Hypergraph association privacy protection method based on Transformer

Abstract

This paper proposes a Transformer-based hypergraph association privacy protection method. This method first sparsifies the original hypergraph structure by constructing a maximum spanning forest, effectively reducing subsequent computational overhead. Secondly, for the sparsified hypergraph, attribute associations and topological association matrices are extracted separately, and multi-view nonlinear interactive fusion is performed using Transformer large model technology to generate an integrated association representation of the hypergraph. Finally, the fused integrated association representation is used as prior knowledge to design a spider web privacy protection model, providing refined interference protection for associated privacy data in the hypergraph. This invention achieves precise defense against high-order network privacy risks.

Description

Technical Field

[0001] The present invention relates to the field of high-order networks and privacy protection technologies, and in particular to a Transformer-based hypergraph association privacy protection method. Background Art

[0002] With the rapid advancement of science and technology, interactions and relationships between entities in an increasing number of fields can be measured and recorded. For example, the impact of mobile phone communications on user friendships in social networks, the interaction between genetic loci in biology and their association with disease variation, and the impact of roads or routes on people's social activity networks in transportation systems. These interactions and relationships between entities can be characterized using high-order networks. Effective analysis and mining of this data can help us gain a deeper understanding of the valuable information contained in their organizational structures.

[0003] In the field of network science, higher-order networks are becoming a research frontier and hotspot. Hypergraphs, a type of higher-order network, are characterized by their ability to simultaneously connect multiple nodes using hyperedges, allowing them to naturally model multivariate relationships and group interactions between entities, more accurately capturing the complex relationships present in higher-order networks.

[0004] Existing hypergraph association representation methods still have significant shortcomings in modeling high-order associations, particularly in integrating attribute and topological information. Most work, such as HEM schemes and PageRank-based hypergraph clustering schemes, solely utilizes node attributes or hypergraph structure, lacking a clear and effective fusion path, thus failing to fully leverage the complementary strengths of both. Some methods, such as NMF and SVD, attempt to achieve integration through matrix factorization. However, these methods involve large-scale matrix operations, resulting in extremely high computational overhead, making them difficult to adapt to large-scale data scenarios and severely restricting their scalability. To jointly model attributes and topology, the AHCKA scheme proposes multi-hop random walks on the attribute graph to propagate node attribute similarity. However, since the attribute graph already directly encodes the maximum cosine similarity between node pairs, further multi-hop propagation may introduce unnecessary noise and even weaken the original similarity semantics. Currently, there is a lack of theoretical evidence or empirical analysis to support the rationale of this mechanism, which limits its effectiveness and reliability in downstream tasks.

[0005] At the same time, the high-order network structure of hypergraphs also poses potential privacy threats. This multi-level, multi-association structure makes it easier for sensitive attribute information of nodes to be leaked due to their high associations with other nodes. Data perturbation technology, as an important means of privacy protection, aims to hide sensitive information while minimizing the accuracy of data analysis by modifying, transforming, or adding noise to the original data. However, existing data perturbation technologies, such as differential privacy and local differential privacy, have significant limitations when addressing the privacy protection issues of complex and correlated background information. While these technologies theoretically provide strong protection capabilities, they ignore the potential correlations between data and the impact of background knowledge on privacy protection. They do not fully consider the potential risk of leakage of sensitive background information associated with users. Therefore, these traditional data perturbation technologies are not suitable for high-order network scenarios with complex correlations and multi-dimensional background knowledge.

[0006] A recent trend in privacy research is to incorporate prior background knowledge into privacy concepts, ultimately providing a better trade-off between utility and privacy. These concepts can be categorized into two main types: those based on average-case privacy guarantees and those based on worst-case privacy guarantees. Average-case privacy guarantees are generally weaker than the protection provided by worst-case privacy guarantees. Therefore, current research focuses on privacy concepts based on worst-case guarantees, which offer stronger protection and are applicable to more scenarios. Privacy technologies that consider contextual background knowledge and are based on worst-case guarantees include Blowfish privacy, Bayesian differential privacy, local information privacy, and attribute privacy. Among them, Blowfish privacy is not comprehensive enough and may not provide adequate protection for local data (such as sensitive information of individual users). Bayesian differential privacy only protects the correlation between tuples, rather than the overall cross-correlation function of each element in the tuple. This poses certain privacy risks, as privacy may be compromised by this cross-correlation. Local information privacy only protects local information, ignoring the privacy of global information. By analyzing the combination of multiple local information, sensitive global information may still be inferred. Furthermore, the distributed architecture it relies on results in excessive overall computational overhead and low data availability due to excessive noise addition. Attribute privacy, when protecting local data, only targets structured data distributed as Bayesian networks. Therefore, existing privacy research often or all suffers from technical limitations.

