Federal learning participant selection method and system

Abstract

The invention discloses a federated learning participant selection method and system, and the method comprises the steps: obtaining local model parameters uploaded by all federated learning client participants through a central server, and constructing a local model feature observation tensor of each client participant; calculating kernel similarity between participants of the client; determining a positive semi-definite similarity kernel matrix; the central server complements missing values of the positive semi-definite similarity kernel matrix to obtain a complete similarity kernel matrix; the central server constructs a timeliness weight diagonal matrix; the central server constructs a likelihood matrix according to the complete similarity kernel matrix and the timeliness weight diagonal matrix; solving the likelihood matrix through a determinant point process to obtain a selected client participant subset; averaging all local model parameters in the selected client participant subset to obtain a global model; and finally, issuing the global model to the selected client participant.

Description

Technical Field

[0001] The present invention relates to the field of machine learning optimization technology, and in particular to a method and system for selecting federated learning participants. Background Art

[0002] In the Industrial Internet of Things (IIoT) environment, heterogeneous data generated by massive distributed devices (such as sensors and edge computing nodes) is highly valuable and can be used to train machine learning models to optimize critical tasks such as equipment monitoring and predictive maintenance. However, due to data privacy regulations and industry security requirements, device data must be stored locally in isolation, creating "data silos" and making traditional centralized training models difficult to directly apply. Federated learning, through its collaborative "data remains static, model moves" paradigm, allows multiple participants to jointly train a global model while protecting local data privacy, making it a key technological path to industrial intelligence.

[0003] Existing federated learning systems rely on explicit participant reporting mechanisms for participant selection, such as screening based on reported local training losses or resource consumption forecasts. However, these dynamic metrics have inherent flaws in industrial environments: data non-independent and identically distributed (IID) prevents local losses from reflecting global data diversity requirements; Furthermore, the real-time execution status reported by participants exhibits lags and discrepancies, leading to the misselection of delayed nodes and slowing global convergence.

[0004] Although previous studies have attempted to optimize selection strategies through dynamic resource monitoring or reputation scoring, these solutions require clients to actively report their real-time status, which not only increases communication overhead but also exposes node vulnerabilities. Therefore, a federated learning framework is urgently needed that can dynamically optimize the selection of participating parties under imperceptible conditions, while also possessing anti-fraud capabilities and adaptability to heterogeneous data, to ensure the efficiency and security of collaborative learning in complex industrial environments. Summary of the Invention To address the shortcomings of existing technologies, this invention provides a method and system for selecting federated learning participants. Addressing the technical pain points of the widespread presence of untrusted nodes and severe data silos in the Industrial Internet of Things, this invention achieves secure and efficient collaborative training through an implicit evaluation mechanism. By constructing a selection mechanism agnostic to client participants and integrating data diversity quantification with missing data compensation techniques, this method effectively suppresses false reporting and improves model robustness in heterogeneous environments.

[0005] On the one hand, a method for selecting federated learning participants is provided, including: The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; calculates the kernel similarity between the client participants based on the local model feature observation tensors of any two client participants; and determines the semi-positive definite similarity kernel matrix based on the kernel similarity between the client participants; The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; The central server generates a calculation time decay coefficient for each client participant's local model based on the update timeliness of all client participants' local models; based on the time decay coefficient, a timeliness weight diagonal matrix is ​​constructed; The central server constructs a likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; solves the likelihood matrix through the determinant point process to obtain the selected subset of client participants; takes the average of all local model parameters in the selected subset of client participants to obtain the global model; finally, sends the global model to the selected client participants.

[0006] In another aspect, a federated learning participant selection system is provided, comprising: a central server and a plurality of client participants; The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; calculates the kernel similarity between the client participants based on the local model feature observation tensors of any two client participants; and determines the semi-positive definite similarity kernel matrix based on the kernel similarity between the client participants; The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; The central server generates a calculation time decay coefficient for each client participant's local model based on the update timeliness of all client participants' local models; based on the time decay coefficient, a timeliness weight diagonal matrix is ​​constructed; The central server constructs a likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; solves the likelihood matrix through the determinant point process to obtain the selected subset of client participants; takes the average of all local model parameters in the selected subset of client participants to obtain the global model; finally, sends the global model to the selected client participants.

