Federal data selection method and device based on proxy verification data set

Abstract

The invention discloses a federal data selection method and device based on a proxy verification data set, and the method comprises the steps: a central server selects participated clients according to the selection probability of the clients, and transmits a prediction model and meta-model parameters to the clients; the client calculates a sample and a client weight by using the meta-model, selects a training sample to update a prediction model according to the sample weight, and uploads the client weight and the updated prediction model to the central server; the client with the highest weight updates the meta-model by using the proxy verification data set, and uploads the updated meta-model to the central server; the client updates the proxy verification data set by using a meta-marginal function; and after receiving all the prediction model parameters, the central server carries out parameter aggregation to obtain an updated prediction model, and updates the client selection probability to carry out the next round of model training. According to the method, the dependence on an extra unbiased verification set is eliminated, and the calculation efficiency is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of federated learning, and particularly relates to a federated data selection method and device based on proxy verification data sets. BACKGROUND

[0002] Federated learning has been widely applied in many practical fields, including healthcare, city computing, and recommendation systems. Unlike traditional solutions that rely on centralized data storage and processing, federated learning builds a predictive model on the server side by aggregating predictive models trained on multiple clients (i.e., data owners), thereby ensuring the privacy of sensitive data. However, this decentralized system architecture also has problems such as slow convergence speed of the predictive model and large resource overhead (i.e., computation and communication cost). This is because each client must perform multiple local computations using private data and frequently synchronize with the central server. This problem is particularly prominent when there are a large number of clients and the quality of data varies greatly.

[0003] As a promising direction, data selection methods can improve the effectiveness and efficiency of federated learning by reducing invalid or redundant training data. They build predictive models by selecting important samples from the entire training dataset, thereby reducing resource overhead. This approach not only speeds up the training process but also does not compromise the accuracy of the model. In the centralized learning scenario, there are many data selection methods. Since these methods require sharing private user data, they cannot be directly applied to the federated learning scenario because they cannot meet the strict efficiency requirements in the federated learning scenario.

[0004] Existing federated data selection methods are mainly divided into client selection and sample selection. Specifically, client selection is coarse-grained and suboptimal because it treats all samples from selected clients as equally important. Sample selection, on the other hand, is fine-grained, scoring each sample and selecting samples based on a pre-defined threshold or benchmark model to reduce resource overhead and further improve performance. Despite this, existing federated sample selection methods still have the following two limitations.

[0005] First, existing federated sample selection methods ignore the fact that the importance of samples in the training process changes. For example, in the early training stage, the model has not yet fully learned the underlying features of the data, so even seemingly simple samples can have an important impact on the model. However, as the model gradually captures key features, the impact of these samples may decrease. Conversely, samples that are consistently misclassified or rarely appear may become more important in subsequent rounds. In addition, the training data owned by data owners may be affected by noise (e.g., label noise), which can affect the learning process of the model and hinder the development of high-quality models.

[0006] Secondly, many federated sample selection methods highly rely on external unbiased validation datasets. On one hand, it is very difficult to obtain comprehensive validation datasets in practical application scenarios due to various factors such as application knowledge and budget constraints. On the other hand, additional information (e.g., average gradient or performance difference) from the validation dataset is shared with the clients to guide the selection process and ensure consistent model convergence. However, these shared parameters increase the risk of privacy leakage. SUMMARY

[0007] To overcome the above technical problems, the application implements and provides a federated data selection method and device based on proxy validation datasets, which eliminates the dependence on additional unbiased validation datasets while considering dynamic data importance.

[0008] According to a first aspect of an embodiment of the application, a federated data selection method based on proxy validation datasets is provided, comprising: S1: a center server selects participating clients according to client selection probabilities, and distributes prediction model parameters and meta-model parameters to the selected clients; S2: the clients calculate samples and client weights using the meta-model, convert sample weights into sample selection probabilities, select training samples according to the sample selection probabilities to update the prediction model, and upload the client weights and updated prediction model parameters to the center server; S3: the client with the highest weight updates the meta-model using a proxy validation dataset, and uploads the updated meta-model parameters to the center server; S4: the clients update the local sample partition number dictionary using a momentum-based meta-marginal function, and update the proxy validation dataset according to the updated local sample partition number dictionary; S5: after the center server receives the updated meta-model parameters and prediction model parameters, it performs parameter aggregation on the prediction model parameters to obtain a globally updated prediction model, normalizes the client weights to update the client selection probabilities; S6: continue to repeat S1-S5 until the convergence condition is reached, and the clients obtain the optimal prediction model to output the final prediction result.