[0007] In view of this, the present invention proposes a Transformer-based hypergraph association privacy protection method; thereby more effectively integrating the topological information and attribute information of the hypergraph to improve the quality of the association representation between nodes; and while achieving high scalability on large-scale hypergraphs, avoiding redundant edges and reducing computational costs; in addition, when performing queries, it can not only prevent the leakage of sensitive association information, but also maximize the availability of query results. Summary of the Invention

[0008] The purpose of the present invention is to propose a Transformer-based hypergraph association privacy protection method that can more effectively protect association privacy in high-order networks. The invention first performs a sparse operation on the original hypergraph structure, while removing redundant connections, retaining the key topological backbone, thereby greatly reducing the computational complexity of subsequent processing to achieve high scalability on large-scale hypergraphs. Secondly, for the sparse hypergraph, the attribute association matrix and the structural association matrix are extracted based on hypercube modeling and hypergraph random walk strategies respectively, and the Transformer large model is used to realize deep nonlinear interaction of multi-view information, generating a more expressive hypergraph fusion association representation. Finally, the integrated association representation is used as association prior knowledge to design a spider web privacy protection mechanism, which performs differentiated perturbation protection on nodes and their associations in the hypergraph that may involve sensitive information, thereby achieving precise defense against high-order network privacy risks.

[0009] To achieve the above objectives, the technical solution of the present invention is: a Transformer-based hypergraph association privacy protection method, comprising the following steps:

[0010] S1. Perform a sparsification operation on the original hypergraph; use the random walk idea to calculate the initial transition matrix of the hypergraph to capture the high-order relationships between nodes during the random walk process; then use the maximum spanning forest to sparsify the hypergraph, remove redundant connections, retain important relationships, and symmetrize the initial transition matrix of the hypergraph to obtain an undirected adjacency matrix; then, iteratively use the Kruskal algorithm to generate a series of non-overlapping maximum spanning forests; finally, restore the undirected edges in the accumulated forests to bidirectional edges to obtain the sparsified hypergraph and transition matrix;

[0011] S2. For the sparse hypergraph, extract the association representation between nodes from the two dimensions of attribute information and topological structure;

[0012] In the attribute dimension, an N-dimensional weighted hypercube based on node attributes is constructed to extract the correlation of node attributes, the weight of each edge in the hypercube is calculated, and the attribute correlation matrix of the sparse hypergraph is generated; is the number of attributes of the node;

[0013] In the topological dimension, the design -Hypergraph random walk strategy is used to capture the topological structure information in the hypergraph, measure the topological correlation between nodes, and construct a topological correlation matrix; and Represent the restart probability and the number of walking steps respectively;

[0014] The attribute association matrix and the topology association matrix are mapped to the embedding space respectively, and the Transformer input sequence is constructed. The attributes and topology are nonlinearly interactively fused through the Transformer to obtain the integrated association representation of the hypergraph.

[0015] S3. When responding to queries from third-party servers, the hypergraph integrated association representation is used as the associated prior knowledge of private data, and a spider web privacy protection model is designed to add noise to the query results to prevent the adversary from obtaining sensitive private data based on prior knowledge. Based on the spider web privacy model, a privacy leakage metric is designed to quantitatively evaluate the degree of leakage of private data, so as to construct a utility-privacy trade-off optimization mechanism to ensure that the availability and analytical utility of data are maximized without exposing sensitive private data.

