on its conditional mutual information; when the conditional mutual information exceeds the preset threshold and passes the statistical significance test, it is determined that the privacy attribute pair has a stable attribute dependency relationship.

[0014] In the above technical solution, preferably, determining the inference direction based on the conditional entropy reduction criterion includes: calculating the first conditional entropy reduction amount and the second conditional entropy reduction amount respectively under the same preset conditional attribute set constraints. The first conditional entropy reduction amount represents the degree of reduction of the uncertainty of the second privacy attribute by introducing the first privacy attribute under the premise of a known conditional attribute set. The second conditional entropy reduction amount represents the degree of reduction of the uncertainty of the first privacy attribute by introducing the second privacy attribute under the premise of a known conditional attribute set. The one with the larger conditional entropy reduction amount and passing the statistical significance test is determined as the starting attribute of the inference direction, and a corresponding directed attribute dependency relationship edge is established in the privacy ontology.

[0015] In the above technical solution, preferably, the privacy ontology further includes semantic extension information, which includes at least fine-grained sub-attributes of privacy attributes, candidate synonyms, and hyponyms and hypernyms; the semantic extension information is used to obtain semantic support weights based on the concept co-occurrence statistics of social network corpora, and the semantic support weights are used for attribute node description and attribute relationship interpretation in the privacy ontology.

[0016] In the above technical solution, preferably, the addition detection and node initialization of attribute nodes for the batch data includes: when it is detected that the batch data contains privacy attribute variables that do not appear in the privacy ontology, a corresponding attribute node is added in the privacy ontology, and the value space, statistical count and probability estimation parameters of the attribute node are initialized; The incremental updates of node statistics include: incremental updates of attribute value frequency, marginal probability distribution and joint statistical information, and updating the statistical basis used for online estimation of information entropy, mutual information and conditional mutual information using cumulative counting or sliding window mechanisms.

[0017] In the above technical solution, preferably, the accumulation of observation evidence of attribute relationships includes: taking the privacy attribute pairs that appear together in the data batch as observation evidence and updating the corresponding joint statistical information; The conditional mutual information of the candidate attribute dependencies is calculated when the accumulated samples within the time slice meet the minimum sample size or stability requirements. At the end of the time slice, if the candidate attribute dependency relationship meets the unified judgment condition, then in the privacy ontology, the addition of non-existent attribute dependency relationships is performed, and the relationship weight update is performed on existing attribute dependency relationships.

[0018] This invention also proposes a privacy ontology construction system for social networks, used to implement the privacy ontology construction method for social networks disclosed in any of the above technical solutions, including: The data acquisition module is used to acquire a social network dataset, which includes multiple user records. Each user record contains multiple privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. The privacy attribute extraction module is used to extract candidate privacy attributes from the social network dataset to form a candidate privacy attribute set; The sensitivity assessment module is used to calculate the information entropy for each candidate privacy attribute in the candidate privacy attribute set based on the probability distribution obtained from the statistics of all user samples, and to calculate the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable; after normalizing the information entropy and the mutual information respectively, a comprehensive sensitivity score is constructed according to a preset weight, and a set of privacy attributes is obtained by filtering based on the comprehensive sensitivity score; The attribute dependency analysis module is used to calculate conditional mutual information for each privacy attribute pair in the privacy attribute set under the constraints of a preset conditional attribute set, and to use the conditional mutual information as the dependency strength between the privacy attribute pairs; when the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pairs is determined based on the conditional entropy reduction criterion, and the dependency strength is normalized to obtain the relation weight. A privacy ontology construction module is used to construct a privacy ontology based on the privacy attribute set. The privacy ontology includes attribute nodes and attribute dependency edges represented by the inference direction and the relation weight. The time-slice incremental update module is used to divide the social network data into multiple time slices according to a preset time granularity, and receive corresponding data batches in each time slice; for the data batches, it performs new attribute node detection and node initialization, incremental update of node statistics, and accumulation of attribute relationship observation evidence; at the end of the time slice, it performs a unified judgment on the candidate attribute dependencies accumulated in the time slice, and when the dependency strength corresponding to the candidate attribute dependency exceeds the preset threshold, it performs an incremental update on the structure of the privacy ontology or the relationship weight.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By extracting candidate privacy attributes from the social network dataset, and calculating the information entropy of the candidate privacy attributes and their mutual information with user identity attribute variables or target sensitive attribute variables based on the probability distribution of all user samples, the two types of indicators are normalized and then a comprehensive sensitivity score is formed according to the preset weight. Based on this, a set of privacy attributes is selected, realizing a unified quantitative measurement and automated screening of privacy attribute sensitivity, and improving the comparability and consistency of screening criteria between different privacy attributes.

