Personalized federal model training method based on meta-learning and multi-level segmentation

Abstract

The invention relates to the technical field of artificial intelligence, and particularly discloses a personalized federal model training method based on meta-learning and multi-level segmentation, which has a multi-level model segmentation architecture, further divides a model into an input layer, a middle layer and an output layer, and realizes finer calculation load distribution. Compared with an existing split Learning mode that the model is divided into a client side and a server side, the method has the advantages that an input layer and an output layer are subjected to localization processing, a middle layer is reserved at the server side, sensitive data transmission is reduced, the calculation efficiency and privacy protection are balanced, meanwhile, the method has the rapid fine adjustment capacity, and the method is suitable for popularization and application. And the input layer and the output layer can dynamically adapt to local data distribution of the client. According to the method, rapid adaptation of the local model is realized through the meta-learning strategy, the problem of data isomerism is solved, and the personalized performance of the model is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and specifically discloses a personalized federated model training method based on meta-learning and multi-level segmentation. Background Art

[0002] Federated learning, as a distributed machine learning technology, aims to achieve co-modeling and improve the performance of AI models on the basis of protecting data privacy, security, and legal compliance.

[0003] In the training tasks of existing federated models, the client needs to send the local model to the server, and then the server aggregates the models received from all clients, and finally sends the aggregated model to the client. Then, by repeating the above process until the model converges or reaches the maximum number of iteration rounds.

[0004] When it is specifically applied in fields such as healthcare and finance that are sensitive to data and have different requirements, due to the global aggregation mechanism of federated learning, there are problems in dealing with the differences in client data distributions. At the same time, when balancing privacy protection and model performance, it often ignores the need to dynamically adapt to the characteristics of local data.

[0005] The present invention provides a personalized federated model training method based on meta-learning and multi-level segmentation to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a personalized federated model training method that combines meta-learning and multi-level model segmentation, and solves the performance bottleneck problem caused by data heterogeneity through a fine-grained segmentation architecture, a meta-learning-driven local fast fine-tuning module, and a hierarchical gradient refinement mechanism.

[0007] To achieve the above objective, the basic solution of the present invention provides a personalized federated model training method based on meta-learning and multi-level segmentation, including the following steps: Step S1: The server obtains the target model and divides it into an input layer, an intermediate layer, and an output layer to obtain a multi-level meta-model. The server initializes the parameters of the multi-level meta-model, and distributes copies of the input layer and the output layer to each client, and retains the intermediate layer on the server; Step S2: Each client respectively constructs a local dataset with actual class labels. Each client respectively downloads the input layer parameters and the output layer parameters, and performs meta-learning fine-tuning with the local data in the local dataset as input samples, and extracts preliminary features and transmits them to the server to obtain the intermediate layer activation values. The local data includes medical image data and corresponding class labels; Step S3: The server calculates the global features based on the intermediate layer activation values and normalizes them, updates the intermediate layer activation values based on the normalized global features, and distributes them to each client; Step S4: Each client generates the predicted classification results of the local data based on the updated intermediate layer activation values, and establishes an intermediate layer loss function based on the predicted classification results and the actual class labels; Step S5: Based on the intermediate layer loss function, each client calculates the gradients of the input layer parameters, intermediate layer parameters, and output layer parameters through backpropagation to update the input layer parameters and output layer parameters; Step S6: Each client uploads the gradients of the intermediate layer parameters to the server, and the server performs weighted average aggregation to update the global intermediate layer parameters and broadcasts them to all clients; Step S7: Repeat steps S2 to S6 for iterative training until the model converges, and execute the target task through the converged model. The target task includes identifying medical image data through the converged model to output the identification results.

[0008] Furthermore, in step S2, the local dataset is constructed as follows: ; In the formula, represents the i th input sample of the n th client, represents the class label of the i th input sample of the n th client, represents the number of input samples in the local dataset of the i th client.

[0009] Furthermore, in step S2, the meta-learning fine-tuning includes performing K steps of gradient descent through the local dataset , where K takes values between 1 and 5, and the larger the local data volume, the larger the value of K.