[0016] Preferably, step S1 is specifically implemented as follows:

[0017] S1.1. Calculate the initial transfer matrix of the hypergraph ; Starting from any node, randomly and uniformly select a hyperedge from all the associated hyperedges of the node, and obtain the transfer matrix based on the transition probability from the node to the hyperedge ; Starting from the selected hyperedge, select another node uniformly and randomly from all the associated nodes in the hyperedge, and obtain the transfer matrix based on the transition probability from the hyperedge to the node , according to the transfer matrix and the transfer matrix Calculate the initial transition matrix of the hypergraph ; Calculate the symmetric matrix , combining the transition probabilities in both directions between nodes, we get an adjacency matrix of an undirected graph ;

[0018] S1.2. Given a maximum number of iterations Residual plot and cumulative graph ; Initially, the residual graph With the adjacency matrix Same, cumulative graph With the adjacency matrix Have the same node set; starting with an empty edge set, in each iteration, a residual graph is generated by the Kruskal algorithm A maximum spanning forest , by looking at the residual graph Subtract the maximum spanning forest from Update the edge set in , take the maximum spanning forest and cumulative graph Update the edge set of ; Repeated iteration Second, in the adjacency matrix Generate a union of a set of maximum spanning forests with disjoint edges ; Get the sparse hypergraph , and obtain the corresponding sparse hypergraph transfer matrix according to the sparse hypergraph W ; When constructing the sparse transfer matrix, for each undirected edge in the forest, its two-direction directed edges are retained, thereby restoring a sparse hypergraph with a simple structure and high information retention.

[0019] Preferably, extracting the association representation from the attribute information dimension specifically includes:

[0020] According to the sparse hypergraph Node attribute information , construct the relationship between any two nodes on the same hyperedge -weighted hypercube, is the number of nodes in the hypercube, satisfying , each node in the hypercube Corresponding to one -dimensional attribute vector: , Representation node In the The value of an attribute dimension;

[0021] Calculate the weight of each edge ; and according to the weight of each edge of the hypercube , calculate any two nodes 、 Attribute correlation ;

[0022] Associate the attributes Perform normalization to obtain the attribute association matrix of the sparse hypergraph :

[0023] ;

[0024] in, is a real number, and n represents the number of nodes in the hypergraph after sparseness.

[0025] Preferably, the weight of each edge Represents the degree of association between attributes, calculated using Hellinger distance:

[0026] ;

[0027] ;

[0028] in, Neighbor nodes and The Hellinger distance between attribute distributions measures the degree of dissimilarity of the distributions and has a value range of ; Represents the dimension index of the node attribute vector; and Representation node and In the The value of the attribute dimension; convert the Hellinger distance into the association degree of each edge , the closer it is to 1, the higher the degree of attribute association is, and the closer it is to 0, the more independent the attributes are.

[0029] Preferably, any two nodes 、 Attribute correlation for 、 The sum of the products of the correlation degrees on each path between them:

[0030] ;

[0031] in, For any two nodes in the hypercube 、 The number of disjoint paths between , is the upper limit of the path length, so The value range of is: , the path is , the tth path , for length, Indicates the path midpoint The previous node of .

[0032] Preferably, extracting the association representation from the topological structure dimension specifically includes:

[0033] Based on sparse hypergraph transfer matrix ,design -Hypergraph random walk strategy to capture the multi-hop topological relationships between nodes in the hypergraph; starting from any node, -Hypergraph Random Walk moves, with probability Terminate at the current node and then jump back to the source node; with probability Move from the current node to the next node;

[0034] exist -Based on the hypergraph random walk strategy, the topological correlation between nodes in the sparse hypergraph is calculated to obtain the topological correlation matrix :

[0035] ;

[0036] in, is the hypergraph transfer matrix after sparse Next, from the node Departure -Hypergraph random walk terminates at a node The probability of By considering up to The random walk of length is used to capture the multi-hop topological similarity, Control the probability of restarting the random walk from the initial node at each step to balance local and global topological information; is the number of node transfer steps, Means just passing by The probability of transferring from one node to another; the sum Describes walks of different lengths (jumping from 0 to The contribution of jump to the overall topological correlation

[0037] Preferably, according to the attribute association matrix of the hypergraph and topological incidence matrix , using Transformer to fuse attributes and topology information, we get an integrated relational representation of the hypergraph, as follows:

[0038] The attribute correlation matrix and the topological correlation matrix are projected into a unified embedding space through the mapping function MLP to achieve consistency and alignment of feature representations between different views:

[0039] ;

[0040] ;