[0020] (2) By calculating the conditional mutual information of privacy attribute pairs under the constraints of the preset conditional attribute set, and using the conditional mutual information as the dependency strength and filtering with a preset threshold, the direct dependency between attributes can be characterized while controlling the interference of other attributes, reducing the probability of misjudging indirect correlation as direct dependency, and improving the stability and reliability of attribute dependency identification.

[0021] (3) By determining the inference direction of privacy attribute pairs based on the conditional entropy reduction criterion and normalizing the dependency strength to obtain the relation weight, the attribute dependency relation edges in the privacy ontology simultaneously possess both directionality and strength representation, supporting the directed analysis of privacy inference paths and the characterization of risk propagation links, thereby enhancing the interpretability and usability of the ontology structure.

[0022] (4) By dividing social network data into time slices according to time granularity, and performing batch detection and initialization of attribute nodes, incremental updates of node statistics and accumulation of observation evidence of attribute relationships on the data within the time slices, and uniformly judging the dependency relationship of candidate attributes and incrementally updating the ontology structure or relationship weight at the end of the time slice, the privacy ontology can maintain consistency and timeliness as the data continues to evolve, thereby improving the engineering implementation capability of privacy risk assessment and governance strategy formulation. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a method for constructing a privacy ontology for social networks, as disclosed in one embodiment of the present invention. Figure 2 This is a schematic diagram of the architecture of a privacy ontology construction method for social networks disclosed in one embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 and Figure 2 As shown, a privacy ontology construction method for social networks provided by the present invention includes: Obtain a social network dataset, organized by user records. Each dataset contains multiple user records, each with several privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. Based on field statistics and attribute extraction rules in the dataset, candidate privacy attributes are extracted to form a candidate privacy attribute set.

[0026] For each candidate privacy attribute in the candidate privacy attribute set, the information entropy is calculated based on the probability distribution obtained from the statistics of all user samples to characterize the uncertainty of the attribute value; at the same time, the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable is calculated to characterize the correlation strength between the candidate privacy attribute and the identity / sensitive attribute.

[0027] After normalizing information entropy and mutual information respectively, a comprehensive sensitivity score is constructed according to preset weights to realize the quantification of sensitivity at the single attribute level. The set of privacy attributes is selected based on the comprehensive sensitivity score, so that the attributes entering ontology modeling have higher privacy relevance and sensitivity, thereby reducing the interference of irrelevant attributes on relation learning.

[0028] For each privacy attribute pair in the privacy attribute set, conditional mutual information is calculated under the constraints of a preset conditional attribute set. This conditional mutual information is used as the dependency strength between the privacy attribute pairs to characterize the information dependency that still exists between the two attributes given the conditional attributes. When the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pair is determined based on the conditional entropy reduction criterion: by comparing the degree of uncertainty reduction of the other attribute after knowing one attribute, the side with the greater information contribution is determined as the starting point for inference; and the dependency strength is normalized to obtain the relation weight, which is used to characterize the strength of the attribute dependency relation edge.

[0029] A privacy ontology is constructed based on a set of privacy attributes. The privacy ontology includes attribute nodes and attribute dependency edges represented by inference direction and relation weight. Privacy attributes are used as attribute nodes, and inference direction and relation weight are used to jointly represent directed attribute dependency edges, thereby forming an interpretable attribute inference structure that supports a unified expression of which attributes are inferred from which attributes and how strong the inference is.

[0030] To adapt to the continuous evolution of social network data, social network data is divided into multiple time slices according to a preset time granularity, and corresponding data batches are received within each time slice.

[0031] For each data set, the following steps are performed in batches: new attribute node detection and initialization, incremental update of node statistics, and accumulation of evidence for attribute relationship observation. At the end of the time slice, a unified judgment is made on the candidate attribute dependencies accumulated within the time slice. When the dependency strength corresponding to the candidate attribute dependency exceeds a preset threshold, the structure or relation weight of the privacy ontology is incrementally updated.

