A personalized federated graph learning method and system based on dynamic structure perception and elastic alignment

Abstract

The application discloses a kind of personalized federated graph learning method and system based on dynamic structure perception and elastic alignment, belongs to federated graph learning technical field.The method introduces self-supervised enhancement dynamic structure learning mechanism in client, relies on variational information bottleneck and graph contrast learning to realize topological denoising and multi-scale feature fusion;Design elastic parameter alignment mechanism and curvature perception weighted aggregation strategy based on Fisher information on the server side, accurately quantify parameter importance and impose anisotropic constraints, effectively alleviate the model drift and catastrophic forgetting caused by non-independent and identically distributed data.The application can inhibit topological noise propagation, protect local personalized knowledge, improve federated training stability and model accuracy, suitable for noisy graph data and distributed heterogeneous scene, with good generalization in biochemistry molecular analysis, encrypted traffic classification and other fields.

Description

Technical Field

[0001] This application relates to the field of federated graph learning technology, specifically to a personalized federated graph learning method and system based on dynamic structure awareness and flexible alignment. Background Technology

[0002] With the exponential growth of graph data generated by distributed devices, Federated Graph Learning (FGL) has become a standard paradigm for breaking down data silos while adhering to privacy regulations such as GDPR. Through collaborative training of Graph Neural Networks (GNNs), without sharing the original graph topology, it has shown great potential in fields such as biochemical molecular discovery, financial risk control, and recommender systems.

[0003] However, traditional mainstream federated graph learning methods still face the following key technical bottlenecks: First, existing methods often make unrealistic assumptions about the completeness of local graph structures. In real-world scenarios, graph topology data is often noisy, sparse, or even task-irrelevant due to limitations in data collection or the injection of privacy-preserving noise. Traditional FGL methods typically perform message passing directly on these low-quality graphs. This approach ignores a crucial issue: high-quality adjacency is a prerequisite for effective GNN learning; incorrect edge propagation can mislead features, leading to degraded local representations. Furthermore, clients face a severe label scarcity problem. Relying solely on sparse monitoring signals is insufficient to effectively guide the reconstruction of complex graph structures. While some centralized methods attempt to optimize the topology, they are prone to overfitting noise in federated settings.

[0004] Secondly, the catastrophic forgetting caused by non-independent and identically distributed data distribution remains an unsolved problem in federated learning. Local optima often deviate significantly from the global optimum due to user preferences or geographical differences; this phenomenon is called model drift. While pioneering work like FedProx introduced neighbor terms to constrain local updates, their isotropic constraint strategy lacks fine-grained awareness of the functional importance of parameters. By imposing the same constraints on all parameters, the model is forced to forget different local knowledge while aligning with the global model.

[0005] Furthermore, existing anti-forgetting mechanisms often overlook the strong coupling between parameter importance and graph topology. In scenarios with noisy structures and heterogeneous distributions, how to elastically protect local personalized knowledge based on structure sensitivity while simultaneously modifying the topological structure remains a challenging problem to be solved. Summary of the Invention

[0006] In view of the technical problems mentioned in the background, the purpose of this invention is to provide a personalized federated graph learning method and system based on dynamic structure awareness and flexible alignment.

[0007] To achieve the objectives of this invention, the technical solution provided by this invention is as follows: First aspect This invention provides a personalized federated graph learning method based on dynamic structure awareness and flexible alignment, comprising: A self-supervised enhanced dynamic structure learning mechanism is introduced locally on the client side, using graph contrast learning as a structure regularization constraint. This mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing adaptive reconstruction of local graph topology. A Fisher-based resilient parameter alignment (FRPA) mechanism is introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to accurately quantify the importance of parameter functions and imposes anisotropic constraints.

