A graph vertical federated data enhancement method and device based on neural pathways

Abstract

The present invention discloses a graph vertical federation data enhancement method and device based on neural pathways. First, a graph vertical federation learning framework with a reward mechanism is constructed. During the training process, the local model neuron activation value is calculated, and the inverse neural pathway of the local model is obtained according to the gradient relationship between the local model neuron activation value and the input of each layer of the model. The graph topology structure that is beneficial to the performance of the central server model is found through the model neural pathway. Then, the local model neuron activation value is calculated, and the inverse neural pathway of the local model is obtained according to the gradient relationship between the local model neuron activation value and the input of each layer of the model. The node features that are beneficial to the performance of the central server model are found through the model neural pathway. Then, based on the found graph topology structure and node features, the local data is enhanced. Finally, the enhanced local data is further used for the training of the graph vertical federation learning model, thereby improving the contribution of local participants in the graph vertical federation model.

Description

Technical Field

[0001] The present invention belongs to the field of network security technology, and specifically relates to a graph vertical federated data enhancement method and device based on neural pathways. Background Art

[0002] Graph neural networks have achieved remarkable results in the field of graph representation learning. They can efficiently process and analyze irregular graph-structured data, greatly facilitating downstream applications such as node classification, graph classification, and link prediction. Graph convolutional models are a classic type of graph neural network. However, existing graph neural networks rely on complete node connectivity information and rich node feature information, which is often difficult to achieve in real life. Different institutions may possess different data. Due to privacy restrictions and competitive relationships, this data is often not directly shared, resulting in "data silos" and hindering the maximum effectiveness of graph neural networks. Graph vertical federated learning was proposed to address this data isolation issue. In graph vertical federated learning, different graph data owners can exchange intermediate results of model computations, jointly maintaining a reliable central model while ensuring data privacy and maintaining local data privacy. This improves the quality of local models and achieves better results than when only local data is available.

[0003] However, in some federated learning frameworks, the central server calculates each participant's contribution based on the quality of their uploaded data and allocates different rewards to each participant based on this contribution metric. This encourages participants to provide higher-quality data, which in turn improves the central server's model's operational capabilities and generates greater revenue. Therefore, improving the contribution of participants in graph vertical federation scenarios through data augmentation, thereby achieving better model prediction capabilities and more rewards, is of great practical significance. Summary of the Invention

[0004] In the scenario where a graph vertical federation model is jointly trained by multiple parties, there may be cases where the data quality of some participants is poor. In order to improve the contribution of these participants in the graph vertical federation model, improve the business capabilities of the central server model, and obtain more benefits, the present invention provides a graph vertical federation data enhancement method based on a neural pathway. The method reversely searches for the graph structure and node features that are most conducive to the performance of the server model through the neural pathway of the local model, and performs data enhancement, thereby improving the contribution of the participant in the graph vertical federation model.

[0005] To achieve the above invention, the technical solution provided by the present invention is:

[0006] A first aspect of an embodiment of the present invention provides a graph vertical federated data enhancement method based on a neural pathway, comprising the following steps:

[0007] Step S1: define a reward mechanism based on a graph vertical federated learning framework composed of several local model participants and a central server;

[0008] Step S2: Calculate the activation value of the local model neurons and obtain the reverse neural pathway of the local model based on the gradient relationship between the activation value and the input of each layer of the local model, and find the graph topology through the reverse neural pathway;

[0009] Step S3, searching for node features of the target modified node that are beneficial to the performance of the central server model based on the reverse neural pathway obtained in step S2;

[0010] Step S4: Based on the graph topology and node features obtained in steps S2 and S3, data enhancement is performed on the local data.

[0011] Furthermore, the local model is specifically a graph neural network, and the local node embedding vector E output by the local model is:

[0012]

[0013]

[0014] Where ρ(·) represents the ReLU activation function, X is the node feature matrix, W1 and W2 are the model parameters of the first and second layers respectively, and A is the adjacency matrix corresponding to the graph structure; I N is an N×N diagonal matrix, where N is the number of nodes;

[0015] is the degree matrix of the node.

