Multi-Party Secure Computation System Based on Medical Artificial Intelligence and Its Modeling Method

Abstract

The present invention discloses a multi-party secure computing system based on medical artificial intelligence and its modeling method. Each subsystem is communicatively connected through the MPC protocol, and the scheduling and management system is communicatively connected to each subsystem through the MPC protocol. Therefore, when an artificial intelligence model needs to be established at a subsystem, the model is first trained locally. Since the artificial intelligence model also needs to apply the resource data of other subsystems, the artificial intelligence model can be forwarded to other subsystems through the MPC protocol communication method by the scheduling and management system. Furthermore, training is carried out in other subsystems, and the model parameters obtained after training are transmitted to the said one subsystem through the MPC protocol communication method. The MPC protocol is adopted throughout the entire information processing process, and the data is transported, loaded, cleaned, and model-trained under encryption. The entire process completes the AI model training while ensuring the privacy and security of information and the non-disclosure of data.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a multi-party secure computing system based on medical artificial intelligence and its modeling method. Background Art

[0002] With the empowerment of artificial intelligence technology in all walks of life, the AI technology in the medical field has made great progress. At present, medical artificial intelligence data has the following characteristics:

[0003] The data resources are large in quantity;

[0004] The data resources generally involve patient privacy;

[0005] The data resources of different hospitals are scattered and there are sharing barriers;

[0006] The data and its annotation are of great value and high cost.

[0007] The above characteristics of medical artificial intelligence data limit the development of medical artificial intelligence.

[0008] At present, AI developers in major hospitals generally adopt the common practices in the industry, that is: manually collecting medical data, manually annotating, modeling, model training, model deployment, etc. However, due to the consideration of data security of each hospital party, their respective data cannot be shared.

[0009] The Chinese patent with the application number CN202111167309.6 and the name of "Medical Data Sharing Method and Device Based on Blockchain" proposes to use blockchain technology to protect medical data. However, this method cannot seamlessly connect to AI model training, and only the data operation records are tamper-proof, and the data silos still cannot be broken. Summary of the Invention

[0010] The main object of the present invention is to propose a multi-party secure computing system based on medical artificial intelligence and its modeling method, aiming to complete AI model training on the premise that each participating party in the hospital does not disclose its own private data information.

[0011] To achieve the above object, the present invention proposes a modeling method for a multi-party secure computing system based on medical artificial intelligence. The multi-party secure computing system based on medical artificial intelligence includes a plurality of subsystems and a scheduling and management system. The plurality of subsystems are correspondingly arranged at a plurality of hospital ends. The plurality of subsystems include a first subsystem and a second subsystem. Each of the subsystems is communicatively connected through an MPC protocol. The scheduling and management system is communicatively connected to each of the subsystems. The modeling method for the multi-party secure computing system based on medical artificial intelligence includes the following steps:

[0012] Step S10: The first subsystem receives a model training request initiated by a staff member at the hospital end where it is located. The model training request includes the specified data resources of the multiple subsystems and an artificial intelligence model. The specified data resources are located in the first subsystem and the second subsystem. The artificial intelligence model is trained using the resource data of the first subsystem to obtain local model parameters, and the model training request is forwarded to the scheduling and management system.

[0013] Step S20: The scheduling and management system forwards the artificial intelligence model in the model training request to the second subsystem.

[0014] Step S30: The second subsystem trains the artificial intelligence model using its specified data resources to obtain remote model parameters, and transmits the remote model parameters to the first subsystem via MPC.

[0015] Optionally, the multiple subsystems further include a third subsystem that does not have the specified data resources.

[0016] After step S30, it includes:

[0017] Step S31: The scheduling and management system obtains the current load rate of the first subsystem and the current load rate of the second subsystem.

[0018] Step S32: Compare the current load rate of the first subsystem with the current load rate of the second subsystem to obtain a difference.

[0019] Step S33: When the difference is greater than a first set value, the scheduling and management system transmits some of the specified data resources of the subsystem with the higher load rate among the first subsystem and the second subsystem to the third subsystem with a load rate lower than a second set value among the multiple subsystems via MPC, and sends the artificial intelligence model to the third subsystem.