[0007] The above technical solution has the following advantages or beneficial effects: The disclosed implicit client participant selection algorithm for federated learning based on data diversity and model freshness is based on an improved determinant point process sampling framework, combined with radial basis kernel similarity calculation, alternating direction matrix completion, and dynamic model staleness weight compensation technology to achieve efficient selection. First, a central server collects model parameters from selected client participants in each round through a secure aggregation protocol, extracts the first-layer neural network weights, and vectorizes them to construct a multidimensional feature tensor library. Using the radial basis function kernel method, the system performs pairwise similarity matching on the feature tensors to generate a nonlinear correlation matrix between client participants. For client participants with insufficient historical participation, low-rank matrix completion technology is introduced to probabilistically infer the missing similarity data. An alternating direction optimization algorithm is used to reconstruct a complete global similarity topology, effectively overcoming the traditional method's dependence on client participants being continuously online. Second, a dynamic assessment system for the timeliness of client participant model updates is established, calculating an exponential time decay factor based on the most recent update timestamp of each node. This factor is encoded as a diagonal weighted matrix that quantifies the negative impact of the client participant's model staleness on global updates. Ultimately, the system spectrally couples the completed similarity matrix with the time decay matrix to form a composite likelihood matrix that combines data diversity with model freshness. Based on the determinant point process probability sampling theory, the system adaptively selects the subset of client participants that maximizes the determinant value in each federated iteration, ensuring that the selected nodes cover both data distribution diversity and model freshness.

[0008] This disclosure proposes a federated learning implicit client participant optimization mechanism, which constructs a nonlinear similarity association matrix based on the client participant model parameters through the radial basis kernel function, reconstructs the global data distribution relationship in combination with dynamic topology completion technology, and generates a time-sensitive weight matrix based on a timestamp-driven exponential decay model to quantify the staleness of the model update. Further integrating the association matrix and the time-sensitive weight, an adaptive selection strategy based on the determinant point process is designed to intelligently screen high-quality client participants with both data diversity and model freshness, and integrate node model updates to optimize the global model. The mechanism operates in a zero-perception state of the client participants. The nodes only need to follow the standard federated protocol interaction, while the server side dynamically optimizes the node combination through explainable screening logic, significantly improving the model convergence efficiency. While effectively alleviating the challenges of non-independent and identically distributed data, this method can suppress interference from malicious nodes, providing an efficient and reliable implicit selection solution for cross-institutional federated collaboration.

[0009] The implicit client participant optimization mechanism for federated learning proposed in this disclosure is driven by the characteristics of multi-source heterogeneous data and achieves efficient convergence of the global model through a dual-track strategy of data quality assessment and model timeliness compensation. This mechanism intelligently screens high-value data nodes during the federated iteration process, and simultaneously builds a compensation channel for the impact of the model staleness attenuation factor on gradient updates, effectively balancing the contradictory relationship between data distribution diversity and parameter update timeliness. Different from traditional explicit screening schemes, data nodes complete federated collaboration without activating the client participant selection module. This design makes it impossible for client participants to parse the screening rules through reverse engineering, fundamentally eliminating attacks such as forging data features and maliciously manipulating selection results, and building a trusted federated computing environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0011] Figure 1 This is a flow chart of an adaptive client participant selection algorithm based on a determinant point process according to an embodiment of the present disclosure.

[0012] Figure 2 In the federated learning framework of the embodiment of the present disclosure, Diagram of the system architecture for dynamic client participant screening in rounds. DETAILED DESCRIPTION

[0013] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0014] Example 1 This embodiment provides a method for selecting federated learning participants; A method for selecting federated learning participants, comprising: S101: The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; calculates the kernel similarity between the client participants based on the local model feature observation tensors of any two client participants; and determines the semi-positive definite similarity kernel matrix based on the kernel similarity between the client participants; S102: The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; S103: The central server generates a calculation time decay coefficient for each client participant's local model based on the update timeliness of all client participants' local models; and constructs a timeliness weight diagonal matrix based on the time decay coefficient. S104: The central server constructs a likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; solves the likelihood matrix through the determinant point process to obtain the selected subset of client participants; averages all local model parameters in the selected subset of client participants to obtain a global model; finally, sends the global model to the selected client participants.