[0016] Furthermore, the central server model is specifically as follows: the local model uploads the local node embedding vector E to the central server model, and the central server model performs splicing processing on the uploaded local node embedding vectors to generate a global node embedding vector E global :

[0017] E global =E1||...||E m

[0018] Among them, || is the concatenation operator, and m is the number of participants;

[0019] Using the global node embedding vector E global Training center server model;

[0020] The training objectives of the central server are:

[0021]

[0022] Among them, {V L} is the set of labeled nodes, |P| is the number of categories of nodes in the graph. Y is the true label of the node. The central server model updates the central server model parameters [θ1...θ l ].

[0023] Furthermore, the reward mechanism is specifically as follows: the central server distributes rewards according to the contribution of each local model, and the contribution of the i-th participant C i The calculation formula is:

[0024]

[0025] Among them, acc i represents the prediction accuracy of the server when only the data of the i-th participant is available, η is the contribution scaling constant, and m is the number of participants;

[0026] The calculation formula for reward distribution based on the contribution of the local model is:

[0027]

[0028] Among them, R total is the total reward value, and m is the number of participants.

[0029] Furthermore, the step S2 is specifically as follows: training the central server model, selecting a local model as the data enhancement party, taking the neuron corresponding to the maximum value of the uploaded local node embedding vector as the starting neuron of the reverse neural pathway, and calculating the gradient of the neuron activation value M with respect to the previous layer input to obtain the neuron gradient matrix g neural , then the gradient matrix of the l-1 layer is:

[0030]

[0031] Among them, M l Activate the maximum value of the l-th layer neuron:

[0032]

[0033] Among them, h l Represents the activation value of the neurons in the first layer of the model. Obtain the important neurons in the previous layer until the input layer neurons are obtained, forming a reverse neural pathway;

[0034] All samples are tested as above to obtain the important input layer neurons of each sample, and the important input neurons of the same type are statistically obtained. The k nodes with the smallest corresponding neuron activation values are selected as the target modified node set {v s}, take the input layer neuron activation value M in The largest node is used as the target to generate node v t ;

[0035] Derivative the adjacency matrix of the input neuron activation value corresponding to the target generation node to obtain the edge gradient matrix g edge :

[0036]

[0037] Select the target generation node v t In g edge The set of K edges with the largest gradient {e d} and the corresponding node set {v d}, (·) T Represents the matrix transpose operation to obtain the graph topology.

[0038] Furthermore, the step S3 is specifically as follows:

[0039] According to the reverse neural pathway obtained in step S2, the target modified node set {v s}、Target generation node v t And the corresponding neuron activation value M in , derive the input node feature matrix of the neuron activation value corresponding to the target generation node to obtain the feature gradient matrix G attri :

[0040]

[0041] Select the feature set with the largest gradient on the target generation node {F d}Node features that need to be added as the target modified node.

[0042] Furthermore, the step S4 is specifically as follows:

[0043] Connect the nodes corresponding to the graph topology to the target modification nodes:

[0044] {v nei}={v ori}∪{v d}

[0045] {e nei}={e ori}∪{e d}

[0046] Among them, {vnei} is the neighbor node set of the target modified node, {v ori} is the original set of neighbor nodes, {e nei} is the edge set of the target modified node, {e ori} is the original edge set;

[0047] Add the node features obtained in step S3 to the features of the target modified node to complete the data enhancement of the local data. The formula is as follows:

[0048] {F}={F ori}∪{F d}

[0049] Among them, {F} is the node feature set of the target modified node after data enhancement, {F ori} is the original node feature set, {F d} Modify the node for the target.

[0050] Furthermore, the enhanced local data is used to train the graph vertical federated learning framework.

[0051] A second aspect of an embodiment of the present invention provides a graph vertical federated data enhancement based on a neural pathway, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned graph vertical federated data enhancement method based on a neural pathway.

[0052] A third aspect of an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the above-mentioned neural pathway-based graph vertical federated data enhancement method is implemented.

[0053] Compared with the prior art, the beneficial effects of the present invention include at least the following: the neural pathway-based graph vertical federated data enhancement method provided by the present invention obtains the reverse neural pathway of the local model by calculating the neuron activation value and according to its gradient relationship with the input of each layer of the model; finds the graph topology structure and node features that are most conducive to the performance of the central server model through the found neural pathway; performs data enhancement on the local data based on the found graph topology structure and node features; finally, continues to use the enhanced local data for the training of the graph vertical federated learning model, thereby improving the contribution of local participants in the graph vertical federated model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of the overall framework of the graph vertical federated data enhancement method based on neural pathways;

[0055] Figure 2Schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0056] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0057] The terms used in this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The singular forms "a," "the," and "the" used in this invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0058] It should be understood that although the terms "first," "second," "third," etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if" as used herein may be interpreted as "when," "when," or "in response to determining."