[0008] Furthermore, the specific steps include the following: Step 1: System initialization and client graph structure construction: The server initializes and distributes the global graph model parameters and global Fisher information matrix; each client node collects local raw data and preprocesses it into unified graph topology structure data; Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning: Based on the local graph topology data, the client uses variational information bottleneck and self-supervised graph contrastive learning mechanism to dynamically denoise and extract features from the local graph structure, and combines the global Fisher information matrix to calculate elastic alignment loss, thus completing the iterative update of the local graph neural network model. Step 3: Local Fisher Information Estimation and Parameter Upload on the Client: After local training is completed, the client uses local data to calculate the Fisher information matrix of the current local model parameters and calculates the trace of the matrix. Then, the updated local model parameters, Fisher information matrix and trace of the matrix are encrypted and uploaded to the server. Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness: The server receives model data uploaded by each active client, performs weighted aggregation using the trace of each client's Fisher information matrix as the effective information content measure, and updates the global model parameters and the global Fisher information matrix. Step 5: Global Model Distribution and Multi-Round Collaborative Optimization: The server distributes the updated global model parameters and global Fisher information matrix to each client as prior knowledge for the next round of training, and returns to Step 2 until the preset convergence condition is met or the maximum number of communication rounds is reached, finally obtaining the trained personalized federated graph model.

[0009] Furthermore, in step 1, the client preprocesses the local raw data and transforms it into a unified graph topology structure. For unstructured time-series data, such as encrypted traffic data, the specific implementation method is as follows: Each data packet in each traffic session is instantiated as an independent node in the graph; the payload length and transmission direction features of the data packet are extracted as node attribute features to construct a node feature matrix; at the same time, directed edges are established between adjacent data packet nodes to explicitly model temporal causal dependencies, and finally a time-aware traffic interaction graph structure is constructed.

[0010] Furthermore, step 2 includes the following sub-steps: Step 2.1: Adaptive feature denoising based on variational information bottleneck: Construct a structure learner network to generate a probability-preserving mask for the original graph features, perform soft mask pruning on the features, and filter out redundant topological noise that is irrelevant to the task by minimizing mutual information to obtain the potential denoised graph representation. Step 2.2: Multi-scale representation fusion: A multi-layer graph isomorphic network GIN is used as the backbone encoder. The output of each layer is regarded as a structural view under different topological ranges. The features of each layer are weighted and fused through an adaptive attention mechanism, so that higher-level features are given greater weight to local regions containing high noise. Step 2.3, Self-supervised auxiliary view construction: Perform random data augmentation on the potential denoised graph structure obtained in Step 2.1 to generate two different augmented views, and extract the graph representations of the two views.

[0011] Furthermore, the total loss function for local model optimization in step 2 consists of three parts: The first part is the supervised classification loss based on local real labels; The second part is the self-supervised graph contrastive learning loss NT-Xent, which is used to maximize the mutual information between the two enhanced view representations in step 2.3 as a structural invariance prior. The third part is the Riemannian manifold elastic alignment regularization term FRPA based on Fisher information. It uses the global Fisher information matrix issued by the server to define the geometric metric of the parameter space and calculates the Mahalanobis distance between the local model parameters and the global model parameters.

[0012] Moreover, the physical mechanism of the Riemannian manifold elastic alignment regularization term FRPA based on the Fisher information matrix in the third part is as follows: the diagonal elements of the global Fisher information matrix serve as importance coefficients in the parameter dimension, imposing strict constraints on high curvature directions to force them to anchor to global consensus; and imposing relaxed constraints on low curvature directions to allow local parameters to adapt individually, thereby avoiding catastrophic forgetting during the optimization process.

[0013] Furthermore, the specific implementation method of step 4 is as follows: The server takes the trace of the Fisher information matrix received from the client as the total amount of effective structural information contained in that client; it divides the client's trace by the sum of the traces of all active clients to obtain the client's aggregate weight; it then uses this weight to perform a weighted summation of the local model parameters of all clients to update the global model parameters, and performs a weighted summation of the local Fisher information matrices to update the global Fisher information matrix. This mechanism can automatically reduce the impact of anomalous or low-quality clients with large gradient variance but little effective information on the global model.

[0014] Second aspect This invention provides a personalized federated graph learning system based on dynamic structure awareness and flexible alignment, used to execute the aforementioned personalized federated graph learning method based on dynamic structure awareness and flexible alignment, including a client and a server: A self-supervised enhanced dynamic structure learning mechanism is introduced locally on the client side, using graph contrast learning as a structure regularization constraint. This mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing adaptive reconstruction of local graph topology. A Fisher-based resilient parameter alignment (FRPA) mechanism is introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to accurately quantify the importance of parameter functions and imposes anisotropic constraints.

