Single-round federated graph learning method and system thereof

Abstract

The application discloses a single-round federated graph learning matrix alignment and global hybrid method and a system thereof.K clients and a server are used to randomly divide an open benchmark graph data set by using a Dirichlet distribution to obtain a private graph data set; each client adopts a training node mask to participate in the node of local proxy loss calculation, trains a condition generator, and uploads the generator parameters through single-round online communication; the server combines a pre-set key pair, sampling noise and corresponding labels, reconstructs M complete pseudo graphs from the generator parameters of each client, forms a global pseudo graph data set, and trains a global GNN model through random global playback.The application relieves the pseudo node feature mode collapse and improves the sparse topology reconstruction quality through the joint alignment of feature distribution, statistical moment, global topology and local edge connection, so that the generated pseudo graph maintains the intra-class diversity and distribution expansion degree, and provides high-quality proxy data for downstream global training.

Description

Technical Field

[0001] This invention specifically relates to a method and system for moment alignment and global blending in single-round federated graph learning. Background Technology

[0002] Federated Graph Learning (FGL) is a technique that allows multiple clients to collaboratively train a graph neural network model without sharing the original graph data. This technique can be applied to scenarios such as social network analysis, recommender systems, medical collaboration networks, paper citation networks, and knowledge graphs.

[0003] One-Shot Federated Graph Learning (OFGL) significantly reduces communication overhead and energy consumption by limiting online communication between the client and server to only one instance (the client uploads only a lightweight proxy model or generator parameters, and the server uses these to synthesize pseudo-data to train a global model before distributing it). This makes it particularly suitable for resource-constrained mobile edge network environments. However, existing one-shot federated graph learning methods still have significant shortcomings in highly heterogeneous environments, mainly due to two core bottlenecks:

[0004] First, on the client side, existing methods often employ prototype alignment, which only matches the mean of node features, to train conditional generators or surrogate models. This leads to excessive homogenization of generated pseudo-node features, resulting in severe mode collapse and failing to preserve the diversity and variance information of intra-class distributions in real data. Furthermore, due to the sparsity and non-Euclidean nature of graph structures, without direct edge-level supervision signals, local topological reconstruction of pseudo-graphs is often inaccurate, exhibiting low structural fidelity. These problems are particularly pronounced under strict single-round constraints because there is no opportunity for iterative refinement.

[0005] Secondly, on the server side, existing methods employ parameter averaging, retention penalties, or soft prediction distillation to fuse knowledge from clients. These methods easily lead the global model into a trade-off of client-specific local optima, resulting in model homogenization and catastrophic forgetting. They fail to effectively fuse the diverse structural and feature knowledge from heterogeneous clients, leading to poor generalization performance of the final model across different client distributions. Summary of the Invention

[0006] Purpose of the invention: To provide a method and system for moment alignment and global blending in single-round federated graph learning, which solves the above-mentioned problems existing in the prior art.

[0007] Technical Solution: A moment alignment and global hybridization method for single-round federated graph learning, comprising the following steps: setting up k clients and one server, initializing each client to store a private graph dataset, randomly partitioning the public benchmark graph dataset using a Dirichlet distribution to obtain k subgraphs, storing these k subgraphs as private graph datasets within the k clients, and using training node masks within each client's subgraph. Nodes participating in the local agent loss calculation are selected, and an extreme heterogeneous mobile edge scenario is constructed in which the distribution of node labels and topology among clients are not independent and identically distributed, in order to test the robustness of the method under highly heterogeneous conditions.

[0008] Each client generates a key and stores it on the server. The public key is then encrypted and sent to the server using TEE secure communication.

[0009] Server-side initialization of the global GNN model Each client has a condition generator. Construct a sparse pseudo-adjacency matrix and calculate the complete pseudo-graph;

[0010] Training node mask At the defined nodes, a composite surrogate objective function is used, simultaneously applying statistical alignment constraints and structural alignment constraints to obtain a composite loss. This composite loss is then converted into generator parameter gradients via backpropagation. The Adam optimizer is used to iteratively update the parameters of the conditional generator until it reaches a stationary state, at which point the iterative update stops, yielding the generator parameters. Each client transmits generator parameters via a single round of online uplink communication. Upload to the server;

[0011] The server depends on the generator parameters. Combined with the key pair pre-set via the TEE secure channel ( Sampling noise and corresponding labels are used to reconstruct the server-side pseudograph. For the j-th pseudograph, the server follows a normal distribution. Independent sampling noise And select one-hot category condition labels as input to the condition generator. ;

[0012] For each client k, the server calculates based on its generator parameters. Refactoring A complete pseudo-image ,in To adopt The sparse pseudo-adjacency matrix obtained from nearest neighbors The server will combine K×M pseudo-images and their corresponding pseudo-labels to form a global pseudo-image dataset. And adopt an extended training budget;

[0013] Before the start of each global training round, the global pseudo-graph dataset is processed. Perform a global random shuffle to obtain an updated pseudo-image dataset. Then, extract a small batch of pseudo-images and pseudo-labels from the updated pseudo-image dataset and calculate the cross-entropy loss between the node-level predictions of the global GNN and the pseudo-labels. Then, the Adam optimizer is used to iteratively update the global model parameters θ. After training is completed, the global model parameters θ are distributed to each client through the downlink to deploy local node classification inference.

