Medical institution data sharing method based on trust degree decentralized federated learning

Abstract

The application discloses a medical institution data sharing method based on trust degree decentralized federated learning, and contains the following steps: all medical institutions are randomly grouped, and representative institutions are selected; the average gradient value and the credibility of each group are calculated; the representative medical institutions calculate the Euclidean distance between themselves and the remaining representative medical institutions; the parameters of the final participating representative medical institutions are selected according to the trust degree / Euclidean distance between the representative institutions; and the average value of the selected institution parameters is calculated to update the global model. The application realizes hierarchical aggregation based on decentralized federated learning, thereby realizing privacy protection and robustness; and experimental results show that in the scene of different scales of malicious medical institutions, the method can achieve good experimental results.

Description

Technical Field

[0001] The present invention relates to the field of privacy protection for medical data sharing, and in particular to methods for federated learning and homomorphic encryption. Specifically, it provides a medical institution data sharing method based on trust-based decentralized federated learning. Background Art

[0002] The explosive growth of technologies such as the internet, big data analytics, and mobile edge computing has driven the development of the "sharing economy." With the generation of massive amounts of medical data and the demands of modern times, the focus of the sharing economy has shifted to the healthcare industry. However, because medical data involves patient privacy and is stored independently in medical institution databases, the use of medical data requires consideration of privacy protection and the issue of "dirty" data during the sharing process. Furthermore, third parties that bring together multiple medical institutions can also pose trust issues, leading to the consideration of decentralized systems for data sharing.

[0003] With the development of medical big data, problems such as medical data security, personal privacy protection, and the difficulty of implementing trusted third parties have become increasingly prominent. The privacy security and robustness of distributed federated learning have received increasing attention. Summary of the Invention

[0004] This invention addresses the problem of providing a trust-based decentralized federated learning method for secure medical data sharing. Building on existing research on decentralized federated learning and combining it with research on the privacy, security, and robustness of centralized federated learning, this approach achieves privacy and security while enhancing the robustness of decentralized federated learning.

[0005] Technical solution:

[0006] A trust-based decentralized federated learning method for sharing medical institution data includes the following steps:

[0007] Step 1: All medical institutions were randomly divided into groups, and one representative medical institution was selected from each group;

[0008] Step 2: Each medical institution in each group trains model parameters based on its local dataset, encrypts the gradient values ​​using a homomorphic encryption algorithm, and sends the encrypted gradient values ​​to the representative medical institution. The representative medical institution calculates the average value of the model parameters within each group and selects the local training dataset as the test set to calculate the credibility;

[0009] Step 3: The representative medical institutions send their parameters and credibility to each other and calculate the Euclidean distance between their own parameters and those of other representative medical institutions;

[0010] Step 4: The representative medical institutions select the parameters of the representative medical institutions that will eventually participate in the aggregation based on the credibility and Euclidean distance;

[0011] Step 5: Calculate the average value of the parameters of the selected representative medical institutions as the parameter value to update the global model.

[0012] Preferably, the method for selecting a representative medical institution in step 1 is as follows:

[0013] All medical institutions are regarded as vertices in an undirected graph, and the communication distances between institutions are regarded as edges in the undirected graph. The shortest path from each medical institution to the remaining institutions is calculated according to the Dijkstra algorithm, and the institution with the shortest path is regarded as the representative medical institution of each group.

[0014] Preferably, the credibility is calculated in step 2 The specific method is:

[0015]

[0016] Where f(·) represents the loss function, f(x;ξ) represents the loss value of the global parameter, x represents the global parameter of the current round, ξ represents the local data sample of the node, η represents the learning rate, and g represents the gradient obtained by local training of a single node. represents the i-th group.

[0017] Preferably, the method for selecting parameters in step 4 is specifically as follows:

[0018]

[0019] in, Indicates the credibility of the corresponding group, represents the Euclidean distance between the gradient of the corresponding group and other representative medical institutions, argmax(·) represents the gradient of the maximum value in the returned array, c represents the number of groups, represents the i-th group.

[0020] Beneficial effects of the present invention

[0021] 1. This paper implements hierarchical aggregation based on distributed federated learning to achieve privacy protection and robustness. Experimental results show that this method can achieve good experimental results in scenarios with malicious medical institutions of different sizes.

[0022] 2. This invention's hierarchical aggregation algorithm, based on decentralized federated learning, ensures both privacy and robustness, and is more resistant to gradient reversal attacks than other methods. Furthermore, it maintains high robustness even when the packet size increases. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the trust-based decentralized federated learning method of the present invention;

[0024] Figure 2 Schematic diagram of the method of the present invention

[0025] Figure 3 This is a schematic diagram of the Gaussian attack result in an embodiment of the present invention.

[0026] Figure 4 Schematic diagram of the Bit-flipattack result in an embodiment of the present invention DETAILED DESCRIPTION

[0027] The present invention will be further described below with reference to the embodiments, but the scope of protection of the present invention is not limited thereto:

[0028] Combine Figure 2 , a schematic diagram of the method of decentralized federated learning based on trust is given. It can be seen that the present invention is mainly divided into five steps, namely, step 1: all medical institutions are randomly grouped and representative institutions are selected; step 2: the average gradient value and credibility of each group are calculated; step 3: the representative medical institution calculates the Euclidean distance between itself and the other representative medical institutions; step 4: the parameters of the representative medical institutions that finally participate in the aggregation are selected based on the trust / Euclidean distance between the representative institutions; step 5: the average value of the parameters of the selected institutions is calculated to update the global model.