[0015] Furthermore, the client and server are specifically used to perform the following steps: Step 1: System initialization and client graph structure construction: The server initializes and distributes the global graph model parameters and global Fisher information matrix; each client node collects local raw data and preprocesses it into unified graph topology structure data; Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning: Based on the local graph topology data, the client uses variational information bottleneck and self-supervised graph contrastive learning mechanism to dynamically denoise and extract features from the local graph structure, and combines the global Fisher information matrix to calculate elastic alignment loss, thus completing the iterative update of the local graph neural network model. Step 3: Local Fisher Information Estimation and Parameter Upload on the Client: After local training is completed, the client uses local data to calculate the Fisher information matrix of the current local model parameters and calculates the trace of the matrix. Then, the updated local model parameters, Fisher information matrix and trace of the matrix are encrypted and uploaded to the server. Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness: The server receives model data uploaded by each active client, performs weighted aggregation using the trace of each client's Fisher information matrix as the effective information content measure, and updates the global model parameters and the global Fisher information matrix. Step 5: Global Model Distribution and Multi-Round Collaborative Optimization: The server distributes the updated global model parameters and global Fisher information matrix to each client as prior knowledge for the next round of training, and returns to Step 2 until the preset convergence condition is met or the maximum number of communication rounds is reached, finally obtaining the trained personalized federated graph model.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention designs a self-supervised and enhanced dynamic structure learning mechanism on the client side.

[0017] To address the shortcomings of existing federated graph learning methods, which directly perform message passing on noisy or missing graph structures, leading to degraded model representation capabilities, this invention constructs a differentiable information gating mechanism through variational information bottlenecks to adaptively shield irrelevant features. Combined with multi-scale representation fusion and graph contrastive learning, it deeply mines the inherent implicit dependencies of graph data in environments with scarce local labels. This invention effectively suppresses the cascading propagation of topological noise and dynamically corrects erroneous connection weights from the data's underlying layers, thereby significantly improving the discriminative power of graph node feature representations and the robustness of the model in complex structural noise environments.

[0018] (2) This invention proposes a flexible parameter alignment mechanism FRPA based on Fisher information.

[0019] To address the shortcomings of existing federated learning methods that often employ coarse-grained constraints in the same direction to force global model alignment when dealing with model drift caused by non-independent and identically distributed data, leading to catastrophic forgetting of local personalized knowledge, this invention utilizes a global Fisher information matrix to precisely quantify the functional importance of model parameters, i.e., loss curvature, and applies anisotropic regularization constraints to the Riemannian manifold parameter space. This invention enables the model to flexibly align global consensus in low-curvature (non-critical) directions, while strictly regulating and protecting local core knowledge in high-curvature (critical) directions. Thus, while ensuring global collaboration, it fundamentally overcomes the catastrophic forgetting problem in collaborative training.

[0020] (3) The present invention introduces an adaptive weighted aggregation strategy with perceived curvature on the server side.

[0021] To address the issue that existing global average aggregation methods, such as FedAvg, are susceptible to interference from extreme outlier nodes or missing high-quality clients, this invention uses the trace of the Fisher information matrix calculated and uploaded by each client as an evaluation metric for the amount of effective structural information it contains. Based on this metric, dynamic weighted aggregation is performed on each local model. This invention can automatically reduce the weights of inferior clients with large gradient variance but low effective information content, essentially constructing a structural smoother at the global level, thus enhancing the stability of the global model when facing extreme class imbalances and dense topological perturbations in real-world scenarios.

[0022] (4) The method of the present invention has excellent detection accuracy and wide application in various scenarios.

[0023] This invention organically combines structure awareness in the data space with flexible alignment in the parameter space, resulting in excellent model recognition accuracy and rapid convergence within a few communication rounds. This significantly reduces the bandwidth overhead of cross-node communication and the overall training time. Furthermore, this invention does not rely on specific domain-specific prior knowledge, making it applicable not only to standard graph data such as biochemical molecules and social networks, but also, when combined with the full-message time-series topology graph construction algorithm provided by this invention, to unstructured data scenarios such as large-scale encrypted traffic classification. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the client-side self-supervised reinforcement dynamic structure learning mechanism in this embodiment of the invention. Figure 3 This is a schematic diagram of the server-side elastic parameter alignment mechanism in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that the acquisition of data and collection of information in this application are legal, compliant, or obtained with the consent of the subject of the data collection.