[0014] Preferably, the private graph dataset of the k-th client. ,in, Represents a partial view. This represents the set of nodes for the k-th client. Represents the set of edges; Let be the node feature matrix, where Represents the d-dimensional node feature matrix. Indicates the number of nodes. It represents the real number field, that is, the set of all real numbers; C represents the node label; C represents the number of node categories. Representing a partial view The binary topological adjacency matrix, The elements in the table are used to indicate whether there are edges between the corresponding nodes, and serve as the actual topological basis for subsequent local structure alignment and edge reconstruction.

[0015] Preferably, a sparse pseudo-adjacency matrix is ​​constructed. The process is as follows:

[0016] Each client has a condition generator. A condition generator is initialized and trained by the client. generator For a regularized multilayer feedforward network, a random noise vector is received. and category one-hot tags The features are used as input to the generator and propagated forward to output the original pseudo-node features. ;

[0017] Each client generates an offset vector independently. , offset vector The 64 components are respectively from the interval Independent, equal-probability sampling is performed on a continuous uniform distribution, and the components are combined in order to obtain... Sampling is performed, and then QR decomposition is applied to the random Gaussian matrix to obtain the orthogonal rotation matrix. In each round of local training, the client bases the offset vectors. and orthogonal rotation matrix Generating original pseudo-node features, performing L2 normalization on these features, calculating the pseudo-adjacency logits matrix, and then constructing a sparse pseudo-adjacency matrix. .

[0018] Preferably, the process of normalizing the pseudo-node features before calculating the pseudo-adjacency logits matrix is ​​as follows:

[0019] First, the original pseudo-node features conduct Normalization yields the unit norm characteristics. ;

[0020] Recalculate pseudo-adjacency matrix ;

[0021] Finally adopted Neighbors The nearest neighbor of each node is selected spatially.

[0022] Preferably, the process of generating a key on the client side is as follows:

[0023] Using offset vector From uniform distribution Sampling, in which, It is a small constant. For noise dimension;

[0024] The orthogonal rotation matrix is ​​obtained by performing QR decomposition on the random Gaussian matrix. ,satisfy ;

[0025] The key pair is injected into the forward propagation path during the generation process, which can be represented as follows: , , , In the formula: This represents a noise vector sampled from the offset multivariate normal distribution; Represents a multivariate normal distribution; This represents the unit covariance matrix with the same dimension as the noise vector. This represents the original pseudo-node feature matrix obtained by outputting the condition generator and undergoing orthogonal rotation; Represents the one-hot label of the category corresponding to the noise vector; Represents an orthogonal matrix; This represents the unit norm pseudo-node feature matrix obtained by L2 normalization of the original pseudo-node features; Represents the L2 norm; This represents a function that performs L2 normalization on the input features; This represents the pseudo-adjacency logits matrix, which is composed of the pairwise inner products of the normalized pseudo-node features. This represents the normalized pseudo-node feature matrix; This represents the transpose of the normalized pseudo-node feature matrix;

[0026] Then adopt Nearest neighbor in pseudo-adjacency Selecting nearest neighbors in matrix space to construct a sparse pseudograph ,

[0027] because It is an orthogonal matrix that preserves the Euclidean distance. In the formula, Let represent any d-dimensional eigenvector; This represents any d-dimensional feature vector with the same dimension as u; This represents the L2 norm of a vector, also known as the Euclidean norm.

[0028] Preferably, generator parameters The setup process is as follows:

[0029] First, when training the generator on the client side, a composite surrogate objective function is used, simultaneously applying statistical alignment constraints and structural alignment constraints to calculate the composite loss, where the statistical alignment constraints and structural alignment constraints are the feature alignment losses. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss ,in,

[0030] Feature alignment loss The Wasserstein distance of the slices is used to align the original pseudo-node features with the real features in the training node mask. The marginal distribution on the surface is captured by multiple random one-dimensional projections to capture the differences in multidimensional distribution;

[0031] Statistical moment alignment loss Simultaneously match the first and second moments, and calculate the sum of squares of the differences between the empirical means and the sum of squares of the differences between the true features and the original pseudo-node features in each dimension;

[0032] Global topology alignment loss Using the fused Gromov-Wasserstein distance, based on the training node mask From the real graph and the pseudo graph respectively, select the corresponding nodes and their induced edges to obtain the mask subgraphs, and jointly analyze the features of the real nodes. Features of the original pseudo-node and actual adjacency relationships and pseudo-adjacency To maintain the high-order topological dependencies of the community structure

[0033] Local edge reconstruction loss Training node mask On the limited pairs of nodes, based on the actual adjacency relationship As a supervisory label, it is used to monitor pseudo-adjacency. Direct edge-level supervision is performed using positive class weighted binary classification cross-entropy.

[0034] Positive class weight It is used to compensate for the class imbalance between real edges and non-edges in sparse graphs, amplify the gradient contribution corresponding to real edges, prevent the generator from degenerating into an empty graph, and supplement the global coarse-grained topological gradient provided by FGW distance with fine-grained edge-level gradients.

[0035] The composite loss is then converted into parameter gradients through backpropagation, and the Adam optimizer is used in conjunction with adaptive rules to calculate the gradients until the generator converges, thus obtaining the generator parameters. .

