A differential privacy federated learning method and system based on dual knowledge transfer

Abstract

The present application discloses a differential privacy federated learning method and system based on dual knowledge transfer, which relates to the field of machine learning, wherein the method includes: training a teacher model using a data subset; determining the teacher voting result of each sample using a weighted voting mechanism; determining the label of an unlabeled teacher data set based on the teacher voting result, and training a student model using a labeled teacher data set; determining the student voting result of each sample using a weighted voting mechanism in a medical server; determining the label of an unlabeled student data set based on the student voting result, and training a global model using a labeled student data set; fine-tuning the global model using an importance sampling mechanism and further self-training to obtain the final global model. The present application introduces a weighted voting mechanism in knowledge transfer, adjusts the model weights according to the individual differences of the teacher model, improves the model utility of federated learning, and also effectively reduces communication overhead.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a differential privacy federated learning method and system based on dual knowledge transfer. Background Art

[0002] Smart healthcare can generate important value in clinical research, public health, public services, and industrial development by conducting machine learning and mining on medical and health big data. However, medical data often involves user privacy information, which makes it difficult for various medical institutions to directly share original medical data, resulting in the formation of data silos between medical institutions, and the inability to bring out the full value of the data. Federated learning, as a distributed machine learning framework, can break down data silos without sharing original data, allowing multiple participants to jointly train machine learning models, and has gradually become an effective solution to data security and privacy issues.

[0003] Although federated learning effectively avoids the risk of data leakage by training locally and uploading model parameters, the contribution of uploaded model parameters may still expose sensitive information of participants, especially when data is unevenly distributed. To solve this problem, differential privacy, as a widely used privacy protection framework, has been introduced into federated learning. It introduces random noise to model updates to prevent the leakage of model sensitive information. However, the application of differential privacy usually affects the final training effect of the model. In addition, the high communication cost in federated learning and the problem of data heterogeneity also need to be solved. Therefore, it is necessary to design a new learning method that can optimize model utility, reduce federated learning communication overhead, and solve the problems caused by data heterogeneity while providing privacy protection. Summary of the invention

[0004] The embodiments of the present application provide a differential privacy federated learning method and system based on dual knowledge transfer, which are used to solve the problems of high communication cost and low model utility when facing heterogeneous data in federated learning in the prior art.

[0005] On the one hand, an embodiment of the present application provides a differential privacy federated learning method based on dual knowledge transfer, including:

[0006] Obtain a data subset for each medical client and train a teacher model using each data subset;

[0007] The weight of each teacher model is calculated to obtain the teacher weight; the teacher model uses a voting mechanism based on the teacher weight to predict the unlabeled teacher data set to determine the teacher voting result of each sample in the unlabeled teacher data set; determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if noise is added, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; determine the label of the unlabeled teacher data set based on the teacher noisy voting result to obtain the labeled teacher data set, use the labeled teacher data set to train the student model, and upload the student model to the medical server;

[0008] In the medical server, the student models are aggregated, and the weight of each student model is calculated to obtain the student weight; the student model uses a voting mechanism based on the student weight to predict the unlabeled student data set to determine the student voting result of each sample in the unlabeled student data set; whether to add noise to the student voting result is determined according to the privacy level of the medical server, and if noise is added, Laplace noise is added to the student voting result to obtain the student noisy voting result; the label of the unlabeled student data set is determined based on the student noisy voting result to obtain the labeled student data set, and the labeled student data set is used to train the global model;

[0009] For the global model, the importance sampling mechanism is used to resample the labeled student dataset to obtain a high model uncertainty dataset, and the high model uncertainty dataset is used to fine-tune the global model to obtain a fine-tuned model; for the fine-tuned model, the importance sampling mechanism is used to resample the labeled student dataset to obtain a low model uncertainty dataset, and the low model uncertainty dataset is used to self-train the fine-tuned model to obtain the final global model.

