An anonymous social graph recovery method and system based on graph variational autoencoder

Abstract

The present invention relates to an anonymous social graph recovery method and system based on graph variational autoencoder. A social graph is constructed for anonymized social networks, in which users are nodes and the associations between users are edge sets, and an adjacency matrix and node set are extracted from the social graph; the adjacency matrix and node set are input into an encoder in a graph variational autoencoder, and the structural features of the nodes in the social graph are extracted by the encoder to form a latent vector space to expand the distance between false edges and real edges; the decoder in the graph variational autoencoder is used to make a rationality judgment on the existence of the edges, and according to the rationality judgment result, the added false edges are identified by setting a threshold, thereby completing the recovery of the anonymous social graph. In view of the defects of existing graph anonymity attacks, the present invention realizes a prototype system for identifying anonymous false edges in social graphs, whose encoding-decoding structure can extract implicit features that distinguish false edges from normal edges, and identify and restore social network graphs with high accuracy.

Description

Technical Field

[0001] The present invention relates to an anonymous social graph recovery method and system based on graph variational autoencoder, belonging to the field of computer software technology. Background Art

[0002] Since graph structures can well express the relationship between things, many social networks use graphs to express the connection between users, namely social graphs. People can understand social networks by analyzing social graphs, so as to complete personalized recommendations, extract user preferences, predict friend relationships, etc. In order to protect user privacy, social graphs need to be appropriately anonymized before being publicly released, that is, identifiers that are strongly related to identity are removed to protect users' personal information from being leaked. In recent years, with the maturity of privacy-safe computing technology, many privacy-preserving algorithms have been proposed, such as differential privacy. Social graph anonymization methods are not simply removing identifiers that are not related to identity, but are combined with privacy-preserving algorithms to produce social graph anonymization methods that meet certain privacy protection criteria.

[0003] Existing graph anonymity mechanisms generally achieve the purpose of anonymity by adding certain pre-designed false links to the original graph. Representative works can be divided into three categories: graph anonymity mechanisms based on random perturbations, graph anonymity mechanisms based on k-anonymity, and graph anonymity mechanisms based on differential privacy. Graph anonymity based on random perturbations relies on randomizing the connections of nodes in the original graph, such as AddDel, which adds k false connections to the original graph and deletes k real connections. Although this type of method is simple in principle and easy to implement, it also has the disadvantage of low anonymity. The graph anonymity method based on k-anonymity utilizes the k-anonymity mechanism commonly used in database systems. The threat model of this type of mechanism assumes that the attacker knows the world friend information of a specific node, and matches the target user with the specific node by comparing the degree or structural information of the nodes in the graph. The k-anonymity mechanism adds false edges to the original graph based on the residual degree of the node, and ensures that each node in the graph has the same degree information or structural information as k-1 other nodes. In this way, the probability of the attacker identifying a specific node will drop to 1 / k. Compared with the graph anonymization method based on random perturbation, this method has greatly improved anonymity, but it lacks the explainability of privacy protection theory. The graph anonymization mechanism based on differential privacy first counts the joint degree information in the graph, then applies differential noise to the joint degree, and reconstructs and restores the anonymous graph according to the perturbed joint degree sequence. The false connections added during reconstruction and restoration change the original graph in a random manner. This method is more cumbersome and time-consuming to implement, but it has high anonymity, and the introduction of differential noise provides explainability in theory.

[0004] Although relatively mature methods have been developed for social graph anonymization, the existing methods have a common flaw, that is, none of the three methods carefully consider the way of adding false edges, which makes the false edges appear "unreasonable" in the anonymous graph, that is, the difference between normal edges and false edges in the graph structure or higher hidden space is large. Attackers can easily use this difference to identify and detect anonymous graphs, and then restore the original graph to carry out de-anonymization attacks, causing serious damage to users' privacy. However, many social network managers have not yet realized the seriousness of the problem. Summary of the invention

[0005] In order to emphasize the above security threats and fill the loopholes in privacy protection so as to propose more effective graph anonymity methods in the future, the present invention proposes an anonymous social graph recovery method based on graph variational autoencoder. The graph variational autoencoder has an encoding-decoding architecture. The encoder can effectively expand the expression distance of the feature vector in the latent embedding space, and the decoder can effectively identify the added false connections, thereby completing the recovery of the anonymous social graph.