[0036] Preferably, the server uses the generator parameters uploaded by each client. Constructing a global pseudograph dataset The process is as follows:

[0037] For the k-th client, the server uses j to represent the sequence number of the pseudo-graph generated by the server, and sets j to a value from 1 to... For the j-th pseudo-graph, the server starts from... The mean and unit covariance matrix The normal distribution of the covariance matrix Independent sampling noise And select one-hot category condition labels as input to the condition generator. ;calculate The original pseudo-node feature matrix of the j-th pseudo-graph is obtained. ;right Performing L2 normalization yields the identity norm pseudo-node feature matrix. Using nearest neighbor parameters Constructing a sparse pseudo-adjacency matrix The server-side reconstruction was obtained. A complete pseudo-image Among them, complete pseudo-image sparse pseudo-adjacency matrix Used to build a global pseudo-graph dataset on the server; the server adopts the following... To determine the nearest neighbor size and obtain richer graph structure information;

[0038] Reconstruct the complete pseudograph for each of the k clients and their corresponding M servers. and its pseudo-labels Directly concatenate to form a global pseudo-graph dataset. ; among them, each All are composed of normalized pseudo-node features With sparse pseudo-adjacency matrix The generator parameters, sparse pseudo-adjacency matrix, and complete pseudograph represent the basis for generating the server-side reconstructed pseudograph, its components, and the final data object, respectively. Their calculation formulas are as follows:

[0039] In the formula, This represents the j-th pseudo-graph generated by the k-th client. This represents the j-th pseudo-tag generated by the k-th client.

[0040] Preferably, when the server constructs the adjacency relationship of the pseudograph, it adopts... Nearest neighbor scale, when the client builds the pseudograph locally. Neighborhood scale, and This allows for the acquisition of richer graph structure information during the server-side training phase.

[0041] Preferably, the process of expanding the training budget on the server side is as follows:

[0042] Based on global pseudomap dataset The amount and diversity of data exceed those of a single client, and the server uses an extended training budget. A global graph neural network model is trained using a random global replay mechanism. Before the start of each global training round, the global pseudo-graph dataset is processed. Perform a global random shuffle, causing the pseudographs for each client to appear interleaved in batches across epochs, to simulate a global mixed distribution. The independent and identically distributed sampling is performed in sequence, and then each mini-batch is extracted from the randomly shuffled updated pseudo-image dataset according to the preset batch size. ,in, This indicates a fake image within this batch. This represents the pseudo-labels corresponding to each pseudo-image, based on the calculation of node-level predictions of the global GNN. With pseudo-tags Cross-entropy loss The calculation formula is as follows:

[0043] In the formula: This represents the node-level prediction of a global GNN; Indicates a pseudo-tag;

[0044] Use the Adam optimizer to optimize global model parameters. After the iterative update, the data is distributed to all clients via the downlink, and each client receives the updated global model parameters. Update the client-side local generator parameters and deploy the global model for local node classification inference.

[0045] A moment alignment and global hybridization system for single-round federated graph learning is provided to implement the moment alignment and global hybridization method for single-round federated graph learning described above. The system includes a client agent generation module, a latent space key conditionalization module, a statistical and structural joint alignment module, a server pseudo-dataset hybridization module, and a random global replay training module.

[0046] The client agent generation module is used to build and train the condition generator. ;

[0047] The latent space key conditionalization module is used for offset vectors. and orthogonal rotation matrix Key conditionalization is applied to the pseudo-node feature generation process;

[0048] The statistical and structural joint alignment module is based on feature alignment loss. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss Four-term alignment loss calculation, optimization condition generator ;

[0049] The server-side pseudo-dataset mixing module is used to construct a global pseudo-dataset. ;

[0050] The random global replay training module is used to train a global graph neural network model under the random global replay mechanism. .

[0051] Beneficial effects: This invention relates to a moment alignment and global hybrid method and system for single-round federated graph learning, which utilizes feature alignment loss. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss This approach simultaneously aligns feature marginal distribution and statistical moments, effectively mitigating the feature pattern collapse problem of pseudo-nodes. It ensures that the generated pseudo-samples maintain intra-class diversity and distribution expansion, significantly improving the statistical fidelity of pseudo-graphs and providing high-quality proxy data for downstream global training.

[0052] By leveraging the joint supervision of global fusion Gromov-Wasserstein topological alignment and positive class reweighted local edge reconstruction, the structural recovery capability of sparse graphs is significantly improved. This provides strong local gradient signals for the generator, avoids topological distortion and empty graph degradation caused by weak supervision, and makes the pseudo-graph close to the real graph in both macroscopic community structure and microscopic edge connectivity.

[0053] In extreme non-IID scenarios, by replacing parameter space averaging or retention penalties with data-level pseudograph hybridization, the server can directly integrate heterogeneous knowledge from different clients in the data space, avoiding model homogenization and catastrophic forgetting, and improving the generalization ability of the global model to diverse client distributions.

