An education big data-oriented high-availability differential privacy graph generation method

Abstract

The application discloses a high-availability differential privacy graph generation method for educational big data, which comprises the following steps: firstly, taking original graph structure data as input, a backbone network in the graph is mined by reverse learning to generate a supposed adjacency matrix; secondly, in order to deepen the topological information in the supposed adjacency matrix, a possibility matrix is generated by using an encoding decoder according to the supposed adjacency matrix obtained in the previous step; thirdly, a corresponding differential privacy graph is generated according to the size of a privacy budget to obtain false edges; and finally, according to the possibility matrix, a certain number of edges are respectively sampled from real edges and false edges of each node in the graph, and the total number is equal to the number of real edges, so that a protected differential privacy adjacency matrix is generated. On the one hand, the important information in the graph is retained by sampling from real edges with high possibility, thereby guaranteeing the availability of the graph; on the other hand, the original graph topological information is effectively protected by sampling from false edges, thereby realizing the privacy protection of the graph and protecting the important feature of node degree distribution.

Description

Technical Field

[0001] This invention belongs to the field of graph privacy protection, and specifically relates to a highly available differential privacy graph generation method for educational big data. Background Technology

[0002] In recent years, with the deepening application of big data technology in the education field, there has been increasing concern about the potential privacy risks arising from the collection, sharing, and dissemination of educational data. Educational big data, as a subset of big data, mainly consists of learners' behavioral data, including but not limited to teaching data, management data, and daily life data. Educational big data comes from diverse sources, and its collection and storage stages aggregate data from various types and information sources, including a large amount of graph-structured data. This data is characterized by its large scale, high dimensionality, and long span, placing high demands on analysis and processing, requiring collaboration with external institutions. On the one hand, the application of educational big data processing technology can provide learners with personalized services, build suitable learning paths, optimize the learning experience, and improve learning efficiency. On the other hand, in the process of analyzing and mining massive and fragmented educational big data, there is a risk of leakage of students' personal privacy. In particular, sensitive information collected by intelligent devices such as sensors, including learners' faces and physical characteristics, which can identify individual learners' behavior, is unique and immutable. Once data leakage or misuse occurs, it could affect learners' personal safety and rights. It is worth noting that when learners' fragmented data is illegally stolen and recombined for related applications, a learner data chain with new value is generated, making learners constantly "monitored," thus posing a risk of learner privacy disclosure. Because the consequences of exposing learners' personal privacy are severe, educational big data requires strong protection for learners' personal privacy. Differential privacy is a privacy protection technology whose main idea is to blur data by adding a certain amount of noise, thereby preventing the identification and inference of personal data. Although differential privacy can protect personal data privacy while still allowing data analysis and utilization, it introduces a certain degree of error and uncertainty, significantly reducing data availability and the accuracy of generated models. With the popularization and promotion of educational big data, how to generate highly available differential privacy graphs while ensuring learner privacy security, and then collaborate with external institutions to analyze and process data, has become an urgent problem to be solved. This is of great significance for accelerating the deployment of educational big data and further implementing educational big data.

[0003] Currently, the main invention patents related to privacy-preserving image generation include:

[0004] Reference 1: A Privacy Protection Method Based on Uncertain Graph of Mobile CDN Network Structure (202010903033.2)

[0005] Reference 2: A privacy protection method for directed graphs in large-scale social networks based on k-kernel (202010543252.4)

[0006] Reference 1 proposes a privacy protection method based on uncertain graphs in mobile CDN network structures. The data transmission records between user devices and smart base stations are segmented according to time periods and simplified into a graph form. An uncertain graph model based on ternary closure and related algorithms are used for processing, thereby achieving the goal of protecting the privacy of users' social networks.

[0007] Reference 2 proposes a privacy protection method for directed graphs in large-scale social networks based on k-cores. Based on the Pregel model and the "node-centric" characteristics of the distributed graph processing system GraphX, it achieves edge privacy protection of the social network while ensuring that the number of node cores remains unchanged through message passing and node value updates between nodes in the social network graph. The unchanged number of node cores ensures that the influence of nodes in the social network remains unchanged, thus protecting the community structure of the social network.