[0032] This enables a closed-loop mechanism where the ontology is updated as data arrives and decisions are made centrally at the end of the time slice, improving the timeliness and stability of the ontology and reducing jitter caused by frequent structural changes.

[0033] In the above implementation, preferably, the normalization processing of information entropy and mutual information adopts extreme value normalization based on the candidate privacy attribute set. Specifically, the minimum and maximum values ​​of the information entropy corresponding to all candidate privacy attributes in the candidate set are taken, and the information entropy of any candidate privacy attribute is normalized according to "(current value - minimum value) / (maximum value - minimum value + minimum positive number)". The mutual information is also normalized according to the extreme values ​​in the candidate set, so that the normalization result falls into the [0,1] interval, and a minimum positive number is introduced to avoid the anomaly caused by the denominator being zero.

[0034] The comprehensive sensitivity score is obtained by weighting and summing the normalized information entropy and normalized mutual information according to preset weights, and the preset weights satisfy the condition that the sum of the weights is 1. Through the above processing, on the one hand, entropy and mutual information of different dimensions and scales are comparable at the same scale, and on the other hand, the scoring results are adaptive to the overall distribution of the candidate set, which facilitates stable screening of privacy attribute sets under different platforms and different data scales. The implementation effect is that the screening threshold is easier to configure and the score fluctuation is smaller when migrating across datasets.

[0035] In the above implementation, preferably, the probability distribution used to calculate information entropy and mutual information is estimated from the statistical results of all user samples in the social network dataset. Specifically, for each candidate privacy attribute, the frequency of each value is counted and converted into a marginal probability distribution; for the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable, the co-occurrence frequency of their value pairs is counted and converted into a joint probability distribution.

[0036] Information entropy is calculated based on the aforementioned marginal probability distribution, reflecting the dispersion and uncertainty of the privacy attribute's value. Mutual information is calculated based on the joint probability distribution and marginal probability distribution, reflecting the degree of information sharing between the privacy attribute and the identity / sensitive attribute. By establishing the probability distribution on the statistics of all user samples, entropy and mutual information become more sensitive to overall data characteristics, reducing scoring bias caused by individual noise. The implementation effect is a more consistent characterization of the coupling degree between sensitive attributes and identity / target sensitive attributes.

[0037] In the above embodiments, preferably, the preset condition attribute set adopts one of two implementation methods: One approach is to pre-set the conditional attribute set as a set of privacy attribute variables other than the privacy attribute pairs to be analyzed, so that the conditional mutual information calculation is carried out under the constraints of the remaining privacy attributes, so as to isolate as much indirect correlation caused by third-party privacy attributes as possible, thereby getting closer to stable dependencies.

[0038] Another approach is to pre-select a subset of conditional attributes from the privacy attribute set based on relevance and comprehensive sensitivity scores. This method converges the condition set to attributes that contribute more to relation determination, reduces the dimensionality of conditional mutual information calculation, lowers computational overhead, and alleviates the sample sparsity problem caused by high-dimensional conditions. Dependency strength estimation is more robust when the data scale is limited or the attribute dimensionality is high.

[0039] In the above implementation, preferably, when making a unified determination of privacy attribute pairs at the end of the time slice, a statistical significance test is performed on the conditional mutual information of the attribute pairs. The unified determination stage first obtains candidate dependencies by comparing the conditional mutual information with a preset threshold, and then performs a significance test on the conditional mutual information results corresponding to the candidate dependencies to eliminate spurious dependencies caused by random fluctuations.

[0040] When the conditional mutual information exceeds a preset threshold and passes the statistical significance test, the privacy attribute is deemed to have a stable attribute dependency, and the process of adding or updating the ontology edge is initiated accordingly. By embedding the significance test into a unified decision at the end of the time slice, the update of the relation structure can be made more prudent, avoiding the triggering of erroneous edges by a batch of short-term abnormal data, resulting in a more stable ontology structure and higher consistency in the interpretation of dependencies.

[0041] In the above implementation, preferably, determining the inference direction based on the conditional entropy reduction criterion includes: for the same privacy attribute pair, calculating a first conditional entropy reduction amount and a second conditional entropy reduction amount respectively under the same preset conditional attribute set constraints; the first conditional entropy reduction amount represents the degree of reduction in uncertainty of the second privacy attribute after introducing the first privacy attribute, and the second conditional entropy reduction amount represents the degree of reduction in uncertainty of the first privacy attribute after introducing the second privacy attribute. Comparing the two reduction amounts under the same condition set makes the inference direction determination comparable.