[0010] On the other hand, the embodiment of the present application also provides a differential privacy federated learning system based on dual knowledge transfer, including:

[0011] A local training module is used to obtain a data subset of each medical client and train a teacher model using each data subset;

[0012] The medical client knowledge transfer module is used to calculate the weight of each teacher model to obtain the teacher weight; the teacher model uses a voting mechanism based on the teacher weight to predict the unlabeled teacher data set to determine the teacher voting result of each sample in the unlabeled teacher data set; determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if noise is added, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; determine the label of the unlabeled teacher data set based on the teacher noisy voting result to obtain the labeled teacher data set, use the labeled teacher data set to train the student model, and upload the student model to the medical server;

[0013] The medical server knowledge transfer module is used to aggregate student models in the medical server, calculate the weight of each student model, and obtain the student weight; the student model uses a voting mechanism based on the student weight to predict the unlabeled student data set to determine the student voting result of each sample in the unlabeled student data set; determine whether to add noise to the student voting result according to the privacy level of the medical server, and if noise is added, add Laplace noise to the student voting result to obtain the student noisy voting result; determine the label of the unlabeled student data set based on the student noisy voting result to obtain the labeled student data set, and use the labeled student data set to train the global model;

[0014] The model optimization module is used to resample the labeled student data set using the importance sampling mechanism for the global model to obtain a high model uncertainty data set, and use the high model uncertainty data set to fine-tune the global model to obtain a fine-tuned model; for the fine-tuned model, the importance sampling mechanism is used to resample the labeled student data set to obtain a low model uncertainty data set, and the low model uncertainty data set is used to self-train the fine-tuned model to obtain the final global model.

[0015] A differential privacy federated learning method and system based on dual knowledge transfer in this application has the following advantages:

[0016] It adopts a dual knowledge transfer mechanism of medical client and medical server, which can extract knowledge from decentralized medical data, achieve high-accuracy learning through only one round of federated learning, and provide stronger privacy protection. The method of this application takes into account the impact of differences in teacher model capabilities caused by the heterogeneity of medical data on knowledge transfer in federated learning, and innovatively introduces a weighted voting mechanism in knowledge transfer, adjusts model weights according to individual differences in teacher models, and improves the model utility of federated learning when facing data heterogeneity problems. In addition, this application uses a global model fine-tuning strategy based on data uncertainty to mitigate the negative impact of noise on model accuracy. A large number of experimental results show that this application can significantly improve model accuracy, significantly improve data utility, and effectively reduce communication overhead while providing strong privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1An architectural diagram of a differentially private federated learning method based on dual knowledge transfer provided in an embodiment of the present application.

[0019] Figure 2 This is a schematic diagram of the change in accuracy for different privacy parameters compared on the cod-rna dataset when the privacy level of this application is medical server-side privacy protection.

[0020] Figure 3 This is a schematic diagram of the change in accuracy for different privacy parameters compared on the cod-rna dataset when the privacy level of this application is medical client privacy protection. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0022] Figure 1 The present invention provides an architecture diagram of a differential privacy federated learning method based on dual knowledge transfer. The present invention provides a differential privacy federated learning method based on dual knowledge transfer, including:

[0023] S100, obtaining a data subset of each medical client, and using each data subset to train a teacher model.

[0024] For example, in the embodiment of the present application, the medical client has n , respectively expressed as , for the i Medical Clients ,in i =1,2,…, n , which has a private dataset . In acquiring each medical client Private datasets After that, the private dataset Divide into multiple, e.g. t non-overlapping subsets of data ,in j =1,2,…, t .

[0025] Respectively t By training on a subset of data, we can get t Teacher Model .

[0026] S110, calculate the weight of each teacher model to obtain the teacher weight; use the teacher model to predict the unlabeled teacher data set using a voting mechanism based on the teacher weight to determine the teacher voting result of each sample in the unlabeled teacher data set; determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if noise is added, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; determine the label of the unlabeled teacher data set based on the teacher noisy voting result to obtain a labeled teacher data set, use the labeled teacher data set to train the student model, and upload the student model to the medical server.