[0008] The aforementioned technical solutions focus on graph privacy protection, aiming to prevent the leakage of private information contained within the graph. However, in reality, due to privacy protection, the information contained in the graph is inevitably reduced, and directly using it for model training leads to a significant drop in accuracy. Research has shown that retaining edges with high influence and excluding edges with low influence during the information aggregation process of graph neural network models can still achieve excellent performance in downstream graph neural network models. Therefore, how to retain the main information of graph data in differential privacy graphs to simultaneously achieve privacy protection and high availability has become an urgent problem to be solved. Summary of the Invention

[0009] This invention aims to provide a highly available differential privacy graph generation method for educational big data. First, using the original graph structure data as input, inverse learning is used to mine the backbone network within the graph, generating a hypothetical adjacency matrix A that assumes similar results for the same downstream tasks. s Secondly, to deepen the topological information obtained from the assumed adjacency matrix, based on the assumed adjacency matrix obtained in the previous step, a probability matrix L is generated using an encoder-decoder, where the elements L... i,jA larger value indicates a higher probability that an edge exists between node i and node j, thus measuring the likelihood of edges existing between different nodes. Secondly, based on the privacy budget, a corresponding differential privacy graph is generated, in which spurious edges constitute the vast majority; directly using such graphs for downstream tasks would severely degrade performance. Finally, based on the probability of edges existing between nodes, each node in the graph samples an equal number of edges from both real and spurious edges to generate a protected differential privacy adjacency matrix. On the one hand, sampling from highly probable real edges preserves important information in the graph, ensuring its usability; on the other hand, sampling from spurious edges effectively protects the original graph topology, achieving privacy protection while preserving the important feature of the overall degree distribution of the graph.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A highly available differential privacy graph generation method for educational big data includes:

[0012] First, using the original graph structure data as input, inverse learning is used to mine the backbone network in the graph, generating a hypothetical adjacency matrix A that assumes similar results for the same downstream tasks. s Secondly, to deepen the topological information obtained from the assumed adjacency matrix, based on the assumed adjacency matrix obtained in the previous step, a probability matrix L is generated using an encoder-decoder, where the elements L... i，j A larger value indicates that there is a greater likelihood of an edge between node i and node j, thus measuring the probability of an edge between different nodes. Finally, based on the size of the privacy budget, a corresponding differential privacy graph is generated, in which false edges account for the vast majority, and directly using them for downstream tasks will lead to a severe performance degradation. Based on the probability of an edge between nodes, each node in the graph samples an edge from the real edges and false edges equal to the number of real edges, generating a protected differential privacy adjacency matrix.

[0013] A further improvement of the present invention is that the method specifically includes the following steps:

[0014] 1) Generation of the assumed adjacency matrix based on inverse learning

[0015] A graph neural network model is trained using the original graph structure data. The backbone network is mined through the loss function and gradient descent in the graph neural network model. The assumed adjacency matrix is ​​learned, and the assumed adjacency matrix is ​​similar to the aggregated information obtained from the original graph structure data during the information aggregation process of the graph neural network.

[0016] 2) Edge Existence Probability Analysis Based on Decoder and Encoder

[0017] A two-layer graph convolutional neural network is trained using an assumed adjacency matrix to classify nodes; the first layer is the encoder, which outputs the hidden embeddings of the nodes, and the second layer is the decoder, which uses the hidden embeddings to calculate the probability matrix as the basis for sampling.

[0018] 3) Differential privacy graph generation based on sampling

[0019] Based on the privacy budget, the number of real and fake edges that a node should have is calculated, and the corresponding number of edges are sampled based on the probability matrix to obtain the protected adjacency matrix.

[0020] A further improvement of this invention is that, in step 1), the generation of the assumed adjacency matrix based on inverse learning specifically includes the following steps:

[0021] Step 1: Training the baseline graphical neural network

[0022] Initially, the adjacency matrix is ​​assumed to be the same size as the original adjacency matrix, with all elements being 0. A baseline graph neural network model is trained on the original graph structure data, and the predicted labels are obtained by comparing with the real labels.