[0041] in, is the first layer weight, is a real number, n is the number of nodes in the sparse hypergraph, is the hidden layer dimension, is the second layer weight, represents the embedding dimension, and is the bias term, is the nonlinear activation function ReLU, represents the mapping function, and Represents the embedded representation of the node in the attribute view and topology view respectively, and each node corresponds to a dimensional embedding vector;

[0042] Will and Splice and construct the input sequence of Transformer :

[0043] ;

[0044] The input sequence is fed into the Transformer encoder, and the information interaction between attributes and topology is realized through the multi-head attention mechanism:

[0045] ;

[0046] Among them, the Transformer encoder output The front The row corresponds to the attribute view, and the The rows correspond to the topology view. For each node , which is represented in the attribute view and the topology view They are:

[0047] ;

[0048] ;

[0049] Generate an integrated association representation of nodes by averaging , and obtain the node-level integrated association representation matrix :

[0050] ;

[0051] ;

[0052] The similarity between node embeddings is calculated by Gaussian kernel, and the integrated correlation matrix of the sparse hypergraph is obtained. :

[0053] ;

[0054] in, Represents the Euclidean distance between nodes; is the kernel bandwidth parameter (scale parameter), the larger it is, the smoother it is, and the smaller it is, the more localized it is. .

[0055] Preferably, the hypergraph integrated association representation is used as the association prior knowledge of the private data to design a spider web privacy protection model; specifically, as follows:

[0056] Design a spider web privacy protection model and use hypergraph integrated association representation as association prior knowledge , the query results are protected based on the spider web privacy model, where the random mechanism of the spider web privacy model satisfies:

[0057]

[0058] in, represents a random mechanism; Indicates the actual query result. Indicates the actual query results The random mechanism applied above; Represents the user's sensitive privacy data; for and The joint probability distribution of Indicates the query result after disturbance protection; For privacy budget, control the noise size and privacy protection strength. When it is small, the ratio is close to 1, which means that the output result is independent of the private data, that is, the output result has nothing to do with the private data, thus preventing attackers from obtaining sensitive private data based on the association; Represents a probability function, which is used to measure the probability of a certain output result under given conditions.

[0059] Preferably, based on the cobweb privacy model, Laplace noise is added to the query results:

[0060] ;

[0061] in, represents the query result after adding noise, In the original dataset The real query results on , represents the sensitivity of the query function, It's a privacy budget. Indicates that from The noise is sampled from a -dimensional Laplace distribution with a location parameter of 0.

[0062] Preferably, based on the spider web privacy model, the degree of leakage of private data is quantified to perform a utility-privacy trade-off; specifically, as follows:

[0063] Based on the spider web privacy model, the degree of privacy leakage of sensitive information Quantification: Assume that the attacker knows a set of prior knowledge , that is, the query result and privacy data The joint probability distribution set between , , , then the attacker can query the result based on the perturbation and association Inferring user's private information The degree of privacy leakage associated with for:

[0064] ;

[0065] in, Indicates that among all possible query outputs , results after protection and privacy value Take the maximum value to measure the worst-case association privacy leakage intensity;

[0066] Based on the spider web privacy protection model and associated privacy leakage, a utility-privacy trade-off is made to minimize the degree of associated privacy leakage and achieve optimal data utility:

[0067] ;

[0068] ;

[0069] in, is the original query statistics result, is the query statistical result after protection with noise, Represents the original query results Query results after adding disturbance The mean squared error (MSE) between them is the utility loss.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1. Transformer is used to fuse the attributes and topological information of the hypergraph, fully exploring the deep nonlinear dependencies between multi-view inputs and strengthening the high-order correlation interactions between nodes, thereby generating an integrated correlation representation with more discriminative and generalizable capabilities, providing a solid representation foundation for subsequent privacy protection.