[0027] Exemplarily, in S110, in order to accurately reflect the ability differences of teacher models, a reputation update mechanism and a contribution update mechanism are introduced when calculating teacher weights. Specifically, the weight of each teacher model is calculated to obtain the teacher weight, including: determining the teacher reputation value and the teacher contribution according to the prediction results of the teacher model for the unlabeled teacher data set; performing normalized weighted summation of the teacher reputation value and the teacher contribution to obtain the teacher weight.

[0028] In an embodiment of the present application, the teacher reputation value is determined based on the consistency between the prediction results of the teacher model for each sample in the unlabeled teacher data set and the average prediction results of the collection of teacher models for the unlabeled teacher data set, and the teacher contribution is determined based on the average prediction probability of the combination of teacher models for the unlabeled teacher data set.

[0029] The reputation update mechanism is used to evaluate the consistency of the teacher model in voting. i Medical Clients Middle j Teacher Model Teacher reputation value According to the teacher model, the unlabeled teacher dataset The consistency between the previous sample prediction and the collective prediction is updated using the formula:

[0030] (1)

[0031] in, represents the initial reputation value, represents the reward penalty factor, represents the privacy parameter, which is used to control the noise scale when achieving differential privacy. Represents the unlabeled teacher dataset The number of samples in r Represents the unlabeled teacher dataset The sample number in Represents the teacher model For the unlabeled teacher dataset Middle r Samples The prediction and i Medical Clients Is the average prediction of the set of all teacher models consistent? ,otherwise .

[0032] The contribution update mechanism measures the marginal contribution of the teacher model through the Shapley value. , teacher model Teachers' contribution The calculation formula is:

[0033] (2)

[0034] in, Represents medical client The collection of all teacher models in The set of all non-empty subsets of , a express A combination of express The number of combinations in Represents the combination of teacher models For samples Predicted as label The average probability of Indicated by combination a and teacher model The teacher model set is composed of Represents the combination of teacher models For samples Predicted as label The average probability of Indicated by combination a The teacher model set is composed of Indicates all tags A collection of Indicates taking the absolute value.

[0035] Next, calculate the teacher model In the entire unlabeled teacher dataset Teachers' contribution , the formula is:

[0036] (3)

[0037] After calculating the teacher model Teacher reputation value and teacher contribution After that, the normalized weighted sum is performed to calculate the teacher model Teacher weight , the formula is:

[0038] (4)

[0039] in, It is a factor that balances the teacher's reputation and contribution. and Respectively, in the medical client In t Teacher Model Teacher reputation value The minimum and maximum values ​​in and Respectively, in the medical client In t Teacher Model Teachers' contribution The minimum and maximum values ​​in .

[0040] In an embodiment of the present application, the teacher model uses a voting mechanism based on teacher weights to predict the unlabeled teacher data set to determine the teacher voting results of each sample in the unlabeled teacher data set, including: each teacher model participates in the voting of multiple teacher candidate labels of the unlabeled teacher data set with the teacher weight as probability, forming a teacher voting result representing the number of votes for each teacher candidate label.

[0041] Specifically, the unlabeled teacher dataset Each sample The label is determined by a voting mechanism based on the teacher's weight. , calculate the number of teacher votes , the formula is:

[0042] (5)

[0043] in, Represents the teacher model For samples The predicted label of Represents the teacher model by The probability of participating in the label in voting. represents the indicator function, if equal ,but The result is 1, otherwise The result is 0.

[0044] The medical client judges the privacy level and decides whether to add noise to the teacher voting results. If the privacy level is medical client privacy protection, The number of teacher labels Add Laplace noise, the adding process is shown in the formula:

[0045] (6)

[0046] in, represents the number of teacher labels after adding Laplace noise, is the Laplace noise, whose scale is determined by the privacy parameter control.

[0047] The label of the unlabeled teacher data set is determined based on the teacher noise-added voting results to obtain a labeled teacher data set, including: selecting the teacher candidate label with the largest number of votes as the final teacher label of the unlabeled teacher data set, adding the final teacher label to the unlabeled teacher data set, and obtaining a labeled teacher data set.

[0048] Specifically, the medical client predicts the teacher voting results for the unlabeled teacher dataset through the teacher model Divide the labels to obtain a labeled teacher dataset , the formula for dividing labels is:

[0049] (7)

[0050] in, Represented as the unlabeled teacher dataset The divided labels, formula (7) indicates that the teacher candidate label with the largest number of votes is selected as the final teacher label of the dataset.

