CN116861152A - Tax data security graph neural network training method based on matrix decomposition - Google Patents

Abstract

The invention discloses a tax data security graph neural network training method based on matrix decomposition, which comprises the following steps: firstly, carrying out safe eigenvalue decomposition on an adjacent matrix part of a tax data graph by using an external server, dividing an obtained eigenvalue decomposition result into a plurality of parts, and carrying out operation on the parts and an eigenvector matrix to generate a plurality of distributable adjacent matrixes; secondly, carrying out differential privacy on the feature matrix part of the tax data graph; thirdly, the tax data has the characteristic matrix after the decomposed adjacency matrix and the differential privacy are distributed to each computing party through a parameter server for model training; and finally, returning the calculation result to the tax data owner by the calculator, and obtaining the target model parameters through integrating and updating by the parameter server. According to the method, the original tax data is safely decomposed in the modes of topology secret sharing and adjacency matrix eigenvalue decomposition, so that the tax data is efficiently analyzed and modeled by means of external computing resources, and the analysis efficiency is improved.

Description

A neural network training method for tax data security graph based on matrix decomposition

Technical field

The invention belongs to the technical field of graph privacy protection methods, and particularly relates to a tax data security graph neural network training method based on matrix decomposition.

Background technique

In recent years, with the rapid development of the national economy and the continuous prosperity of the market economy, tax data has become increasingly complex. Tax data is often represented as a graph structure data type, reflecting individual tax information and social relationship information. Therefore, graph neural networks can effectively model graph-structured data in tax data and deeply mine the information contained therein. Tax data modeling is the basic work for intelligent processing of tax data and a key prerequisite for realizing tax big data. However, the increasing scale of tax data and the large amount of private information contained in it hinder its analysis and utilization. Traditional data protection focuses on discrete data points, making individual data unable to be identified and utilized through technical means. However, graph-structured data contains not only node information, but also rich and important topological information, making it difficult for traditional data protection methods to fully protect it. Different from traditional data protection methods, current privacy protection research is committed to making data "available and invisible", that is, protecting private information from being leaked without affecting the use of data. For graph-structured data, privacy protection research focuses on protecting graph topology information to avoid leakage of sensitive information, which can more effectively ensure the security of graph-structured data than traditional methods. Existing tax data modeling is often inefficient due to the huge scale of tax data and the limited computing power of tax agencies. It is often inefficient for tax agencies to complete the modeling tasks themselves, and there is an urgent need to use external computing resources to improve efficiency; but at the same time, tax data It contains a large amount of sensitive information, and the consequences of exposure are very serious. It is not allowed to directly use the computing power of external organizations to process relevant data. On the one hand, tax data needs to be processed securely to avoid the leakage of private information, and on the other hand, the processed data needs to be processed. can be modeled correctly. With the increasing amount of taxpayer data, the scale of tax data is increasing, and the content is becoming increasingly complex. How to ensure the security of tax data while getting rid of the constraints of local computing power and using external computing power to efficiently train graph neural network models for tax data has become a problem. It has become an urgent problem to be solved, which is of great significance for accelerating tax data processing and further realizing tax big data.

At present, there is no relevant research that proposes corresponding solutions for the tax data privacy protection graph neural network training method. The main invention patents related to tax data protection include:

Document 1: A blockchain-based tax information processing method and system (202011290032.1)

Document 2: An enterprise batch clustering method and system based on multi-dimensional features (202211142876.0)

Document 1 designs a tax information processing method and system based on blockchain. Using blockchain, the tax agency is used as the tax node of the blockchain to manage the blocks and divide different channels according to the business organization. Each channel Link the tax node and the corresponding business organization node, and use the tax node to broadcast the tax certificate information to the business organization node in the corresponding channel according to user authorization, so that the business organization node obtains the tax certificate information.

Document 2 designs a batch clustering method and system for enterprises based on multi-dimensional features. It collects tax data, news data and public opinion data from multiple target enterprises in the tax field to be clustered, and analyzes the collected data to generate feature data. , and construct a graph structure based on the characteristic data, and use the graph structure as the input of the optimal graph neural network clustering model to obtain the clustering results of the target enterprises to be clustered.