[0072] 2. Design a spider web privacy protection model that combines hypergraph integrated association representation. By integrating the high-order association representation of attributes and topological structures to model prior knowledge, a correlation privacy leakage measurement method and a utility-privacy joint optimization mechanism are constructed, achieving high-precision privacy protection and high-availability query support for hypergraph private data. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Schematic diagram of the overall process of an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0075] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0076] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0077] like Figure 1 As shown, the present invention provides a Transformer-based hypergraph association privacy protection method, which includes the following steps:

[0078] S1. Hypergraph sparsification: Calculate the initial hypergraph transfer matrix and use the maximum spanning forest to sparsify the hypergraph to remove redundant connections and retain key topological information. The details are as follows:

[0079] First calculate the initial transfer matrix of the hypergraph In order to maintain the high-order information in the hypergraph, combined with the idea of ​​random walk, the initial transfer matrix The calculation is : From any node Starting from the node A hyperedge is randomly selected from all associated hyperedges , the transfer matrix is ​​recorded as ; From the super edge Starting from the edge A node is uniformly randomly selected from all associated nodes in , the transfer matrix is ​​recorded as ;

[0080] According to the initial transfer matrix of the hypergraph Computing symmetric matrices , combined with the transition probabilities in both directions between nodes, we get an adjacency matrix of an undirected graph, recorded as ; Subsequently, Kruskal's algorithm is applied, which aims to find the maximum spanning forest on an undirected graph, specifically, given a maximum number of iterations , and , Represents a positive integer; by maintaining two graphs: residual graph and cumulative graph To proceed, initially, and same, and has the same set of nodes, but starts with an empty set of edges. In each iteration, Kruskal's algorithm generates A maximum spanning forest ; Then, by Subtract Update the edge set in and by taking and Update the edge set of , the process repeats times, in Generate a union of a set of maximum spanning forests with disjoint edges ,Will After removing the edges in They are all maximum spanning forests generated independently on different residual graphs, and the union of these forests is recorded as ; Finally, for Each undirected edge in , in the sparse transfer matrix Keep a pair of directed edges in and .

[0081] S2. Hypergraph Integrated Association Representation: Starting from the two dimensions of attribute information and topological structure, we use a large Transformer-like model to perform nonlinear deep fusion to obtain a high-quality hypergraph integrated association representation. The details are as follows:

[0082] First, calculate the attribute association matrix, according to the sparse hypergraph Node attribute information , construct the relationship between any two nodes on the same hyperedge -weighted hypercube, , the weight of each edge of the hypercube Represents the degree of association between attributes, calculated using Hellinger distance:

[0083] ;

[0084] ;

[0085] in, Neighbor nodes and Hellinger distance between attribute distributions, Represents the dimension index of the node attribute vector, , and Representation node and In the The value of an attribute dimension;

[0086] According to the weight of each edge of the hypercube , calculate any node of the hypergraph , Attribute correlation:

[0087]

[0088] in, For any two nodes in the hypercube 、 The number of disjoint paths between , the tth path , for length; Indicates the path midpoint The previous node of

[0089] Normalize the attribute correlation degree to obtain the attribute correlation matrix between the nodes of the sparse hypergraph :

[0090]

[0091] in, is a real number, and n represents the number of nodes in the hypergraph after sparseness.

[0092] Secondly, the topological association matrix is ​​calculated based on the sparse hypergraph transfer matrix. ,use -Hypergraph random walk to capture the multi-hop topological relationships between nodes in the hypergraph, where and Represent the restart probability and the number of walking steps respectively:

[0093] (1) Based on probability End at the current node and then jump back to the source node ;

[0094] (2) Based on probability , based on the sparse hypergraph transfer matrix From the current node Transfer to node ;

[0095] use - Hypergraph random walk calculates the topological correlation between nodes and obtains the topological correlation matrix :

[0096]

[0097] in, is the hypergraph transfer matrix after sparse Next, from the node Departure -Hypergraph random walk terminates at a node The probability of By considering up to Random walks of varying lengths are used to capture multi-hop topological similarity; Control the probability of restarting the random walk from the initial node at each step; is the number of node transfer steps, Means just passing by The probability of transferring from one node to another; the sum Describes the contribution of walks of different lengths to the overall topological correlation; the parameter It can be infinite, but in practice it is set to a constant for effective approximation. It strikes a balance between computational cost and effectiveness and is therefore set as the default.