[0010] Furthermore, in step S2, the expression of the single-step gradient descent update process is as follows: ; ; In the formula, is the fine-tuning learning rate of client i , is the input layer parameter of client i , is the output layer parameter of client i , is clienti The updated input layer parameters for the client i The updated output layer parameters for the client i The training loss The calculation process is as follows: ; ; ; ; In the formula, represents the input layer function of the input layer parameters i on the client ; represents the output layer function of the output layer parameters i on the client ; represents the middle layer function of the middle layer parameters ; In the formula, , and respectively represent the input layer result, middle layer result: middle layer activation value, and output layer result corresponding to the i th input sample on the client n [[ID=5"3]];

[0011] Furthermore, in step S3, the global feature is expressed as: ; In the formula, is the number of clients, represents the middle layer result: middle layer activation value corresponding to the i th input sample on the client n ; ]> is the weight of the client i ; represents the number of input samples in the local dataset of the i th client; ; The expression of the normalized global feature is as follows: ; The expression of the updated middle layer activation value is as follows: ; In the formula, is the global update weight.

[0012] Further, in step S4, the predicted classification result is expressed as follows: ; In the formula, represents the output layer function of the output layer parameters i on the client ; The intermediate layer loss function is calculated according to the following formula: ; In the formula, the second term is the L2 regularization term, is the regularization hyperparameter, is the actual class label.

[0013] Further, in step S5, the calculation formulas for the gradients of the input layer parameters , the intermediate layer parameters and the output layer parameters are shown as follows respectively: ; ; ; In the formula, represents the input layer result corresponding to the i th input sample on the client n .

[0014] Further, in step S6, the server receives the intermediate layer gradients uploaded by all clients, and aggregates them through weighted average to obtain the global gradient: ; The server uses the aggregated global gradient to update the global intermediate layer parameters : ; In the formula, is the learning rate of server i.

[0015] The principle and effect of this solution are as follows: 1. Compared with the prior art, the present invention has a multi-level model segmentation architecture, and further divides the model into an input layer, an intermediate layer, and an output layer, realizing a more refined calculation load distribution. Compared with the existing Split Learning method that divides the model into two parts: client - server, the present invention localizes the input layer and the output layer, retains the intermediate layer on the server side, which not only reduces the transmission of sensitive data but also balances the computational efficiency and privacy protection.

[0016] 2. Compared with the prior art, the present invention has the ability of rapid fine-tuning, enabling the input layer and the output layer to dynamically adapt to the local data distribution of the client. The present invention realizes the rapid adaptation of the local model through a meta-learning strategy, solves the problem of data heterogeneity, and significantly improves the personalized performance of the model.

[0017] 3. Compared with the prior art, the present invention is also equipped with a hierarchical gradient update and aggregation mechanism: the client is responsible for updating the input layer and the output layer, and the server aggregates the gradients of the intermediate layer to update the intermediate layer, and iterates until the final input layer, output layer, and intermediate layer are obtained to converge the model, and the target task is executed through the converged model. This way of gradient processing by dividing weights optimizes the communication efficiency (reduces redundant data transmission), and at the same time takes into account both global shared features and local personalized needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application, and those skilled in the art can also obtain other drawings without creative efforts based on these drawings.

[0019] Figure 1 FIG. shows a flowchart of a personalized federated model training method based on meta-learning and multi-level segmentation proposed in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, describe in detail the specific implementation manners, structures, features, and their effects according to the present invention.

[0021] A personalized federated model training method based on meta-learning and multi-level segmentation, as shown in the embodiments Figure 1 below, includes the following steps: Step S1: The server obtains the target model and divides it into an input layer, an intermediate layer, and an output layer to obtain a multi-level meta-model. The server initializes the parameters of the multi-level meta-model, and distributes copies of the input layer and the output layer to each client, and retains the intermediate layer on the server. In this embodiment, a pneumonia diagnosis model is used as the target model.

[0022] Step S2: Each client respectively constructs a local dataset with actual class labels , and each client respectively downloads the input layer parameters and output layer parameters of the multi-level meta-model, and uses the local data in the local dataset as input samples to perform meta-learning fine-tuning on the input layer parameters and output layer parameters, and extracts preliminary features and transmits them to the server to obtain intermediate layer activation values.

[0023] The local dataset is constructed according to the following formula: ; In the formula, represents the i th input sample of the n th client, represents the class label of the i th input sample of the n th client, represents the number of input samples in the local dataset of the i th client.

[0024] In this embodiment, the local dataset is a private dataset, which refers to the non-public data stored locally by each client, including local data: medical image data and corresponding class labels. Protecting client privacy by establishing a local dataset. When training the federated model, the client only uploads intermediate features (not raw data), avoiding direct sharing of raw data and reducing the risk of privacy leakage.

[0025] In this embodiment, each client specifically represents the client used by a medical institution participating in model training. For the client used by a medical institution, the label distribution of the local data it uses can also be used to represent the proportion of medical image data containing cancer cell images in X-ray films and the proportion of medical image data not containing cancer cell images in X-ray films.