Among the above technical solutions, Document 1 focuses on the storage protection of tax data and applies blockchain technology to ensure data security. However, it does not consider the application of protected data, and the query and use efficiency of data is low. Document 2 is good at collecting data. The tax data graph is modeled on the premise of tax data, and the tax data is analyzed using graph neural network. Although good analysis results are obtained, the privacy protection of tax data is not considered in the entire process, which may bring certain security risks. However, in reality, limited by the computing power of tax agencies, the processing efficiency of existing tax data is low. At the same time, limited by the sensitivity of tax data, it is not possible to directly borrow the computing power of external organizations to analyze and process relevant data. Therefore, how to efficiently train a graph neural network model for tax data while ensuring the security of tax data has become an urgent problem to be solved.

Contents of the invention

The present invention aims to provide a tax data security graph neural network training method based on matrix decomposition. First, use an external server to perform secure eigenvalue decomposition on the adjacency matrix part of the tax data graph, divide the obtained eigenvalue decomposition results into multiple parts, and perform operations with the eigenvector matrix to generate multiple distributable adjacency matrices; secondly, , carry out differential privacy on the feature matrix part of the tax data graph; thirdly, the tax data owner distributes the decomposed adjacency matrix and the differentially private feature matrix through the parameter server to each computing party for model training; finally, the computing party will The calculation results are returned to the tax data owner, and the target model parameters are obtained through integration and update by the parameter server.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A tax data security graph neural network training method based on matrix decomposition, including:

First, use an external server to perform secure eigenvalue decomposition on the adjacency matrix of the tax data graph, divide the obtained eigenvalue decomposition results into multiple parts, and perform operations with the eigenvector matrix to generate partial secrets of multiple distributable adjacency matrices. ; Secondly, carry out differential privacy on the feature matrix of the tax data graph; thirdly, the tax data has the partial secret of the decomposed adjacency matrix and the differentially private feature matrix distributed to each computing party through the parameter server for model training; finally , the calculation party returns the calculation results to the tax data owner, and obtains the target model parameters through integration and update by the parameter server.

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

1) Adjacency matrix secret sharing based on eigenvalue decomposition

For the adjacency matrix of the tax data graph, use an external server to perform secure eigenvalue decomposition; randomly divide the eigenvalues into corresponding parts according to the number of calculation parties, and the eigenvalue decomposition result and the eigenvector matrix operation result can be published Part of the secret of the adjacency matrix;

2) Feature matrix protection based on differential privacy

For the feature matrix of the tax data graph, the differential privacy method is used and the Laplacian mechanism is applied to protect it;

3) Model training and integration based on parameter server

Distribute the partial secrets of the decomposed adjacency matrix and the differentially private feature matrix to each computing party. Each computing party trains the graph convolutional neural network model based on the distributed data, sends, collects and integrates model parameters through the parameter server to obtain the target model parameters.

A further improvement of the present invention is that in step 1), the adjacency matrix secret sharing based on eigenvalue decomposition includes:

Step1: Safe matrix eigenvalue decomposition

For the adjacency matrix A of the tax data graph, through multiple iterations of QR decomposition, a sufficiently accurate numerical solution of eigenvalue decomposition is obtained:

where t is the iteration round, Q _t and R _t are the QR decomposition results of A _t in round t respectively; after k iterations, the eigenvalue diagonal matrix Λ=A _k and the eigenvector matrix X=Q ₁ ...Q ₁ , the original adjacency matrix A=XΛX ^-1 ;

Step2: Topology secret sharing

For the obtained eigenvalue diagonal matrix Λ, the eigenvalues are randomly divided into multiple groups in the form of multiple diagonal matrices. When divided into two groups, the specific steps are as follows:

Generate a random diagonal 01 matrix S, where the diagonal elements obey the following rules:

Generate new diagonal matrices Λ ₁ and Λ ₂ as follows:

Where In represents the n-dimensional unit matrix, × _h represents the Hadamard product, that is, the multiplication of corresponding elements of the matrix;

Use the newly generated diagonal matrices Λ ₁ and Λ ₂ to generate new matrices A1 and A2 as follows:

A ₁ and A ₂ have the following properties:

In the GNN model, the graph topology is expressed in the form of an adjacency matrix, and the power of the adjacency matrix can reflect the information transfer process of the GNN model.