[0027] This embodiment provides a personalized federated graph learning method based on dynamic structure awareness and flexible alignment, including: First, a self-supervised, enhanced dynamic structure learning mechanism is introduced locally on the client side. Utilizing graph contrastive learning as a structure regularization constraint, this mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing for adaptive reconstruction of local graph topology. This effectively suppresses the propagation of topological noise and significantly enhances the discriminability and robustness of graph representations.

[0028] Secondly, a Fisher-based resilient parameter alignment (FRPA) mechanism was introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to precisely quantify the importance of parameter functions, i.e., loss curvature, and imposes anisotropic constraints. This promotes the flexible alignment of non-critical parameters with the global model while strictly protecting core parameters associated with local knowledge.

[0029] like Figure 1 As shown, this embodiment specifically includes the following steps: Step 1: System initialization and client graph structure construction; The server initializes the global graph model parameters during the 0th round of communication. Random values ​​are assigned, and the global Fisher information matrix is ​​initialized simultaneously. And distribute them to all participating clients; Each client collects local raw data and constructs a graph topology. ;in, For a set of nodes, Let be the set of edges. The node feature matrix, It is an adjacency matrix.

[0030] Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning; such as... Figure 2 Specifically, it includes the following: Step 2.1: Adaptive feature denoising based on variational information bottleneck; Using a lightweight structure learner to process input features Perform dimensional feature importance assessment and generate a probability-preserving mask. : ; in, For learnable variational parameters, For activation function, The Sigmoid function maps the output to the (0,1) interval, representing the probability of retaining the feature dimension. Based on the generated mask The original features are soft-pruned to obtain a potential denoised image representation. : ; It should be noted that the meaning of this mathematical formula is to construct a differentiable information gate, which filters out redundant topological noise that is irrelevant to downstream tasks by minimizing mutual information, and implicitly corrects the incorrect connection weights in the adjacency matrix.

[0031] Step 2.2: Multi-scale representation fusion; A multi-layer graph isomorphic network (GIN) is used as the backbone encoder, and the following definition is made. For the first The feature representations output by each layer are fused using an adaptive attention mechanism: ; ; in, For the first Attention weights for layer features For attention scoring function, For graph-level pooling functions, This is the final graph-level representation vector; The total number of layers in the network; For the first Feature representation of layer output; It should be noted that this logical judgment gives dynamic weights to shallow features with low noise and deep features containing global context, acting as a structural smoother.

[0032] Step 2.3: Calculate the total loss and update the local model; The potential denoised image structure is enhanced to obtain two views. and Extraction representation Calculate the contrastive learning loss : ; in, For cosine similarity, For temperature hyperparameters, For the negative sample set, It is an exponential function; The negative sample representation vector; Subsequently, combined with supervised classification of losses and based on the global Fisher matrix Riemannian manifold elastic alignment regularization term Ultimately, the total local loss was obtained. : ; in, To balance the hyperparameters; Riemannian manifold elastic alignment regularization term , For the client Local parameters; This represents the transpose operation; the client updates parameters via gradient descent. , The learning rate for local training on the client side; This is the gradient operator.

[0033] Among them, in the regularization term It serves as the importance coefficient for each parameter dimension, with tighter constraints in the high curvature direction and looser constraints in the low curvature direction.

[0034] Step 3: Estimation of local Fisher information and uploading of parameters on the client side; like Figure 3 As shown, after local training is completed, the client calculates the diagonal approximation of the Fisher information matrix of the local model parameters. , its first The physical meaning of each diagonal element is the second moment of the loss function with respect to the weight gradient: ; Then, the trace of the matrix is ​​calculated. This is considered as the total amount of valid structural information contained in the client.

[0035] Finally, the client will update the parameters locally. Fisher Information Matrix and traces The data is encrypted and uploaded to the server.