[0009] According to a second aspect of an embodiment of the application, a federated data selection device based on proxy validation datasets is provided, comprising: a center server client selection module configured to select participating clients according to client selection probabilities by a center server, and distribute prediction model parameters and meta-model parameters to the selected clients; a local prediction model parameter updating module, configured to calculate samples and client weights by using the meta model, convert the sample weights into sample selection probabilities, select training samples according to the sample selection probabilities to update the prediction model, and upload the client weights and the updated prediction model parameters to the central server; a local meta model parameter updating module, configured to update the meta model by using the proxy validation dataset by the client with the highest weight, and upload the updated meta model parameters to the central server; a local proxy validation dataset updating module, configured to update the local sample partition number dictionary by using the momentum-based meta marginal function by the client, and update the proxy validation dataset according to the updated local sample partition number dictionary; a central server parameter aggregation module, configured to, after the central server receives the updated meta model parameters and the prediction model parameters, perform parameter aggregation on the prediction model parameters to obtain a globally updated prediction model, and update the client selection probabilities by normalizing the client weights; a local prediction module, configured to continue to repeat the central server client selection module-central server parameter aggregation module until a convergence condition is reached, the client obtains an optimal prediction model, and finally output a prediction result.

[0010] According to a third aspect of the embodiments of the present application, a federated data selection method based on a proxy validation dataset is provided, applied to a central server, and including: A1: The central server selects participating clients according to the client selection probabilities, and distributes the prediction model parameters and the meta model parameters to the selected clients, so that the clients calculate samples and client weights by using the meta model, convert the sample weights into sample selection probabilities, select training samples according to the sample selection probabilities to update the prediction model, and upload the client weights and the updated prediction model parameters to the central server; the client with the highest weight updates the meta model by using the proxy validation dataset, and uploads the updated meta model parameters to the central server; the client updates the local sample partition number dictionary by using the momentum-based meta marginal function, and updates the proxy validation dataset according to the updated local sample partition number dictionary; A2: After the central server receives the updated meta model parameters and the prediction model parameters, the central server performs parameter aggregation on the prediction model parameters to obtain globally updated prediction model parameters, and updates the client selection probabilities by normalizing the client weights; A3: Continue to repeat A1-A2 until a convergence condition is reached, the client obtains an optimal prediction model, and finally output a prediction result.

[0011] According to a fourth aspect of the embodiments of the present application, a federated data selection method based on a proxy validation dataset is provided, applied to a client, and including: B1: The client receives the prediction model parameters and the meta-model parameters sent by the central server; B2: The client calculates the sample and the client weight using the meta-model, converts the sample weight into a sample selection probability, selects a training sample according to the sample selection probability to update the prediction model, and uploads the client weight and the updated prediction model parameters to the central server, so that after the central server receives the updated meta-model parameters and the prediction model parameters, the prediction model parameters are aggregated to obtain global updated prediction model parameters, and the client weight is normalized to update the client selection probability; B3: The client with the highest weight updates the meta-model using the proxy validation dataset, and uploads the updated meta-model parameters to the central server; B4: The client updates the local sample partition number dictionary using the momentum-based meta-marginal function, and updates the proxy validation dataset according to the updated local sample partition number dictionary; B5: Continue to repeat B1-B4 until the convergence condition is reached, and the client obtains the optimal prediction model, thereby outputting the final prediction result.

[0012] According to a fifth aspect of an embodiment of the present application, an electronic device is provided, comprising: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of the first aspect.