[0029] Combine Figure 1 ,A medical institution data sharing method based on trust and decentralized federated learning,including the following steps:

[0030] Step 1: All medical institutions are randomly divided into groups, and one representative medical institution is selected from each group. The specific method for selecting representative medical institutions is as follows:

[0031] All medical institutions are regarded as vertices in an undirected graph, and the communication distances between institutions are regarded as edges in the undirected graph. The shortest path from each medical institution to the remaining institutions is calculated according to the Dijkstra algorithm, and the institution with the shortest path is regarded as the representative medical institution of each group.

[0032] Step 2: Each medical institution in each group trains the model parameters based on its local data set, encrypts the gradient value using the homomorphic encryption algorithm, and sends the encrypted gradient value to the representative medical institution. The representative medical institution calculates the average value of the model parameters in each group and selects the local training data set as the test set to calculate the credibility; Calculate credibility The specific method is:

[0033]

[0034] Where f(·) represents the loss function, f(x;ξ) represents the loss value of the global parameter, x represents the global parameter of the current round, ξ represents the local data sample of the node, η represents the learning rate, and g represents the gradient obtained by local training of a single node. represents the i-th group.

[0035] Step 3: The representative medical institutions send their parameters and credibility to each other and calculate the Euclidean distance between their own parameters and those of other representative medical institutions;

[0036] Step 4: The representative medical institutions select the parameters of the representative medical institutions that will eventually participate in the aggregation based on the credibility and Euclidean distance. The specific method for selecting parameters is as follows:

[0037]

[0038] in, Indicates the credibility of the corresponding group, represents the Euclidean distance between the gradient of the corresponding group and other representative medical institutions, argmax(·) represents the gradient of the maximum value in the returned array, c represents the number of groups, represents the i-th group.

[0039] Step 5: Calculate the average value of the parameters of the selected representative medical institutions as the parameter value to update the global model.

[0040] Example 1

[0041] The performance of the method of the present invention was tested using the MNIST public dataset, and the results are as follows:

[0042] MNIST consists of 60,000 training examples and 10,000 test images. Each image in the dataset represents an Arabic numeral from 0 to 9. The images contain only grayscale information and are sized 28*28. That is, each image has 784 (28*28) columns of data, and the digit is located in the center of the entire image.

[0043] The test was conducted on the MNIST dataset, using Gaussian and Bit-flip as attack methods, setting the learning rate to 0.05 and the number of malicious nodes to 4. In addition, other aggregation algorithms were used for comparison to observe their results.

[0044] Gaussianattack results are as follows Figure 3 As shown, the Bit-flipattack results are as follows Figure 4 The solid line represents the experimental results of the present invention, and the dotted line represents the experimental results of other aggregation algorithms.

[0045] Depend on Figure 3 、 Figure 4 It can be seen that the convergence rate of the present invention is similar to that of other aggregation algorithms, but the convergence accuracy of the present invention is higher and more stable. Therefore, the method of the present invention can achieve better results than other aggregation algorithms.

[0046] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A trust-based decentralized federated learning method for sharing medical institution data, comprising the following steps: Step 1: All medical institutions were randomly divided into groups, and one representative medical institution was selected from each group; Step 2: Each medical institution in each group trains model parameters based on its local dataset, encrypts the gradient values ​​using a homomorphic encryption algorithm, and sends the encrypted gradient values ​​to the representative medical institution. The representative medical institution calculates the average value of the model parameters within each group and selects the local training dataset as the test set to calculate the credibility; Step 3: The representative medical institutions send their parameters and credibility to each other and calculate the Euclidean distance between their own parameters and those of other representative medical institutions; Step 4: The representative medical institutions select the parameters of the representative medical institutions that will eventually participate in the aggregation based on the credibility and Euclidean distance; Step 5: Calculate the average value of the parameters of the selected representative medical institutions as the parameter value to update the global model.

2. The method according to claim 1, characterized in that The specific method for selecting representative medical institutions in step 1 is as follows: All medical institutions are regarded as vertices in an undirected graph, and the communication distances between institutions are regarded as edges in the undirected graph. The shortest path from each medical institution to the remaining institutions is calculated according to the Dijkstra algorithm, and the institution with the shortest path is regarded as the representative medical institution of each group.

3. The method according to claim 2, characterized in that Calculate the credibility in step 2 The specific method is: Where f(·) represents the loss function, f(x, ξ) represents the loss value of the global parameter, x represents the global parameter of the current round, ξ represents the local data sample of the node, η represents the learning rate, and g represents the gradient obtained by local training of a single node. represents the i-th group.

4. The method according to claim 1, wherein The specific method for selecting parameters in step 4 is: in, Indicates the credibility of the corresponding group, represents the Euclidean distance between the gradient of the corresponding group and other representative medical institutions, argmax(·) represents the gradient of the maximum value in the returned array, c represents the number of groups, represents the i-th group.

Images