[0015] Furthermore, based on the local model feature observation tensors of any two client participants, the kernel similarity between the client participants is calculated, specifically using a radial basis kernel function for calculation.

[0016] Furthermore, in S101, the central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters, specifically including: The first-layer network parameters of each client participant model are extracted as the core feature carrier: If the first layer of the client participant model is a fully connected layer, then directly extract its weight matrix as the basic parameter feature; or, If the first layer of the client participant model is a convolutional layer, the three-dimensional convolution kernel is expanded into a feature vector along the spatial dimension and stacked in the order of the output channel dimension to form a two-dimensional parameter matrix; or, If the first-layer network of the client participant model is a Transformer module, its core components are parameter deconstructed and a joint parameter representation is formed through matrix expansion and vertical splicing; the core components include: the query matrix, key matrix, value matrix in the multi-head attention mechanism, and the double-layer weight matrix of the feedforward neural network.

[0017] The first-level features of all client participants are standardized and then a one-dimensional feature observation tensor is formed through a linear dimensionality reduction operation.

[0018] Furthermore, the step S101: calculating the kernel similarity between any two client participants based on the local model feature observation tensors of the client participants; and determining a semi-positive definite similarity kernel matrix based on the kernel similarity between the client participants, specifically includes: Assuming the client party The feature observation tensor of , assuming that the client party The feature observation tensor of , the kernel similarity between the two parties Defined as:

[0019] in, It is the width parameter of the function, which controls the radial range of the function.

[0020] The pairwise similarity calculation results are organized according to the matrix topology to form a kernel matrix reflecting the Euclidean distance distribution in the global parameter space between models. :

[0021] in, is the total number of client participants. The spectral characteristics of the matrix reveal the similarity topological relationship of the client participant model in the data space.

[0022] It should be understood that the similarity calculation stage adopts a kernel fusion strategy: a feature space mapping is constructed based on the radial basis kernel function (RBFKernel), and the kernel similarity between the observation tensor of a specific client participant and the observation tensors of the other parties is calculated respectively, and the nonlinear matching characteristics of the parameter distribution pattern are captured through kernel techniques.

[0023] In the iterative process of federated learning, the server adopts a dynamic client participant selection mechanism (only aggregating the model updates of some nodes in each round), resulting in the generated kernel matrix The kernel matrix exhibits dynamic sparsity. This sparsity is primarily manifested by insufficient overlap between client participant combinations between non-consecutive rounds, resulting in a sparse and low-rank distribution of effective similarity calculation units in the kernel matrix. This characteristic fundamentally conflicts with the mathematical premise of the determinant point process sampling algorithm, which requires the kernel matrix to be full-rank and semi-positive definite. Consequently, existing sparse kernel matrices cannot directly meet the DPP technical requirements for matrix reversibility and eigenvalue distribution.

[0024] Therefore, this embodiment converts the sparse kernel matrix into Convert to a full rank semi-positive definite matrix .

[0025] Furthermore, the S102: the central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix, including: using an alternating direction multiplier algorithm to estimate the missing values.

[0026] Furthermore, the missing value estimation is performed using the alternating direction multiplier algorithm, specifically including: First define The partial observation PSD matrix of , ,in, represents the total number of participants, is the matrix rank constraint.

[0027] Then, create an index set Matrix projection of:

[0028] in, is the known element coordinate set; The goal is to recover a complete PSD matrix. By introducing nuclear norm relaxation, the matrix completion problem is transformed into a nuclear norm optimization problem: , in express No. b Large singular values, is the noise tolerance threshold.

[0029] Because the kernel matrix The semi-positive definite matrix characteristics of , the final trace norm model with semi-positive definite constraints is constructed as:

[0030]

[0031] Construct a nuclear norm regularized linear least squares model with semi-positive definite constraints:

[0032] in, is the noise level parameter.

[0033] By introducing the splitting variable and impose constraints , we get the equivalent optimization problem:

[0034] The augmented Lagrangian function of the equivalent optimization problem is defined as:

[0035] in is the Lagrange multiplier parameter, is the penalty coefficient.

[0036] Use alternating direction multipliers on variables and Perform alternating minimization in the The update rule in the iteration is:

[0037]

[0038]

[0039] in, is a constant term.