[0036] Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness; specifically including the following: Step 4.1: Calculate the aggregate weight of each client. : ; in, This refers to the set of active clients participating in the aggregation in the current round. For the client The effective amount of information; For the client The effective amount of information; Step 4.2: Utilizing weights Update the global model parameters and the global Fisher information matrix respectively: , ; in, This represents the number of clients participating in this global model update.

[0037] It should be noted that this aggregation strategy automatically reduces the weights of clients with large gradient variance but low effective information content, such as those containing a large amount of abnormal noise graph data, to ensure the stability of the global model in a dual heterogeneous environment.

[0038] Step 5: Global model distribution and multi-round collaborative optimization; The server will aggregate the new global model. With the global Fisher information matrix The data is broadcast to all clients, serving as the geometric spatial metric benchmark and geometric prior knowledge for the next round of training. This process is repeated iteratively until the set maximum number of communication rounds is reached. The output is a finally converged personalized federated graph model.

[0039] To better demonstrate the beneficial effects of the present invention, this embodiment conducts comparative experiments based on seven benchmark datasets, as shown in Table 1, including DAPP, NCI1, MUTAG, PROTEINS, DD, COLLAB, and IMDB-BINARY. These datasets represent graph structure data and complex scenarios in different fields. NCI1, MUTAG, PROTEINS, and DD are biochemical datasets used for molecular graph classification; COLLAB and IMDB-BINARY are social network datasets representing dense topological analysis scenarios; and DAPP is a real-world industrial-grade large-scale encrypted traffic dataset with extreme class imbalance and high-intensity structural noise. These data are highly heterogeneous, and training the model under the federated learning framework requires overcoming local topological noise and the non-independent and identically distributed (Non-IID) problem.

[0040] Table 1

[0041] For the distributed classification task prediction of the above dataset, this embodiment adopts the personalized federated graph learning method based on dynamic structure awareness and elastic alignment proposed in this invention.

[0042] In the experimental setup, all methods used the same 3-layer GraphCNN architecture. To accommodate different data sizes, the hidden layer dimension was set to 128 for the DAPP dataset and 64 for the other datasets. Experiments were conducted using a Dirichlet distribution. The data was segmented to simulate a severe client-side tag heterogeneity (Non-IID) scenario in the real world.

[0043] The comparison methods cover general federated learning and state-of-the-art federated graph learning methods, including: FedAvg, FedProx, Ditto, GCFL+, FedStar, FedAGHN, SCFGL, and the method of this invention.

[0044] As shown in Table 2, the method of this invention achieved excellent performance on all seven datasets. Particularly on the large-scale industrial dataset DAPP with extreme structural noise and data imbalance, the accuracy of the method reached 96.81%, an improvement of over 8% compared to the existing best baseline method, demonstrating the superior ability of the adaptive feature denoising mechanism in this invention to filter out task-irrelevant topological noise. This proves that the elastic parameter alignment mechanism FRPA based on Fisher information and the adaptive weighted aggregation strategy based on perceived curvature proposed in this invention can effectively suppress model drift caused by data heterogeneity, successfully overcome the catastrophic forgetting problem in federated collaborative training, and ensure the stability and extremely high generalization ability of the model.

[0045] Table 2

[0046] In other embodiments, a personalized federated graph learning system based on dynamic structure awareness and flexible alignment is also provided for executing the aforementioned personalized federated graph learning method based on dynamic structure awareness and flexible alignment, including a client and a server: A self-supervised enhanced dynamic structure learning mechanism is introduced locally on the client side, using graph contrast learning as a structure regularization constraint. This mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing adaptive reconstruction of local graph topology. A Fisher-based resilient parameter alignment (FRPA) mechanism is introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to accurately quantify the importance of parameter functions and imposes anisotropic constraints.