[0042] The attribute with the larger conditional entropy reduction and passing the statistical significance test is identified as the starting attribute for the inference direction, and a corresponding directed attribute dependency edge is established in the privacy ontology. Determining the side with stronger information contribution as the inference starting point by controlling the reduction magnitude ensures consistency between the relationship direction and the inference logic. Constraining the direction selection through significance testing ensures that the edge direction better meets the causal explanation requirements of attribute inference, reducing the probability of direction determination being affected by noise and minimizing explanatory conflicts caused by inconsistent bidirectional edges.

[0043] In the above embodiments, preferably, the privacy ontology further includes semantic extension information, which includes at least fine-grained sub-attributes of privacy attributes, candidate synonyms, and hypernyms and hyponyms. Specifically, based on social network corpora, concepts related to each privacy attribute are extracted from text fragments, and the co-occurrence frequency of concepts and privacy attributes in the corpus is statistically analyzed. Semantic support weights are calculated based on the co-occurrence statistics, and these weights are used for attribute node descriptions and attribute relationship interpretations in the privacy ontology.

[0044] By introducing fine-grained sub-attributes and synonym / hyper-hyper-attribute concepts at the node level, the semantic coverage of privacy attributes is more complete; by using semantic support weights to characterize the intensity of extended concepts, node descriptions can be supported by traceable corpus; the semantic expression of the ontology is more refined, and explanations that are closer to corpus evidence can be given when interpreting attribute dependency edges.

[0045] In the above implementation, preferably, the addition detection and node initialization of attribute nodes for data batches includes: after receiving the data batch within a time slice, comparing the privacy attribute variables appearing in the data batch with the existing attribute node set of the ontology; when it is detected that the data batch contains privacy attribute variables that do not appear in the privacy ontology, adding the corresponding attribute node in the privacy ontology, and initializing the value space, statistical count and probability estimation parameters of the attribute node, so that the added node has a unified data statistical basis before entering the subsequent entropy, mutual information and conditional mutual information calculation.

[0046] When incrementally updating node statistics, the frequency of attribute values, marginal probability distributions, and joint statistical information are updated incrementally. Furthermore, the statistical basis used for online estimation of information entropy, mutual information, and conditional mutual information is updated using either cumulative counting or a sliding window mechanism. Cumulative counting allows statistics to reflect long-term overall trends; the sliding window mechanism makes statistics more sensitive to recent distribution changes. The ontology can both maintain a memory of long-term stable relationships and respond promptly to distribution drift, improving the timeliness of dynamic completion.

[0047] In the above implementation, preferably, the accumulation of observed evidence of attribute relationships includes: for each data batch, taking the commonly occurring privacy attribute pairs as observed evidence and updating the corresponding joint statistical information, so that the statistical basis of candidate attribute dependencies continues to accumulate within the time slice. When the accumulated samples within the time slice meet the minimum sample size or stability requirements, the conditional mutual information of the candidate attribute dependencies is calculated to form an estimate of the dependency strength within that time slice.

[0048] At the end of the time slice, if the candidate attribute dependencies meet the unified judgment criteria, then new attribute dependency edges that do not exist are added in the privacy ontology, and relationship weights of existing attribute dependency edges are updated. Through the process of batch arrival, evidence accumulation, and final judgment, relationship learning can be based on sample sufficiency, avoiding misjudgments caused by insufficient early samples; newly added edges are more reliable, weight updates are smoother, and the identification of stable dependencies between newly emerging attributes and existing attributes is more timely.

[0049] According to the privacy ontology construction method for social networks disclosed in the above embodiments, during implementation, for any attribute... The information entropy formula is used to measure its distinctiveness and uniqueness; ; Among them, A j Let v be the j-th candidate privacy attribute variable. k For attribute A j The kth possible value.

[0050] Meanwhile, mutual information is used to assess the inferability of this attribute to user identity or other sensitive attributes; Where U represents user identity or other sensitive attribute variables; v represents attribute A. j A certain value of ; u is a certain value of attribute U.

[0051] Based on the above two types of indicators, information entropy and mutual information are first normalized to the [0,1] interval to obtain H. norm (a) MI norm (a).

[0052] Information entropy normalization across all candidate attributes A1,…,A n Above, the normalized result of information entropy is defined as: Where ε is a very small positive number, used to avoid the case where the denominator is zero.