[0013] The technical scheme provided by the embodiments of the present application can include the following beneficial effects: Compared with the prior art, the improvement of the present application is that a federated sample selection framework based on proxy validation dataset and importance perception is proposed, which selects according to the importance change of the sample, and at the same time eliminates the dependence on external validation dataset; a momentum-based meta-marginal function is proposed, which is used to identify influential samples as proxy validation dataset to provide adaptive rewards for sample selection; a meta-model updating mechanism is introduced to ensure efficient and robust federated sample selection; compared with the current federated data selection method, the downstream task precision is improved by 5.39%, and the calculation efficiency is also high. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A flowchart of a federated data selection method based on a proxy validation dataset is provided for the embodiments of the present application.

[0015] Figure 2 A training block diagram of a federated data selection method based on a proxy validation dataset is provided for the embodiments of the present application.

[0016] Figure 3 The average calculation cost and accuracy comparison provided by the CIFAR-10 dataset of the embodiment of the present application.

[0017] Figure 4 The block diagram of the federated data selection device based on proxy validation dataset provided by the embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the specific contents of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application. The contents not described in detail in the embodiments of the present application are the prior art known by those skilled in the art.

[0019] The existing federated data selection methods are mainly divided into client selection and sample selection. Specifically, the client selection is coarse-grained and suboptimal because it regards all samples from the selected client as equally important. The sample selection is fine-grained, which selects samples according to a pre-defined threshold or a benchmark model by scoring each sample, thereby reducing resource overhead and further improving performance. However, the existing federated sample selection methods still have the following two limitations. First, the existing federated sample selection methods ignore that the importance of samples in the training process is variable. Second, many federated sample selection methods are highly dependent on external unbiased validation datasets.

[0020] In order to overcome the above technical problems, the embodiments of the present application provide a federated data selection method and device based on proxy validation dataset, which eliminates the dependence on additional unbiased validation datasets while considering the importance of dynamic data.

[0021] The federated data selection method based on proxy validation dataset provided by the present application can be applied to almost all scenarios of federated learning prediction tasks, such as image and text classification tasks. In these scenarios, the client needs to locally maintain a local sample weight dictionary, use a momentum-based meta-marginal function to select a proxy validation dataset as an external signal to update the meta-model; the client calculates the sample weight using the meta-model, selects the data of the sample according to the weight, and updates the prediction model using the selected sample; the center server updates the client selection probability while aggregating the model parameters, and uses the selection probability for the next round of client selection. The method is proved to be effective and efficient through experimental verification.

[0022] REFERENCE Figure 1 and Figure 2The federal data selection method based on proxy verification datasets provided in this application includes the following steps: S1: The central server selects participating clients based on the client selection probability and distributes the prediction model parameters and meta-model parameters to these selected clients. Specifically, the prediction model The model structure is not fixed; different base models can be selected based on downstream tasks; the meta-model It consists of a multilayer perceptron network, with the input being the first... i Sample x i loss value The output is the corresponding sample weight. w i , represented as ,in These are the parameters of the prediction model. It is a predicted value. It is a loss function. y i For the first i Sample x i The tag value, These are the model parameters of the meta-model; the reason why the meta-model is a multilayer perceptron network is that the simple network structure can realize the mapping from loss values ​​to sample weights, thereby providing a basis for the selection of clients and samples.

[0023] S2: The client uses the meta-model to calculate sample and client weights, converts the sample weights into sample selection probabilities, selects training samples based on the sample selection probabilities to update the prediction model, and uploads the client weights and updated prediction model parameters to the central server; this step may include the following sub-steps: S21: Calculate the sample and client weights using the meta-model, and convert the sample weights into corresponding selection probabilities; Specifically, the first i Sample x i Input into the prediction model The predicted value is obtained from the data, and this predicted value can be expressed as: ,in, This represents the parameters of the prediction model; the loss value is obtained by comparing the predicted value with the actual value. ， Represented as The loss values ​​of the samples are input into the meta-model to obtain the corresponding sample weights. w i , represented as ,in, This represents the meta-model parameters. The first... iThe sample weight of each sample is converted into a corresponding sample selection probability, i.e. This is because the greater the weight, the greater the importance of the sample in the current round of collaborative training of the model, and therefore the greater the probability of the sample being selected. According to the sample selection probability, a high-quality selected sample set can be constructed.

