Construction method of multi-party graph trusted learning architecture and application of construction method in propagation prediction

Abstract

The invention discloses a multi-party graph credible learning architecture construction method and application thereof in propagation prediction, and provides a multi-party graph credible computing method by constructing a new multi-layer graph structure, so that multi-party graph learning can be safely and effectively performed, and the multi-party graph credible learning architecture can be applied to propagation prediction. And moreover, the method shows outstanding effects in downstream tasks of various graph neural networks such as disease prediction and network information spreading, and lays a solid foundation for subsequent multi-party graph credible learning research.

Description

Technical Field:

[0001] The present invention relates to the technical fields of machine learning and information networks, and particularly relates to a graph machine learning method based on multi-party secure computing, which can ensure cooperative computing among multiple graph data owners under the conditions of ensuring data security and privacy, and can be applied to fields such as disease spread prediction and network information spread prediction. Background Art:

[0002] In recent years, the research on graph representation learning has attracted great attention and had a profound impact on daily life and society. By analyzing the complex relationships and structures embedded in networks, graph representation learning can drive social progress and enhance our understanding of multiple fields, including healthcare and disease control, environmental protection, urban planning, etc. However, traditional graph research models have some limitations, such as the inability to capture complex relationships, scalability challenges, limited flexibility, sensitivity to noise and incomplete data, and lack of interpretability. These constraints may hinder their effectiveness in solving the inherent complexity in graph-structured data.

[0003] The widespread application of graph neural networks (GNNs) has revealed valuable insights and significantly expanded our understanding of various fields. Notably, GNNs are a powerful application tool with obvious advantages compared to traditional machine learning methods. However, due to privacy restrictions, regulatory limitations, and research or commercial competition, GNN models only partially solve large-scale graph tasks.

[0004] Federated learning (FL) has become a new type of machine learning method that enables customers to collaboratively train a shared model under the guidance of a central server while solving the problem of local data confidentiality. By maintaining locally fixed data and globally shared parameters, FL has been identified as a promising solution for training GNNs on isolated graph data. Therefore, federated graph learning (FGL) has received increasing attention in solving the above challenges related to GNNs. However, FGL requires the model parameters or gradients of each distributed institution to be uploaded to the central server. This framework means that distributed institutions must share the same structure of the GNN model, thus reducing fault tolerance. Therefore, the owner of the central server has greater power. In addition, exchanging gradients inadvertently leaks privacy information, thus increasing the risk of data leakage.

[0005] Meanwhile, Swarm Learning (SL) combines edge computing with a blockchain-based peer-to-peer network to achieve coordination without a central coordinator. SL promotes parameter sharing through the Swarm network and independently constructs private data models at each site, thereby providing strong security measures to ensure data privacy and confidentiality. However, SL also has its limitations. For example, all nodes in SL must use the same machine learning model, and SL is not applicable to non-I.I.D (independent and identically distributed) data, such as graph data with complex structures. Therefore, neither FGL nor SL can create flexible and diverse models for each participating institution with different network structures. A key issue that emerges is how to fairly coordinate the network structures and data of different organizations without compromising privacy and the need for a central database.

[0006] The collection of multi-institutional graph data is a complex process, but due to privacy concerns, laws and regulations, and business restrictions, the data cannot be fully shared. The present invention hopes that multiple data owners (institutions) can achieve data sharing, interoperability, computing, and modeling without exposing the data itself, ultimately generating value beyond the data itself while ensuring data security and privacy, that is, not leaking to other participating parties.

[0007] Epidemic spread prediction is an important research direction in fields such as epidemiology and social communication. Using methods such as mathematical models, network analysis, and machine learning to study the spread patterns and trends of information or diseases among the population helps predict the development trends of epidemics, public opinions, etc., formulate effective prevention and control strategies, and optimize the allocation of public health resources. These prediction models and methods play an important role in epidemic monitoring, vaccination strategies, public opinion monitoring, etc., providing support for social health and public safety. However, due to the significant impact of local disease control data in each region on the local geographical location and the need for privacy protection of disease control data in each region, how to perform multi-party secure epidemic spread prediction safely and effectively is an issue to be solved. How to better utilize existing technologies to construct a secure and efficient multi-party secure computing model, thereby being able to reasonably and effectively predict the spread of epidemics and being applicable to various downstream tasks, can better serve the field of graph neural network research. Summary of the Invention:

[0008] The technical solution measures of the present invention are as follows: The present invention provides a novel multi-party graph trusted learning architecture. The method includes:

[0009] 1. Obtain the topological structure and propagation data information of the network. The graph data owned by each institution is called a "local graph", and each institution deploys its own GNN model locally;

[0010] 2. Abstract each organization into a separate node, and construct a "global graph" based on whether there is a link between the nodes corresponding to different organizations in the local graph; specifically, if there is an edge between the local graphs corresponding to two organizations, then there is an edge between the nodes corresponding to the two organizations in the global graph.

[0011] 3. The nodes abstracted from the subgraph of each institution are connected to each node in the local graph. The resulting graph is called an "integrated graph". The integrated graph integrates global and local information, so it can better capture the inherent correlation of the data while ensuring the security and privacy of each institution's data.

[0012] 4. Based on three types of graphs, a multi-party graph trusted learning architecture is established;

[0013] 5. For each organization, a number of neighboring nodes (organizations) are randomly selected as homomorphic encryption computing nodes. These homomorphic encryption computing nodes will embed the nodes after homomorphic computing and send them back to the target organization that initiated the computing service request;

[0014] 6. The target organization fuses the embedding learned by the local model with the acquired neighbor embedding;

[0015] 7. The target organization uses the fused embedding to perform downstream tasks and update model parameters;

[0016] The present invention proposes a multi-party propagation prediction model based on the multi-party graph trusted learning architecture.

[0017] 1. For each institution, the communication data of each region within the institution (such as daily epidemic data and public opinion communication data in social networks) is used as the time series features of each node in the local graph to build a local communication data graph network.

[0018] 2. Use multi-scale convolutions to extract the temporal features of local data and transmit the learned embedding to the homomorphic encryption computing node;

[0019] 3. Obtain the embedded information of the neighboring organization after homomorphic calculation from the homomorphic encryption computing node;

[0020] 4. Fuse the temporal features of local data with the embedded information of neighbors to obtain an integrated graph;

[0021] 5. Use graph neural network (GNN) on the integrated graph to learn cross-regional spatial features;

[0022] 6. Use the decoder to decode the learned embedding, calculate the loss function and update the parameters;

[0023] 7. Obtain the final propagation dynamics prediction results.

[0024] The algorithm model proposed by the present invention provides services based on a multi-party graph trusted learning system platform, which is successively processed and calculated at three system levels:

[0025] 1. Architecture layer: This layer provides basic library support for the system. First, the data support of the entire system is carried out through the function libraries of the basic architecture layer. Specifically, it includes basic scientific computing libraries such as numpy, scipy, and torch; security modules include RSA encryption library, zerorpc, and PHE. At the same time, the architecture layer will package the required modules into a Docker virtual environment, and deploy the service through the image file to be relatively separated from the physical machine.

[0026] 2. Computation layer: The computation layer is the core functional layer of multi-party graph trusted learning, which mainly includes: obtaining local embeddings and embeddings of neighbor institutions. The specific steps for obtaining local embeddings are as follows:

[0027] (1) The local institution initializes local parameters and models;

[0028] (2) Train the local embedding model;

[0029] (3) Learn to obtain local embeddings;

[0030] (4) Encrypt the local embeddings through the method of homomorphic encryption;

[0031] (5) Transmit the ciphertext after homomorphic encryption to the homomorphic encryption computing node, and the homomorphic encryption computing node will send the embeddings obtained through homomorphic calculation to the target institution that initiates the service request to support its operation of the integrated model;

[0032] On the basis of obtaining local embeddings in step (3), the computation layer can also initiate a service request to obtain the embeddings of neighbor institutions. By fusing local embeddings and embeddings of neighbor institutions, the internal correlation of data can be better captured. The specific steps for obtaining the embeddings of neighbor institutions from the homomorphic encryption computing node are as follows:

[0033] (A) Use a decentralized P2P network to achieve communication security;

[0034] (B) Randomly select several (greater than 0 and less than the total number of neighbor institutions) neighbor institutions as homomorphic encryption computing nodes according to the topological structure of the global graph;

[0035] Remote Procedure Call (RPC) refers to the mutual communication between processes running on different machines. A process running on a certain machine can call the service on a remote machine as if it were accessing a local service without knowing the underlying communication details. The remote procedure call RPC technology simplifies the construction and use of the framework of the present invention.