[0040] By refactoring the optimization items, No. The iterative update can be expressed as:

[0041] Among them, the auxiliary matrix Defined as ; For the observation area and unobserved areas The separation process will The optimization problem is decomposed into two sub-problems, and the corresponding variable update rules are derived; for Observable entries of ; for The unobservable items of ; in, ; Through multiple rounds of iteration, the matrix completion algorithm will Convert to .

[0042] The matrix completion adopted in this patent ensures the physical interpretability of the matrix, and the implementation process is compatible with the dynamic client selection mechanism in federated learning.

[0043] Furthermore, the S103: the central server generates a calculation time series attenuation coefficient of each client participant's local model according to the update timeliness of all client participants' local models, specifically including: Based on the timeliness of the client participant's local model update, define the client participant In the The time decay coefficient of the round The recursive update rule is:

[0044] Among them, Boolean variables Represents the client party In the The selection state of the round, when When , it means the client party Selected to participate in model aggregation in this round.

[0045] The update mechanism of the recursive update rule is: (1) Reset rules: If the client party In the The round is selected (i.e. ),That Reset to the baseline value 1, reflecting that the client has completed its contribution to the global model through local model updates 2; (2) Incremental rule: If the client is not selected (i.e. ),That The value of the previous round is incremented by 1 to quantify the cumulative number of rounds in which the client has not been selected.

[0046] By continuously tracking the consecutive rounds in which each client did not participate in aggregation through recursive update rules, the global model convergence efficiency is ensured when the average AoU value of each client falls below a set threshold. Clients with AoU values ​​above the set threshold are given priority in subsequent training rounds. By dynamically adjusting the selection probability, the problem of accumulated deviation between the local model and the global model caused by long-term non-participation in aggregation is suppressed.

[0047] Furthermore, the step S103: constructing a timeliness weight diagonal matrix based on the time series attenuation coefficient specifically includes: The timeliness weight diagonal matrix is A round is defined as:

[0048] The update timeliness matrix and the similarity kernel matrix perform a dot product operation to generate a dynamically weighted similarity evaluation matrix, and further optimize client selection through a determinant point process sampling mechanism, thereby suppressing the update of outdated models while ensuring that the clients participating in the aggregation have both high model relevance and data distribution diversity.

[0049] Get all federated learning client participants Uploaded model parameters, use the uploaded model first-layer parameter matrix for likelihood local data statistics, and construct the local model feature observation tensor ; Calculate client participants based on radial basis function The similarity between the observation tensor of the local model and the tensor of the models of other client participants, and generate a semi-positive definite similarity kernel matrix ; Estimating the similarity kernel matrix based on the Alternating Direction Method of Multipliers (ADMM) matrix completion algorithm The similarity between the missing local models in the kernel matrix is ​​obtained ; According to the client participants Update the timeliness index (Age of Update, AoU), generating the calculation time series decay coefficient of each local model , construct the timeliness weight diagonal matrix ; The kernel matrix and timeliness matrix Coupling is the likelihood matrix , solve the optimal subset through the determinantal point process (DPP) , select the subset of client participants with the most data diversity ,in The number of client participants selected for each round. Based on the results of the client participant selection algorithm, a subset of the selected client participants is selected. Perform weighted aggregation to obtain the global model under the new parameters.

[0050] Figure 1 This is a flow chart of the adaptive client participant selection algorithm based on the determinant point process of the embodiment of the present disclosure. The algorithm takes the first-layer parameters of the client local model as input, and constructs a similarity kernel matrix in the multi-dimensional feature space through the Gaussian radial basis kernel function. During the construction process, the algorithm specifically integrates a data freshness evaluation module, which dynamically constructs a timeliness weight diagonal matrix by calculating the time decay factor, and finally generates a likelihood matrix. In the eigenvalue decomposition stage, the system deconstructs the likelihood matrix and extracts the principal component eigenvectors that represent the diversity of the data. Based on the dynamically constructed feature space, the algorithm adopts a progressive screening strategy: after initializing the empty set, the comprehensive selection probability of each node is calculated according to the probability density distribution of the eigenvector, and the client that maximizes the determinant value is iteratively selected to join the set until the preset client participant size threshold is met.