[0026] After each client downloads the input layer parameters and output layer parameters of the meta-model, the input layer parameters and output layer parameters downloaded to client i are respectively denoted as and . Meta-learning fine-tuning includes performing K steps of gradient descent through the local dataset , where K takes values between 1 and 5, and the larger the local data volume, the larger the value of K.

[0027] Specifically, the expression of the single-step gradient descent update process is as follows: ; ; In the formula, is the fine-tuning learning rate of client i , is the updated input layer parameter of client i , is the updated output layer parameter of client i . For the client i training loss, the calculation process is as follows: ; ; ; ; In the formula, represents the input layer function of the input layer parameters i on the client ; represents the output layer function of the output layer parameters i on the client ; represents the middle layer function of the middle layer parameters ; In the formula, , and respectively represent the input layer result, middle layer result, middle layer activation value, and output layer result corresponding to the i th input sample on the client n .

[0028] Step S3: The server calculates the global feature based on the middle layer activation value and normalizes it, updates the middle layer activation value based on the normalized global feature, and distributes it to each client.

[0029] Specifically, the global feature is calculated as follows: ; In the formula, is the number of clients, is the weight of the client i , which is set according to its local data volume, and the expression is as follows: ; The server further normalizes the global feature to obtain the normalized global feature : ; Next, the normalized global feature is used to update the middle layer activation value to obtain the updated middle layer activation value : ; In the formula, For global weight update.

[0030] Step S4: The client generates a predicted classification result based on the received updated intermediate layer activation values and according to the output layer parameters and establishes an intermediate layer loss function based on the predicted classification result and the actual class label. .

[0031] Among them, the predicted classification result is calculated based on the updated intermediate layer activation values, and the formula is as follows: ; The intermediate layer loss function is calculated according to the following formula: ; In the formula, the second term is the L2 regularization term, is the regularization hyperparameter used to prevent overfitting, and is the actual class label.

[0032] Step S5: Based on the intermediate layer loss function , the client i calculates the gradients of the input layer parameters , the intermediate layer parameters and the output layer parameters through backpropagation, and the client updates the local parameters based on the obtained gradients.

[0033] Specifically, the calculation formulas for the gradients of the input layer parameters , the intermediate layer parameters and the output layer parameters are as follows respectively: ; ; ; In the formula, represents the gradient of the input layer parameter , represents the gradient of the intermediate layer parameter , represents the gradient of the output layer parameter .

[0034] The client updates the local parameters based on the obtained gradients according to the following formula: ; ; Among them, is the client​i Training learning rate.

[0035] Step S6: Each client uploads the gradients of the intermediate layer parameters to the server side. The server side performs weighted average aggregation to obtain the aggregated gradients of the intermediate layer parameters, and the server side updates the global intermediate layer parameters based on the aggregated gradients of the intermediate layer parameters and broadcasts them to all clients.

[0036] The server receives the intermediate layer gradients uploaded by all clients and aggregates them through weighted average to obtain the global gradient: ; The server uses the aggregated global gradient to update the global intermediate layer parameters : ; In the formula, is the learning rate of server side i.

[0037] Step S7: Repeat steps S2 to S6 for iterative training until the model converges, and execute the target task through the converged model. The target task includes: identifying medical image data through the converged model to output the identification result.

[0038] This step ensures that the global model can learn shared features from the local data distributions of all clients, while avoiding directly sharing the original data and protecting privacy.

[0039] The present invention has a multi-level model segmentation architecture, further dividing the model into an input layer, an intermediate layer, and an output layer, achieving a more refined computational load distribution. Compared with the existing Split Learning method that divides the model into two parts: client - server, the present invention localizes the input layer and the output layer, retains the intermediate layer on the server side, which not only reduces the transmission of sensitive data but also balances the computational efficiency and privacy protection.

[0040] The present invention has the ability of fast fine-tuning, enabling the input layer and the output layer to dynamically adapt to the local data distribution of the client. The present invention realizes the fast adaptation of the local model through the meta-learning strategy, solves the problem of data heterogeneity, and significantly improves the personalized performance of the model.

[0041] The present invention is also equipped with a hierarchical gradient update and aggregation mechanism: the client is responsible for updating the input layer and the output layer, the server side aggregates the intermediate layer gradients to update the intermediate layer, and iterates until the final input layer, output layer, and intermediate layer are obtained to converge the model, and the target task is executed through the converged model. This way of gradient processing with divided weights optimizes the communication efficiency (reducing redundant data transmission), while taking into account both global shared features and local personalized needs.