[0023] Step 2: Calculate the loss function

[0024] Using a pre-trained baseline graph neural network model, and taking the assumed adjacency matrix and the feature matrix from the original graph structure data as input, predicted labels based on the assumed adjacency matrix are obtained. Optimize the assumed adjacency matrix using the following loss function:

[0025]

[0026]

[0027] Where N is the number of nodes, y i To predict labels based on the original graph structure data, The true label is α, which is a hyperparameter controlling the proportion of the two losses; ||A s || F This is a regularization term that guarantees the sparsity of the assumed adjacency matrix;

[0028] Step 3: Optimize the assumed adjacency matrix

[0029] To control the range of values ​​for elements in the assumed adjacency matrix, projective gradient descent is employed:

[0030]

[0031]

[0032] Where t is the number of iterations, η is the learning rate, and g is the gradient.

[0033] A further improvement of the present invention is that, in Step 2 of step 1), L GNN It consists of two parts: the loss between the predicted labels of the current model and the predicted labels based on the original graph structure data, and the loss between the predicted labels of the current model and the true labels.

[0034] A further improvement of this invention is that, in step 2), the edge existence probability analysis based on the decoder encoder specifically includes the following steps:

[0035] Step 1: The encoder generates latent node embedding vectors.

[0036] Using the assumed adjacency matrix and feature matrix as input, a two-layer node classification graph convolutional neural network model is trained; where the first layer is regarded as an encoder, and the output is the latent node embedding vector Z, which represents the node features in a low-dimensional vector.

[0037] Step 2: Decoder generates probability matrix

[0038] The latent node embedding vectors Z and Z T Multiply the nodes, where each element is the dot product of the corresponding node embedding vectors. Larger element values ​​indicate greater similarity between the two nodes, and a higher probability that an edge exists between them. The probability matrix L is obtained using a decoder.

[0039] L = sigmod(ZZ) T ).

[0040] A further improvement of the present invention is that, in Step 2), the elements of the probability matrix L take values ​​in the range of (0,1), which can intuitively reflect whether there is an edge between nodes, and gives an assessment of the possibility of edge existence based on the assumed adjacency matrix.

[0041] A further improvement of this invention is that, in step 3), the generation of the differential privacy graph based on sampling specifically includes the following steps:

[0042] Step 1: Temperature Protection

[0043] In some scenarios, the degree of a node also contains important privacy information. Therefore, the privacy budget is divided into two parts, ∈1 and ∈2, for protecting the node degree and edge respectively. The Laplace mechanism is used to protect the degree of the node.

[0044]

[0045] Where Lap(λ) represents Laplacian noise with a mean of 0 and a scale of λ, and [·] represents rounding;

[0046] Step 2: Calculate the number of samples

[0047] Based on the privacy budget ∈2 setting for edge protection, the edge difference privacy representation is as follows: for an element a in the original adjacency matrix i,j The adjacency matrix after differential privacy is as follows:

[0048]

[0049] Where p and q are the probabilities of reversing the edge and keeping the edge, respectively;

[0050] Assume node i has a degree of m. i This indicates that the i-th row of the adjacency matrix contains m. i 1 and (Nm) i (0) ; After differential privacy, let the degree of node i be m′. i It is expressed by the following formula:

[0051]

[0052] Where m′ i,1 Let m' represent the actual number of edges. i,2 The number of spurious edges is represented; the degree m of node i is reconstructed by randomly sampling the differentially private graph. i This preserves the sparsity of the graph; for each edge, the probability of being sampled is m. i / m′ i ,have:

[0053]

[0054] Among them, E i,1 and E i,2 Let represent the number of real edges sampled and the number of fake edges sampled, respectively, both determined by the privacy budget ∈ , the total number of nodes N, and the degree m of node i. i Jointly decided; when the privacy budget ∈ 2 becomes smaller, E i,1 It will decrease accordingly, while E i,2 It will increase, but the sum of the two will always remain at m. i ;