A further improvement of the present invention is that in Step 1), a third-party server is set up to perform safe decomposition of the adjacency matrix A: the data owner generates a sparse random 01 matrix P, calculates and uploads A′=PAP to the third-party server ^-1 , the third-party server calculates the eigenvalue decomposition of A′ according to the above iterative solution process and returns the calculation results X′, Λ to the data owner. A′=X′ΛX′ ^-1 , the data owner calculates X=P ^-1 X′, get the matrix decomposition result.

A further improvement of the present invention is that in Step 2 of step 1), for the two-layer GCN, the node embedding is affected by its neighbors within the two-hop range. The neighbors within the two ranges are represented by the square of the adjacency matrix A ² , and A ² can effectively Indicate the connection relationship of the graph and the information transfer between nodes; record the number of nodes n, decompose the original adjacency matrix into k matrices, and each decomposed matrix contains eigenvalues, missing/> eigenvalues, on the premise of obtaining all eigenvalues, the probability of correct arrangement

A further improvement of the present invention is that when n=100 and k=2, p≈3.3×10 ^-65 .

A further improvement of the present invention is that in step 2), the feature matrix protection based on differential privacy includes:

Step1: Privacy budget and global sensitivity calculation

Apply the Laplacian mechanism to perform differential privacy protection on the feature matrix X of the tax data graph, and calculate the global sensitivity Δ _f according to the set privacy budget ∈:

Δ _f =max _{D, D′} {|h＝h′|}

Among them, D and D′ are a pair of adjacent data, h and h are the results of random queries for D and D′ respectively; let Set the Laplacian noise distribution to be added as follows:

The above Laplacian mechanism satisfies ∈-differential privacy, that is:

Pr[M(D)＝y]≤e ^∈ Pr[M(D′)＝y]

Where M is the processing mechanism applied;

Step2: Noise injection

Insert the Laplacian noise generated in the previous step into the feature matrix X of the tax data graph to obtain the privacy-preserving feature matrix X′.

A further improvement of the present invention is that in step 3), the model training and integration based on the parameter server includes:

Step1: Data distribution

The adjacency matrix A of the tax data graph is decomposed into {A _k ...}, k=1, 2, ..., and the feature matrix X of the tax data graph is processed through differential privacy to obtain Each calculation side obtains A _k and X′ as inputs to the GNN model;

Step2: Model training based on privacy-preserving data

The graph convolutional neural network model is selected for training. Calculator k locally has a two-layer GCN model and is trained on the data assigned to itself; the first layer input is the node feature matrix X and the adjacency matrix Ak. After information transfer and After aggregation, output node hidden feature matrix:

H _k,1 =f(A _k X'W _k,1 )

The input of the second layer is the output H _k,1 of the first layer and the adjacency matrix A _k , and the output node hidden feature matrix H _k,2 is used for node classification or other downstream tasks;

H _k,2 =f(A _k H _k,1 W _k,2 )

After training, the calculating party uploads the model parameters W _k,1 , W _k,2 to the parameter server held by the data owner, and at the same time pulls the updated model parameters from the parameter server; the parameter server collects the model parameters uploaded by each participant Finally, the model parameters are integrated using the model averaging method in the distributed machine learning method to obtain new model parameters.

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

(1) This invention performs privacy protection processing on the adjacency matrix and the special certificate matrix of the tax data graph respectively. Through topological secret sharing and adjacency matrix eigenvalue decomposition, the topological information of the graph is protected from being known by the calculating party. Through differential privacy, way to protect node characteristic information, and the present invention ensures the security of sensitive information in the entire process.

(2) The present invention uses topological secret sharing and adjacency matrix eigenvalue decomposition to safely decompose original tax data, and then uses external computing resources to achieve efficient analysis and modeling of tax data, improving analysis efficiency.

Description of the drawings

Figure 1 is the overall framework flow chart.

Figure 2 is a flow chart of adjacency matrix secret sharing based on eigenvalue decomposition.