[0036] (D) Secure computing process: A variety of encryption technologies are used in this process. Serialization and RSA encryption technologies together ensure the security of data communication. Homomorphic encryption technology directly encrypts the original text and then performs various operations on the ciphertext, and finally obtains the ciphertext of the result, which can ensure the privacy security during the calculation process, that is, the plaintext will not be leaked.

[0037] (E) Finally, obtain the embeddings of neighbor institutions from the homomorphic encryption computing nodes.

[0038] By fusing the local embeddings and the embeddings of neighbor institutions, the internal correlations of the data can be better captured, providing support for performing various downstream tasks at the application layer.

[0039] 3. Application layer: Apply the embedding results of the computational fusion to downstream tasks. For example, it can be applied to fields such as disease spread prediction, network information spread prediction, node classification, link prediction, etc., while ensuring the customization of downstream models.

[0040] The beneficial effects of the present invention are as follows: By constructing a new multi-layer graph structure and proposing a multi-party graph trusted computing method, multi-party graph learning can be carried out safely and effectively, and outstanding effects are shown in various downstream tasks of graph neural networks such as disease prediction and network information dissemination, laying a solid foundation for subsequent research on multi-party graph trusted learning. Description of the drawings:

[0041] Figure 1 It is a schematic diagram of the multi-party graph trusted learning architecture in the present invention;

[0042] Figure 2 It is a schematic diagram of the multi-party propagation prediction model in the present invention;

[0043] Figure 3 It is a schematic diagram of the multi-party graph trusted learning system platform architecture in the present invention;

[0044] Figure 4 It is a schematic diagram of the prediction result of the infectious disease transmission dynamics in Embodiment 1 of the present invention;

[0045] Figure 5 It is a diagram of the algorithm implementation process in Embodiment 3 of the present invention. Detailed implementation manners:

[0046] Further illustrate the construction method and application of a multi-party graph trusted learning architecture of the present invention in combination with the drawings.

[0047] The present invention provides a novel multi-party graph trusted learning architecture. The method includes:

[0048] 1. Obtain the topological structure and propagation data information of the network. The graph (network) data owned by each institution is called the "local graph", and each institution deploys its own GNN model locally;

[0049] 2. Abstract each institution into a separate node, and construct a "global graph" based on whether there are links between the nodes corresponding to different institutions in the local graphs. Specifically, if there is an edge between the local graphs corresponding to two institutions, then there is an edge between the nodes corresponding to these two institutions in the global graph.

[0050] 3. Connect the nodes abstracted from the subgraphs of each institution to each node in the local graph. The formed graph is called the "integrated graph". The integrated graph integrates global and local information, so it can better capture the internal correlations of the data;

[0051] 4. Based on the three graphs, establish a multi-party graph trusted learning architecture;

[0052] 5. For each institution, randomly select a certain number of neighbor nodes (institutions) as homomorphic encryption computing nodes. These homomorphic encryption computing nodes will send back the node embeddings after homomorphic calculation to the target institution that sent the request;

[0053] 6. The target institution fuses the embeddings learned by the local model and the obtained neighbor embeddings;

[0054] 7. Use the fused embeddings to perform downstream tasks and update the model parameters;

[0055] The present invention proposes a multi-party propagation prediction model based on this multi-party graph trusted learning architecture.

[0056] 1. For each institution, use the propagation data of each region within the institution (such as daily epidemic data, public opinion propagation data in the social network) as the temporal features of each node in the local graph, and construct a local propagation data graph network.