[0047] Furthermore, the client and server are specifically used to perform the following steps: Step 1: System initialization and client graph structure construction: The server initializes and distributes the global graph model parameters and global Fisher information matrix; each client node collects local raw data and preprocesses it into unified graph topology structure data; Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning: Based on the local graph topology data, the client uses variational information bottleneck and self-supervised graph contrastive learning mechanism to dynamically denoise and extract features from the local graph structure, and combines the global Fisher information matrix to calculate elastic alignment loss, thus completing the iterative update of the local graph neural network model. Step 3: Local Fisher Information Estimation and Parameter Upload on the Client: After local training is completed, the client uses local data to calculate the Fisher information matrix of the current local model parameters and calculates the trace of the matrix. Then, the updated local model parameters, Fisher information matrix and trace of the matrix are encrypted and uploaded to the server. Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness: The server receives model data uploaded by each active client, performs weighted aggregation using the trace of each client's Fisher information matrix as the effective information content measure, and updates the global model parameters and the global Fisher information matrix. Step 5: Global Model Distribution and Multi-Round Collaborative Optimization: The server distributes the updated global model parameters and global Fisher information matrix to each client as prior knowledge for the next round of training, and returns to Step 2 until the preset convergence condition is met or the maximum number of communication rounds is reached, finally obtaining the trained personalized federated graph model.

[0048] Further, step 1 specifically includes the following: The server initializes the global graph model parameters during the 0th round of communication. Random values ​​are assigned, and the global Fisher information matrix is ​​initialized simultaneously. And distribute them to all participating clients; Each client collects local raw data and constructs a graph topology. ;in, For a set of nodes, Let be the set of edges. The node feature matrix, It is an adjacency matrix.

[0049] Furthermore, step 2 specifically includes the following: Step 2.1: Adaptive feature denoising based on variational information bottleneck; Using a lightweight structure learner to process input features Perform dimensional feature importance assessment and generate a probability-preserving mask. : ; in, For learnable variational parameters, For activation function, The Sigmoid function maps the output to the (0,1) interval, representing the probability of retaining the feature dimension. Based on the generated mask The original features are soft-pruned to obtain a potential denoised image representation. : ; Step 2.2: Multi-scale representation fusion; A multi-layer graph isomorphic network (GIN) is used as the backbone encoder, and the following definition is made. For the first The feature representations output by each layer are fused using an adaptive attention mechanism: ; ; in, For the first Attention weights for layer features For attention scoring function, For graph-level pooling functions, This is the final graph-level representation vector; The total number of layers in the network; For the first Feature representation of layer output; Step 2.3: Calculate the total loss and update the local model; The potential denoised image structure is enhanced to obtain two views. and Extraction representation Calculate the contrastive learning loss : ; in, For cosine similarity, For temperature hyperparameters, For the negative sample set, It is an exponential function; The negative sample representation vector; Subsequently, combined with supervised classification of losses and based on the global Fisher matrix Riemannian manifold elastic alignment regularization term Ultimately, the total local loss was obtained. : ; in, To balance the hyperparameters; Riemannian manifold elastic alignment regularization term , For the client Local parameters; This represents the transpose operation; the client updates parameters via gradient descent. , The learning rate for local training on the client side; This is the gradient operator.

[0050] Furthermore, step 4 specifically includes the following: Step 4.1: Calculate the aggregate weight of each client. : ; in, This refers to the set of active clients participating in the aggregation in the current round. For the client The effective amount of information; For the client The effective amount of information; Step 4.2: Utilizing weights Update the global model parameters and the global Fisher information matrix respectively: , ; in, This represents the number of clients participating in this global model update.

[0051] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.

Claims

1. A personalized federated graph learning method based on dynamic structure awareness and flexible alignment, characterized in that, include: A self-supervised enhanced dynamic structure learning mechanism is introduced locally on the client side, using graph contrast learning as a structure regularization constraint. This mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing adaptive reconstruction of local graph topology. A Fisher-based resilient parameter alignment (FRPA) mechanism is introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to accurately quantify the importance of parameter functions and imposes anisotropic constraints.