Figure 3 is a flow chart of model training and integration based on parameter server.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although 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 provide a thorough understanding of the disclosure, and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, as long as there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

As shown in Figure 1, the invention provides a tax data security graph neural network training method based on matrix decomposition, including the following steps:

1) Adjacency matrix secret sharing based on eigenvalue decomposition

For the adjacency matrix of the tax data graph, use an external server to perform secure eigenvalue decomposition; randomly divide the eigenvalues into corresponding parts according to the number of calculation parties, and the eigenvalue decomposition result and the eigenvector matrix operation result can be published Part of the secret of the adjacency matrix; the secret sharing of the adjacency matrix based on eigenvalue decomposition includes:

Step1: Safe matrix eigenvalue decomposition

For the adjacency matrix A of the tax data graph, through multiple iterations of QR decomposition, a sufficiently accurate numerical solution of eigenvalue decomposition is obtained:

where t is the iteration round, Q _t and R _t are the QR decomposition results of A _t in round t respectively; after k iterations, the eigenvalue diagonal matrix Λ=A _k and the eigenvector matrix X=Q ₁ ...Q ₁ , the original adjacency matrix A=XΛX ^-1 ;

Step2: Topology secret sharing

For the obtained eigenvalue diagonal matrix Λ, the eigenvalues are randomly divided into multiple groups in the form of multiple diagonal matrices. When divided into two groups, the specific steps are as follows:

Generate a random diagonal 01 matrix S, where the diagonal elements obey the following rules:

Generate new diagonal matrices Λ ₁ and Λ ₂ as follows:

Where In represents the n-dimensional unit matrix, × _h represents the Hadamard product, that is, the multiplication of corresponding elements of the matrix;

Use the newly generated diagonal matrices Λ ₁ and Λ ₂ to generate new matrices A1 and A2 as follows:

A ₁ and A ₂ have the following properties:

In the GNN model, the graph topology is expressed in the form of an adjacency matrix, and the power of the adjacency matrix can reflect the information transfer process of the GNN model. Securely decompose the adjacency matrix A by setting up a third-party server: the data owner generates a sparse random 01 matrix P, calculates and uploads A′=PAP ^-1 to the third-party server, and the third-party server calculates A′ according to the above iterative solution process The eigenvalues are decomposed and the calculation results X′ and Λ are returned to the data owner, with A′=X′ΛX′ ^-1 . The data owner calculates X=P ^-1 X′ and obtains the matrix decomposition result. For a two-layer GCN, the node embedding is affected by its neighbors within the two-hop range. The neighbors within the two ranges are represented by the square A ² of the adjacency matrix. A ² can effectively indicate the connection relationship of the graph and the information transfer between nodes. Due to the properties of A ₁ and A ₂ , they are distributed to different calculation parties to calculate exponentiation, and the exponentiation result of the original data is obtained by recycling the integration results. On the other hand, for the calculation side, the decomposed A ₁ and A ₂ are numerically very different from the original adjacency matrix with a value of 01, and contain a large number of decimals, making it impossible to identify the existence of any edge numerically. For the calculation side, only part of the eigenvalues of the original adjacency matrix can be derived from the allocated matrix, which is not enough to restore the original adjacency matrix. Even if the calculation side obtains all the eigenvalues, it needs to arrange the eigenvalues in the correct order to restore the original adjacency matrix, but the probability of correct arrangement is very small. Record the number of nodes n, and decompose the original adjacency matrix into k matrices. Each decomposed matrix contains eigenvalues, missing/> eigenvalues, on the premise of obtaining all eigenvalues, the probability of correct arrangement/> In the case where all eigenvalues are not obtained, the probability of the calculation party recovering the original adjacency matrix is much smaller than the above probability p.