[0020] Step S34: The third subsystem trains the artificial intelligence model using the some of the specified data resources to obtain local model parameters, and transmits the local model parameters to the first subsystem via MPC.

[0021] Optionally, before step S33, it further includes:

[0022] Step S35: Obtain the current load rate of the third subsystem.

[0023] Step S36: When the load rate of the third subsystem is lower than the second set value, then enter step S33.

[0024] Optionally, the third subsystem is set to multiple;

[0025] After step S35, it further includes:

[0026] Step S36, when the load rate of the third subsystem is greater than the second set value and less than the third set value, the scheduling and management system splits a part of the specified data resources of the party with the larger load rate in the first subsystem and the second subsystem according to the proportion of the load rate of each third subsystem, and transmits the split multiple parts of the specified data resources to the corresponding multiple third subsystems through MPC respectively, and transmits the artificial intelligence model to each third subsystem through MPC respectively;

[0027] Step S37, each third subsystem trains the artificial intelligence model respectively according to the corresponding split specified data resources to obtain local model parameters, and transmits the local model parameters to the first subsystem through MPC.

[0028] Optionally, the artificial intelligence model includes one of a machine learning model, a deep learning model, and a reinforcement learning model.

[0029] Optionally, the data resources include medical images or medical text information.

[0030] The present invention also provides a multi-party secure computing system based on medical artificial intelligence, characterized in that the multi-party secure computing system based on medical artificial intelligence includes a plurality of subsystems and a scheduling and management system, the plurality of subsystems are correspondingly arranged at a plurality of hospital ends, the plurality of subsystems include a first subsystem and a second subsystem, and each subsystem is communicatively connected through the MPC protocol, and the scheduling and management system is communicatively connected to each subsystem through the MPC protocol.

[0031] In the technical solution provided by the present invention, each subsystem is communicatively connected through the MPC protocol, and the scheduling and management system is communicatively connected to each subsystem through the MPC protocol. Therefore, when an artificial intelligence model needs to be established at a first subsystem, model training is first performed locally. Because the artificial intelligence model also needs to apply the resource data of other subsystems, the artificial intelligence model can be forwarded to other subsystems through the MPC protocol communication method by the scheduling and management system. Furthermore, training is performed in other subsystems, and the model parameters obtained after training are transmitted to the first subsystem through the MPC protocol communication method. The entire information processing process uses the MPC protocol, and data is transported, loaded, cleaned, and model-trained under encryption. The training result can be obtained while ensuring the privacy security of information and the non-disclosure of data, and encryption transmission, operation, and analysis and processing can be realized to complete the AI model training. Description of the Drawings

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the structures shown in these drawings.

[0033] Figure 1 It is a schematic flowchart of an embodiment of the modeling method of the multi-party secure computing system based on medical artificial intelligence provided by the present invention;

[0034] Figure 2 It is a schematic framework diagram of an embodiment of the multi-party secure computing system based on medical artificial intelligence provided by the present invention;

[0035] Figure 3 It is a specific schematic flowchart of an embodiment of the modeling method of the multi-party secure computing system based on medical artificial intelligence provided by the present invention;

[0036] Figure 4 It is a schematic framework diagram of the multi-party secure computing mechanism among various hospital systems in the multi-party secure computing system based on medical artificial intelligence provided by the present invention.

[0037] The realization of the purpose, functional characteristics and advantages of the present invention will be further described with reference to the embodiments and the drawings. Specific Embodiments

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0039] It should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative position relationship and movement conditions between components in a specific posture (as shown in the drawings). If the specific posture changes, the directional indications will also change accordingly.