[0057] 2. Use a multi-scale convolution network (Multi-Scale Convolutions) to extract the temporal features of the local data, and transmit the learned embeddings to the homomorphic encryption computing nodes;

[0058] 3. Obtain the embedding information of neighbor institutions after homomorphic calculation from the homomorphic encryption computing nodes;

[0059] 4. Fuse the temporal features of the local data and the embedding information of the neighbors to obtain an integrated graph;

[0060] 5. Use a graph neural network (GNN) on the integrated graph to learn cross-regional spatial features;

[0061] 6. Use the decoder to decode the learned embedding, calculate the loss function and update the parameters;

[0062] 7. Obtain the final propagation dynamics prediction results.

[0063] The algorithm model proposed in this invention provides services based on the multi-party graph trusted learning system platform, which is divided into three system levels for processing and calculation:

[0064] 1. Architecture layer: This layer provides basic library support for the system. First, the function library of the infrastructure layer is used to support the data of the entire system. Specifically, it includes basic scientific computing libraries such as numpy, scipy, and torch; security modules include RSA encryption library, zerorpc, and PHE. At the same time, the architecture layer will package the required modules into a Docker virtual environment, and deploy services through image files to separate them from physical machines.

[0065] 2. Computation layer: The computation layer is the core functional layer of multi-party graph trust learning, which mainly includes: obtaining local embedding and neighboring organization embedding. The specific steps of obtaining local embedding are as follows:

[0066] (1) The local organization initializes local parameters and models;

[0067] (2) Train the local embedding model;

[0068] (3) Learn to obtain local embedding;

[0069] (4) Encrypt the local embedding using homomorphic encryption;

[0070] (5) The homomorphically encrypted ciphertext is transmitted to the homomorphically encrypted computing node, and the homomorphically encrypted computing node sends the embedding obtained by the homomorphic computing to the target organization that initiated the service request to support it in performing the operation of the integrated model;

[0071] Based on the local embedding obtained in step (3), the computation layer can also be used to obtain the embedding of neighboring organizations after homomorphic computation. By fusing the local embedding and the embedding of neighboring organizations, the intrinsic correlation of the data can be better captured. The specific steps for obtaining the embedding of neighboring organizations are as follows:

[0072] (A) Using decentralized P2P networks to achieve secure communication;

[0073] (B) According to the topological structure of the global graph, a number of neighboring organizations (greater than 0 and less than the total number of neighboring organizations) are randomly selected as homomorphic encryption computing nodes;

[0074] (C) Remote Procedure Call (RPC) refers to the communication between processes running on different machines. A process running on one machine can call a service on a remote machine without knowing the underlying communication details, just as if it were accessing a local service. The Remote Procedure Call (RPC) technology simplifies the construction and use of the framework of the present invention.

[0075] (D) Secure computing process: A variety of encryption technologies are used in this process. Serialization and RSA encryption technologies jointly ensure data communication security. Homomorphic encryption technology directly encrypts the original text and then performs various operations on the ciphertext, and finally obtains the ciphertext of the result, which can ensure privacy security during the calculation process, that is, the plaintext will not be leaked.

[0076] (E) Finally, the embedding obtained by homomorphic calculation of the neighbor institution is obtained from the homomorphic encryption calculation node.

[0077] By fusing the local embedding and the embedding of the neighbor institution, the internal correlation of data across regions can be better captured, providing support for various downstream tasks at the application layer.

[0078] 3. Application layer: Apply the embedding result of the computational fusion to downstream tasks. For example, it can be applied to fields such as disease spread prediction, network information spread prediction, node classification, and link prediction, ensuring the customization of downstream models.

[0079] Example 1

[0080] The present invention is tested in the prediction of epidemic dynamics and traditional graph learning tasks. The method proposed by the present invention is named CNL, and the test results are as Figure 4 . Figure 4 It represents the prediction of the number of influenza infections in the United States. We divide it into 4 institutions according to geographical location, namely the Midwest institution, the West institution, the South institution, and the Northeast institution. The Pearson correlation coefficient (PCC) is used as the evaluation index, which represents the correlation between the predicted value and the true value, and the larger the value, the more accurate. Among them Figure 4 (a) represents the result of disease prediction using only local data, while Figure 4 (b) shows the result obtained after integrating the data of different institutions using the integrated graph of the CNL framework. (b) shows that the PCC on the four institutions has improved, and it can be clearly seen that the integrated graph model of CNL is effective. It can be seen from the figure that the result of link prediction using the integrated graph is better than the result of local prediction. These results further prove the effectiveness of the framework.