[0054] By leveraging the latent spatial key conditionalization mechanism, the privacy protection of uploaded generator parameters is enhanced without increasing online communication overhead or sacrificing alignment utility, ensuring that generator parameters are uploaded only once to the server. It is suitable for mobile edge federated graph learning tasks that are resource-constrained, privacy-sensitive, and have highly heterogeneous client graph data. Attached Figure Description

[0055] Figure 1 This is a system block diagram of the present invention;

[0056] Figure 2 This is a diagram of the overall architecture of the present invention;

[0057] Figure 3 This is a verification diagram of the privacy protection effect under the center offset guessing and uniform distribution guessing attacks of the present invention. Detailed Implementation

[0058] like Figures 1 to 3 As shown, this invention provides a technical solution: a moment alignment and global blending method for single-round federated graph learning, comprising the following steps:

[0059] Set up k clients and one server. Initialize each client to store a private graph dataset, where the private graph dataset of the k-th client is... ,in, Represents a partial view. This represents the set of nodes for the k-th client. Represents the set of edges; Let be the node feature matrix, where Represents the d-dimensional node feature matrix. Indicates the number of nodes. It represents the real number field, that is, the set of all real numbers; Indicates the number of node categories; Representing a partial view The binary topological adjacency matrix, The elements in the dataset are used to indicate whether there are edges between corresponding nodes, serving as the true topological basis for subsequent local structure alignment and edge reconstruction. The public benchmark graph dataset is randomly divided into k subgraphs using a Dirichlet distribution. These k subgraphs are stored as private graph datasets within k clients. Within each client's subgraph, a training node mask is used. Nodes participating in the local agent loss calculation are selected, and an extreme heterogeneous mobile edge scenario is constructed in which the distribution of node labels and topology among clients are not independent and identically distributed, in order to test the robustness of the method under highly heterogeneous conditions.

[0060] Each client generates a key and stores it on the server. The public key is then encrypted and sent to the server using TEE (Trusted Execution Environment). The client-side key generation process is as follows:

[0061] Using offset vector From uniform distribution Sampling, in which, It is a small constant. For noise dimension;

[0062] The orthogonal rotation matrix is ​​obtained by performing QR decomposition on the random Gaussian matrix. ,satisfy ;

[0063] The key pair is injected into the forward propagation path during the generation process, which can be represented as follows: , , , In the formula: This represents a noise vector sampled from the offset multivariate normal distribution; Represents a multivariate normal distribution; This represents the unit covariance matrix with the same dimension as the noise vector. This represents the original pseudo-node feature matrix obtained by outputting the condition generator and undergoing orthogonal rotation; Represents the one-hot label of the category corresponding to the noise vector; Represents an orthogonal matrix; This represents the unit norm pseudo-node feature matrix obtained by L2 normalization of the original pseudo-node features; Represents the L2 norm; This represents a function that performs L2 normalization on the input features; This represents the pseudo-adjacency logits matrix, which is composed of the pairwise inner products of the normalized pseudo-node features. This represents the normalized pseudo-node feature matrix; This represents the transpose of the normalized pseudo-node feature matrix;

[0064] Then adopt Nearest neighbor in pseudo-adjacency Nearest neighbor selection in matrix space to construct a sparse pseudograph, due to It is an orthogonal matrix that preserves the Euclidean distance. In the formula, Let represent any d-dimensional eigenvector; This represents any d-dimensional feature vector with the same dimension as u; This represents the L2 norm of the vector, also known as the Euclidean norm. Therefore, it does not disrupt subsequent statistical and structural alignment processes, while significantly increasing the difficulty of naive inversion attacks after the generator parameters are intercepted.

[0065] Initialize the global GNN model on the server. Each client has a condition generator. A sparse pseudo-adjacency matrix is ​​constructed, and the complete pseudo-graph is calculated. In this embodiment, a 3-layer multilayer perceptron (MLP) is used with a hidden dimension of 256, an activation function of LeakyReLU (negative slope of 0.2), and an output layer dimension equal to the feature dimension. Generator parameters Random initialization, noise dimension Each client has a condition generator. Construct a sparse pseudo-adjacency matrix and calculate the complete pseudo-graph. The process of constructing the sparse pseudo-adjacency matrix is ​​as follows:

[0066] Each client has a condition generator. A condition generator is initialized and trained by the client. generator For a regularized multilayer feedforward network, a random noise vector is received. and category one-hot tags The features are used as input to the generator and propagated forward to output the original pseudo-node features. ;

[0067] Each client generates an offset vector independently. , offset vector The 64 components are respectively from the interval Independent, equal-probability sampling is performed on a continuous uniform distribution, and the components are combined in order to obtain... Sampling is performed, and then QR decomposition is applied to the random Gaussian matrix to obtain the orthogonal rotation matrix. In each round of local training, the client bases the offset vectors. and orthogonal rotation matrix The process of generating original pseudo-node features, performing L2 normalization on these features, calculating the pseudo-adjacency logits matrix, constructing a sparse pseudo-adjacency matrix, and then normalizing the pseudo-node features before calculating the pseudo-adjacency logits matrix is ​​as follows:

[0068] First, the original pseudo-node features conduct Normalization yields the unit norm characteristics. ;

[0069] Then calculate the pseudo-adjacency logits matrix. ;

[0070] Finally adopted The nearest neighbor is selected for each node in the logits space.