[0053] Mutual information normalization is defined as follows: Then, a weighted fusion method was used to construct a comprehensive sensitivity score: Score(a) = w1•H norm (a)+w2•MI norm (a), where w1 and w2 are preset weights and w1+w2=1. Finally, the data is sorted from high to low according to Score(a), and high-sensitivity privacy attributes are selected by combining a threshold or Top-K strategy.

[0054] For any user u, each attribute A i Each corresponds to a value A iu However, the various probability distributions P used for subsequent calculations of information entropy and mutual information... Ai P Ai,Aj The values ​​are not derived from a single user, but are estimated from statistical results of all user samples in the social network dataset.

[0055] First, let the set of privacy attributes for all users be: A = {A1, A2} 2， …, A n}; Among them, A i Let A represent the i-th privacy attribute variable, such as age, gender, education level, occupation type, marital status, income, etc. i The formula for calculating information entropy is: Among them, a i For attribute A i A specific value that can be taken; Attribute A is obtained through statistics from all user samples. i Take the value a i Empirical probability; H(A) i ) is attribute A i The uncertainty measure.

[0056] Secondly, to measure attribute A i With attribute A j To determine the dependencies between attributes, conditional mutual information is introduced as an indicator of attribute dependency strength. This is used to control the remaining attribute sets. After (or a subset thereof), calculate: Among them, a i ,a j For attribute A i A j The specific value of z is a single joint value of the condition attribute set Z.

[0057] Conditional mutual information describes the relationship between attribute A and other privacy attributes, given that other privacy attributes are known. i The observation of attribute A j The degree of uncertainty reduction. When I(A) i A j When |Z) is large, it indicates that attribute A i For attribute A j It exhibits strong inferability, and there is a significant privacy dependency between the two.

[0058] Attribute dependency strength uses conditional mutual information The representation is used to characterize the degree of information dependency between attributes A and B after controlling for the set of conditions C. For direction determination, "conditional entropy reduction" is used as the criterion: under the same set of conditions C, the following calculations are performed respectively. and .like Furthermore, through statistical significance testing, it is considered that A has a stronger inferential power over B in the sense of privacy inference, denoted as a potential inferential relationship. A → B This direction does not rely on strong causal assumptions, but rather reflects which attribute significantly reduces the uncertainty of another attribute given other known information, thus better aligning with real-world scenarios involving privacy breaches and attribute inference. It's important to note that the "direction" of the attribute relationship is determined by the conditional entropy reduction criterion, while the "weight" of the relationship is still given by the normalized result of the corresponding conditional mutual information. This ensures both interpretability of the direction and comparability of the strengths of different attribute relationships. Finally, for attribute pairs that pass the statistical significance test, a set of privacy-preserving attribute dependencies is constructed, and its relationship weights are defined as follows: in, This means normalizing the dependency strength of different attribute relationships to the [0,1] interval. The larger the weight, the stronger the dependency of attribute A during the privacy inference process. i The leakage of attribute A j The more significant the potential exposure impact, the better.

[0059] Further analysis of data from representative social media platforms (such as Weibo) is conducted, focusing on finer-grained sub-attributes, candidate synonyms, and hyponyms / hypernyms under each privacy attribute. Relevant text fragments are extracted from the original text, and co-occurrence frequency statistics are performed. Based on this, semantic support weights for the sub-classification nodes are constructed. Let attribute A be... i The set of candidate concepts extracted is C. i ={c i1 ,c i2 ,…,c im}, let cooc(c ij A i () is the concept c ij With attribute A i The weight of a subclass node is defined as its normalized co-occurrence ratio within the same attribute candidate set, based on the number of times it co-occurs in the corpus. ; Here, ε is a smoothing term to prevent the denominator from being zero.

[0060] Without changing the core relation weights determined in Module 2 (obtained by conditional mutual information normalization), the semantic support weights... It is only used to supplement and verify node descriptions, synonym / hyper-hyper-relationship candidates and relation explanation information in the privacy ontology, thereby improving the interpretability and semantic consistency of the ontology.

[0061] Social network data is divided into multiple discrete time slices according to a preset time granularity, denoted as: ; Among them, T t This represents the t-th time slice. In any time slice T... t Within this context, data arrives in batches and serves as the update unit for a single ontology completion. This is defined within a time slice. The data batch arriving within the period is , where each X i For a user record, it represents a set of attribute value pairs: Among them, a i v is a privacy attribute variable. i The corresponding values ​​are assigned, and the batch data of each time slice collectively drives the update of the ontology, rather than processing each item individually.