[0051] Figure 2 In the federated learning framework of the embodiment of the present disclosure, Diagram of the architecture of the dynamic client participant screening system. This architecture, centered around global model iteration, constructs a full-chain federated learning mechanism encompassing "data perception - association modeling - intelligent screening - targeted aggregation." The system uses a similarity calculation module deployed on client participants to capture the gradient distribution characteristics of each client's first-layer network parameters in real time and construct a cross-device weight similarity matrix based on the radial basis kernel function. The server then reconstructs the global similarity matrix using a matrix completion module. The timeliness calculation module dynamically fuses historical update records with current observations using a time decay factor to form a likelihood matrix with enhanced timeliness. During the screening phase, a k-determinant point process is used to perform spectral decomposition on the completed matrix, maximizing the diversity of the client subset sampled. Finally, a differentiated model fusion strategy is implemented through a targeted weight edit box, ensuring that the selected clients both characterize the data distribution characteristics and mitigate the negative impact of outdated updates on the global model.

[0052] The construction of a trusted federated computing environment is achieved through the following technological innovations: First, a Gaussian kernel space mapping technique is used to establish a similarity measurement system for cross-client model parameters. This is combined with a matrix completion mechanism driven by the alternating direction multiplier method to effectively solve the problem of correlation matrix reconstruction in sparse networks. A time-sensitive decay operator is then integrated to dynamically quantify the timeliness of model updates through an exponential weight decay function. Finally, a multi-objective optimization framework is constructed based on a determinant point process, encoding data distribution diversity and model freshness indicators into a probability mass function to achieve optimal subset selection of client participants under global convergence constraints. This embodiment provides a client participant decision-making solution for the federated learning system that is both robust and verifiable.

[0053] The likelihood matrix is ​​constructed based on the similarity kernel matrix and the timeliness weight diagonal matrix. -The determinant point process solves the optimal subset, selects the subset of client participants with the most data diversity, and aggregates the global model based on the results.

[0054] Furthermore, the S104: the central server constructs a likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; specifically includes: In a certain round of federated learning, the kernel similarity matrix and timeliness weight matrix Perform weighted Hadamard product operations to generate dynamically adjusted likelihood matrices ,in is the total number of client participants, is defined as a temporally weighted positive semidefinite kernel matrix:

[0055] The definition of the likelihood matrix contains a dual adjustment mechanism: by suppressing The matrix suppresses the weight of clients whose AoU is higher than the set threshold and The matrix preserves the similarity characteristics of the model parameters.

[0056] The goal is to select the cardinality from A subset of clients , such that:

[0057] in, express The corresponding subset The principal submatrix of the rows and columns of .

[0058] Next, proceed -Calculation of the determinant point process, where The number of client participants selected in each round.

[0059] Furthermore, the step S104: solving the likelihood matrix through a determinant point process to obtain a selected subset of client participants specifically includes: S104-1: Yes Perform eigenvalue decomposition:

[0060] in are the eigenvalues ​​in descending order, For the corresponding eigenvectors, forming an orthogonal basis set .

[0061] Secondly, based on the eigenvalue construction - Feature Set : The first step is to calculate the normalized probability distribution:

[0062] The second step is to generate a Index collection ,satisfy:

[0063] The third step is to extract the corresponding feature vector:

[0064] S104-2: Iterative sampling of feature vectors: right Perform iterative sampling, that is, construct an orthogonal basis projection space and select clients one by one: The first step is to initialize the auxiliary set: let the remaining candidate set , the set of selected client participants .

[0065] The second step is to cycle through the to Round: (1) Calculate each candidate client participant The projection probability of :

[0066] in, For client parties The one-hot encoded vector of .

[0067] (2) According to probability distribution Sampling client participants and join .

[0068] (3) From Remove All eigenvectors of the same vector:

[0069] S104-3: Output selected client participant subset .

[0070] Furthermore, averaging all local model parameters in the selected subset of client participants to obtain a global model specifically includes: Finally, a weighted aggregation process is performed based on the weights to obtain the global model: ,in For client parties The local model.