2. The personalized federated graph learning method based on dynamic structure awareness and flexible alignment according to claim 1, characterized in that, Specifically, the steps include the following: Step 1: System initialization and client graph structure construction: The server initializes and distributes the global graph model parameters and global Fisher information matrix; each client node collects local raw data and preprocesses it into unified graph topology structure data; Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning: Based on the local graph topology data, the client uses variational information bottleneck and self-supervised graph contrastive learning mechanism to dynamically denoise and extract features from the local graph structure, and combines the global Fisher information matrix to calculate elastic alignment loss, thus completing the iterative update of the local graph neural network model. Step 3: Local Fisher Information Estimation and Parameter Upload on the Client: After local training is completed, the client uses local data to calculate the Fisher information matrix of the current local model parameters and calculates the trace of the matrix. Then, the updated local model parameters, Fisher information matrix and trace of the matrix are encrypted and uploaded to the server. Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness: The server receives model data uploaded by each active client, performs weighted aggregation using the trace of each client's Fisher information matrix as the effective information content measure, and updates the global model parameters and the global Fisher information matrix. Step 5: Global Model Distribution and Multi-Round Collaborative Optimization: The server distributes the updated global model parameters and global Fisher information matrix to each client as prior knowledge for the next round of training, and returns to Step 2 until the preset convergence condition is met or the maximum number of communication rounds is reached, finally obtaining the trained personalized federated graph model.

3. The personalized federated graph learning method based on dynamic structure awareness and flexible alignment according to claim 2, characterized in that, Step 1 specifically includes the following: The server initializes the global graph model parameters during the 0th round of communication. Random values ​​are assigned, and the global Fisher information matrix is ​​initialized simultaneously. And distribute them to all participating clients; Each client collects local raw data and constructs a graph topology. ;in, For a set of nodes, Let be the set of edges. The node feature matrix, It is an adjacency matrix.

4. The personalized federated graph learning method based on dynamic structure awareness and flexible alignment according to claim 3, characterized in that, Step 2 specifically includes the following: Step 2.1: Adaptive feature denoising based on variational information bottleneck; Using a lightweight structure learner to process input features Perform dimensional feature importance assessment and generate a probability-preserving mask. : ； in, For learnable variational parameters, For activation function, The Sigmoid function maps the output to the (0,1) interval, representing the probability of retaining the feature dimension. Based on the generated mask The original features are soft-pruned to obtain a potential denoised image representation. : ； Step 2.2: Multi-scale representation fusion; A multi-layer graph isomorphic network (GIN) is used as the backbone encoder, and the following definition is made. For the first The feature representations output by each layer are fused using an adaptive attention mechanism: ； ； in, For the first Attention weights for layer features For attention scoring function, For graph-level pooling functions, This is the final graph-level representation vector; The total number of layers in the network; For the first Feature representation of layer output; Step 2.3: Calculate the total loss and update the local model; The potential denoised image structure is enhanced to obtain two views. and Extraction representation Calculate the contrastive learning loss : ； in, For cosine similarity, For temperature hyperparameters, For the negative sample set, It is an exponential function; The negative sample representation vector; Subsequently, combined with supervised classification of losses and based on the global Fisher matrix Riemannian manifold elastic alignment regularization term Ultimately, the total local loss was obtained. : ； in, To balance the hyperparameters; Riemannian manifold elastic alignment regularization term , For the client Local parameters; This represents the transpose operation; the client updates parameters via gradient descent. , The learning rate for local training on the client side; This is the gradient operator.

5. A personalized federated graph learning method based on dynamic structure awareness and flexible alignment according to claim 4, characterized in that, Step 4 specifically includes the following: Step 4.1: Calculate the aggregate weight of each client. : ； in, This refers to the set of active clients participating in the aggregation in the current round. For the client The effective amount of information; For the client The effective amount of information; Step 4.2: Utilizing weights Update the global model parameters and the global Fisher information matrix respectively: ， ； in, This represents the number of clients participating in this global model update.

6. A personalized federated graph learning system based on dynamic structure awareness and flexible alignment, used to execute the personalized federated graph learning method based on dynamic structure awareness and flexible alignment as described in any one of claims 1-5, characterized in that, Including client and server sides: A self-supervised enhanced dynamic structure learning mechanism is introduced locally on the client side, using graph contrast learning as a structure regularization constraint. This mechanism mines the inherent multi-scale dependencies of graph data under unsupervised constraints, allowing adaptive reconstruction of local graph topology. A Fisher-based resilient parameter alignment (FRPA) mechanism is introduced on the server side to mitigate catastrophic forgetting. FRPA utilizes the Fisher information matrix to accurately quantify the importance of parameter functions and imposes anisotropic constraints.