[0062] When a batch of data Upon arrival, the present invention sequentially executes the following completion process: 1) Attribute node existence detection and structural completion For each attribute 'a' in the batch data i : If a i If it already exists in the current privacy ontology, then it is considered an existing attribute node; If a i If a property does not exist in the ontology, it is identified as a new privacy attribute, and a corresponding attribute node is added to the ontology. Simultaneously, it initializes the value space, statistical counts, and related probability estimation parameters. This process enables dynamic expansion of the attribute layer structure, ensuring that new attributes can be introduced with new data batches.

[0063] 2) Attribute values ​​and numerical updates of statistics For existing or newly introduced attribute nodes, this invention incrementally updates their statistical information based on the current input data, including: the frequency of attribute values; the marginal probability distribution p(a i ); the joint distribution p(a) formed with other attributes i ,a j The aforementioned statistics are updated through cumulative counting or a sliding window mechanism, providing a data foundation for subsequent online estimation of information entropy, mutual information, and conditional mutual information. This process corresponds to updating the relevant values ​​of nodes in the ontology without altering the structure.

[0064] 3) Accumulation of attribute relationship evidence and updating of candidate relationships For any pair of attributes that appear together in the input data (a) i ,a j This invention treats this as evidence of an attribute relationship observation and updates the joint statistical information between the two accordingly. In time slice T... t Within the time slice, as multiple input data points arrive, the joint distribution and conditional distribution of attribute pairs are gradually accumulated. When the accumulated statistical samples meet the minimum sample size or stability requirements, the following calculations are performed within the current time slice based on the conditional mutual information form defined in Module 2: ; C is a predefined set of conditional attributes. This value is used as attribute a. i With a j A measure of the strength of potential dependencies between them.

[0065] In time slice T t At the end, a unified determination is made of the attribute relationships obtained by gradually accumulating individual data within the time slice: 1) If a certain attribute pair (a i ,a j If the conditional mutual information between the two exceeds a preset threshold and passes the statistical significance test, then a stable privacy dependency relationship is considered to exist between them. 2) If the relationship does not yet exist in the ontology, then add the corresponding attribute relationship edge in the privacy ontology; 3) If the relationship already exists, only update its relationship weight value.

[0066] 4) The determination of the relationship direction still adopts the conditional entropy reduction criterion: and The inference direction is determined by comparing the magnitudes of the two factors mentioned above, thereby introducing a directed relation 'a' into the ontology. i →a j or a j →a i .

[0067] This invention also proposes a privacy ontology construction system for social networks, used to implement the privacy ontology construction method for social networks disclosed in any of the above embodiments, including: The data acquisition module is used to acquire social network datasets. The social network datasets include multiple user records. Each user record contains multiple privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. The privacy attribute extraction module is used to extract candidate privacy attributes from social network datasets to form a candidate privacy attribute set; The sensitivity assessment module is used to calculate the information entropy for each candidate privacy attribute in the candidate privacy attribute set based on the probability distribution obtained from the statistics of all user samples, and to calculate the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable; after normalizing the information entropy and mutual information respectively, a comprehensive sensitivity score is constructed according to the preset weight, and the privacy attribute set is obtained by screening based on the comprehensive sensitivity score. The attribute dependency analysis module is used to calculate conditional mutual information for each privacy attribute pair in the privacy attribute set under the constraints of a preset conditional attribute set, and to use the conditional mutual information as the dependency strength between privacy attribute pairs; when the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pair is determined based on the conditional entropy reduction criterion, and the dependency strength is normalized to obtain the relation weight. The privacy ontology construction module is used to construct a privacy ontology based on a set of privacy attributes. The privacy ontology includes attribute nodes and attribute dependency edges represented by inference direction and relation weight. The time-slice incremental update module is used to divide social network data into multiple time slices according to a preset time granularity and receive corresponding data batches within each time slice. For each data batch, it performs new attribute node detection and node initialization, incremental updates of node statistics, and accumulation of attribute relationship observation evidence. At the end of the time slice, it makes a unified judgment on the candidate attribute dependencies accumulated within the time slice. When the dependency strength corresponding to the candidate attribute dependency exceeds a preset threshold, it performs an incremental update on the structure or relation weight of the privacy ontology.