[0055] Step 3: Generate a differential privacy adjacency matrix

[0056] Step 2) provides the possibility matrix L, which reflects the plausibility of an edge existing between any two nodes; therefore, based on L, sample E. i,1 real edges and E i,2 In the process of identifying false edges, the edge with the maximum value in L is selected, i.e., the edge with the highest probability of existence. For a specific node i, based on the i-th row of L, where each element corresponds to an edge, the element values ​​corresponding to the real edges are sorted, and the edge with the highest probability of existence, E, is selected. i,1Add the edge to the new edge set; sort the element values ​​corresponding to the spurious edges and select the one with the highest probability of existence, E. i,2 Each edge is added to a new edge set; after repeating the edge selection process for each node, a highly available differential privacy adjacency matrix is ​​generated.

[0057] A further improvement of this invention is that, in Step 2 of step 3), due to the sparsity of the graph, N is much larger than m. i Directly using differential privacy will make m′ i Compared to m i It is much larger, which causes the graph to become denser.

[0058] The present invention has at least the following beneficial technical effects:

[0059] (1) The present invention provides a highly available differential privacy graph generation method, which effectively protects privacy information in educational big data. At the same time, compared with the simple differential privacy method, the differential privacy graph generated by the present invention still has high availability in downstream tasks.

[0060] (2) This invention mines the backbone network of the graph through reverse learning and provides further analysis of the possibility of edge existence through encoder and decoder. Unlike the method of directly using the original graph structure data, this invention can eliminate the interference of irrelevant edges, extract the backbone network of the graph, retain the key information in the graph, delete a large number of unimportant edges, effectively protect privacy while ensuring the availability of data.

[0061] (3) Based on the differential privacy method, this invention combines edge existence probability analysis. Under the premise of keeping the degree distribution of important feature nodes unchanged, it reasonably allocates the ratio of real edges and false edges according to the privacy budget and constructs a more reasonable adjacency matrix through sampling method. Compared with the differential privacy method, it has the same privacy protection effect and higher usability. Attached Figure Description

[0062] Figure 1 This is a flowchart of the overall framework.

[0063] Figure 2 The flowchart shows the process of generating the assumed adjacency matrix based on inverse learning.

[0064] Figure 3 This is a graph for edge existence probability analysis based on the decoder encoder. Detailed Implementation

[0065] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] like Figure 1 As shown, the present invention provides a highly available differential privacy graph generation method for educational big data, comprising the following steps:

[0067] 1) Generation of the assumed adjacency matrix based on inverse learning

[0068] A graph neural network model is trained using the original graph structure data. The backbone network is then mined using a specific loss function and gradient descent within the graph neural network model to learn a hypothetical adjacency matrix. This hypothetical adjacency matrix is ​​assumed to be similar to the aggregated information obtained from the original graph structure data during the information aggregation process of the graph neural network. The generation of the hypothetical adjacency matrix based on inverse learning specifically includes the following steps:

[0069] Step 1: Training the baseline graphical neural network

[0070] The initialization assumes that the adjacency matrix is ​​the same size as the original adjacency matrix, with all elements being 0. A baseline graph neural network model is trained on the original graph structure data, and the predicted labels are obtained by comparing with the real labels.

[0071] Step 2: Calculate the loss function

[0072] Using the baseline graph neural network model trained in the previous step, and taking the assumed adjacency matrix and the feature matrix from the original graph structure data as input, we obtain predicted labels based on the assumed adjacency matrix. Optimize the assumed adjacency matrix using the following loss function:

[0073]

[0074]

[0075] Where N is the number of nodes, y i To predict labels based on the original graph structure data, The labels are the true values, and α is a hyperparameter controlling the proportion of the two loss components. L GNNIt consists of two parts: the loss between the predicted labels of the current model and the predicted labels based on the original graph structure data, and the loss between the predicted labels of the current model and the true labels. ||A s || F This is a regularization term that guarantees the sparsity of the assumed adjacency matrix.

