Cross-device ECG federated privacy classification framework method, system and storage medium

Abstract

The present invention discloses a cross-device ECG federal privacy classification framework method, system and storage medium, and relates to the field of federated learning technology. The specific steps include the following: the server sends initial model parameters to the client; the client trains the initial model using a training data set to obtain basic type recognition model parameters, and sends them to the server in a differential manner; the server receives the basic type recognition model parameters for clustering to obtain clustering results; the clustering results are weighted averaged to obtain the final update parameters of the server model, and the training of the initial model is completed; the ECG data is recognized by the trained initial model, and the respective ECG data type recognition results are output on each client. The present invention can improve data recognition accuracy while protecting user data privacy during ECG health monitoring.

Description

Technical Field

[0001] The present invention relates to the field of federated learning technology, and more specifically to a cross-device ECG federated privacy classification framework method, system and storage medium. Background Art

[0002] With the increasing incidence of cardiovascular diseases and the development of IoT software and hardware devices, the demand for long-term home ECG health monitoring devices has increased rapidly. At the same time, the user privacy and security issues exposed in the ECG health monitoring process have become increasingly prominent. Although the centralized ECG diagnosis model can avoid user data leakage, it is unable to incorporate more rare types of data into the model in a timely manner. In order to balance the recognition and diagnosis performance of the model and the privacy attributes, the existing best ECG privacy protection classification scheme is a collaborative learning strategy, including Federated Learning (FL) and Split Learning (SL). Studies have shown that under the 1D-CNN model, the convolutional layer output transmitted from the client to the server using split learning will lead to user data leakage. In contrast, federated learning only transmits the trained model parameters, avoiding the migration of user data, and thus can better avoid privacy leakage caused by user data exposure.

[0003] However, when using the federated learning framework in the ECG type recognition task, cross-device user data will show extreme Non-IID performance, resulting in a sharp decline in the prediction ability of the federated model for rare types of clients. Existing work cannot adapt to the scenario of extreme Non-IID user data because the parameters from different clients to the server are treated equally, and it is impossible to achieve an effective knowledge transfer mechanism between different clients and the global perception ability of a single client.

[0004] Therefore, for those skilled in the art, how to improve data recognition accuracy while protecting user data privacy during ECG health monitoring is an urgent problem to be solved. Summary of the invention

[0005] In view of this, the present invention provides a cross-device ECG federal privacy classification framework method, system and storage medium, which can protect user data privacy while improving data recognition accuracy during ECG health monitoring.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions: On the one hand, a cross-device ECG federated privacy classification framework method is provided, and the specific steps include the following:

[0007] S1, the server sends the initial model parameters to the client;

[0008] S2. The client trains the initial model using the training data set to obtain basic type recognition model parameters, and sends the parameters to the server in a differential manner;

[0009] S3, the server receives the basic type recognition model parameters and performs clustering to obtain clustering results; performs weighted averaging on the clustering results to obtain update parameters of the final server model, and completes the training of the initial model;

[0010] S4. Recognize the ECG data through the trained initial model, and output the respective ECG data type recognition results on each client.

[0011] Optionally, the training data set is the sum of hierarchical shared data and client private ECG data.

[0012] By adopting the above technical solution, the following beneficial technical effects are achieved: the method of expanding the client data set by adding globally shared data can make the data distribution across devices more similar, so as to improve the deviation of data distribution in space and solve the Non-IID problem existing in federated learning.

[0013] Optionally, the specific process of step S2 is:

[0014] The client receives the initial model parameters sent by the server;

[0015] Training an initial model based on the training data set to obtain local parameters;

[0016] The differential parameters are calculated according to the local parameters and the initial model parameters and sent to the client. The calculation formula of the differential parameters is as follows:

[0017] Among them, diff i is the difference parameter, is a local parameter, w t are the initial model parameters.

[0018] Optionally, the specific process of step S3 is:

[0019] The client receives the differential parameters and performs clustering using Kmeans, and the calculation formula is:

[0020] Where i is the index number of the client, F1 is the Kmeans clustering function, C k is the kth cluster;

[0021] Update the clustering results and get the updated parameters. The calculation formula is: Among them, L(C k) represents the number of clients belonging to type k, and K represents the total number of types of all clients.

[0022] Optionally, the initial model is a CNN deep learning model.