[0081] Example 2

[0082] The present invention supports the deployment of multiple model structures, and the test results are shown in Table 1. The Pearson correlation coefficient PCC and the root mean square error RMSE are used as evaluation indicators. The larger the PCC, the better the performance, and the smaller the RMSE value, the better the performance. Different models can be deployed under the CNL framework. The results show that the performance of the integrated GNN model is better than that of the local GNN model, which proves the scalability and generality of the CNL framework.

[0083] Table 1 Propagation dynamics prediction performance of different models under the CNL framework (the plum blossom symbol represents the local GNN model, the triangle symbol ▲ represents the integrated GNN model, and the diamond symbol ◆ represents the global GNN model)

[0084]

[0085] Example 3

[0086] As Figure 5 shown, in this example, an epidemic propagation dynamics prediction task is carried out on a network consisting of 6 machines, specifically as follows:

[0087] S1: Obtain the topological structure and propagation data information of the network. The graph data owned by each institution is called "local graph";

[0088] S2: Abstract each institution into a separate node, and construct a "global graph" for the nodes corresponding to different institutions according to whether there are links between the local graphs; specifically, if there is an edge between the local graphs corresponding to two institutions, there is an edge between the nodes corresponding to these two institutions in the global graph.

[0089] S3: Integrate the node representing the institution in the global graph with the local graph of the institution to form a new graph structure - the integrated graph. The integrated graph integrates global and local information, so it can better capture the internal correlation of the data;

[0090] S4: Use the connections between institutions in the global graph for the trusted computing process. Among them, the gray institution is an arbitrarily selected homomorphic encryption computing institution, the black represents the target institution that initiates the computing service request, and the white represents the institution that only plays the role of information transmission. Integrate the epidemic population change data of each node in the target institution subgraph as the feature of the institution. The gray institution integrates and calculates the vectors transmitted by each institution and returns the result to the target institution; for example, in the figure, institution A is the target institution, and institutions C and D are selected as homomorphic encryption computing institutions. The neighbor institutions of A will randomly send their embeddings to C or D, and finally the computing institution will send the homomorphic calculation result to A;

[0091] S5: The target institution obtains the information of the neighbor institutions and updates the global graph neural network model;

[0092] S6: Update the integrated graph neural network model deployed on the target institution;

[0093] S7: Use the decoder to decode the learned embedding to obtain the predicted propagation data; calculate the loss function and update the parameters. The decoder consists of a linear layer and an activation layer;

[0094] S8: Obtain the final propagation prediction data.

[0095] The above are only the preferred embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

Claims

1. A method for constructing a multi-party graph trusted learning architecture, characterized in that It includes the following steps: Step S1: Obtain the topological structure and propagation data information of the network. The graph data owned by each institution is called the local graph, and each institution deploys its own graph neural network (GNN) model locally; Step S2: Abstract each institution into a separate node, and construct a global graph based on whether there are links between the nodes corresponding to different institutions according to their local graphs; Step S3: Connect the nodes abstracted from the subgraphs of each institution to each node in the local graph. The formed graph is called the integrated graph. The integrated graph integrates global and local information, so it can better capture the internal correlations of the data while ensuring the privacy and security of the data of each institution; Step S4: Establish a multi-party graph trusted learning architecture based on the three graphs; Step S5: For each institution, randomly select a certain number of neighbor nodes as homomorphic encryption calculation nodes, and these homomorphic encryption calculation nodes will send back the node embeddings after homomorphic calculation to the target institution that sends the request; Step S6: The target institution fuses the embeddings learned by the local model and the obtained neighbor embeddings; Step S7: The target institution uses the fused embeddings to perform downstream tasks and update the model parameters.

2. The construction method of a multi-party graph trusted learning architecture according to claim 1, wherein: The institution is the data owner.

3. The construction method of a multi-party graph trusted learning architecture according to claim 1, characterized in that: In Step 2, if there is an edge between the local graphs corresponding to two institutions, then there is an edge between the nodes corresponding to these two institutions in the global graph.