[0059] The following describes in detail the neural pathway-based graph vertical federated data enhancement of the present invention in conjunction with the accompanying drawings. The features of the following embodiments and implementations may be combined with each other unless they conflict.

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is described in further detail below with reference to the accompanying drawings.

[0061] Reference Figure 1 , an embodiment of the present invention proposes a graph vertical federated data enhancement method based on neural pathways, comprising the following steps:

[0062] 1) Build a graph-based federated learning model with an incentive mechanism;

[0063] like Figure 1 As shown in the figure, the graph vertical federated learning framework is jointly constructed by each training participant and a central server model.

[0064] It includes the following sub-steps:

[0065] 1.1) Build a local model based on graph neural network;

[0066] The training participant's local end uses a graph neural network as a local model. This embodiment of the present invention takes one of the participants and a two-layer graph convolutional network model as an example. The local node embedding vector E output by its local model is:

[0067]

[0068] Where ρ(·) represents the ReLU activation function, X is the node feature matrix, W1 and W2 are the parameters of the first and second layers of the graph convolution model respectively, and A is the adjacency matrix corresponding to the graph structure. N is an N×N diagonal matrix (N is the number of nodes in the graph), representing the self-connection of the nodes. is the degree matrix of the nodes in the graph.

[0069] During training, each participant uploads the local node embedding vector E to the central server model for training. The central server model then uses the gradient information fed back by the central server model to update the local graph convolutional model parameters W1 and W2. The remaining participants build and train their local models in the same manner.

[0070] 1.2) Build a central server model;

[0071] The central server model concatenates the local node embedding vectors uploaded by each training participant to generate a global node embedding vector E global :

[0072] E global =E1||...||E m (2)

[0073] Among them, || is the concatenation operator and m is the number of participants.

[0074] According to the global node embedding vector E global Train the central server model. Here, the central server model takes a multi-layer perceptron with l layers as an example and outputs the confidence Y':

[0075] Y'=f(θ l ·ρ(...ρ(θ1·E global ))) (3)

[0076] Among them, θ l is the model parameter of the lth layer, f(·) is the softmax function, and ρ(·) represents the ReLU activation function.

[0077] The training objectives of the central server are:

[0078]

[0079] Among them, {V L} is the set of labeled nodes, |P| is the number of categories of nodes in the graph. Y is the true label of the node. The central server model updates the central server model parameters [θ1...θ l ].

[0080] 1.3) Establish a reward mechanism based on contribution;

[0081] The central server feeds back gradient and confidence information to the local models of each training participant, and distributes rewards to the training participants based on their contribution. The contribution calculation formula for the i-th participant is:

[0082]

[0083] Among them, acc i represents the prediction accuracy of the server when only the data of the i-th participant is available, η is the contribution scaling constant, and m is the number of participants.

[0084] The calculation formula for reward distribution based on the contribution of training participants is:

[0085]

[0086] Among them, R total is the total reward value, and m is the number of participants.

[0087] 2) By calculating the local model neuron activation value and its gradient relationship with the input of each layer of the model, the reverse neural pathway of the local model is obtained. The graph topology that is most conducive to the performance of the central server model is found through the model neural pathway. Specifically:

[0088] First, after the central server model is trained, a local model is selected as the data augmentation party. The data augmentation party takes the neuron corresponding to the maximum value of the uploaded local node embedding vector as the starting neuron of the reverse neural pathway, and calculates the gradient of the neuron activation value M with respect to the previous layer input to obtain the neuron gradient matrix g. neural , then the gradient matrix of the l-1 layer is:

[0089]

[0090] Among them, M l Activate the maximum value of the l-th layer neuron:

[0091]

[0092] Among them, h lRepresents the activation value of the neurons in the first layer of the model. According to formulas (7) and (8), the important neurons in the previous layer are obtained until the input layer neurons are obtained, forming a reverse neural pathway.