[0051] Each medical client Based on the labeled teacher dataset Train the student model , and upload it to the medical server.

[0052] S120, in the medical server, aggregates from n Student Model of a Medical Client , calculate the weight of each student model to obtain the student weight; the student model uses a voting mechanism based on the student weight to predict the unlabeled student data set to determine the student voting result of each sample in the unlabeled student data set; determine whether to add noise to the student voting result according to the privacy level of the medical server, and if noise is added, add Laplace noise to the student voting result to obtain the student noisy voting result; determine the label of the unlabeled student data set based on the student noisy voting result to obtain the labeled student data set, and use the labeled student data set to train the global model.

[0053] Exemplarily, in S120, a reputation update mechanism and a contribution update mechanism are also introduced when calculating student weights. Specifically, the weight of each student model is calculated to obtain the student weight, including: determining the student reputation value and the student contribution according to the prediction results of the student model for the unlabeled student data set; performing normalized weighted summation of the student reputation value and the student contribution to obtain the student weight.

[0054] In an embodiment of the present application, the student reputation value is determined based on the consistency between the prediction results of the student model for each sample in the unlabeled student data set and the average prediction results of the collection of student models for the unlabeled student data set, and the student contribution is determined based on the average prediction probability of the combination of student models for the unlabeled student data set.

[0055] The reputation update mechanism is used to evaluate the consistency of student models in voting. i Medical Clients Uploaded student models Student reputation value According to the student model, the unlabeled student dataset The consistency between the previous sample prediction and the collective prediction is updated using the formula:

[0056] (8)

[0057] in, Represents an unlabeled student dataset The number of samples in u Represents an unlabeled student dataset The sample number in Represents the student model For the unlabeled student dataset Middle u Samples Is the prediction of the student model consistent with the average prediction of the set of all student models? ,otherwise .

[0058] The contribution update mechanism measures the marginal contribution of the student model through the Shapley value. , student model Student contribution The calculation formula is:

[0059] (9)

[0060] in, Represents the collection of all student models on the medical server side The set of all non-empty subsets of b express A combination of express The number of combinations in Represents the combination of student models For samples Predicted as label The average probability of Indicated by combination b and student models The student model set consists of Represents the combination of student models For samples Predicted as label The average probability of Indicated by combination b The student model set is composed of Indicates all tags A collection of .

[0061] Next, calculate the student model In the entire unlabeled student dataset Student contribution , the formula is:

[0062] (10)

[0063] After calculating the student model Student reputation value and student contribution After that, the normalized weighted sum is performed to calculate the student model Student weight , the formula is:

[0064] (11)

[0065] in, It is also a factor that balances student reputation and student contribution. and Respectively nStudent Model Student reputation value The minimum and maximum values ​​in and Respectively n Student Model Student contribution The minimum and maximum values ​​in .

[0066] In an embodiment of the present application, the student model uses a voting mechanism based on student weights to predict the unlabeled student data set to determine the student voting results of each sample in the unlabeled student data set, including: each student model participates in the voting of multiple student candidate labels of the unlabeled student data set with the student weight as probability, forming a student voting result representing the number of votes for each student candidate label.

[0067] Specifically, the unlabeled student dataset Each sample The labels are determined by a voting mechanism based on student weights. , calculate the number of student votes , the formula is:

[0068] (12)

[0069] in, Represents the student model For samples The predicted label of Represents the student model by The probability of participating in the label in voting.

[0070] The medical server determines the privacy level and decides whether to add noise to the student voting results. If the privacy level is medical server-side privacy protection, then for each sample Number of student labels Add Laplace noise, the adding process is shown in the formula:

[0071] (13)

[0072] in, Represents the number of student labels after adding Laplace noise.

[0073] The label of the unlabeled student data set is determined based on the student noise-added voting results to obtain the labeled student data set, including: selecting the student candidate label with the largest number of votes as the final student label of the unlabeled student data set, adding the final student label to the unlabeled student data set, and obtaining the labeled student data set.