[0098] Finally, the Transformer model is used to fuse attribute and topology information, enhance the hypergraph integrated association representation, and project the association matrices of attribute and topology views into a unified embedding space through the mapping function MLP:

[0099] ;

[0100] ;

[0101] in, is the first layer weight, is the hidden layer dimension, is the second layer weight, represents the embedding dimension, and is the bias term, is the nonlinear activation function ReLU; represents the mapping function, and Represents the embedded representation of the node in the attribute view and topology view respectively, and each node corresponds to a dimensional embedding vector;

[0102] Secondly, and Perform splicing to construct the input sequence of Transformer:

[0103] ;

[0104] The input sequence is fed into the Transformer encoder, and the information interaction between attributes and topology is realized through the multi-head attention mechanism:

[0105]

[0106] Among them, the output Center front The row corresponds to the attribute view, and the The rows correspond to the topology view. For each node , which is represented in the attribute view and the topology view They are:

[0107] ;

[0108] ;

[0109] Generate an integrated association representation of nodes by averaging , and obtain the node-level integrated association representation matrix :

[0110] ;

[0111]

[0112] Finally, the similarity between node embeddings is calculated by Gaussian kernel to obtain the integrated correlation matrix of the sparse hypergraph :

[0113]

[0114] in, Represents the Euclidean distance between nodes; is the kernel bandwidth parameter, .

[0115] S3. Spider-web privacy protection model: Taking the hypergraph integrated association representation as prior knowledge, a spider-web privacy model is designed to achieve accurate and efficient association privacy protection of hypergraph sensitive information. The details are as follows:

[0116] When responding to queries from third-party servers, hypergraphs are integrated with associated representations. Associated prior knowledge as secret data , designing a spider web privacy model ( -Cobweb Privacy, -CP), which relaxes the definition of differential privacy by leveraging contextual knowledge and improves its utility by explicitly modeling the adversary's knowledge, ensuring that even if the query results change, the adversary cannot infer the value of the secret variable by controlling the association between the query results and the secret data.

[0117] Given a random mechanism , a secret dataset and a set of data evolution scenarios . , ,by, , then if and only if When the following formula is satisfied, satisfy -Spiderweb Privacy;

[0118]

[0119] in, Indicates the actual query result. Indicates the actual query results The random mechanism applied above; Represents the user's sensitive privacy data; for and The joint probability distribution of Indicates the query result after disturbance protection; To budget privacy, control the noise level and privacy protection strength; Represents a probability function, which is used to measure the probability of a certain output result under given conditions;

[0120] -CP operation means that any output , the belief that the secret variable takes any particular value does not significantly increase or decrease relative to the prior distribution known to the adversary. Budget for privacy, when When it is small, the ratio is close to 1, which means that the output result is independent of the secret data, that is, the output structure is independent of the secret data, thereby preventing attackers from obtaining sensitive secret data based on association.

[0121] Based on the spiderweb privacy framework, add appropriate Laplace noise to the query results:

[0122]

[0123] in, Indicates the query result after adding noise; In the original dataset The real query results on ; represents the sensitivity of the query function; the scale parameter of the Laplace distribution is ; Indicates that from -dimensional Laplace distributed sampling noise.

[0124] Based on the spider web privacy protection, the degree of privacy leakage of sensitive information Quantification: Assume that the attacker knows a set of prior knowledge , that is, the query result and privacy data The joint probability distribution set between , , , then the attacker can query the result based on the perturbation and association Inferring user's private information The degree of privacy leakage associated with for:

[0125]

[0126] in, Indicates that among all possible query outputs , results after protection and privacy value Take the maximum value to measure the worst-case association privacy leakage intensity;

[0127] The problem of associated privacy leakage under the spider web privacy protection model can be transformed into a trade-off between utility and privacy, minimizing the degree of associated privacy leakage and achieving optimal data utility. The problem of associated privacy leakage under the spider web privacy mechanism can be transformed into a trade-off between utility and privacy:

[0128]

[0129]

[0130] in, is the original query statistics result, is a statistical result with noise. Represents the original query results Query results after adding disturbance The mean square error (MSE) between them, that is, the utility loss:

[0131]

[0132] Therefore, maximizing utility is equivalent to minimizing the mean squared error (MSE). Ultimately, the hypergraph data query results processed with spiderweb privacy protection can effectively protect sensitive information from being identified while maintaining the practicality of the data for subsequent third-party server analysis tasks.