[0076] Step 3: Optimize the assumed adjacency matrix

[0077] To control the range of values ​​for elements in the assumed adjacency matrix, projective gradient descent is employed:

[0078]

[0079]

[0080] Where t is the iteration number, η is the learning rate, and g is the gradient. The generated assumed adjacency matrix reduces a large number of unimportant edges compared to the adjacency matrix of the original graph structure data, effectively preserving key topological information, and still achieves similar performance in downstream tasks.

[0081] 2) Edge Existence Probability Analysis Based on Decoder and Encoder

[0082] A two-layer graph convolutional neural network is trained using an assumed adjacency matrix to classify nodes. The first layer is the encoder, which outputs the hidden embeddings of the nodes. The second layer is the decoder, which uses the hidden embeddings to calculate the probability matrix as the basis for sampling. The edge existence probability analysis based on the decoder and encoder specifically includes the following steps:

[0083] Step 1: The encoder generates latent node embedding vectors.

[0084] Using the assumed adjacency matrix and feature matrix as input, a two-layer convolutional neural network model for node classification graphs is trained. The first layer can be viewed as an encoder, whose output is the latent node embedding vector Z, representing node features as a low-dimensional vector.

[0085] Step 2: Decoder generates probability matrix

[0086] The latent node embedding vectors Z and Z T The nodes are multiplied, where each element is the dot product of the corresponding node embedding vectors. Larger element values ​​indicate greater similarity between the two nodes, and a higher probability that an edge exists between them. The probability matrix L is obtained using the decoder.

[0087] L = sigmod(ZZ) T )

[0088] The elements in the probability matrix take values ​​in the range (0,1), which can intuitively reflect whether there are edges between nodes. Based on the assumption of the adjacency matrix, an assessment of the probability of edge existence is given.

[0089] 3) Differential privacy graph generation based on sampling

[0090] Based on the privacy budget, the number of real and fake edges that a node should have is calculated, and a corresponding number of edges are sampled based on the probability matrix to obtain the protected adjacency matrix. The generation of the differential privacy graph based on sampling specifically includes the following steps:

[0091] Step 1: Temperature Protection

[0092] In certain scenarios, the degree of a node also contains important privacy information. Therefore, this method divides the privacy budget ∈ into two parts, ∈1 and ∈2, for protecting the node degree and edge, respectively. This method uses the Laplace mechanism to protect the degree of nodes:

[0093]

[0094] Where Lap(λ) represents Laplacian noise with a mean of 0 and a scale of λ, and [·] denotes rounding. Differential privacy of degree can effectively prevent the degree of a node from being leaked.

[0095] Step 2: Calculate the number of samples

[0096] Based on the privacy budget ∈2 setting for edge protection, edge differential privacy can be expressed as: for an element a in the original adjacency matrix i,j The adjacency matrix after differential privacy is as follows:

[0097]

[0098] Where p and q are the probabilities of reversing the edge and keeping the edge, respectively.

[0099] Assume node i has a degree of m. i This indicates that the i-th row of the adjacency matrix contains m. i 1 and (Nm) i (0) After differential privacy, let the degree of node i be m′. i It can be expressed by the following formula:

[0100]

[0101] Where m′ i,1 Let m' represent the actual number of edges. i,2 This represents the number of spurious edges. Due to the sparsity of graphs, N is often much larger than m. i Directly using differential privacy will make m′ i Compared to m i The degree m of node i is much larger, resulting in a denser graph. This method reconstructs the degree m of node i by randomly sampling the differentially privacy-enhanced graph. iThis effectively preserves the sparsity of the graph. For each edge, the probability of being sampled is m. i / m′ i ,have:

[0102]

[0103] Among them, E i,1 and E i,2 Let represent the number of real edges sampled and the number of fake edges sampled, respectively, both determined by the privacy budget ∈ , the total number of nodes N, and the degree m of node i. i Joint decision. When the privacy budget ∈2 decreases, E i,1 It will decrease accordingly, while E i,2 It will increase, but the sum of the two will always remain at m. i .