2) Feature matrix protection based on differential privacy

For the feature matrix of the tax data graph, the differential privacy method is used and the Laplacian mechanism is applied to protect it; the feature matrix protection based on differential privacy includes:

Step1: Privacy budget and global sensitivity calculation

Apply the Laplacian mechanism to perform differential privacy protection on the feature matrix X of the tax data graph, and calculate the global sensitivity Δ _f according to the set privacy budget ∈:

Δ _f =max _{D, D′} {|h＝h′|}

Among them, D and D′ are a pair of adjacent data, h and h are the results of random queries for D and D′ respectively; let Set the Laplacian noise distribution to be added as follows:

The above Laplacian mechanism satisfies ∈-differential privacy, that is:

Pr[M(D)＝y]≤e ^∈ Pr[M(D′)＝y]

Where M is the processing mechanism applied;

Step2: Noise injection

Insert the Laplacian noise generated in the previous step into the feature matrix X of the tax data graph to obtain the privacy-preserving feature matrix X′.

3) Model training and integration based on parameter server

Distribute the partial secrets of the decomposed adjacency matrix and the differentially private feature matrix to each computing party. Each computing party trains the graph convolutional neural network model based on the distributed data, sends, collects and integrates model parameters through the parameter server to obtain the target model parameters. As shown in Figure 3, model training and integration based on parameter server include:

Step1: Data distribution

The adjacency matrix A of the tax data graph is decomposed into {A _k ...}, k=1, 2, ..., and the feature matrix X of the tax data graph is processed through differential privacy to obtain Each calculation side obtains A _k and X′ as inputs to the GNN model;

Step2: Model training based on privacy-preserving data

The graph convolutional neural network model is selected for training. Calculator k locally has a two-layer GCN model and is trained on the data assigned to itself; the first layer input is the node feature matrix X and the adjacency matrix A _k . After information transfer And the hidden feature matrix of the output node after aggregation:

H _k,1 =f(A _k X'W _k,1 )

The input of the second layer is the output H _k,1 of the first layer and the adjacency matrix A _k , and the output node hidden feature matrix H _k,2 is used for node classification or other downstream tasks;

H _k,2 =f(A _k H _k,1 W _k,2 )

After training, the calculating party uploads the model parameters W _k,1 , W _k,2 to the parameter server held by the data owner, and at the same time pulls the updated model parameters from the parameter server; the parameter server collects the model parameters uploaded by each participant Finally, the model parameters are integrated using the model averaging method in the distributed machine learning method to obtain new model parameters.

Example

Select the local tax data graph from 2017 to 2019 in the national tax of a certain region, which contains 2786 nodes, 5728 edges, the node feature dimension is 1289, and the label dimension is 6. The present invention will be described in further detail below with reference to the accompanying drawings, experimental cases and specific implementations. All technologies implemented based on the content of the present invention belong to the scope of the present invention.

As shown in Figure 1, in the specific implementation of the present invention, the tax data privacy protection graph neural network training method based on matrix decomposition and differential privacy includes the following steps:

Step 1. Adjacency matrix secret sharing based on eigenvalue decomposition

The tax data graph contains a large number of adjacency matrices, and the secret sharing method can effectively prevent the calculation party from guessing the adjacency matrix. The adjacency matrix secret sharing implementation process is shown in Figure 2, which specifically includes the following steps:

S101. Adjacency matrix eigenvalue decomposition

First, a random 01 matrix P with a size of 2786×2786 is generated locally, and then the masked matrix adjacency matrix A′=PAP ^-1 is calculated and uploaded to the third-party server. The third-party server will upload the masked matrix adjacency matrix A′ for eigenvalue decomposition, and transmit the decomposition results X′ and Λ back to the local. The decomposition results X′ and Λ are further processed locally to obtain the tax data graph adjacency matrix decomposition result X=P ^-1 X′P, Λ.

S102. Topology secret sharing

There are two calculation sides in this embodiment, so with the help of a randomly generated diagonal 01 matrix S of size 2786×2786, Λ is decomposed into Λ ₁ =S× _h Λ, Λ ₁ =(I ₂₇₈₆ -S)× _h Λ, and thus obtain new matrices A ₁ =XΛ ₁ X ^-1 and A ₂ =XΛ ₂ X ^-1 . Divide A ₁ and A ₂ to two calculating parties respectively.

Step 2. Feature matrix protection based on differential privacy

Using differential privacy, the feature matrix can be effectively protected simply and effectively.