[0071] In response to the technical pain points of the widespread existence of untrusted nodes and serious data silos in the Industrial Internet of Things, an implicit selection scheme for fusion model parameter similarity analysis and timeliness compensation is proposed. The method includes: constructing a radial basis kernel similarity matrix through the first-layer model parameters of the client, and using the alternating direction multiplier method to achieve dynamic topology matrix completion to reconstruct the global data distribution; establishing an exponential decay weight model based on the update timeliness index to quantify the impact of the node model obsolescence on the aggregation process; coupling the completion kernel matrix and the timing decay matrix to construct a composite likelihood matrix, and screening the optimal client participant subset with both data diversity and model freshness based on the determinant point process sampling mechanism. The present invention realizes the screening of the local model of the client participant in the zero-perception state of the client participant, improves the model convergence efficiency in the non-independent and identically distributed data scenario, and at the same time reduces the negative impact of obsolete updates on the global model through the timeliness weight compensation mechanism, providing a safe and efficient federated learning solution for complex environments such as the Industrial Internet of Things.

[0072] In actual industrial scenarios, such as cross-regional power grid load forecasting scenarios, substations distributed in different cities need to collaborate to train short-term electricity consumption models through federated learning. The local data of each substation has strong spatiotemporal heterogeneity (for example, the load characteristics of industrial areas and residential areas are significantly different), and edge nodes are often offline due to network fluctuations. This embodiment can quantize the data distribution pattern through the feature observation tensor constructed by the first-layer parameters, and use radial basis kernel similarity to identify substation combinations with complementary characteristics; for the missing values ​​of the kernel matrix caused by node offline, the ADMM matrix completion technology is used to reconstruct the global similarity topology; at the same time, the freshness of the global model is dynamically guaranteed based on the AoU time series attenuation coefficient. While ensuring the diversity of data distribution, this mechanism significantly suppresses the negative impact of updating outdated models, and does not require the client to actively report the operating status throughout the process, which complies with the power system safety specifications and edge computing resource constraints.

[0073] Example 2 This embodiment provides a federated learning participant selection system, including: a central server and several client participants; The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; calculates the kernel similarity between the client participants based on the local model feature observation tensors of any two client participants; and determines the semi-positive definite similarity kernel matrix based on the kernel similarity between the client participants; The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; The central server generates a calculation time decay coefficient for each client participant's local model based on the update timeliness of all client participants' local models; based on the time decay coefficient, a timeliness weight diagonal matrix is ​​constructed; The central server constructs a likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; solves the likelihood matrix through the determinant point process to obtain the selected subset of client participants; takes the average of all local model parameters in the selected subset of client participants to obtain the global model; finally, sends the global model to the selected client participants.

[0074] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for selecting federated learning participants, characterized by: include: The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; Based on the local model feature observation tensors of any two client participants, the kernel similarity between the client participants is calculated; based on the kernel similarity between the client participants, a semi-positive definite similarity kernel matrix is ​​determined; The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; The central server generates the calculation time series attenuation coefficient of each client participant’s local model based on the update timeliness of all client participants’ local models; Based on the time series decay coefficient, a diagonal matrix of timeliness weights is constructed; The central server constructs the likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; The likelihood matrix is ​​solved through a determinant point process to obtain a selected subset of client participants; all local model parameters in the selected subset of client participants are averaged to obtain a global model; finally, the global model is sent to the selected client participants.

2. A method for selecting federated learning participants according to claim 1, characterized in that: The central server obtains the local model parameters uploaded by all federated learning client participants and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters, including: The first-layer network parameters of each client participant model are extracted as the core feature carrier: If the first layer of the client participant model is a fully connected layer, then directly extract its weight matrix as the basic parameter feature; or, If the first layer of the client participant model is a convolutional layer, the three-dimensional convolution kernel is expanded into a feature vector along the spatial dimension and stacked in the order of the output channel dimension to form a two-dimensional parameter matrix; or, If the first layer of the client-side model is a Transformer module, its core components are parameterized and a joint parameter representation is formed through matrix expansion and vertical concatenation. The core components include the query matrix, key matrix, and value matrix in the multi-head attention mechanism, and the two-layer weight matrix of the feedforward neural network. The first-level features of all client participants are standardized and then a one-dimensional feature observation tensor is formed through a linear dimensionality reduction operation.