[0104] Step 3: Generate a differential privacy adjacency matrix

[0105] Step 2) provides the possibility matrix L, which reflects the plausibility of an edge existing between any two nodes. Therefore, based on L, E is sampled. i,1 real edges and E i,2 In the process of identifying false edges, the edge with the maximum value in L is selected, i.e., the edge with the highest probability of existence. For a specific node i, based on the i-th row of L, where each element corresponds to an edge, the element values ​​corresponding to the real edges are sorted, and the edge with the highest probability of existence, E, is selected. i,1 Add the edge to the new edge set; similarly, sort the element values ​​corresponding to the spurious edges and select the one with the highest probability of existence, E. i,2 Each edge is added to a new edge set. After repeating the edge selection process for each node, a highly available differential privacy adjacency matrix is ​​generated.

[0106] Example

[0107] A partial teaching data graph of a school from 2017 to 2019 was selected, containing 3327 nodes, 4732 edges, a node feature dimension of 3703, and a label dimension of 6. The invention will be further described in detail below with reference to the accompanying drawings, experimental examples, and specific implementation methods. All technologies implemented based on the content of this invention fall within the scope of this invention.

[0108] like Figure 1 As shown, in a specific implementation of the present invention, a highly available differential privacy graph generation method for educational big data includes the following steps:

[0109] Step 1. Generation of the assumed adjacency matrix based on inverse learning

[0110] Educational data graphs contain a large amount of sensitive information, and direct use poses a risk of privacy exposure. Generating a hypothetical adjacency matrix can help extract key information and reduce privacy breaches. The hypothetical adjacency matrix generation process is as follows: Figure 2 Specifically, it includes the following steps:

[0111] S101. Training the baseline graph neural network

[0112] Using a local educational data graph as input, a two-layer graph convolutional neural network (GCN) model is selected as the baseline graph neural network model for training on a node classification task. The final model parameters are denoted as f. θ The first layer input consists of the node feature matrix X and the adjacency matrix A. k After information transmission and aggregation, the output node hidden feature matrix is ​​as follows:

[0113] H1=f(AX′W1)

[0114] The second layer input is the first layer output H. k,1 With adjacency matrix A k Output node hidden feature matrix H k,2 Used for node classification:

[0115] H2 = f(AH1W2)

[0116] After training, retain the model's predicted label y for the local teaching data graph. i .

[0117] S102. Calculate the loss function

[0118] First, initialize the adjacency matrix with a size of 3327×3327 and set all elements to 0, denoted as A. s Using the baseline model from the previous step, and taking the assumed adjacency matrix and the feature matrix X of the local teaching data graph as input, we obtain the baseline model's predicted label for the assumed adjacency matrix. Secondly, the hyperparameters α = 0.5 and β = 0.0001 are set, and the predicted labels y of the original graph structure data are calculated based on the model. i and real labels Set the loss function as For A s Optimize.

[0119] S103. Optimize the assumed adjacency matrix

[0120] To control the range of values ​​for elements in the assumed adjacency matrix, projective gradient descent is used to limit the gradient update values ​​to the range [0,1].

[0121]

[0122] Set the learning rate η t =0.1, iterative optimization assumes adjacency matrix Perform 100 iterations.

[0123] Step 2. Edge Existence Probability Analysis Based on Decoder and Encoder

[0124] The assumed adjacency matrix obtained through inverse learning reduces a large number of unimportant edges while preserving the main topological information, effectively reducing privacy leaks. The preserved information can be further enhanced using an encoder and decoder. The edge existence probability analysis implementation process is as follows: Figure 3 Specifically, it includes the following steps:

[0125] S201. The encoder generates latent node embedding vectors.

[0126] The first layer of the baseline model in step 1 is used as an encoder to output the node embedding vector Z, which can simultaneously reflect the node information and topological structure of the educational data graph.

[0127] S202. Decoder generates probability matrix

[0128] Calculate L = sigmod(ZZ) T We obtain the probability matrix L, which serves as the basis for evaluating whether edges exist between nodes.