[0068] The privacy ontology construction system for social networks disclosed in the above embodiments has modules whose functions correspond to the steps of the privacy ontology construction method for social networks disclosed in the above embodiments. During implementation, the above embodiments are referred to for operation, and will not be repeated here.

[0069] The privacy ontology construction method and system for social networks disclosed in the above embodiments define the relationships between privacy attributes, enabling the derivation of other related privacy attributes from the measurement of one privacy attribute. This network of relationships ensures that privacy measurements are no longer isolated but are comprehensively evaluated based on the correlations between attributes, enhancing the depth and accuracy of privacy measurements.

[0070] Furthermore, an attribute dependency analysis method based on information theory is employed to characterize the dependency strength between privacy attributes using information entropy, mutual information, and conditional mutual information. Stable key associations are then screened through significance testing, thereby quantifying the inferability between attributes at the statistical level. Further, inference direction modeling is combined with comparing the degree of conditional entropy reduction to determine the directional relationship of "inferring the target attribute from the known attribute," forming an interpretable privacy inference chain to support the analysis and measurement of privacy leakage risk propagation.

[0071] By employing a time-evolution-based dynamic completion mechanism for privacy-preserving ontology, social network data is continuously updated in time slices. Upon detecting newly emerging privacy attributes, these attributes are introduced into the privacy ontology as new nodes. Simultaneously, based on conditional mutual information under time conditions and time-window weighted dependency strength, stable dependencies between new and existing attributes are automatically identified, and the direction of these dependencies is determined, thereby achieving incremental updates to the ontology structure and relation weights. This mechanism enables the privacy ontology to continuously expand and adaptively maintain itself as data evolves, avoiding the failure of static ontology in long-term applications and improving the completeness, stability, and interpretability of privacy relation modeling.

[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a privacy ontology for social networks, characterized in that, include: Obtain a social network dataset, which includes multiple user records. Each user record contains multiple privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. Candidate privacy attributes are extracted from the social network dataset to form a candidate privacy attribute set; For each candidate privacy attribute in the candidate privacy attribute set, the information entropy is calculated based on the probability distribution obtained from the statistics of all user samples, and the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable is calculated. After normalizing the information entropy and the mutual information respectively, a comprehensive sensitivity score is constructed according to a preset weight, and a set of privacy attributes is obtained by filtering based on the comprehensive sensitivity score; For each privacy attribute pair in the privacy attribute set, conditional mutual information is calculated under the constraints of a preset conditional attribute set, and the conditional mutual information is used as the dependency strength between the privacy attribute pairs. When the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pair is determined based on the conditional entropy reduction criterion, and the dependency strength is normalized to obtain the relation weight; A privacy ontology is constructed based on the set of privacy attributes. The privacy ontology includes attribute nodes and attribute dependency edges represented by the inference direction and the relation weight. The social network data is divided into multiple time slices according to a preset time granularity, and the corresponding data batch is received in each time slice; for the data batch, the addition detection and initialization of attribute nodes, incremental updates of node statistics, and accumulation of attribute relationship observation evidence are performed. At the end of the time slice, the candidate attribute dependencies accumulated within the time slice are uniformly determined. When the dependency strength corresponding to the candidate attribute dependency exceeds the preset threshold, the structure of the privacy ontology or the relation weight is incrementally updated.

2. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, The normalization processing of the information entropy and the mutual information adopts extreme value normalization based on the candidate privacy attribute set, so that the normalization result falls into the [0,1] interval, and a very small positive number is introduced to avoid the denominator being zero; the comprehensive sensitivity score is the weighted sum of the normalized information entropy and the normalized mutual information according to the preset weights, and the weight sum of the preset weights is 1.

3. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, The probability distribution used to calculate the information entropy and the mutual information is estimated from the statistical results of all user samples in the social network dataset. The probability distribution includes at least the marginal probability distribution of the candidate privacy attribute and the joint probability distribution of the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable.

4. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, The preset conditional attribute set is a set of privacy attribute variables excluding the privacy attribute pairs to be analyzed, or a subset of conditional attributes selected from the privacy attribute set based on correlation and comprehensive sensitivity scores.

5. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, When making a unified determination on privacy attribute pairs at the end of the time slice, a statistical significance test is performed on their conditional mutual information; when the conditional mutual information exceeds the preset threshold and passes the statistical significance test, it is determined that the privacy attribute pair has a stable attribute dependency relationship.

6. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, Determining the inference direction based on the conditional entropy reduction criterion includes: calculating the first conditional entropy reduction amount and the second conditional entropy reduction amount respectively under the same preset conditional attribute set constraints. The first conditional entropy reduction amount represents the degree to which the introduction of the first privacy attribute reduces the uncertainty of the second privacy attribute under the premise of a known conditional attribute set. The second conditional entropy reduction amount represents the degree to which the introduction of the second privacy attribute reduces the uncertainty of the first privacy attribute under the premise of a known conditional attribute set. The attribute with the larger conditional entropy reduction amount and passing the statistical significance test is determined as the starting attribute of the inference direction, and a corresponding directed attribute dependency edge is established in the privacy ontology.

7. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, The privacy ontology further includes semantic extension information, which includes at least fine-grained sub-attributes of privacy attributes, candidate synonyms, and hypernyms and hyponyms. The semantic extension information is used to calculate semantic support weights based on the concept co-occurrence statistics of social network corpora, and the semantic support weights are used for attribute node descriptions and attribute relationship interpretations in the privacy ontology.

8. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, The process of adding and initializing attribute nodes for the batch data includes: when it is detected that the batch data contains privacy attribute variables that do not appear in the privacy ontology, a corresponding attribute node is added in the privacy ontology, and the value space, statistical count and probability estimation parameters of the attribute node are initialized. The incremental updates of node statistics include: incremental updates of attribute value frequency, marginal probability distribution and joint statistical information, and updating the statistical basis used for online estimation of information entropy, mutual information and conditional mutual information using cumulative counting or sliding window mechanisms.

9. The method for constructing a privacy ontology for social networks according to claim 1, characterized in that, Accumulating observational evidence of attribute relationships includes: using privacy attribute pairs that co-occur in a data batch as observational evidence and updating the corresponding joint statistics; The conditional mutual information of the candidate attribute dependencies is calculated when the accumulated samples within the time slice meet the minimum sample size or stability requirements. At the end of the time slice, if the candidate attribute dependency relationship meets the unified judgment condition, then in the privacy ontology, the addition of non-existent attribute dependency relationships is performed, and the relationship weight update is performed on existing attribute dependency relationships.

10. A privacy ontology construction system for social networks, characterized in that, A method for constructing a privacy ontology for social networks as described in any one of claims 1 to 9 includes: The data acquisition module is used to acquire a social network dataset, which includes multiple user records. Each user record contains multiple privacy attribute variables and their corresponding values, as well as user identity attribute variables or target sensitive attribute variables and their corresponding values. The privacy attribute extraction module is used to extract candidate privacy attributes from the social network dataset to form a candidate privacy attribute set; The sensitivity assessment module is used to calculate the information entropy for each candidate privacy attribute in the candidate privacy attribute set based on the probability distribution obtained from the statistics of all user samples, and to calculate the mutual information between the candidate privacy attribute and the user identity attribute variable or the target sensitive attribute variable; after normalizing the information entropy and the mutual information respectively, a comprehensive sensitivity score is constructed according to a preset weight, and a set of privacy attributes is obtained by filtering based on the comprehensive sensitivity score; The attribute dependency analysis module is used to calculate conditional mutual information for each privacy attribute pair in the privacy attribute set under the constraints of a preset conditional attribute set, and to use the conditional mutual information as the dependency strength between the privacy attribute pairs; when the dependency strength exceeds a preset threshold, the inference direction of the privacy attribute pairs is determined based on the conditional entropy reduction criterion, and the dependency strength is normalized to obtain the relation weight. A privacy ontology construction module is used to construct a privacy ontology based on the privacy attribute set. The privacy ontology includes attribute nodes and attribute dependency edges represented by the inference direction and the relation weight. The time-slice incremental update module is used to divide the social network data into multiple time slices according to a preset time granularity, and receive corresponding data batches in each time slice; for the data batches, it performs new attribute node detection and node initialization, incremental update of node statistics, and accumulation of attribute relationship observation evidence; at the end of the time slice, it performs a unified judgment on the candidate attribute dependencies accumulated in the time slice, and when the dependency strength corresponding to the candidate attribute dependency exceeds the preset threshold, it performs an incremental update on the structure of the privacy ontology or the relationship weight.

Images