4. Application of a multi-party graph trusted learning architecture according to any one of claims 1-3 in propagation prediction, characterized in that: Based on this multi-party graph trusted learning architecture, a multi-party propagation prediction model is proposed.

5. The application of a multi-party graph trusted learning architecture according to claim 4 in propagation prediction, characterized in that: It includes the following steps: Step S1: For each institution, use the propagation data of each region within the institution as the temporal features of each node in the local graph to construct a local propagation data graph network; Step S2: Use a multi-scale convolutional network to extract the temporal features of the local data, and encrypt the learned embeddings and transmit them to the homomorphic encryption calculation nodes; Step S3: Obtain the embedding information of neighbor institutions after homomorphic calculation from the homomorphic encryption calculation nodes; Step S4: Fuse the temporal features of the local data and the embedding information of neighbor institutions to obtain an integrated graph; Step S5: Use a graph neural network (GNN) on the integrated graph to learn cross-regional spatial features; Step S6: Use a decoder to decode the learned embeddings, calculate the loss function, and update the parameters; Step S7: Obtain the final propagation dynamics prediction result.

6. Application of a multi-party graph trusted learning architecture in propagation prediction according to any one of claims 1-3, characterized in that: A system platform based on multi-party graph trusted learning includes three system levels for processing and computing in sequence: the architecture layer, the computing layer, and the application layer; The architecture layer provides basic library support for the system. First, it provides data support for the entire system through the function library of the basic architecture layer. At the same time, the architecture layer packages the required modules into a Docker virtual environment and deploys the service through the image file to be relatively separated from the physical machine; The computing layer is the core functional layer of multi-party graph trusted learning, which mainly includes: obtaining local embeddings and embeddings of neighbor institutions. By fusing local embeddings and embeddings of neighbor institutions, it can better capture the internal correlations of the data and provide support for the application layer to perform various downstream tasks; The application layer is used to apply the calculation and fusion embedding results to downstream tasks to ensure the customization of downstream models.

7. The application of a multi-party graph trusted learning architecture according to claim 6 in propagation prediction, characterized in that, The function libraries specifically include numpy, scipy and torch basic scientific computing libraries; the security modules include RSA encryption library, zerorpc and PHE.

8. The application of a multi-party graph trusted learning architecture according to claim 6 in propagation prediction, characterized in that, The specific steps to obtain local embedding are as follows: Step S1, the local mechanism initializes local parameters and models; Step S2: training a local embedding model; Step S3, learning to obtain local embedding; Step S4: Encrypt the local embedding by homomorphic encryption; Step S5: The homomorphically encrypted ciphertext is transmitted to the homomorphically encrypted computing node, and the homomorphically encrypted computing node sends the embedding obtained by the homomorphic computing to the target organization that initiated the service request to support the target organization in performing the operation of the integrated model; Based on the local embedding obtained in step S3, the computing layer can also be used to obtain the embedding of neighboring organizations from the homomorphic encryption computing node. By fusing the local embedding and the embedding information of the neighboring organizations, the intrinsic correlation of the data can be better captured.

9. The application of a multi-party graph trusted learning architecture in propagation prediction according to claim 6, wherein The specific steps to obtain the neighbor organization embedding from the homomorphic encryption computing node are as follows: Step S1, using a decentralized P2P network to achieve secure communication; Step S2: randomly select a number of neighboring organizations whose number is greater than 0 and less than the total number of neighboring organizations as homomorphic encryption computing nodes according to the topological structure of the global graph; Step S3, Remote Procedure Call, i.e., Remote Procedure Call (RPC), refers to the mutual communication between processes running on different machines. A process running on a certain machine can call a service on a remote machine just like accessing a local service without knowing the underlying communication details. The Remote Procedure Call (RPC) technology simplifies the construction and use of the framework of the present invention. Step S4, secure computing process: A variety of encryption technologies are used in this process. Serialization and RSA encryption technologies jointly ensure data communication security. Homomorphic encryption technology directly encrypts the original text, and then performs various operations on the ciphertext to finally obtain the resulting ciphertext, which can ensure privacy security during the computing process, that is, the plaintext will not be leaked; Step S5: Finally, the embedding of the neighboring organization is obtained from the homomorphic encryption computing node.

Images