[0129] Step 3. Generation of differential privacy graph based on sampling

[0130] To further protect privacy, let the privacy budget ∈ = 1. The privacy budget is divided into two parts to protect the degree and edges of the educational data graph, denoted as ∈1 = 0.2∈ and ∈2 = 0.8∈. First, for degree protection, based on the allocated privacy budget ∈1, the corresponding Laplace noise is calculated, and the degree of the nodes after differential privacy is generated. Secondly, for edge protection, a differential privacy adjacency matrix is ​​generated based on the edge privacy budget ∈2. For each node, some real edges are reversed, and a large number of spurious edges are added. To maintain the sparsity of the educational data graph, the number of real edges and spurious edges to be sampled for each node is calculated based on the edge privacy budget ∈2. Finally, based on the probability matrix L, for node i, the edge E with the highest probability of existence is sampled. i,1 real edges and E i,2 False edges, totaling By using edges, a highly available differential privacy adjacency matrix can be obtained.

[0131] Specifically, in this embodiment, under strict privacy budget (∈=1) requirements, the generated differential privacy graph can still achieve an accuracy of 78.06% in downstream tasks, which is very similar to the performance of the original teaching data graph.

[0132] Those skilled in the art will readily understand that the above description is merely an embodiment of the method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A highly available differential privacy graph generation method for educational big data, characterized in that, include: First, using the original graph structure data as input, inverse learning is used to mine the backbone network in the graph, generating a hypothetical adjacency matrix that assumes similar results for the same downstream tasks. ; Secondly, to deepen the topological information obtained from the assumed adjacency matrix, a probability matrix is ​​generated using an encoder-decoder based on the assumed adjacency matrix obtained in the previous step. , of which elements The larger the value, the more likely the node is to be affected. With nodes The likelihood of an edge existing between nodes is measured by considering the probability of an edge existing between different nodes. Finally, based on the privacy budget, a differential privacy graph is generated, in which spurious edges constitute the vast majority, and directly using them for downstream tasks would lead to a severe performance degradation. Based on the probability of an edge existing between nodes, each node in the graph samples an equal number of edges from both real and spurious edges to generate a protected differential privacy adjacency matrix. The method specifically includes the following steps: 1) Generation of the assumed adjacency matrix based on inverse learning A graph neural network model is trained using the original graph structure data. The backbone network is mined through the loss function and gradient descent in the graph neural network model. The assumed adjacency matrix is ​​learned, and the assumed adjacency matrix is ​​similar to the aggregated information obtained from the original graph structure data during the information aggregation process of the graph neural network. 2) Edge Existence Probability Analysis Based on Decoder and Encoder A two-layer graph convolutional neural network is trained using an assumed adjacency matrix to classify nodes; the first layer is the encoder, which outputs the hidden embeddings of the nodes, and the second layer is the decoder, which uses the hidden embeddings to calculate the probability matrix as the basis for sampling. 3) Differential privacy graph generation based on sampling Based on the privacy budget, the number of real and fake edges that a node should have is calculated, and the corresponding number of edges are sampled based on the probability matrix to obtain the protected adjacency matrix.

2. The method for generating a highly available differential privacy graph for educational big data according to claim 1, characterized in that, In step 1), the generation of the assumed adjacency matrix based on inverse learning specifically includes the following steps: Step 1: Training the baseline graphical neural network Initially, the adjacency matrix is ​​assumed to be the same size as the original adjacency matrix, with all elements being 0. A baseline graph neural network model is trained on the original graph structure data, and the predicted labels are obtained by comparing with the real labels. Step 2: Calculate the loss function Using a pre-trained baseline graph neural network model, and taking the assumed adjacency matrix and the feature matrix from the original graph structure data as input, we obtain predicted labels based on the assumed adjacency matrix. The assumed adjacency matrix is ​​optimized using the following loss function: in Number of nodes To predict labels based on the original graph structure data, For real labels, The hyperparameter is used to control the proportion of the two losses; This is a regularization term that guarantees the sparsity of the assumed adjacency matrix; Step 3: Optimize the assumed adjacency matrix To control the range of values ​​for elements in the assumed adjacency matrix, projective gradient descent is employed: in, For iteration rounds, For learning rate, For gradient.