[0024] The client weight is the average of all sample weights, i.e. wherein n k represents the total number of samples owned by the local client. The reason for using the average of all sample weights as the client weight is that the number of samples of each client can be different, and the average of the sample weights reduces the influence of the different number of samples on the calculation result of the client weight.

[0025] S22: Select training samples to update the prediction model according to the sample selection probability, and upload the client weight and the updated prediction model parameters to the center server.

[0026] Specifically, a fixed number of training samples are selected according to the sample selection probability to form a selected sample set, denoted as , and the local T l round of prediction model training is performed to update the prediction model parameters, specifically as follows: wherein t l is the current local training round, t l , the value of T l -1], represents a step hyperparameter, y is a label value, represents the number of samples | in the selected sample set.

[0027] S3: The client with the highest weight updates the meta-model using the proxy validation dataset, and uploads the updated meta-model parameters to the center server; this step can include the following sub-steps: S31: Select the client with the highest weight according to the weight of each client; Specifically, in each round of collaborative training, the client with the highest weight k is selected to update the meta-model using its proxy validation dataset, specifically as follows: wherein represents the weight of the k th client, represents the weight of thei weight of a sample, K represents the set of all client indices. This is because the update of the meta-model relies on a high-quality proxy validation dataset, and the higher the client weight, the higher the quality of the client's local dataset, which can provide a high-quality proxy validation dataset, S32: The selected client with the highest weight updates the meta-model using the proxy validation dataset and uploads the updated meta-model parameters to the central server.

[0028] Specifically, the selected client updates the meta-model 𝝎 using an online optimization strategy through a single optimization cycle, as follows: where, t c represents the number of cooperative training rounds, represents two step hyperparameters, n and m are the batch sizes of the training dataset and the proxy validation dataset, respectively, and represent the sample and the corresponding label in the proxy validation dataset, respectively. This is because traditional meta-learning methods require second-order backpropagation, which has a time overhead three times that of single-step gradient updates. Using an online optimization strategy to update the meta-model through a single optimization cycle can ensure efficiency.

[0029] S4: The client updates the local sample frequency dictionary based on the momentum-based meta-margin function and updates the proxy validation dataset according to the updated local sample frequency dictionary; this step can include the following sub-steps: S41: The client independently maintains a local sample frequency dictionary and updates the local sample frequency dictionary using a momentum-based meta-margin function; Specifically, each client k independently maintains a local sample frequency dictionary , which can be represented as: where, represents the k th sample and the corresponding label of the n k th client, represents the i th sample and the corresponding label of the t c the sample frequency of the cooperative training round, when t c =0, . The calculation process of the sample frequency is as follows.

[0030] First, the meta-marginal score is calculated using the loss difference between two consecutive collaboration training rounds (i.e. t c and t c+1 , which is given by: where denotes the client's k predicted model parameters based on the model parameters and the selected samples, i.e. t c +1 collaboration round.

[0031] The meta-marginal score is used because it focuses on those training samples with low recognition. Even if these samples have high current loss values, they will still be selected preferentially in subsequent updates as long as the loss value can be significantly reduced. In contrast, samples with small loss values will have low meta-marginal values, so they will not be included in the proxy validation set because the prediction model has already learned the characteristics of these samples well. In addition, for noisy data that deviates from the target data distribution, their loss values will not be minimized, and may even increase after updating. In this case, this mechanism can effectively reduce the priority of these samples to avoid their selection. To further utilize a series of historical model states to obtain a robust estimate for each sample, a momentum-based meta-marginal function is used to fully utilize the historical sample scores, which is given by: where is the momentum factor, which is a hyperparameter that controls the weight ratio between the historical model state and the current state in the estimation; and the local sample score dictionary is updated according to the calculated estimate.

[0032] S42: The client updates the proxy validation dataset according to the local sample score dictionary.

[0033] Specifically, it is also believed that the higher the calculated sample score, the more important the sample is to the current model training. Therefore, the momentum-based meta-marginal score is maximized to obtain the updated proxy validation dataset, which is given by: where the proxy validation dataset contains a fixed number of data samples, which is a hyperparameter; denotes the local dataset of the k th client.