[0074] Specifically, the medical server predicts student voting results for the unlabeled student dataset using the student model Divide the labels to obtain a labeled student dataset , the formula for dividing labels is:

[0075] (14)

[0076] in, Represented as an unlabeled student dataset The divided labels, formula (14) means that the student candidate label with the largest number of votes is selected as the final student label of the dataset.

[0077] Medical Server Leverages Labeled Student Dataset Train the global model F .

[0078] S130, for the global model, the labeled student data set is resampled using the importance sampling mechanism to obtain a high model uncertainty data set, and the global model is fine-tuned using the high model uncertainty data set to obtain a fine-tuned model; for the fine-tuned model, the labeled student data set is resampled using the importance sampling mechanism to obtain a low model uncertainty data set, and the fine-tuned model is self-trained using the low model uncertainty data set to obtain the final global model.

[0079] For example, for the global model F The labeled student data set is resampled using an importance sampling mechanism, including: determining a high sampling weight according to the uncertainty of a global model, and resampling the labeled student data set according to the high sampling weight.

[0080] Specifically, for each labeled student dataset Samples in , global model F Uncertainty and high sampling weights Calculated by the following two formulas:

[0081] Uncertainty calculate:

[0082] (15)

[0083] High sampling weight calculate:

[0084] (16)

[0085] in, Represents the global model For samples Belongs to the tag The predicted probability of .

[0086] According to the high sampling weight , in the labeled student dataset Medium Sampling samples form a high model uncertainty dataset , global model use Fine-tune to get the fine-tuned model .

[0087] The importance sampling mechanism is used to resample the labeled student dataset for the fine-tuning model, including: calculating low sampling weights according to the uncertainty of the fine-tuning model, and resampling the labeled student dataset according to the low sampling weights.

[0088] Specifically, for the labeled student dataset Each sample , fine-tune the model Uncertainty and low sampling weights It is calculated as follows:

[0089] (17)

[0090] (18)

[0091] According to the low sampling weight, in the labeled student dataset Medium Sampling samples form a low model uncertainty dataset , fine-tune the model By dataset Self-training to obtain the final global model .

[0092] The embodiment of the present application also provides a differential privacy federated learning system based on dual knowledge transfer, the system comprising:

[0093] A local training module is used to obtain a data subset of each medical client and train a teacher model using each data subset;

[0094] The medical client knowledge transfer module is used to calculate the weight of each teacher model to obtain the teacher weight; the teacher model uses a voting mechanism based on the teacher weight to predict the unlabeled teacher data set to determine the teacher voting result of each sample in the unlabeled teacher data set; determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if noise is added, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; determine the label of the unlabeled teacher data set based on the teacher noisy voting result to obtain the labeled teacher data set, use the labeled teacher data set to train the student model, and upload the student model to the medical server;

[0095] The medical server knowledge transfer module is used to aggregate student models in the medical server, calculate the weight of each student model, and obtain the student weight; the student model uses a voting mechanism based on the student weight to predict the unlabeled student data set to determine the student voting result of each sample in the unlabeled student data set; determine whether to add noise to the student voting result according to the privacy level of the medical server, and if noise is added, add Laplace noise to the student voting result to obtain the student noisy voting result; determine the label of the unlabeled student data set based on the student noisy voting result to obtain the labeled student data set, and use the labeled student data set to train the global model;

[0096] The model optimization module is used to resample the labeled student data set using the importance sampling mechanism for the global model to obtain a high model uncertainty data set, and use the high model uncertainty data set to fine-tune the global model to obtain a fine-tuned model; for the fine-tuned model, the importance sampling mechanism is used to resample the labeled student data set to obtain a low model uncertainty data set, and the low model uncertainty data set is used to self-train the fine-tuned model to obtain the final global model.