[0023] On the other hand, a cross-device ECG federated privacy classification framework system is provided, including a data acquisition module, a model training module, a differential parameter calculation module, and a model update module; wherein,

[0024] The data acquisition module is used to acquire initial model parameters;

[0025] The model training module is used to train the initial model to obtain local parameters according to the initial model parameters and the training data set;

[0026] The differential parameter calculation module is used to calculate the differential parameters according to the local parameters and the initial model parameters;

[0027] The model updating module is used to cluster the differential parameters and update the model parameters.

[0028] Finally, a computer storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the cross-device ECG federal privacy classification framework method are implemented.

[0029] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a cross-device ECG federated privacy classification framework method, system and storage medium, which have the following beneficial effects:

[0030] 1) The optimal algorithm FedAVG in existing federated learning is improved, and a federated aggregation framework FedCluster is proposed to increase the contribution of rare clients to the federated model and appropriately reduce the contribution of common clients to the federated model. The model parameters trained by a single client are used to realize knowledge transfer between clients, thus achieving a global optimization effect.

[0031] 2) A hierarchical global shared data generation method is proposed. Through a series of customized parameters, the scale of the shared data set and the selection ratio of different types of data are constrained to adapt to the general distribution of ECG data, which not only improves the recognition ability of the local model, but also improves the diagnostic performance of the federated model.

[0032] 3) It can more efficiently realize knowledge sharing between different clients and the global perception ability of a single client, improve the overall recognition accuracy of ECG data types and the recognition accuracy of rare types of clients, and meet the basic practical requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 It is a data interaction diagram between the client and the server of the present invention;

[0036] Figure 3 It is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Embodiment 1 of the present invention discloses a cross-device ECG federated privacy classification framework method, such as Figure 1 As shown, the specific steps include the following:

[0039] S1, the server sends the initial model parameters to the client;

[0040] S2. The client trains the initial model using the training data set to obtain basic type recognition model parameters, and sends them to the server in a differential manner;

[0041] S3. The server receives the basic type recognition model parameters and performs clustering to obtain clustering results; the clustering results are weighted averaged to obtain the updated parameters of the final server model, thus completing the training of the initial model;

[0042] S4. Recognize the ECG data through the trained initial model, and output the respective ECG data type recognition results on each client.

[0043] Furthermore, in this embodiment, a public standard ECG database MIT-BIH is selected as the data set used in the experiment, the first 20 user records are used as the public data part, and the last 24 user records are used as the private data of the federated client that needs to be privacy protected.

[0044] Furthermore, the training dataset is the sum of the hierarchical shared data and the client's private ECG data.

[0045] To solve the Non-IID problem in federated learning, adding globally shared data to expand the client dataset can make the data distribution across devices more similar, thereby improving the deviation of data distribution in space. The hierarchical shared data used to expand the training data of a single client should have the following characteristics: a) globally shared; b) unrelated to the data generated by all clients; c) not sensitive to privacy, that is, these data do not require privacy protection.

[0046] The existing shared data methods do not consider the shared data generation method in the unbalanced data set. Therefore, in this embodiment, a novel hierarchical global shared data generation method is proposed for all clients. In order to make the selected global shared data (G) meet the above three characteristics a), b) and c) at the same time, the following two conditions need to be met:

[0047] 1. Determine the size of hierarchical shared data

[0048] Define a scaling factor α to quantify the relationship between the total data of all candidate clients (also called candidate set, which is part of the public data set, derived from the total data of the public data set MIT-BIH arrhythmia database 100~124) D and the scale of G, using the formula To express.

[0049] 2. Determine the proportion of various types of data in G

[0050] Define three proportional factors β, γ and θ, which represent the proportion of type N, type S and type V data in the global shared data, respectively, which can be expressed as G = β × N + γ × S + θ × V. Since the total amount of type F and type Q data in the candidate set is too small (13 and 7 respectively), they should all be selected into the global shared data.

[0051] Before step S1, each user terminal collects electrocardiogram data respectively, and the data recorded in the public data set MIT-BIH arrhythmia database 100~124 is distributed to each user terminal according to the ratio determined in the present invention, and combined with the private data collected by each user terminal to form a training data set; after the data is collected and the training data set is formed, steps S1-S4 are performed, and finally the type corresponding to each electrocardiogram data is output on each user terminal.

[0052] The process of data interaction between the client and the server is as follows Figure 2 As shown, further, the specific process of step S2 is:

[0053] The client receives the initial model parameters sent by the server;

[0054] Train the initial model based on the training data set to obtain local parameters;

[0055] The differential parameters are calculated based on the local parameters and the initial model parameters and sent to the client. The calculation formula of the differential parameters is as follows:

[0056] Among them, diff i is the difference parameter, is a local parameter, w t are the initial model parameters.