7. A personalized federated graph learning system based on dynamic structure awareness and flexible alignment according to claim 6, characterized in that, The client and server are specifically used to perform the following steps: Step 1: System initialization and client graph structure construction: The server initializes and distributes the global graph model parameters and global Fisher information matrix; each client node collects local raw data and preprocesses it into unified graph topology structure data; Step 2: The client performs local training based on self-supervised reinforcement dynamic structure learning: Based on the local graph topology data, the client uses variational information bottleneck and self-supervised graph contrastive learning mechanism to dynamically denoise and extract features from the local graph structure, and combines the global Fisher information matrix to calculate elastic alignment loss, thus completing the iterative update of the local graph neural network model. Step 3: Local Fisher Information Estimation and Parameter Upload on the Client: After local training is completed, the client uses local data to calculate the Fisher information matrix of the current local model parameters and calculates the trace of the matrix. Then, the updated local model parameters, Fisher information matrix and trace of the matrix are encrypted and uploaded to the server. Step 4: Server-side adaptive weighted aggregation based on parameter curvature awareness: The server receives model data uploaded by each active client, performs weighted aggregation using the trace of each client's Fisher information matrix as the effective information content measure, and updates the global model parameters and the global Fisher information matrix. Step 5: Global Model Distribution and Multi-Round Collaborative Optimization: The server distributes the updated global model parameters and global Fisher information matrix to each client as prior knowledge for the next round of training, and returns to Step 2 until the preset convergence condition is met or the maximum number of communication rounds is reached, finally obtaining the trained personalized federated graph model.

8. A personalized federated graph learning system based on dynamic structure awareness and flexible alignment according to claim 7, characterized in that, Step 1 specifically includes the following: The server initializes the global graph model parameters during the 0th round of communication. Random values ​​are assigned, and the global Fisher information matrix is ​​initialized simultaneously. And distribute them to all participating clients; Each client collects local raw data and constructs a graph topology. ;in, For a set of nodes, Let be the set of edges. The node feature matrix, It is an adjacency matrix.

9. A personalized federated graph learning system based on dynamic structure awareness and flexible alignment according to claim 7, characterized in that, Step 2 specifically includes the following: Step 2.1: Adaptive feature denoising based on variational information bottleneck; Using a lightweight structure learner to process input features Perform dimensional feature importance assessment and generate a probability-preserving mask. : ； in, For learnable variational parameters, For activation function, The Sigmoid function maps the output to the (0,1) interval, representing the probability of retaining the feature dimension. Based on the generated mask The original features are soft-pruned to obtain a potential denoised image representation. : ； Step 2.2: Multi-scale representation fusion; A multi-layer graph isomorphic network (GIN) is used as the backbone encoder, and the following definition is made. For the first The feature representations output by each layer are fused using an adaptive attention mechanism: ； ； in, For the first Attention weights for layer features For attention scoring function, For graph-level pooling functions, This is the final graph-level representation vector; The total number of layers in the network; For the first Feature representation of layer output; Step 2.3: Calculate the total loss and update the local model; The potential denoised image structure is enhanced to obtain two views. and Extraction representation Calculate the contrastive learning loss : ； in, For cosine similarity, For temperature hyperparameters, For the negative sample set, It is an exponential function; The negative sample representation vector; Subsequently, combined with supervised classification of losses and based on the global Fisher matrix Riemannian manifold elastic alignment regularization term Ultimately, the total local loss was obtained. : ； in, To balance the hyperparameters; Riemannian manifold elastic alignment regularization term , For the client Local parameters; This represents the transpose operation; the client updates parameters via gradient descent. , The learning rate for local training on the client side; This is the gradient operator.

10. A personalized federated graph learning system based on dynamic structure awareness and flexible alignment according to claim 7, characterized in that, Step 4 specifically includes the following: Step 4.1: Calculate the aggregate weight of each client. : ； in, This refers to the set of active clients participating in the aggregation in the current round. For the client The effective amount of information; For the client The effective amount of information; Step 4.2: Utilizing weights Update the global model parameters and the global Fisher information matrix respectively: ， ； in, This represents the number of clients participating in this global model update.

Images