[0097] refer to Figure 2 , on the cod-rna dataset, for different privacy parameters, the experimental results of the accuracy changing with the privacy parameters are shown. The algorithms compared with the method proposed in this application include Fed-PATE (federated learning based on collective private aggregation of teachers) and FedKT (federated learning based on knowledge transfer), where Fed-PATE is a strategy that directly applies PATE (collective private aggregation of teachers) to federated learning, and FedKT is a cross-domain learning method based on Fed-PATE, which uses partitioning and unanimous voting for federated knowledge transfer. Figure 2 FedDKT (Federated Learning Based on Dual Knowledge Transfer) in this paper represents the dual knowledge transfer differential privacy federated learning algorithm proposed in this application, which combines knowledge transfer, differential privacy and personalized teacher model technology. In the experiment, the unlabeled dataset ratio of the cod-rna dataset is set to 0.5%. Figure 2It shows that with the increase of privacy parameters, the accuracy of all methods is improved, and under most privacy parameters, the accuracy of this application is higher than that of the other two methods, indicating that compared with the other two methods, the method of this application can better balance privacy protection and model utility.

[0098] refer to Figure 3 , is the accuracy rate with the privacy parameter under the privacy protection condition of medical client on the cod-rna dataset. Experiments with variations. Medical client privacy protection provides stronger privacy protection than medical server privacy protection. The methods compared with this application are Fed-PATE and FedKT. In the experiment, the unlabeled dataset ratio of the cod-rna dataset is set to 0.5%. Figure 3 It shows that the accuracy performance of this application is still better than other methods under most privacy parameters, which further illustrates that the method of this application can better balance privacy protection and model utility in federated learning.

[0099] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0100] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A differential privacy federated learning method based on dual knowledge transfer, characterized in that: include: Obtaining a data subset of each medical client, and training a teacher model using each of the data subsets; Calculate the weight of each teacher model to obtain the teacher weight; The teacher model predicts the unlabeled teacher data set using a voting mechanism based on the teacher weight to determine a teacher voting result for each sample in the unlabeled teacher data set; Determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if adding noise, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; Determine the label of the unlabeled teacher data set based on the teacher noise voting result, obtain a labeled teacher data set, use the labeled teacher data set to train a student model, and upload the student model to a medical server; In the medical server, the student models are aggregated, and the weight of each student model is calculated to obtain a student weight; The student model predicts the unlabeled student data set using a voting mechanism based on the student weights to determine the student voting result of each sample in the unlabeled student data set; determining whether to add noise to the student voting result according to the privacy level of the medical server, and if adding noise, adding Laplace noise to the student voting result to obtain the student noisy voting result; Determine the label of the unlabeled student data set based on the student noise-added voting results, obtain a labeled student data set, and use the labeled student data set to train a global model; For the global model, the labeled student dataset is resampled using an importance sampling mechanism to obtain a high model uncertainty dataset, and the global model is fine-tuned using the high model uncertainty dataset to obtain a fine-tuned model; for the fine-tuned model, the labeled student dataset is resampled using an importance sampling mechanism to obtain a low model uncertainty dataset, and the fine-tuned model is self-trained using the low model uncertainty dataset to obtain a final global model.

2. According to claim 1, a differential privacy federated learning method based on dual knowledge transfer is characterized in that: When acquiring the data subsets, firstly, the private data set of each medical client is acquired, and then the private data set is divided into a plurality of non-overlapping data subsets.

3. The differential privacy federated learning method based on dual knowledge transfer according to claim 1, characterized in that: Calculate the weight of each teacher model to obtain the teacher weight, including: Determining a teacher reputation value and a teacher contribution according to a prediction result of the teacher model on the unlabeled teacher data set; Performing normalized weighted summation on the teacher reputation value and the teacher contribution to obtain the teacher weight; Calculate the weight of each student model to obtain the student weight, including: Determining a student reputation value and a student contribution according to a prediction result of the student model on the unlabeled student data set; The student reputation value and the student contribution are normalized and weighted to obtain the student weight.

4. The differential privacy federated learning method based on dual knowledge transfer according to claim 3, characterized in that: The teacher reputation value is determined according to the consistency between the prediction result of the teacher model for each sample in the unlabeled teacher data set and the average prediction result of the set of teacher models for the unlabeled teacher data set, and the teacher contribution is determined according to the average prediction probability of the combination of the teacher models for the unlabeled teacher data set; The student reputation value is determined based on the consistency between the prediction results of the student model for each sample in the unlabeled student dataset and the average prediction results of the set of student models for the unlabeled student dataset, and the student contribution is determined based on the average prediction probability of the combination of student models for the unlabeled student dataset.