3. A method for selecting federated learning participants according to claim 1, characterized in that: Calculate the kernel similarity between any two client participants based on their local model feature observation tensors; Based on the kernel similarity between client participants, a semi-positive definite similarity kernel matrix is ​​determined, specifically including: Assuming the client party The feature observation tensor of , assuming that the client party The feature observation tensor of , the kernel similarity between the two parties Defined as: ； in, is the width parameter of the function; The pairwise similarity calculation results are organized according to the matrix topology to form a kernel matrix reflecting the Euclidean distance distribution in the global parameter space between models. : ； in, The total number of client parties.

4. A method for selecting federated learning participants according to claim 1, characterized in that: The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix, including: using an alternating direction multiplier algorithm to estimate the missing values.

5. A method for selecting federated learning participants according to claim 1, characterized in that: The central server generates a calculation time series attenuation coefficient for each client participant's local model based on the update timeliness of all client participants' local models, specifically including: Based on the timeliness of the client participant's local model update, define the client participant In the The time decay coefficient of the round The recursive update rule is: ； Among them, Boolean variables Represents the client party In the The selection state of the round, when When , it means the client party Selected to participate in model aggregation in this round.

6. A method for selecting federated learning participants according to claim 1, characterized in that: Based on the time series decay coefficient, a timeliness weight diagonal matrix is ​​constructed, which includes: The timeliness weight diagonal matrix is A round is defined as: 。 7. A method for selecting federated learning participants according to claim 1, characterized in that: The central server constructs the likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; Specifically include: In a certain round of federated learning, the kernel similarity matrix and timeliness weight matrix Perform weighted Hadamard product operations to generate dynamically adjusted likelihood matrices ,in is the total number of client participants, is defined as a temporally weighted positive semidefinite kernel matrix: ； The definition of the likelihood matrix contains a dual adjustment mechanism: by suppressing The matrix suppresses the weight of clients whose AoU is higher than the set threshold and The matrix preserves the similarity characteristics of the model parameters; The goal is to select the cardinality from A subset of clients , such that: ； in, express The corresponding subset The principal submatrix of the rows and columns of .

8. The method for selecting federated learning participants according to claim 1, wherein: Solve the likelihood matrix through the determinant point process to obtain the selected client participant subset; specifically, S104-1: Yes Perform eigenvalue decomposition: ； in are the eigenvalues ​​in descending order, For the corresponding eigenvectors, forming an orthogonal basis set ; Secondly, based on the eigenvalue construction - Feature Set : The first step is to calculate the normalized probability distribution: ； The second step is to generate a Index collection ,satisfy: ； The third step is to extract the corresponding feature vector: ； S104-2: Iterative sampling of feature vectors: right Perform iterative sampling, that is, construct an orthogonal basis projection space and select clients one by one: The first step is to initialize the auxiliary set: let the remaining candidate set , the set of selected client parties ; The second step is to cycle through the to Round: (1) Calculate each candidate client participant The projection probability of : ； in, For client parties The one-hot encoded vector of ; (2) According to probability distribution Sampling client participants and join ; (3) From Remove All eigenvectors of the same vector: ； S104-3: Output selected client participant subset .

9. A method for selecting federated learning participants according to claim 1, characterized in that: The method of averaging all local model parameters in the selected subset of client participants to obtain a global model specifically includes: Perform a weighted aggregation process based on the weights to obtain the global model: ,in For client parties The local model.

10. A federated learning participant selection system, characterized by: include: A central server and several client parties; The central server obtains the local model parameters uploaded by all federated learning client participants, and constructs the local model feature observation tensor of each client participant based on the uploaded local model first-layer parameters; Based on the local model feature observation tensors of any two client participants, the kernel similarity between the client participants is calculated; based on the kernel similarity between the client participants, a semi-positive definite similarity kernel matrix is ​​determined; The central server completes the missing values ​​of the semi-positive definite similarity kernel matrix to obtain a complete similarity kernel matrix; The central server generates the calculation time series attenuation coefficient of each client participant’s local model based on the update timeliness of all client participants’ local models; Based on the time series decay coefficient, a diagonal matrix of timeliness weights is constructed; The central server constructs the likelihood matrix based on the complete similarity kernel matrix and the timeliness weight diagonal matrix; The likelihood matrix is ​​solved through a determinant point process to obtain a selected subset of client participants; all local model parameters in the selected subset of client participants are averaged to obtain a global model; finally, the global model is sent to the selected client participants.

Images