3. The method for generating a highly available differential privacy graph for educational big data according to claim 2, characterized in that, In Step 2 of Step 1), It consists of two parts: the loss between the predicted labels of the current model and the predicted labels based on the original graph structure data, and the loss between the predicted labels of the current model and the true labels.

4. The method for generating a highly available differential privacy graph for educational big data according to claim 2, characterized in that, Step 2), the edge existence probability analysis based on the decoder encoder specifically includes the following steps: Step 1: The encoder generates latent node embedding vectors. Using the assumed adjacency matrix and feature matrix as input, a two-layer convolutional neural network model for node classification graphs is trained; the first layer is treated as an encoder, and its output is the latent node embedding vector. Node features are represented by low-dimensional vectors; Step 2: Decoder generates probability matrix Latent node embedding vector and The nodes are multiplied, where each element is the dot product of the corresponding node embedding vectors. Larger element values ​​indicate greater similarity between the two nodes, and a higher probability that an edge exists between them. A probability matrix is ​​then obtained using a decoder. : 。 5. The method for generating a highly available differential privacy graph for educational big data according to claim 4, characterized in that, In Step 2 of Step 2, the probability matrix The elements in the matrix take values ​​in the range of (0, 1), which can intuitively reflect whether there are edges between nodes. Based on the assumption of an adjacency matrix, an assessment of the probability of edge existence is given.

6. The method for generating a highly available differential privacy graph for educational big data according to claim 5, characterized in that, Step 3) involves the following steps for generating the differential privacy graph based on the sampling: Step 1: Temperature Protection In certain scenarios, the degree of a node also contains important privacy information; therefore, privacy budgeting should be considered. Divided into two parts , These are used to protect node degree and edge degree, respectively; the Laplace mechanism is used to protect node degree. in This indicates a mean of 0 and a scale of 0. Laplace noise, [ ] indicates rounding; Step 2: Calculate the number of samples Based on the privacy budget of the border protection The edge difference privacy representation is set as follows: for each element in the original adjacency matrix... The adjacency matrix after differential privacy is as follows: in These are the probabilities of reversing an edge and keeping an edge, respectively. Assuming nodes The degree is This indicates that in the adjacency matrix, the first... The line contains 1 and A zero; after differential privacy, record the node. The degree is It is expressed by the following formula: in This represents the actual number of edges. This indicates the number of spurious edges; nodes are reconstructed by randomly sampling the differentially private graph. degree This preserves the sparsity of the graph; for each edge, the probability of being sampled is... ,have: in, and These represent the number of real edges sampled and the number of fake edges sampled, respectively, both determined by the privacy budget. Total number of nodes and nodes degree Joint decision; when privacy budget When it shrinks, It will decrease accordingly, and It will increase, but the sum of the two will always remain at 1. ; Step 3: Generate a differential privacy adjacency matrix Step 2) The probability matrix given This reflects the rationale for the existence of an edge between any two nodes; therefore, with As a basis, sampling real edges and In the process of creating a false edge, choose to... The edge with the maximum value in the interval is the edge with the highest probability; for a specific node... In other words, with The The behavior is based on the fact that each element corresponds to an edge, and the element values ​​corresponding to the actual edges are sorted, with the one having the highest probability of existence being selected. Add the edge to the new edge set; sort the element values ​​corresponding to the spurious edges and select the ones with the highest probability of existence. Each edge is added to a new edge set; after repeating the edge selection process for each node, a highly available differential privacy adjacency matrix is ​​generated.

7. The method for generating a highly available differential privacy graph for educational big data according to claim 6, characterized in that, In Step 2 of Step 3), due to the sparsity of the graph, Much larger Directly using differential privacy would make compared to It is much larger, which causes the graph to become denser.

Images