[0093] All training samples are tested in the above steps to obtain the important input layer neurons of each sample, and statistically analyze the labels of the training nodes to obtain the important input neurons of the same type of nodes, and select the k nodes with the smallest corresponding neuron activation values as the target modified node set {v s}, take the input layer neuron activation value M in The largest node is used as the target to generate node v t .

[0094] Derivative the adjacency matrix of the input neuron activation value corresponding to the target generation node to obtain the edge gradient matrix g edge :

[0095]

[0096] Select the target generation node v t In g edge The set of K edges with the largest gradient {e d} and the corresponding node set {v d}, (·) T Represents the matrix transpose operation, resulting in a graph topology that is beneficial to the performance of the central server model.

[0097] 3) Based on the reverse neural pathway obtained in step 2), searching for node features of the target modification node that are most conducive to the performance of the training center server model through the model neural pathway;

[0098] According to the reverse neural pathway obtained in step 2), the target modified node set {v s}、Target generation node v t And the corresponding neuron activation value M in . Derivative the input node feature matrix of the neuron activation value corresponding to the target generation node to obtain the feature gradient matrix G attri :

[0099]

[0100] Select the feature gradient matrix G attri The feature set with the largest gradient {F d}Node features that need to be added as target modification nodes, the node features of the target modification nodes are beneficial to improving the performance of the central server model.

[0101] 4) Based on the graph topology and node features found in steps 2) and 3), perform data enhancement on the local data;

[0102] According to step 2) and step 3), a graph topology structure and node features of a target modification node that are beneficial to the performance of the central server model, as well as the target modification node, are obtained.

[0103] Connect the nodes corresponding to the graph topology to the target modification nodes:

[0104] {v nei}={v ori}∪{v d} (11)

[0105] {e nei}={e ori}∪{e d} (12)

[0106] Among them, {v nei} is the neighbor node set of the target modified node, {v ori} is the original set of neighbor nodes, {e nei} is the edge set of the target modified node, {e ori} is the original edge set.

[0107] Add the node features obtained in step 3) to the features of the target modified node:

[0108] {F}={F ori}∪{F d} (13)

[0109] Among them, {F} is the node feature set of the target modified node after data enhancement, {F ori} is the original node feature set, which ultimately completes the data enhancement of local data.

[0110] 5) The enhanced local data is further used for training the graph vertical federated learning framework;

[0111] After step 4), the local graph structure data and node feature data after data enhancement are obtained. In the subsequent training, the enhanced data is used to continue training the graph vertical federation model to improve the contribution of the data enhancement participants in the graph vertical federation model.

[0112] Corresponding to the aforementioned embodiment of the graph vertical federated data enhancement method based on neural pathways, the present invention also provides an embodiment of a graph vertical federated data enhancement device based on neural pathways.

[0113] See also Figure 2 , an embodiment of the present invention provides a graph vertical federated data enhancement device based on a neural pathway, comprising one or more processors for implementing the graph vertical federated data enhancement method based on a neural pathway in the above embodiment.

[0114] The embodiment of the graph vertical federation data enhancement device based on neural pathways of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 2 As shown, it is a hardware structure diagram of any device with data processing capability where the vertical federated data enhancement device based on neural pathways of the present invention is located, except Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0115] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0116] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0117] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the graph vertical federated data enhancement method based on neural pathways in the above embodiment is implemented.

[0118] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium may also be any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0119] In summary, the neural pathway-based graph vertical federated data enhancement method provided by the present invention obtains the reverse neural pathway of the local model by calculating the neuron activation value and according to its gradient relationship with the input of each layer of the model; finds the graph topology structure and node features that are most conducive to the performance of the central server model through the found neural pathway; performs data enhancement on the local data based on the found graph topology structure and node features; finally, the enhanced local data is continued to be used for the training of the graph vertical federated learning model, thereby improving the contribution of local participants in the graph vertical federated model.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A graph vertical federated data enhancement method based on neural pathways, characterized by: The following steps are involved: Step S1: define a reward mechanism based on a graph vertical federated learning framework composed of several local model participants and a central server; Step S2: Calculate the activation value of the local model neurons and obtain the reverse neural pathway of the local model based on the gradient relationship between the activation value and the input of each layer of the local model, and find the graph topology through the reverse neural pathway; Step S3, searching for node features of the target modified node that are beneficial to the performance of the central server model based on the reverse neural pathway obtained in step S2; Step S4: Based on the graph topology and node features obtained in steps S2 and S3, data enhancement is performed on the local data.

2. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The local model is specifically a graph neural network, and the local node embedding vector E output by the local model is: Where ρ(·) represents the ReLU activation function, X is the node feature matrix, W1 and W2 are the model parameters of the first and second layers respectively, and A is the adjacency matrix corresponding to the graph structure; I N is an N×N diagonal matrix, where N is the number of nodes; is the degree matrix of the node.

3. The neural pathway-based graph vertical federated data enhancement method according to claim 2, characterized in that: The central server model is specifically as follows: the local model uploads the local node embedding vector E to the central server model, and the central server model performs splicing processing on the uploaded local node embedding vectors to generate a global node embedding vector E global : HAVE BEEN global =E1||...||E m Among them, || is the concatenation operator, and m is the number of participants; Using the global node embedding vector E global Training center server model; The training objectives of the central server are: Among them, {V L } is a set of labeled nodes, |P| is the number of categories of nodes in the graph; Y is the true label of the node; the central server model updates the central server model parameters [θ1...θ l ].

4. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The reward mechanism is as follows: the central server distributes rewards according to the contribution of each local model, and the contribution of the i-th participant C i The calculation formula is: Among them, acc i represents the prediction accuracy of the server when only the data of the i-th participant is available, η is the contribution scaling constant, and m is the number of participants; The calculation formula for reward distribution based on the contribution of the local model is: Among them, R total is the total reward value, and m is the number of participants.

5. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The step S2 is specifically as follows: training the central server model, selecting a local model as the data enhancement party, taking the neuron corresponding to the maximum value of the uploaded local node embedding vector as the starting neuron of the reverse neural pathway, and calculating the gradient of the neuron activation value M with respect to the previous layer input to obtain the neuron gradient matrix g neural , then the gradient matrix of the l-1 layer is: Among them, M l Activate the maximum value of the l-th layer neuron: Among them, h l Represents the activation value of the neurons in the first layer of the model; obtains the important neurons in the previous layer until the input layer neurons are obtained, forming a reverse neural pathway; Traverse all samples, obtain the important input layer neurons of each sample, count the important input neurons in the same node, and take the k nodes with the smallest corresponding neuron activation value as the target modified node set {v s }, take the input layer neuron activation value M in The largest node is used as the target to generate node v t ; Derivative the adjacency matrix of the input neuron activation value corresponding to the target generation node to obtain the edge gradient matrix g edge : Select the target generation node v t In g edge The set of K edges with the largest gradient {e d } and the corresponding node set {v d }, (·) T Represents the matrix transpose operation to obtain the graph topology.

6. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The step S3 is specifically as follows: According to the reverse neural pathway obtained in step S2, the target modified node set {v s }、Target generation node v t And the corresponding neuron activation value M in , derive the input node feature matrix of the neuron activation value corresponding to the target generation node to obtain the feature gradient matrix G attri : Select the feature set with the largest gradient on the target generation node {F d }Node features that need to be added as the target modified node.

7. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The step S4 is specifically as follows: Connect the nodes corresponding to the graph topology to the target modification nodes: {v nei }={v ori }∪{v d } {And nei }={and ori }∪{e d } Among them, {v nei } is the neighbor node set of the target modified node, {v ori } is the original set of neighbor nodes, {e nei } is the edge set of the target modified node, {e ori } is the original edge set; Add the node features obtained in step S3 to the features of the target modified node to complete the data enhancement of the local data. The formula is as follows: {F}={F ori }∪{F d } Among them, {F} is the node feature set of the target modified node after data enhancement, {F ori } is the original node feature set, {F d } Modify the node for the target.

8. The neural pathway-based graph vertical federated data enhancement method according to claim 1, characterized in that: The augmented local data is used to train the graph vertical federated learning framework.

9. A graph vertical federated data augmentation system based on neural pathways, comprising a memory and a processor, wherein: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the neural pathway-based graph vertical federated data enhancement method described in any one of claims 1-8 above.

10. A computer-readable storage medium having a computer program stored thereon, wherein: When the program is executed by a processor, the graph vertical federated data enhancement method based on neural pathways is implemented.

Images