[0006] The present invention does not rely on the user's personal attribute characteristics, but analyzes the overall structure of the anonymous social graph to identify modified false edges. The present invention proposes a method for expressing the characteristics of nodes in an anonymous social graph, introducing node characteristics into the latent vector space, and providing an effective data basis for false edge discrimination. In response to the defects of existing graph anonymity attacks, the present invention implements a prototype system for identifying anonymous false edges in social graphs. Its encoding-decoding structure can extract implicit features that distinguish false edges from normal edges, and identify and restore social network graphs with high accuracy.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A method for recovering an anonymous social graph based on graph variational autoencoder, comprising the following steps:

[0009] Construct a social graph for the anonymized social network, where users are nodes and the connections between users are edge sets, and extract the adjacency matrix and node set from the social graph;

[0010] The adjacency matrix and the node set are input into the encoder of the graph variational autoencoder, and the encoder is used to extract the structural features of the nodes in the social graph and construct a latent vector space to expand the distance between the false edges and the real edges.

[0011] The decoder in the graph variational autoencoder is used to make a rational judgment on the existence of edges. According to the rational judgment results, a threshold is set to identify the added false edges, thereby completing the recovery of the anonymous social graph.

[0012] Furthermore, the encoder includes two layers of graph convolutional neural networks, the first layer of graph convolutional neural networks is used to learn the variational mean, and the second layer of graph convolutional neural networks is used to learn the variational variance to construct a variational normal distribution to ensure that the intermediate vector matrix generated by the encoder satisfies the normal distribution.

[0013] Furthermore, the decoder generates a reconstruction matrix according to the dot product of each pair of intermediate vectors and sets a threshold.

[0014] Furthermore, the encoder and the decoder are trained by the following steps:

[0015] The adjacency matrix and node set are input into the encoder to generate a variational normal distribution;

[0016] According to the variational normal distribution, latent feature vectors are randomly sampled and input into the decoder to generate a reconstruction matrix. The value of the reconstruction matrix is ​​modified based on the threshold, and the value greater than the threshold is judged as 1, and the value less than the threshold is judged as 0. The modified reconstruction matrix forms a reconstruction adjacency matrix.

[0017] The cross entropy of the two adjacency matrices is used as the loss function to update the weight parameters of the encoder and decoder, and the classification effect is verified using the validation set;

[0018] If the accuracy and recall on the validation set meet the settings, the current graph variational autoencoder model is the optimal classifier; otherwise, the model is retrained until the preset threshold is met to obtain the optimal classifier.

[0019] Furthermore, the output of the optimal classifier corresponds to the rationality of the existence of each edge in the anonymous graph, determines whether a certain edge is a false edge according to a threshold, and restores the original graph.

[0020] Further, the determining whether a certain edge is a false edge according to a threshold value includes: if it is greater than the threshold value, it is considered to be a real edge; if it is less than the threshold value, it is considered to be a false edge.

[0021] An anonymous social graph recovery system based on graph variational autoencoder using the above method comprises:

[0022] A graph variational autoencoder training module is used to train the encoder and decoder in the graph variational autoencoder; the encoder extracts the structural features of the nodes in the social graph to form a latent vector space to expand the distance between the false edges and the real edges; the decoder performs a rationality judgment on the existence of the edges, and identifies the added false edges by setting a threshold according to the rationality judgment result;

[0023] The anonymous social graph recovery module is used to restore the anonymous social graph using the trained graph variational autoencoder.

[0024] The key points of the present invention are:

[0025] 1. Aiming at the defect that the existing recovery methods cannot utilize the node structure characteristics, and taking advantage of the loopholes of the existing anonymity methods in edge set selection, an anonymous false edge identification method based on graph variational autoencoder is proposed to effectively restore the anonymized social network.