[0057] Furthermore, the specific process of step S3 is as follows:

[0058] The client receives the differential parameters and uses Kmeans for clustering. The calculation formula is:

[0059] Where i is the index number of the client, F1 is the Kmeans clustering function,

[0060] Update the clustering results and get the updated parameters. The calculation formula is: in, represents the number of clients belonging to type k, and K represents the total number of types of all clients.

[0061] The update process adopts the method of first averaging within the class and then averaging between classes, that is, first averaging the model parameters of the clients that are clustered into the same class to obtain the value of the cluster center, and then averaging the cluster centers based on the average in the previous step. Since only the difference in model parameters is transmitted between the client and the server, it can protect user privacy by avoiding direct exposure of client data, and at the same time significantly improve the communication efficiency of the system.

[0062] Furthermore, the initial model can arbitrarily select a deep learning classification model that performs well in a local or centralized situation. In this embodiment, the initial model is an RCNN deep learning model.

[0063] Embodiment 2 of the present invention discloses a cross-device ECG federated privacy classification framework system, such as Figure 3 As shown, it includes a data acquisition module, a model training module, a differential parameter calculation module, and a model update module; wherein,

[0064] A data acquisition module, used to obtain initial model parameters;

[0065] A model training module is used to train the initial model to obtain local parameters according to the initial model parameters and the training data set;

[0066] A differential parameter calculation module, used for calculating differential parameters according to local parameters and initial model parameters;

[0067] The model updating module is used to cluster the differential parameters and update the model parameters.

[0068] Finally, a computer storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a cross-device ECG federal privacy classification framework method are implemented.

[0069] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0070] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cross-device ECG federated privacy classification framework method, characterized in that: Applied to the server, the specific steps include: The server sends initial model parameters to the client; The server receives the differential parameter sent by the client; the calculation formula of the differential parameter is as follows: Among them, diff i is the difference parameter, is a local parameter, w t are the initial model parameters; The server clusters the differential parameters to obtain clustering results; performs weighted averaging on the clustering results to obtain update parameters of the final server model, and completes the training of the initial model; clustering is performed using Kmeans, and the calculation formula is: Where i is the index number of the client, F1 is the Kmeans clustering function, and the clustering result is updated to obtain the updated parameters. The calculation formula is: in, represents the number of clients belonging to type k, and K represents the total number of types of all clients; The server recognizes the ECG data through the trained initial model, and outputs the respective ECG data type recognition results on each client.

2. A cross-device ECG federated privacy classification framework method, characterized in that: Applied to the client, the specific steps include: S1, the client receives the initial model parameters sent by the server; S2, the client trains the initial model using the training data set in combination with the initial model parameters to obtain basic type recognition model parameters, and obtains differential parameters based on the basic type recognition model parameters and sends them to the server; The specific process of step S2 is: The client receives the initial model parameters sent by the server; Training an initial model based on the training data set to obtain local parameters; The differential parameters are calculated according to the local parameters and the initial model parameters and sent to the client. The calculation formula of the differential parameters is as follows: Among them, diff i is the difference parameter, is a local parameter, w t are the initial model parameters; The specific process of training the initial model is as follows: The client receives the differential parameters and performs clustering using Kmeans, and the calculation formula is: Where i is the index number of the client, F1 is the Kmeans clustering function, C k is the kth cluster; Update the clustering results and get the updated parameters. The calculation formula is: Among them, L(C k ) represents the number of clients belonging to type k, and K represents the total number of types of all clients.

3. A cross-device ECG federated privacy classification framework method according to claim 2, characterized in that: The training data set is the sum of hierarchical shared data and client private ECG data.

4. A cross-device ECG federated privacy classification framework method according to claim 1, characterized in that: The initial model is a CNN deep learning model.

5. A cross-device ECG federated privacy classification framework system, applying any one of claims 1-4, characterized in that: It includes data acquisition module, model training module, differential parameter calculation module and model update module; among them, The data acquisition module is used to acquire initial model parameters; The model training module is used to train the initial model to obtain local parameters according to the initial model parameters and the training data set; The differential parameter calculation module is used to calculate the differential parameters according to the local parameters and the initial model parameters; The model updating module is used to cluster the differential parameters and update the model parameters.

6. A computer storage medium, characterized in that: The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of a cross-device ECG federal privacy classification framework method as described in any one of claims 1-4.

Images