[0034] S5: After the center server receives the updated meta-model parameters and the prediction model parameters, the prediction model parameters are aggregated to obtain a global updated prediction model, and the client weight is normalized to update the client selection probability; this step can include the following sub-steps: S51: After the center server receives all the updated prediction model parameters, the parameters are aggregated to obtain updated prediction model parameters; Specifically, after the center server receives all the selected client updated prediction model parameters, the model parameters of all selected clients are averaged to obtain updated prediction model parameters, and the specific formula is as follows: wherein, represents the number of selected clients.

[0035] S52: The center server normalizes the client weight to update the client selection probability.

[0036] Specifically, after the center server receives the average weight of all clients, the updated client selection probability is obtained by normalization, so as to select the clients in the next round, and the specific formula is as follows: wherein, K represents the number of all clients, is the weight of the k th client in the t c +1 collaborative training round.

[0037] S6: Continue to repeat S1-S5 until the convergence condition is reached, and the client obtains the optimal prediction model, thereby outputting the final prediction result.

[0038] Specifically, the above five steps (i.e., S1-S5) are repeated for multiple collaborative training rounds until the convergence condition is reached, and the convergence condition includes reaching the maximum collaborative training round or the model accuracy reaching the artificial set accuracy requirement.

[0039] The federated data selection method based on the proxy validation dataset of the application is implemented on an Ubuntu 18.04 system running on an Intel core 2.80 GHz server with 192 GB of memory, and the number of clients is set to 10 in advance, the selection rate of the client is 0.4, the selection rate of the sample is 0.6, the maximum collaborative training rounds are 100 (MNIST, CIFAR-10 and AGNews dataset) and 200 (CIFAR-100 dataset), the heterogeneity degree is 0.6 (representing the number of sampled categories per client) and 4 scenarios (IID (1), IID (2), Non-IID (1), Non-IID (2)), and the average performance of the application embodiment on the four datasets. For model performance evaluation, the accuracy is used to measure the performance of the classification task, and the accuracy measures the proportion of correctly classified samples.

[0040] The performance of the federated data selection method based on the proxy validation dataset (i.e. FedSelect) proposed by the application is analyzed by experimental test, and the results are shown in Table 1 and Figure 3 It can be seen that for the four scenarios, FedSelect has a classification accuracy higher than the existing federated client-level data selection methods (FedAvg, SFedAvg and DELTA) and federated sample-level data selection methods (FedMR, FedBary, FedLoss, FedGrad and FedCSS, which need to rely on external unbiased validation datasets) on all indicators and datasets, with a maximum accuracy improvement of 5.39%, and also performs excellently in computational efficiency. This is because the method introduces a proxy validation dataset and a momentum-based meta-margin function, which can dynamically evaluate the importance of each sample, thereby reducing the dependence on additional validation datasets and effectively addressing the problem of noisy data. In addition, an efficient meta-model updating mechanism is also introduced.

[0041] Table 1: Comparison of average accuracy of all methods on all data and scenarios; Corresponding to the foregoing embodiments of the federated data selection method based on the proxy validation dataset, the application also provides embodiments of the federated data selection method based on the proxy validation dataset. Referring to Figure 4 The device comprises: A central server client selection module for selecting participating clients according to client selection probabilities and distributing prediction model parameters and meta-model parameters to the selected clients. a local prediction model parameter updating module, configured to cause the client to calculate samples and client weights by using the meta model, convert the sample weights into sample selection probabilities, select training samples according to the sample selection probabilities to update the prediction model, and upload the client weights and the updated prediction model parameters to the central server; a local meta model parameter updating module, configured to cause the client with the highest weight to update the meta model by using the proxy verification dataset, and upload the updated meta model parameters to the central server; a local proxy verification dataset updating module, configured to cause the client to update a local sample weight dictionary by using a momentum-based meta marginal function, and update the proxy verification dataset according to the updated local sample weight dictionary; a central server parameter aggregation module, configured to cause the central server to, after receiving the updated meta model parameters and the prediction model parameters, perform parameter aggregation on the prediction model parameters to obtain a global updated prediction model, and update client selection probabilities by normalizing the client weights; a local prediction module, configured to continue to repeat the central server client selection module-central server parameter aggregation module until a convergence condition is reached, the client obtains an optimal prediction model, and finally a prediction result is output.