[0133] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A Transformer-based hypergraph association privacy protection method, characterized by: The steps include: S1, perform a sparsification operation on the original hypergraph; S2. For the sparse hypergraph, extract the association representation between nodes from the two dimensions of attribute information and topological structure; In the attribute dimension, an N-dimensional weighted hypercube based on node attributes is constructed to extract the correlation of node attributes, the weight of each edge in the hypercube is calculated, and the attribute correlation matrix of the sparse hypergraph is generated; is the number of attributes of the node; In the topological dimension, the design -Hypergraph random walk strategy is used to capture the topological structure information in the hypergraph, measure the topological correlation between nodes, and construct a topological correlation matrix; and Represent the restart probability and the number of walking steps respectively; The attribute association matrix and the topology association matrix are mapped to the embedding space respectively, and the Transformer input sequence is constructed. The attributes and topology are nonlinearly interactively fused through the Transformer to obtain the integrated association representation of the hypergraph. S3. When responding to queries from third-party servers, the hypergraph integrated association representation is used as the associated prior knowledge of private data. A spider web privacy protection model is designed to add noise to the query results to prevent the adversary from obtaining sensitive private data based on prior knowledge. Based on the spider web privacy model, the degree of private data leakage is quantified to construct a utility-privacy trade-off optimization mechanism.

2. The Transformer-based hypergraph association privacy protection method according to claim 1, characterized in that: The step S1 is specifically implemented as follows: S1.

1. Calculate the initial transfer matrix of the hypergraph ; Starting from any node, randomly and uniformly select a hyperedge from all the associated hyperedges of the node, and obtain the transfer matrix based on the transition probability from the node to the hyperedge ; Starting from the selected hyperedge, select another node uniformly and randomly from all the associated nodes in the hyperedge, and obtain the transfer matrix based on the transition probability from the hyperedge to the node , according to the transfer matrix and the transfer matrix Calculate the initial transition matrix of the hypergraph ; Calculate the symmetric matrix , combining the transition probabilities in both directions between nodes, we get an adjacency matrix of an undirected graph ; S1.

2. Given a maximum number of iterations , residual plot and cumulative graph ; Initially, the residual graph With the adjacency matrix Same, cumulative graph With the adjacency matrix Have the same node set; starting with an empty edge set, in each iteration, a residual graph is generated by the Kruskal algorithm A maximum spanning forest , by looking at the residual graph Subtract the maximum spanning forest from Update the edge set in , take the maximum spanning forest and cumulative graph Update the edge set of ; Repeated iteration Second, in the adjacency matrix Generate a union of a set of maximum spanning forests with disjoint edges ; Get the sparse hypergraph , and obtain the corresponding sparse hypergraph transfer matrix according to the sparse hypergraph W .

3. The Transformer-based hypergraph association privacy protection method according to claim 1, characterized in that: Extracting association representations from the attribute information dimension, specifically including: According to the sparse hypergraph Node attribute information , construct the relationship between any two nodes on the same hyperedge -weighted hypercube, is the number of nodes in the hypercube, satisfying , each node in the hypercube Corresponding to one -dimensional attribute vector: , Representation node In the The value of an attribute dimension; Calculate the weight of each edge ; and according to the weight of each edge of the hypercube , calculate any two nodes 、 Attribute correlation ; Associate the attributes Perform normalization to obtain the attribute association matrix of the sparse hypergraph : ； in, is a real number, and n represents the number of nodes in the hypergraph after sparseness.

4. The Transformer-based hypergraph association privacy protection method according to claim 3, characterized in that: The weight of each edge Represents the degree of association between attributes, calculated using Hellinger distance: ； ； in, Neighbor nodes and The Hellinger distance between attribute distributions measures the degree of dissimilarity of the distributions and has a value range of ; Represents the dimension index of the node attribute vector; and Representation node and In the The value of the attribute dimension; convert the Hellinger distance into the association degree of each edge , , the closer it is to 1, the higher the degree of attribute association is, and the closer it is to 0, the more independent the attributes are.

5. The Transformer-based hypergraph association privacy protection method according to claim 4, characterized in that: Any two nodes 、 Attribute correlation for 、 The sum of the products of the correlation degrees on each path between them: ； in, For any two nodes in the hypercube 、 The number of disjoint paths between , is the upper limit of the path length, ; The path is , the tth path , for length, Indicates the path midpoint The previous node of .