[0071] Then apply the node mask A composite surrogate objective function is employed, with both statistical alignment and structural alignment constraints applied for composite computation to obtain the composite loss. This composite loss is then converted into parameter gradients via backpropagation, and further computation is performed using the Adam optimizer in conjunction with adaptive rules until the generator converges, yielding the generator parameters. Each client transmits generator parameters via a single round of online uplink communication. The data is uploaded to the server, which then processes it according to the generator parameters. Generate a parametric pseudomap, where the generator parameters are... The setup process is as follows:

[0072] First, when training the generator on the client side, a composite surrogate objective function is used, simultaneously applying statistical alignment constraints and structural alignment constraints to calculate the composite loss, where the statistical alignment constraints and structural alignment constraints are the feature alignment losses. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss ,in,

[0073] Feature alignment loss The Wasserstein distance of the slices is used to align the original pseudo-node features with the real features in the training node mask. The marginal distribution on the surface is captured by multiple random one-dimensional projections to capture the differences in multidimensional distribution;

[0074] Statistical moment alignment loss Simultaneously matching the first and second moments, the sum of squares of the differences between the empirical means and the sum of squares of the differences in standard deviations between the real features and the original pseudo-node features in each dimension is calculated. This prevents pattern collapse caused by mere mean alignment, ensuring that the generated pseudo-features not only approach the center of the real features but also maintain a similar degree of distribution spread as the real data, thereby preserving intra-class diversity.

[0075] Global topology alignment loss Using the fused Gromov-Wasserstein distance, based on the training node mask From the real graph and the pseudo graph respectively, select the corresponding nodes and their induced edges to obtain the mask subgraphs, and jointly analyze the features of the real nodes. Features of the original pseudo-node and actual adjacency relationships and pseudo-adjacency To maintain the high-order topological dependencies of the community structure

[0076] Local edge reconstruction loss Training node mask On the limited pairs of nodes, based on the actual adjacency relationship As a supervisory label, it is used to monitor pseudo-adjacency. Direct edge-level supervision is performed using positive class weighted binary classification cross-entropy.

[0077] Positive class weight It is used to compensate for the class imbalance between real edges and non-edges in sparse graphs, amplify the gradient contribution corresponding to real edges, prevent the generator from degenerating into an empty graph, and supplement the global coarse-grained topological gradient provided by FGW distance with fine-grained edge-level gradients.

[0078] The composite loss is then converted into parameter gradients through backpropagation, and the Adam optimizer is used in conjunction with adaptive rules to calculate the gradients until the generator converges, thus obtaining the generator parameters. .

[0079] The server uses the generator parameters uploaded by each client. Constructing a global pseudograph dataset The process is as follows:

[0080] For the k-th client, the server uses j to represent the sequence number of the pseudo-graph generated by the server, and sets j to a value from 1 to... For the j-th pseudo-graph, the server starts from... The mean and unit covariance matrix The normal distribution of the covariance matrix Independent sampling noise And select one-hot category condition labels as input to the condition generator. ;calculate The original pseudo-node feature matrix of the j-th pseudo-graph is obtained. ;right Performing L2 normalization yields the identity norm pseudo-node feature matrix. Using nearest neighbor parameters Constructing a sparse pseudo-adjacency matrix The server-side reconstruction was obtained. A complete pseudo-image Among them, complete pseudo-image sparse pseudo-adjacency matrix Used to build a global pseudo-graph dataset on the server; the server adopts the following... To determine the nearest neighbor size and obtain richer graph structure information;

[0081] Reconstruct the complete pseudograph for each of the k clients and their corresponding M servers. and its pseudo-labels Directly concatenate to form a global pseudo-graph dataset. ; among them, each All are composed of normalized pseudo-node features With sparse pseudo-adjacency matrix The generator parameters, sparse pseudo-adjacency matrix, and complete pseudograph represent the basis for generating the server-side reconstructed pseudograph, its components, and the final data object, respectively. Their calculation formulas are as follows:

[0082] In the formula, This represents the j-th pseudo-graph generated by the k-th client. This represents the j-th pseudo-tag generated by the k-th client.

[0083] The process of expanding the training budget on the server side is as follows:

[0084] Based on global pseudomap dataset The amount and diversity of data exceed those of a single client, and the server uses an extended training budget. A global graph neural network model is trained using a random global replay mechanism. Before the start of each global training round, the global pseudo-graph dataset is processed. Perform a global random shuffle, causing the pseudographs for each client to appear interleaved in batches across epochs, to simulate a global mixed distribution. The independent and identically distributed sampling is performed in sequence, and then each mini-batch is extracted from the randomly shuffled updated pseudo-image dataset according to the preset batch size. ,in, This indicates a fake image within this batch. This represents the pseudo-labels corresponding to each pseudo-image, based on the calculation of node-level predictions of the global GNN. With pseudo-tags Cross-entropy loss The calculation formula is as follows:

[0085] In the formula: This represents the node-level prediction of a global GNN; Indicates a pseudo-tag;

[0086] Use the Adam optimizer to optimize global model parameters. After the iterative update, the data is distributed to all clients via the downlink, and each client receives the updated global model parameters. Update the client-side local generator parameters and deploy the global model for local node classification inference.

[0087] A moment alignment and global hybridization system for single-round federated graph learning is provided to implement the moment alignment and global hybridization method for single-round federated graph learning described above. The system includes a client agent generation module, a latent space key conditionalization module, a statistical and structural joint alignment module, a server pseudo-dataset hybridization module, and a random global replay training module.