[0042] As described above, it is only the preferred embodiment of the present invention, and there is no limitation to the present invention in any form. Although the present invention has been disclosed as above with the preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to form equivalent embodiments by using the technical content disclosed above within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A personalized federated model training method based on meta - learning and multi - level segmentation, characterized in that, It includes the following steps: Step S1: The server obtains the target model and divides it into an input layer, an intermediate layer, and an output layer to obtain a multi-level meta-model. The server initializes the parameters of the multi-level meta-model and distributes copies of the input layer and the output layer to each client, while retaining the intermediate layer on the server; Step S2: Each client constructs a local dataset with actual class labels respectively. Each client downloads the input layer parameters and the output layer parameters respectively, and uses the local data in the local dataset as input samples for meta-learning fine-tuning, extracts preliminary features and transmits them to the server to obtain intermediate layer activation values. The local data includes medical image data and corresponding class labels; Step S3: The server calculates the global features based on the intermediate layer activation values and performs normalization processing on them, updates the intermediate layer activation values based on the normalized global features and distributes them to each client; Step S4: Each client generates a predicted classification result of the local data based on the updated intermediate layer activation values, and establishes an intermediate layer loss function based on the predicted classification result and the actual class labels; Step S5: Based on the intermediate layer loss function, each client calculates the gradients of the input layer parameters, the intermediate layer parameters, and the output layer parameters through backpropagation to update the input layer parameters and the output layer parameters; Step S6: Each client uploads the gradients of the intermediate layer parameters to the server, and the server performs weighted average aggregation to update the global intermediate layer parameters and broadcasts them to all clients; Step S7: Repeat steps S2 to S6 for iterative training until the model converges, and execute the target task through the converged model. The target task includes identifying medical image data through the converged model to output the identification result.

2. The personalized federated model training method based on meta-learning and multi-level segmentation according to claim 1, wherein, In step S2, the local data set is constructed according to the following formula: ； In the formula, represents the i -th input sample of the n -th client, represents the class label of the i -th input sample of the n -th client, represents the number of input samples in the local dataset of the i -th client.

3. The personalized federated model training method based on meta-learning and multi-level segmentation according to claim 2, characterized in that In step S2, the meta-learning fine-tuning includes using the local dataset to perform K steps of gradient descent, where K ranges from 1 to 5, and the value of K is larger when the local data volume is larger.

4. The personalized federated model training method based on meta-learning and multi-level segmentation according to claim 3, wherein, In step S2, the expression of the single-step gradient descent update process is as follows: ； ； Wherein, is the fine-tuning learning rate of the client i , is the input layer parameter of the client i , is the output layer parameter of the client i , is the updated input layer parameter of the client i , is the updated output layer parameter of the client i , is the training loss of the client i , The calculation process of is as follows: ； ； ； ； In the formula, represents the input layer parameters i on the client and the input layer function, represents the output layer parameters i on the client and the output layer function, represents the intermediate layer parameters and the intermediate layer function; In the formula, , and respectively represent the input layer result, the middle layer result: the middle layer activation value, and the output layer result corresponding to the i th input sample on the client n .

5. A personalized federated model training method based on meta-learning and multi-level segmentation according to claim 1, characterized in that In step S3, the global features are expressed as: ； Wherein, is the number of clients, represents the client i the n intermediate layer result corresponding to the input sample on the i weight of the client represents the i number of input samples in the local dataset of the Normalized global feature The expression is as follows: ； Updated intermediate layer activation value The expression is as follows: ； In the formula, is the global update weight.

6. The personalized federated model training method based on meta - learning and multi - level segmentation according to claim 5, wherein, In step S4, the predicted classification result has the following expression: ； In the formula, represents the client i output layer parameters of the output layer function; The intermediate layer loss function is calculated according to the following formula: ； In the formula, the second term is the L2 regularization term, is the regularization hyperparameter, is the actual class label.

7. A personalized federated model training method based on meta-learning and multi-level segmentation according to claim 6, characterized in that In step S5, the calculation formulas for the gradients of the input layer parameters , the intermediate layer parameters and the output layer parameters are as follows respectively: ； ； ； In the formula, represents the result of the input layer corresponding to the i th input sample on the n client.

8. A personalized federated model training method based on meta-learning and multi-level segmentation according to claim 7, characterized in that In step S6, the server receives the intermediate layer gradients uploaded by all clients , and aggregates them through weighted averaging to obtain the global gradient: ; The server updates the global intermediate layer parameters using the aggregated global gradients : ； wherein, is the learning rate of the server side i.

Images