Specifically, in this embodiment, the privacy budget ∈ = 10, the corresponding Laplacian noise is calculated and inserted into the feature matrix X to obtain the privacy-protecting feature matrix X′.

Step 3. Model training and integration based on parameter server

By utilizing the secret sharing of the adjacency matrix and the feature matrix after differential privacy, the calculation party cannot reversely deduce the information in the original tax data graph during the calculation process. With the help of the parameter server, the correct training of the model can be completed.

Specifically, in this embodiment, calculation method k trains a two-layer GCN model, and the model parameters are recorded as Wi _{, 1} and Wi _{, 2} . _Based on the allocated data A _i and

H _k,1 =f(A _k X'W _k,1 )

The input of the second layer is the output H _k,1 of the first layer and the adjacency matrix A _k , and the output node hidden feature matrix H _k,2 can be used for node classification or other downstream tasks.

H _k,2 =f(A _k H _k,1 W _k,2 )

And after training, W _{k, 1} and W _{k, 2} are uploaded to the parameter server. After both calculation methods are uploaded, the parameter server uses the model averaging method to integrate the model parameters to obtain And the updated parameters W ₁ and W ₂ are sent to the calculation party again for the next iteration, with a total of 100 iterations to obtain the final model parameters. The accuracy of the final model on the original tax data map reached 81.6%, which was only 3 percentage points lower than the accuracy of 84.1% of the model trained directly on the tax data map. However, the modeling speed was greatly accelerated with the help of external computing power.

It is easy for those skilled in the art to understand that the above descriptions are only method embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be regarded as should be included within the protection scope of the present invention.

Claims (9)