[0088] The client agent generation module is used to build and train the condition generator. ;

[0089] The latent space key conditionalization module is used for offset vectors. and orthogonal rotation matrix Key conditionalization is applied to the pseudo-node feature generation process;

[0090] The statistical and structural joint alignment module is based on feature alignment loss. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss Four-term alignment loss calculation, optimization condition generator ;

[0091] The server-side pseudo-dataset mixing module is used to construct a global pseudo-dataset. ;

[0092] The random global replay training module is used to train a global graph neural network model under the random global replay mechanism. .

[0093] The experimental testing method in this embodiment is as follows. The experiments involved in this invention were conducted on a high-performance computing device equipped with an Intel Xeon Platinum 8352V processor and an NVIDIA RTX 4090 GPU (24GB of video memory). The software platform used Python 3.10 and the PyTorch deep learning framework (with the PyTorch Geometric 2.4 graphics neural network library). All models were trained on a single GPU using the Adam optimizer. The experiments were repeated 5 times, and the average results were taken to ensure statistical reliability.

[0094] This experiment evaluates the proposed method on seven real-world graph datasets: Cora, CiteSeer, PubMed, Chameleon, Amazon Photo, Coauthor-CS, and Ogbn-Arxiv. (Number of clients) Data according to Dirichlet( Non-IID segmentation was performed, and concentration parameter coverage was achieved. To simulate a variety of scenarios ranging from extremely heterogeneous to near-IID; the number of synthesized pseudo-images per client. , , Positive class weight Loss weight The local training rounds are 100, and the global extended training budget is 5 times the original budget and at least 500 rounds. This invention is comprehensively compared with methods such as FedAvg, FedProx, FedNova, FedPub, FedTAD, DENSE, FedCVAE, FedSD2C, OASIS, and GHOST. Furthermore, robustness is evaluated under different degrees of heterogeneity, and the privacy protection effect of the potential spatial key conditionalization mechanism is verified under two attacks: center offset guessing and uniform distribution guessing.

[0095] Table 1. Comparison Results of Classification Performance

[0096]

[0097] Table 1 above shows the classification performance comparison results of the method of this invention. As can be seen from the table, this invention (MAGMA) achieved the highest accuracy (Acc) and macro-average F1 score on all seven datasets, especially in the extreme non-independent and identically distributed (Dirichlet) settings. Under these conditions, the performance of the method is significantly improved compared to the strongest baseline method (approximately +10.40% improvement over OASIS on CiteSeer and approximately +6.48% improvement over GHOST on Cora), and the standard deviation is smaller, indicating more stable convergence. Traditional multi-round federated graph learning methods suffer severe performance degradation under single-round constraints, while pure single-round federated learning methods perform poorly due to the lack of topology-aware supervision.

[0098] Table 2 Comparison Results of Classification Performance

[0099]

[0100] Table 2 above shows the results of the method of the present invention under different degrees of non-independent identical distribution (Dirichlet) The table below compares the performance and stability of the methods with existing methods. As the degree of heterogeneity increases ( (As the value decreases), the performance of all methods shows a downward trend, but the method of this invention shows the smallest decrease and still maintains a significant advantage in extremely heterogeneous scenarios, which fully demonstrates the strong robustness of the statistical and structural alignment mechanism and the random global replay strategy to highly heterogeneous mobile edge networks.

[0101] Figure 3 This diagram demonstrates the privacy protection effectiveness of the latent spatial key conditionation mechanism of this invention under center offset guessing and uniform distribution guessing attacks, showing a comparison between reconstruction error and signal energy. In both simulated passive uplink eavesdropping attack scenarios, the feature reconstruction error of the proposed method is significantly higher than that of the keyless baseline method, and the reconstruction error is close to or reaches the same level as the signal energy. This verifies that the key conditionation mechanism combining the offset vector and the orthogonal rotation matrix can effectively resist naive inversion attacks, while the model utility is almost unaffected due to the distance-preserving property of orthogonal transformation.

[0102] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for moment alignment and global blending in single-round federated graph learning, characterized in that, Includes the following steps: Set up k clients and one server. Initially, each client stores a private graph dataset. The public benchmark graph dataset is randomly partitioned using a Dirichlet distribution to obtain k subgraphs. These k subgraphs are stored as private graph datasets within the k clients. Within each client's subgraph, a training node mask is used. Nodes participating in the local agent loss calculation are selected, and an extreme heterogeneous mobile edge scenario is constructed in which the distribution of node labels and topology among clients are not independent and identically distributed, in order to test the robustness of the method under highly heterogeneous conditions. Each client generates a key and stores it on the server. The public key is then encrypted and sent to the server using TEE secure communication. Server-side initialization of the global GNN model Each client has a condition generator. Construct a sparse pseudo-adjacency matrix and calculate the complete pseudo-graph; Training node mask At the defined nodes, a composite surrogate objective function is used, simultaneously applying statistical alignment constraints and structural alignment constraints to obtain a composite loss. This composite loss is then converted into generator parameter gradients via backpropagation. The Adam optimizer is used to iteratively update the parameters of the conditional generator until it reaches a stationary state, at which point the iterative update stops, yielding the generator parameters. Each client transmits generator parameters via a single round of online uplink communication. Upload to the server; The server depends on the generator parameters. Combined with the key pair pre-set via the TEE secure channel ( Sampling noise and corresponding labels are used to reconstruct the server-side pseudograph. For the j-th pseudograph, the server follows a normal distribution. Independent sampling noise And select one-hot category condition labels as input to the condition generator. ; For each client k, the server calculates based on its generator parameters. Refactoring A complete pseudo-image ,in The sparse pseudo-adjacency matrix constructed by the server for the j-th pseudo-graph generated by the k-th client is used. Nearest neighbor data is obtained by combining K×M pseudo-images and their corresponding pseudo-labels into a global pseudo-image dataset. And adopt an extended training budget; Before the start of each global training round, the global pseudo-graph dataset is processed. Perform a global random shuffle to obtain an updated pseudo-image dataset. Then, extract a small batch of pseudo-images and pseudo-labels from the updated pseudo-image dataset and calculate the cross-entropy loss between the node-level predictions of the global GNN and the pseudo-labels. Then, the Adam optimizer is used to iteratively update the global model parameters θ; After training is complete, the global model parameters θ are distributed to each client via the downlink to deploy local node classification inference.