[0042] For the device embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The device embodiment described above is only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. According to actual needs, some or all of the modules can be selected to achieve the purpose of the scheme of the present application. Those skilled in the art can understand and implement it without creative labor.

[0043] The embodiment of the present application also provides a federated data selection method based on a proxy verification dataset, applied to a central server, comprising: A1: The central server selects participating clients according to client selection probabilities, and distributes prediction model parameters and meta model parameters to the selected clients, so that the clients calculate samples and client weights by using the meta model, convert the sample weights into sample selection probabilities, select training samples according to the sample selection probabilities to update the prediction model, and upload the client weights and the updated prediction model parameters to the central server; the client with the highest weight updates the meta model by using the proxy verification dataset, and uploads the updated meta model parameters to the central server; the client updates a local sample weight dictionary by using a momentum-based meta marginal function, and updates the proxy verification dataset according to the updated local sample weight dictionary; A2: After the center server receives the updated meta-model parameters and prediction model parameters, the prediction model parameters are aggregated to obtain global updated prediction model parameters, and the client weight is normalized to update the client selection probability; A3: Continue to repeat A1-A2 until the convergence condition is reached, and the client obtains the optimal prediction model, thereby outputting the final prediction result.

[0044] The details of A1-A3 have been described in detail before, and will not be repeated here.

[0045] The embodiment of the application also provides a federated data selection method based on proxy verification data set, applied to a client, comprising: B1: The client receives the prediction model parameters and meta-model parameters sent by the center server; B2: The client calculates the sample and the client weight using the meta-model, converts the sample weight into a sample selection probability, and then selects the training sample according to the sample selection probability to update the prediction model, and uploads the client weight and the updated prediction model parameters to the center server, so that the center server receives the updated meta-model parameters and prediction model parameters, aggregates the prediction model parameters to obtain global updated prediction model parameters, and normalizes the client weight to update the client selection probability; B3: The client with the highest weight updates the meta-model using the proxy verification data set, and uploads the updated meta-model parameters to the center server; B4: The client updates the local sample partition number dictionary using the momentum-based meta-marginal function, and updates the proxy verification data set according to the updated local sample partition number dictionary; B5: Continue to repeat B1-B4 until the convergence condition is reached, and the client obtains the optimal prediction model, thereby outputting the final prediction result.

[0046] The details of B1-B5 have been described in detail before, and will not be repeated here.

[0047] Correspondingly, the application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the federated incomplete table data prediction method based on missing complementarity as described above.

[0048] Correspondingly, the application also provides a computer readable storage medium having computer instructions stored thereon, which instructions are executed by a processor to implement the federated incomplete table data prediction method based on missing complementarity as described above.

[0049] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

Claims

1. A federated data selection method based on a proxy validation dataset, characterized in that, include: S1: The central server selects participating clients based on the client selection probability and distributes the prediction model parameters and meta-model parameters to these selected clients. S2: The client uses the meta-model to calculate the sample and client weights, converts the sample weights into sample selection probabilities, selects training samples to update the prediction model based on the sample selection probabilities, and uploads the client weights and updated prediction model parameters to the central server. S3: The client with the highest weight uses the proxy to verify the dataset and update the metamodel, and then uploads the updated metamodel parameters to the central server. S4: The client updates the local sample score dictionary using a momentum-based meta-marginal function and updates the proxy validation dataset based on the updated local sample score dictionary; S5: After receiving the updated meta-model parameters and prediction model parameters, the central server performs parameter aggregation on the prediction model parameters to obtain the globally updated prediction model, and normalizes the client weights to update the client selection probability. S6: Continue repeating S1-S5 until the convergence condition is met, the client obtains the optimal prediction model, and thus outputs the final prediction result.

2. The method according to claim 1, characterized in that, The model structure of the prediction model is not fixed, and different base models are selected according to the downstream task; the meta-model consists of a multilayer perceptron network, with the loss value of the sample as the input and the weight of the sample as the output.