[0026] 2. The associations between users in popular social networks are represented by social graphs, with users as nodes and the associations between users as edge sets, forming a graph structure. The graph encoder is used to extract the structural features of the nodes (structural features refer to the neighborhood of the nodes and the features in the graph structure), forming a latent vector space, expanding the distance between false edges and real edges in this space, and revealing the essential difference between the two to the greatest extent. The decoder uses the node vector features to reconstruct the associations between nodes to form a reconstruction matrix, and uses the reconstruction matrix to make a reasonable judgment on the existence of the edges.

[0027] 3. According to the rationality measurement results, the anonymous social graph is restored and reconstructed by setting a threshold.

[0028] 4. Use subsequent real-world datasets for evaluation and verification to ensure the accuracy and generalization ability of the graph model.

[0029] The present invention has the following characteristics and advantages:

[0030] 1. The prototype system can accurately identify anonymous false edges, thereby restoring the original social relationship. The feature expression of the node in the structural space is learned, and according to the distance calculation result of the feature and the set threshold, it can be determined whether a certain edge is a false connection.

[0031] 2. The rationality of each edge in the graph can be measured, thereby quantitatively measuring the anonymity strength of the social graph, providing a new perspective for the evaluation of graph anonymity methods.

[0032] 3. The graph variational autoencoder does not need to retrain the encoder. The variational normal distribution is used in the process of generating the intermediate latent layer vector space. The new reconstruction matrix can be generated by random sampling in a specific distribution.

[0033] 3. The entire process does not require the attribute characteristics of the node, nor does it require decryption operations on the user, thus protecting the user's privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The flowchart of the method of the present invention is shown in FIG. Wherein VGAE represents a graph variational autoencoder. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, and to make the purposes, features and advantages of the present invention more obvious and easy to understand, the technical core of the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] The scheme of the present invention comprises the following steps:

[0037] First, select the anonymous algorithm and social network to be attacked, and use the anonymous algorithm to anonymize the social network. The social network data is divided into two parts, one for training and the other for verification testing.

[0038] Step S101, constructing a graph structure, i.e., a social graph, for the anonymized social network, with users as nodes and relationships between users as edge sets, and extracting an adjacency matrix and a node set from the social graph.

[0039] Step S102, construct an encoder, which includes a two-layer encoding network, and initialize the weight coefficients of the two layers, the first layer is used to learn the variational mean, and the second layer is used to learn the variational variance, so as to construct a variational normal distribution to ensure that the intermediate vector matrix satisfies the normal distribution. Among them, the two-layer encoding network is specifically a two-layer graph convolutional neural network, and the intermediate vector matrix refers to the vector generated by the encoder. Figure 1 In the figure, “first weight” refers to the weight of the first layer of convolutional neural network, and “second weight” refers to the weight of the second layer of convolutional neural network.

[0040] Step S103, construct a decoder, generate a reconstruction matrix based on the dot product of the intermediate vectors, and set a threshold. The decoder is a two-layer graph convolutional neural network.

[0041] Step S201, input the adjacency matrix and node set of step S101 into the encoder to generate a variational normal distribution.

[0042] Step S202: According to the variational normal distribution in step S201, latent feature vectors are randomly sampled and input into the decoder to generate a reconstruction matrix. Combined with the threshold in step S103, the value of the reconstruction matrix is ​​modified, and the value greater than the threshold is judged as 1, and the value less than the threshold is judged as 0. The modified reconstruction matrix forms a reconstruction adjacency matrix.

[0043] Step S203, using the cross entropy of the two adjacency matrices in steps S101 and S202 as the loss function, updating the weight parameters of the encoder and decoder, and using the validation set to verify the classification effect.

[0044] Step S203: If the accuracy and recall on the validation set meet the settings, the current graph variational autoencoder model is the optimal classifier. Otherwise, return to step 202, change the algorithm and parameters, and retrain the model until the preset threshold is met to obtain the optimal classifier.

[0045] Step S305, iterating the classifier according to the result of step S203, and saving the optimal classifier.