2. The method for moment alignment and global blending in single-round federated graph learning according to claim 1, characterized in that, The private graph dataset of the kth client ,in, Represents a partial view. This represents the set of nodes for the k-th client. Represents the set of edges; Let be the node feature matrix, where Represents the d-dimensional node feature matrix. Indicates the number of nodes. Represents the real number field; that is, the set of all real numbers. C represents the node label; C represents the number of node categories. Representing a partial view The binary topological adjacency matrix, The elements in the table are used to indicate whether there are edges between the corresponding nodes, and serve as the actual topological basis for subsequent local structure alignment and edge reconstruction.

3. The method for moment alignment and global blending in single-round federated graph learning according to claim 1, characterized in that, Constructing a sparse pseudo-adjacency matrix The process is as follows: Each client has a condition generator. A condition generator is initialized and trained by the client. generator For a regularized multilayer feedforward network, a random noise vector is received. and category one-hot tags The features are used as input to the generator and propagated forward to output the original pseudo-node features. ; Each client generates an offset vector independently. , offset vector The 64 components are respectively from the interval Independent, equal-probability sampling is performed on a continuous uniform distribution, and the components are combined in order to obtain... Sampling is performed, and then QR decomposition is applied to the random Gaussian matrix to obtain the orthogonal rotation matrix. In each round of local training, the client bases the offset vectors. and orthogonal rotation matrix Generating original pseudo-node features, performing L2 normalization on these features, calculating the pseudo-adjacency logits matrix, and then constructing a sparse pseudo-adjacency matrix. .

4. The method for moment alignment and global blending in single-round federated graph learning according to claim 3, characterized in that, The process of normalizing the pseudo-node features and then calculating the pseudo-adjacency logits matrix is ​​as follows: First, the original pseudo-node features conduct Normalization yields the unit norm characteristics. ; Recalculate pseudo-adjacency matrix ; Finally adopted Neighbors The nearest neighbor of each node is selected spatially.

5. The method for moment alignment and global blending in single-round federated graph learning according to claim 1, characterized in that, The process of generating a key on the client side is as follows: Using offset vector From uniform distribution Sampling, in which, It is a small constant. For noise dimension; The orthogonal rotation matrix is ​​obtained by performing QR decomposition on the random Gaussian matrix. ,satisfy ; The key pair is injected into the forward propagation path during the generation process, which can be represented as follows: , , , In the formula: This represents a noise vector sampled from the offset multivariate normal distribution; Represents a multivariate normal distribution; This represents the unit covariance matrix with the same dimension as the noise vector. This represents the original pseudo-node feature matrix obtained by outputting the condition generator and undergoing orthogonal rotation; Represents the one-hot label of the category corresponding to the noise vector; Represents an orthogonal matrix; This represents the unit norm pseudo-node feature matrix obtained by L2 normalizing the original pseudo-node features; Represents the L2 norm; This represents a function that performs L2 normalization on the input features; This represents the pseudo-adjacency logits matrix, which is composed of the pairwise inner products of the normalized pseudo-node features. This represents the normalized pseudo-node feature matrix; This represents the transpose of the normalized pseudo-node feature matrix; Then adopt Nearest neighbor in pseudo-adjacency Selecting nearest neighbors in matrix space to construct a sparse pseudograph , because It is an orthogonal matrix that preserves the Euclidean distance. In the formula, Let represent any d-dimensional eigenvector; This represents any d-dimensional feature vector with the same dimension as u; This represents the L2 norm of a vector, also known as the Euclidean norm.