3. The method according to claim 1, characterized in that, S2 specifically includes: S21: Calculate the sample and client weights using the meta-model, and convert the sample weights into corresponding selection probabilities; S22: Select training samples based on the sample selection probability to update the prediction model, and upload the client weights and the updated prediction model parameters to the central server.

4. The method according to claim 1, characterized in that, S3 specifically includes: S31: Select the client with the highest weight based on the weight of each client; S32: The client with the highest selected weight uses the proxy to validate the dataset to update the metamodel and uploads the updated metamodel parameters to the central server.

5. The method according to claim 1, characterized in that, S4 specifically includes: S41: The client independently maintains a local sample score dictionary and uses a momentum-based metamarginal function to update the local sample score dictionary; S42: The client updates the proxy validation dataset based on the local sample score dictionary.

6. The method according to claim 1, characterized in that, S5 specifically includes: S51: After receiving all the updated prediction model parameters, the central server performs parameter aggregation to obtain the updated prediction model parameters. S52: The central server normalizes the client weights and updates the client selection probability.

7. A federated data selection device based on a proxy verification dataset, characterized in that, include: The central server client selection module is used by the central server to select participating clients based on the client selection probability, and to distribute the prediction model parameters and meta-model parameters to these selected clients. The local prediction model parameter update module is used by the client to calculate the sample and client weights using the meta-model, convert the sample weights into sample selection probabilities, select training samples to update the prediction model according to the sample selection probabilities, and upload the client weights and updated prediction model parameters to the central server. The local metamodel parameter update module is used by the client with the highest weight to update the metamodel using the proxy verification dataset and upload the updated metamodel parameters to the central server. The local proxy verification dataset update module is used by the client to update the local sample score dictionary using a momentum-based meta-marginal function, and update the proxy verification dataset based on the updated local sample score dictionary. The central server parameter aggregation module is used to aggregate the prediction model parameters after the central server receives the updated meta-model parameters and prediction model parameters to obtain the globally updated prediction model, and to normalize the client weights to update the client selection probability. The local prediction module is used to continue repeating the process of selecting the central server from the client and then aggregating the central server parameters until the convergence condition is met. The client then obtains the optimal prediction model and outputs the final prediction result.

8. A federated data selection method based on a proxy validation dataset, characterized in that, Applied to the central server, including: A1: The central server selects participating clients based on client selection probabilities and distributes prediction model parameters and meta-model parameters to these selected clients. This allows clients to use the meta-model to calculate sample and client weights, convert the sample weights into sample selection probabilities, select training samples based on these probabilities to update the prediction model, and then upload the client weights and updated prediction model parameters to the central server. The client with the highest weight updates its meta-model using the proxy validation dataset and uploads the updated meta-model parameters to the central server. The client updates its local sample score dictionary using a momentum-based meta-marginal function and updates the proxy validation dataset based on the updated local sample score dictionary. A2: After receiving the updated meta-model parameters and prediction model parameters, the central server performs parameter aggregation on the prediction model parameters to obtain the globally updated prediction model parameters, and normalizes the client weights to update the client selection probability. A3: Continue repeating A1-A2 until the convergence condition is met, the client obtains the optimal prediction model, and thus outputs the final prediction result.

9. A federated data selection method based on a proxy validation dataset, characterized in that, Applied to the client side, including: B1: The client receives the prediction model parameters and meta-model parameters sent by the central server; B2: The client uses the meta-model to calculate the sample and client weights, converts the sample weights into sample selection probabilities, selects training samples to update the prediction model based on the sample selection probabilities, and uploads the client weights and updated prediction model parameters to the central server. After receiving the updated meta-model parameters and prediction model parameters, the central server performs parameter aggregation on the prediction model parameters to obtain the globally updated prediction model parameters, and normalizes the client weights to update the client selection probability. B3: The client with the highest weight uses the proxy to verify the dataset and update the metamodel, and then uploads the updated metamodel parameters to the central server; B4: The client updates the local sample score dictionary using a momentum-based meta-marginal function and updates the proxy validation dataset based on the updated local sample score dictionary; B5: Continue repeating B1-B4 until the convergence condition is met, the client obtains the optimal prediction model, and thus outputs the final prediction result.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6, 8, and 9.

Images