6. The Transformer-based hypergraph association privacy protection method according to claim 1, characterized in that: Extracting association representation from the topological structure dimension, specifically including: Based on sparse hypergraph transfer matrix ,design -Hypergraph random walk strategy to capture the multi-hop topological relationships between nodes in the hypergraph; starting from any node, -Hypergraph Random Walk Step move, with probability Terminate at the current node and then jump back to the source node; with probability Move from the current node to the next node; exist -Based on the hypergraph random walk strategy, the topological correlation between nodes in the sparse hypergraph is calculated to obtain the topological correlation matrix : ； in, is the hypergraph transfer matrix after sparse Next, from the node Departure -Hypergraph random walk terminates at a node The probability of By considering up to The random walk of length is used to capture the multi-hop topological similarity, Control the probability of restarting the random walk from the initial node at each step; is the number of node transfer steps, Means just passing by The probability of transferring from one node to another; the sum The contribution of walks of different lengths to the overall topological connectivity is described.

7. The Transformer-based hypergraph association privacy protection method according to claim 1, characterized in that: According to the attribute association matrix of the hypergraph and topological incidence matrix , using Transformer to fuse attributes and topology information, we get an integrated relational representation of the hypergraph, as follows: The attribute association matrix and the topological association matrix are projected into a unified embedding space through the mapping function MLP: ； ； in, is the first layer weight, is a real number, n is the number of nodes in the sparse hypergraph, is the hidden layer dimension, is the second layer weight, represents the embedding dimension, and is the bias term, is the nonlinear activation function ReLU, represents the mapping function, and Represents the embedded representation of the node in the attribute view and topology view respectively, and each node corresponds to a dimensional embedding vector; Will and Splice and construct the input sequence of Transformer : ； The input sequence is fed into the Transformer encoder, and the information interaction between attributes and topology is realized through the multi-head attention mechanism: ； Among them, the Transformer encoder output The front The row corresponds to the attribute view, and the The rows correspond to the topology view. For each node , which is represented in the attribute view and the topology view They are: ； ； Generate an integrated association representation of nodes by averaging , and obtain the node-level integrated association representation matrix : ； ； The similarity between node embeddings is calculated by Gaussian kernel, and the integrated correlation matrix of the sparse hypergraph is obtained. : ； in, Represents the Euclidean distance between nodes; is the kernel bandwidth parameter, .

8. The Transformer-based hypergraph association privacy protection method according to claim 1, characterized in that: The hypergraph integrated association representation is used as the association prior knowledge of private data to design a spider web privacy protection model, which is as follows: Design a spider web privacy protection model and use hypergraph integrated association representation as association prior knowledge , the query results are protected based on the spider web privacy model, where the random mechanism of the spider web privacy model satisfies: in, represents a random mechanism; Indicates the actual query result. Indicates the actual query results The random mechanism applied above; Represents the user's sensitive privacy data; for and The joint probability distribution of Indicates the query result after disturbance protection; To budget privacy, control the noise level and privacy protection strength; Represents a probability function, which is used to measure the probability of a certain output result under given conditions.

9. The Transformer-based hypergraph association privacy protection method according to claim 8, characterized in that: Based on the cobweb privacy model, Laplace noise is added to the query results: ； in, Indicates the query result after adding noise; In the original dataset The real query results on ; represents the sensitivity of the query function; It’s the privacy budget; Indicates from -dimensional Laplace distributed sampling noise.

10. The Transformer-based hypergraph association privacy protection method according to claim 9, characterized in that: Based on the spiderweb privacy model, the degree of privacy data leakage is quantified to build a utility-privacy trade-off optimization mechanism, as follows: Based on the spider web privacy model, the degree of privacy leakage of sensitive information Quantification: Assume that the attacker knows a set of prior knowledge , that is, the query result and privacy data The joint probability distribution set between , , , then the attacker can query the result based on the perturbation and association Inferring user's private information The degree of privacy leakage associated with for: ； in, Indicates that among all possible query outputs , results after protection and privacy value Take the maximum value to measure the worst-case association privacy leakage intensity; Utility-privacy trade-off based on the spiderweb privacy protection model and associated privacy leakage: ； ； in, is the original query statistics result, is the query statistical result after protection with noise, Represents the original query results Query results after adding disturbance The mean square error between , i.e., the utility loss.

Images