[0046] After the above steps, a system is finally generated that can restore anonymous social networks without retraining the encoder. The system inputs the anonymous graph to be restored, extracts nodes, converts the adjacency matrix, and then inputs it into the above optimal classifier. The output of the classifier corresponds to the rationality of the existence of each edge in the anonymous graph. It determines whether an edge is a false edge according to the threshold and restores the original graph. Among them, determining whether an edge is a false edge according to the threshold means: if it is greater than the threshold, it is considered to be a real edge; if it is less than the threshold, it is considered to be a false edge.

[0047] Based on the same inventive concept, another embodiment of the present invention provides an anonymous social graph recovery system based on graph variational autoencoder using the above method, which includes:

[0048] A graph variational autoencoder training module is used to train the encoder and decoder in the graph variational autoencoder; the encoder extracts the structural features of the nodes in the social graph to form a latent vector space to expand the distance between the false edges and the real edges; the decoder performs a rationality judgment on the existence of the edges, and identifies the added false edges by setting a threshold according to the rationality judgment result;

[0049] The anonymous social graph recovery module is used to restore the anonymous social graph using the trained graph variational autoencoder.

[0050] Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smart phone, etc.), which includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the method of the present invention.

[0051] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, CD), which stores a computer program. When the computer program is executed by a computer, it implements the various steps of the method of the present invention.

[0052] The above-mentioned embodiments only express the implementation methods of the present invention, and the description is relatively specific, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A method for anonymous social graph recovery based on graph variational autoencoder, characterized in that: The following steps are involved: Construct a social graph for the anonymized social network, where users are nodes and the connections between users are edge sets, and extract the adjacency matrix and node set from the social graph; The adjacency matrix and the node set are input into the encoder of the graph variational autoencoder, and the encoder is used to extract the structural features of the nodes in the social graph and construct a latent vector space to expand the distance between the false edges and the real edges. The decoder in the graph variational autoencoder is used to make a reasonable judgment on the existence of edges. According to the reasonable judgment results, a threshold is set to identify the added false edges, thereby completing the recovery of the anonymous social graph. The encoder and the decoder are trained using the following steps: The adjacency matrix and node set are input into the encoder to generate a variational normal distribution; According to the variational normal distribution, latent feature vectors are randomly sampled and input into the decoder to generate a reconstruction matrix. The value of the reconstruction matrix is ​​modified based on the threshold, and the value greater than the threshold is judged as 1, and the value less than the threshold is judged as 0. The modified reconstruction matrix forms a reconstruction adjacency matrix. The cross entropy of the two adjacency matrices is used as the loss function to update the weight parameters of the encoder and decoder, and the classification effect is verified using the validation set; If the accuracy and recall on the validation set meet the settings, the current graph variational autoencoder model is the optimal classifier; otherwise, retrain the model until it meets the preset threshold and obtains the optimal classifier; The output of the optimal classifier corresponds to the rationality of the existence of each edge in the anonymous graph, determines whether a certain edge is a false edge according to a threshold, and restores the original graph.

2. The method according to claim 1, characterized in that The encoder includes two layers of graph convolutional neural networks, the first layer of graph convolutional neural networks is used to learn variational mean, and the second layer of graph convolutional neural networks is used to learn variational variance to construct a variational normal distribution, ensuring that the intermediate vector matrix generated by the encoder satisfies the normal distribution.

3. The method according to claim 2, characterized in that The decoder generates a reconstruction matrix according to the dot product of each pair of intermediate vectors and sets a threshold.

4. The method according to claim 1, characterized in that: The determining whether a certain edge is a false edge according to the threshold value includes: if it is greater than the threshold value, it is considered to be a real edge; if it is less than the threshold value, it is considered to be a false edge.

5. An anonymous social graph recovery system based on graph variational autoencoder using the method described in any one of claims 1 to 4, characterized in that: include: A graph variational autoencoder training module is used to train the encoder and decoder in the graph variational autoencoder; the encoder extracts the structural features of the nodes in the social graph to form a latent vector space to expand the distance between the false edges and the real edges; the decoder performs a rationality judgment on the existence of the edges, and identifies the added false edges by setting a threshold according to the rationality judgment result; The anonymous social graph recovery module is used to restore the anonymous social graph using the trained graph variational autoencoder.

6. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method described in any one of claims 1 to 4 is implemented.

Images