1. A tax data security graph neural network training method based on matrix decomposition, characterized by including: First, use an external server to perform secure eigenvalue decomposition on the adjacency matrix of the tax data graph, divide the obtained eigenvalue decomposition results into multiple parts, and perform operations with the eigenvector matrix to generate partial secrets of multiple distributable adjacency matrices. ; Secondly, carry out differential privacy on the feature matrix of the tax data graph; thirdly, the tax data has the partial secret of the decomposed adjacency matrix and the differentially private feature matrix distributed to each computing party through the parameter server for model training; finally , the calculation party returns the calculation results to the tax data owner, and obtains the target model parameters through integration and update by the parameter server. 2. A tax data security graph neural network training method based on matrix decomposition according to claim 1, characterized in that the method specifically includes the following steps: 1) Adjacency matrix secret sharing based on eigenvalue decomposition For the adjacency matrix of the tax data graph, use an external server to perform secure eigenvalue decomposition; randomly divide the eigenvalues into corresponding parts according to the number of calculation parties, and the eigenvalue decomposition result and the eigenvector matrix operation result can be published Part of the secret of the adjacency matrix; 2) Feature matrix protection based on differential privacy For the feature matrix of the tax data graph, the differential privacy method is used and the Laplacian mechanism is applied to protect it; 3) Model training and integration based on parameter server Distribute the partial secrets of the decomposed adjacency matrix and the differentially private feature matrix to each computing party. Each computing party trains the graph convolutional neural network model based on the distributed data, sends, collects and integrates model parameters through the parameter server to obtain the target model parameters. 3. A tax data security graph neural network training method based on matrix decomposition according to claim 2, characterized in that in step 1), the adjacency matrix secret sharing based on eigenvalue decomposition includes: Step1: Safe matrix eigenvalue decomposition For the adjacency matrix A of the tax data graph, through multiple iterations of QR decomposition, a sufficiently accurate numerical solution of eigenvalue decomposition is obtained: where t is the iteration round, Q _t and R _t are the QR decomposition results of A _t in round t respectively; after k iterations, the eigenvalue diagonal matrix Λ=A _k and the eigenvector matrix X=Q ₁ ...Q ₁ , the original adjacency matrix A=XΛX ^-1 ; Step2: Topology secret sharing For the obtained eigenvalue diagonal matrix Λ, the eigenvalues are randomly divided into multiple groups in the form of multiple diagonal matrices. When divided into two groups, the specific steps are as follows: Generate a random diagonal 01 matrix S, where the diagonal elements obey the following rules: Generate new diagonal matrices Λ ₁ and Λ ₂ as follows: Where I _n represents the n-dimensional unit matrix, × h represents the Hadamard product, that is, the multiplication of corresponding elements of the matrix; Use the newly generated diagonal matrices Λ ₁ and Λ ₂ to generate new matrices A ₁ and A ₂ as follows: A ₁ and A ₂ have the following properties: In the GNN model, the graph topology is expressed in the form of an adjacency matrix, and the power of the adjacency matrix can reflect the information transfer process of the GNN model. 4. A tax data security graph neural network training method based on matrix decomposition according to claim 3, characterized in that in Step 1 of step 1), secure decomposition of the adjacency matrix A is performed by setting up a third-party server: Data The owner generates a sparse random 01 matrix P, calculates and uploads A′=PAP ^-1 to the third-party server. The third-party server calculates the eigenvalue decomposition of A′ according to the above iterative solution process and returns the calculation results X′ and Λ to the data The owner has A′=X′ΛX′ ^-1 . The data owner calculates X=P ^-1 X′ and obtains the matrix decomposition result. 5. A tax data security graph neural network training method based on matrix decomposition according to claim 3, characterized in that, in Step 2 of step 1), for a two-layer GCN, the node embedding is affected by its neighbors within the two-hop range. Influence, the neighbors within two ranges are represented by the square of the adjacency matrix A ^2. A ² can effectively indicate the connection relationship of the graph and the information transfer between nodes; record the number of nodes n, decompose the original adjacency matrix into k matrices, and after decomposition Each matrix of contains eigenvalues, missing eigenvalues, on the premise of obtaining all eigenvalues, the probability of correct arrangement/> 6. A tax data security graph neural network training method based on matrix decomposition according to claim 5, characterized in that when n=100, k=2, p≈3.3×10 ^-6 . 7. A tax data security graph neural network training method based on matrix decomposition according to claim 3, characterized in that in step 2), the feature matrix protection based on differential privacy includes: Step1: Privacy budget and global sensitivity calculation Apply the Laplacian mechanism to perform differential privacy protection on the feature matrix X of the tax data graph, and calculate the global sensitivity Δ _f according to the set privacy budget ∈: Δ _f =max _D,D′ {|h＝h′|} Among them, D and D′ are a pair of adjacent data, h and h are the results of random queries for D and D′ respectively; let Set the Laplacian noise distribution to be added as follows: Step2: Noise injection Insert the Laplacian noise generated in the previous step into the feature matrix X of the tax data graph to obtain the privacy-preserving feature matrix X′. 8. A tax data security graph neural network training method based on matrix decomposition according to claim 7, characterized in that, in Step 2 of step 2), the Laplacian mechanism satisfies ∈-differential privacy, that is: Pr[M(D)＝y]≤e ^∈ Pr[M(D′)＝y] Where M is the processing mechanism applied. 9. A tax data security graph neural network training method based on matrix decomposition according to claim 7, characterized in that in step 3), the parameter server-based model training and integration includes: Step1: Data distribution The adjacency matrix A of the tax data graph is decomposed into {A _k ...}, k=1,2,..., and the feature matrix X of the tax data graph is processed with differential privacy to obtain Each calculation side obtains A _k and X′ as inputs to the GNN model; Step2: Model training based on privacy-preserving data The graph convolutional neural network model is selected for training. Calculator k locally has a two-layer GCN model and is trained on the data assigned to itself; the first layer input is the node feature matrix X and the adjacency matrix A _k . After information transfer And the hidden feature matrix of the output node after aggregation: H _k,1 =f(A _k X'W _k,1 ) The input of the second layer is the output H _k,1 of the first layer and the adjacency matrix A _k , and the output node hidden feature matrix H _k,2 is used for node classification or other downstream tasks; H _k,2 =f(A _k H _k,1 W _k,2 ) After training, the calculating party uploads the model parameters W _k,1 and W _k,2 to the parameter server held by the data owner, and at the same time pulls the updated model parameters from the parameter server; the parameter server collects the model parameters uploaded by each participant Finally, the model parameters are integrated using the model averaging method in the distributed machine learning method to obtain new model parameters.