6. The method for moment alignment and global blending in single-round federated graph learning according to claim 1, characterized in that, Generator parameters The setup process is as follows: First, when training the generator on the client side, a composite surrogate objective function is used, simultaneously applying statistical alignment constraints and structural alignment constraints to calculate the composite loss, where the statistical alignment constraints and structural alignment constraints are the feature alignment losses. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss ,in, Feature alignment loss The Wasserstein distance of the slices is used to align the original pseudo-node features with the real features in the training node mask. The marginal distribution on the surface is captured by multiple random one-dimensional projections to capture the differences in multidimensional distribution; Statistical moment alignment loss Simultaneously match the first and second moments, and calculate the sum of squares of the differences between the empirical means and the sum of squares of the differences between the true features and the original pseudo-node features in each dimension; Global topology alignment loss Using the fused Gromov-Wasserstein distance, based on the training node mask From the real graph and the pseudo graph respectively, select the corresponding nodes and their induced edges to obtain the mask subgraphs, and jointly analyze the features of the real nodes. Features of the original pseudo-node and actual adjacency relationships and pseudo-adjacency To maintain the high-order topological dependencies of the community structure Local edge reconstruction loss Training node mask On the limited pairs of nodes, based on the actual adjacency relationship As a supervisory label, it is used to detect false adjacencies. Direct edge-level supervision is performed using positive class weighted binary classification cross-entropy. Positive class weight It is used to compensate for the class imbalance between real edges and non-edges in sparse graphs, amplify the gradient contribution corresponding to real edges, prevent the generator from degenerating into an empty graph, and supplement the global coarse-grained topological gradient provided by FGW distance with fine-grained edge-level gradients. The composite loss is then converted into parameter gradients through backpropagation, and the Adam optimizer is used in conjunction with adaptive rules to calculate the gradients until the generator converges, thus obtaining the generator parameters. .

7. The method for moment alignment and global blending in single-round federated graph learning according to claim 4, characterized in that, The server uses the generator parameters uploaded by each client. Constructing a global pseudograph dataset The process is as follows: For the k-th client, the server uses j to represent the sequence number of the pseudo-graph generated by the server, and sets j to a value from 1 to... ; For the j-th pseudo-graph, the server starts from... The mean and unit covariance matrix The normal distribution of the covariance matrix Independent sampling noise And select one-hot category condition labels as input to the condition generator. ;calculate The original pseudo-node feature matrix of the j-th pseudo-graph is obtained. ;right Performing L2 normalization yields the identity norm pseudo-node feature matrix. Using nearest neighbor parameters Constructing a sparse pseudo-adjacency matrix The server-side reconstruction was completed. A complete pseudo-image Among them, complete pseudo-image sparse pseudo-adjacency matrix Used to build a global pseudo-graph dataset on the server; the server adopts the following... To determine the nearest neighbor size and obtain richer graph structure information; Reconstruct the complete pseudograph for each of the k clients and their corresponding M servers. and its pseudo-labels Directly concatenate to form a global pseudo-graph dataset. ; among them, each All are composed of normalized pseudo-node features With sparse pseudo-adjacency matrix The generator parameters, sparse pseudo-adjacency matrix, and complete pseudograph represent the basis for generating the server-side reconstructed pseudograph, its components, and the final data object, respectively. Their calculation formulas are as follows: In the formula, This represents the j-th pseudo-graph generated by the k-th client. This represents the j-th pseudo-tag generated by the k-th client.

8. The method for moment alignment and global blending in single-round federated graph learning according to claim 7, characterized in that, When the server constructs the adjacency relationship of the pseudo-graph, it adopts... Nearest neighbor scale, when the client builds the pseudograph locally. Neighborhood scale, and This allows for the acquisition of richer graph structure information during the server-side training phase.

9. The method for moment alignment and global blending in single-round federated graph learning according to claim 1, characterized in that, The process of expanding the training budget on the server side is as follows: Based on global pseudomap dataset The amount and diversity of data exceed those of a single client, and the server uses an extended training budget. A global graph neural network model is trained using a random global replay mechanism. Before the start of each global training round, the global pseudo-graph dataset is processed. Perform a global random shuffle, causing the pseudographs for each client to appear interleaved in batches across epochs, to simulate a global mixed distribution. The independent and identically distributed sampling is performed in sequence, and then each mini-batch is extracted from the randomly shuffled updated pseudo-image dataset according to the preset batch size. ,in, This indicates a fake image within this batch. This represents the pseudo-labels corresponding to each pseudo-image, based on the calculation of node-level predictions of the global GNN. With pseudo-tags Cross-entropy loss The calculation formula is as follows: In the formula: This represents the node-level prediction of a global GNN; Indicates a pseudo-tag; Use the Adam optimizer to optimize global model parameters. After the iterative update, the data is distributed to all clients via the downlink, and each client receives the updated global model parameters. Update the client-side local generator parameters and deploy the global model for local node classification inference.

10. A moment alignment and global blending system for single-round federated graph learning, used to implement the moment alignment and global blending method for single-round federated graph learning as described in any one of claims 1-9, characterized in that, The module includes a client-side agent generation module, a latent space key conditionalization module, a statistical and structural joint alignment module, a server-side pseudo-dataset mixing module, and a random global replay training module. The client agent generation module is used to build and train the condition generator. ; The latent space key conditionalization module is used for offset vectors. and orthogonal rotation matrix Key conditionalization is applied to the pseudo-node feature generation process; The statistical and structural joint alignment module is based on feature alignment loss. Statistical moment alignment loss Global topology alignment loss and local edge reconstruction loss Four-term alignment loss calculation, optimization condition generator ; The server-side pseudo-dataset mixing module is used to construct a global pseudo-dataset. ; The random global replay training module is used to train a global graph neural network model under the random global replay mechanism. .

Images