5. The differential privacy federated learning method based on dual knowledge transfer according to claim 1, characterized in that: The teacher model predicts the unlabeled teacher data set using a voting mechanism based on the teacher weight to determine the teacher voting result of each sample in the unlabeled teacher data set, including: Each of the teacher models participates in voting for a plurality of teacher candidate labels of the unlabeled teacher data set with the teacher weight as probability, to form the teacher voting result indicating the number of votes for each of the teacher candidate labels; The student model predicts the unlabeled student data set using a voting mechanism based on the student weights to determine a student voting result for each sample in the unlabeled student data set, including: Each of the student models participates in voting for multiple student candidate labels of the unlabeled student data set with the student weight as probability, forming the student voting result representing the number of votes for each of the student candidate labels.

6. The differential privacy federated learning method based on dual knowledge transfer according to claim 5, characterized in that: Determining the label of the unlabeled teacher data set based on the teacher noise-added voting result to obtain a labeled teacher data set includes: Select the teacher candidate label with the largest number of votes as the final teacher label of the unlabeled teacher data set, and add the final teacher label to the unlabeled teacher data set to obtain the labeled teacher data set; Determining the label of the unlabeled student data set based on the student noise-added voting result to obtain a labeled student data set includes: The student candidate label with the largest number of votes is selected as the final student label of the unlabeled student data set, and the final student label is added to the unlabeled student data set to obtain the labeled student data set.

7. The differential privacy federated learning method based on dual knowledge transfer according to claim 1, characterized in that: Resampling the labeled student data set using an importance sampling mechanism for the global model includes: A high sampling weight is determined according to the uncertainty of the global model, and the labeled student data set is resampled according to the high sampling weight.

8. The differential privacy federated learning method based on dual knowledge transfer according to claim 1, characterized in that: Resampling the labeled student dataset using an importance sampling mechanism for the fine-tuning model includes: A low sampling weight is determined according to the uncertainty of the fine-tuning model, and the labeled student data set is resampled according to the low sampling weight.

9. A system using a differential privacy federated learning method based on dual knowledge transfer as described in any one of claims 1 to 8, characterized in that: include: A local training module, used for obtaining a data subset of each medical client, and training a teacher model using each of the data subsets; A medical client knowledge transfer module, used for calculating the weight of each teacher model to obtain a teacher weight; The teacher model predicts the unlabeled teacher data set using a voting mechanism based on the teacher weight to determine a teacher voting result for each sample in the unlabeled teacher data set; Determine whether to add noise to the teacher voting result according to the privacy level of the medical client, and if adding noise, add Laplace noise to the teacher voting result to obtain the teacher noisy voting result; Determine the label of the unlabeled teacher data set based on the teacher noise voting result, obtain a labeled teacher data set, use the labeled teacher data set to train a student model, and upload the student model to a medical server; A medical server knowledge transfer module, used for aggregating the student models in the medical server, calculating the weight of each student model, and obtaining a student weight; The student model predicts the unlabeled student data set using a voting mechanism based on the student weights to determine the student voting result of each sample in the unlabeled student data set; determining whether to add noise to the student voting result according to the privacy level of the medical server, and if adding noise, adding Laplace noise to the student voting result to obtain the student noisy voting result; Determine the label of the unlabeled student data set based on the student noise-added voting results, obtain a labeled student data set, and use the labeled student data set to train a global model; A model optimization module is used to resample the labeled student data set using an importance sampling mechanism for the global model to obtain a high model uncertainty data set, and use the high model uncertainty data set to fine-tune the global model to obtain a fine-tuned model; resample the labeled student data set using an importance sampling mechanism for the fine-tuned model to obtain a low model uncertainty data set, and use the low model uncertainty data set to self-train the fine-tuned model to obtain a final global model.

Images