[0040] In addition, if there are descriptions such as "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text includes three parallel scenarios. Taking "A and / or B" as an example, it includes Scenario A, or Scenario B, or the scenario where both A and B are satisfied simultaneously. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0041] The present invention provides a modeling method for a multi-party secure computing system based on medical artificial intelligence. The multi-party secure computing system based on medical artificial intelligence includes multiple subsystems and a scheduling and management system. The multiple subsystems are correspondingly arranged at multiple hospital terminals. The multiple subsystems include a first subsystem and a second subsystem. Each of the subsystems is communicatively connected through an MPC protocol. The scheduling and management system is communicatively connected to each of the subsystems; as Figure 1 shown, the modeling method for the multi-party secure computing system based on medical artificial intelligence includes the following steps:

[0042] Step S10: The first subsystem receives a model training request initiated by a staff member at its corresponding hospital terminal. The model training request includes the specified data resources of the multiple subsystems and an artificial intelligence model. The specified data resources are correspondingly located in the first subsystem and the second subsystem. The artificial intelligence model is trained using the resource data of the first subsystem to obtain local model parameters, and the model training request is forwarded to the scheduling and management system;

[0043] Step S20: The scheduling and management system forwards the artificial intelligence model in the model training request to the second subsystem;

[0044] Step S30: The second subsystem trains the artificial intelligence model using its specified data resources to obtain foreign model parameters, and transmits the foreign model parameters to the first subsystem through MPC.

[0045] Secure Multi-party Computation (MPC) emerged as the times require. It can achieve the sharing of private data while protecting personal privacy information. MPC means that a group of mutually distrustful participants can perform collaborative computing while protecting personal privacy.

[0046] In the technical solution provided by the present invention, each of the subsystems is communicatively connected through the MPC protocol, and the scheduling management system is communicatively connected to each of the subsystems through the MPC protocol. Therefore, when an artificial intelligence model needs to be established at a subsystem, such as the first subsystem, model training is first performed locally. Since the artificial intelligence model also needs to apply the resource data of other subsystems, such as the second subsystem, the artificial intelligence model can be forwarded to other subsystems through the MPC protocol communication method by the scheduling management system. Furthermore, training is performed in other subsystems, and the model parameters obtained after training are transmitted to the one subsystem through the MPC protocol communication method. The MPC protocol is used throughout the entire information processing process, and data is transported, loaded, cleaned, and model-trained under encryption. The training results can be obtained while ensuring the privacy and security of information and the non-disclosure of data, and encrypted transmission, operation, and analysis and processing can be achieved to complete the AI model training.

[0047] In an embodiment of the present invention, the multiple subsystems further include a third subsystem, and the third subsystem does not have the specified data resources, that is, the resource data required by the artificial intelligence model exists in some subsystems and does not exist in some subsystems. For example, in the embodiment of the present invention, the first subsystem and the second subsystem have the resource parameters required by the artificial intelligence model, while the third subsystem does not. That is, normally, the third subsystem does not participate in model training.

[0048] In the embodiment of the present invention, after step S30, it includes:

[0049] Step S31, the scheduling management system obtains the current load rate of the first subsystem and the current load rate of the second subsystem;

[0050] Step S32, comparing the current load rate of the first subsystem with the current load rate of the second subsystem to obtain a difference;

[0051] Step S33, when the difference is greater than the first set value, the scheduling management system transmits a part of the specified data resources of the party with the higher load rate in the first subsystem and the second subsystem to the third subsystem with a load rate lower than the second set value among the multiple subsystems through MPC, and sends the artificial intelligence model to the third subsystem;

[0052] Step S34, the third subsystem trains the artificial intelligence model through the part of the specified data resources to obtain local model parameters, and transmits the local model parameters to the one subsystem through MPC.

[0053] When the difference in the load rates of the first subsystem and the second subsystem is large, it indicates that there will be an obvious difference in their computing efficiencies, which will further lead to a relatively large time difference in the final results, and further affect the final model training efficiency. Therefore, in this embodiment, when the difference is large, the scheduling and management system will transfer a part of the specified data resources of the party with the higher load rate among the first subsystem and the second subsystem to the third subsystem with a load rate lower than the second set value among the multiple subsystems through MPC, and send the artificial intelligence model to the third subsystem. The third subsystem trains the artificial intelligence model with the part of the specified data resources to obtain local model parameters, and transfers the local model parameters to a subsystem through MPC, so that the progress of model training of each subsystem can be kept as consistent as possible to improve the model training efficiency.

[0054] Further, in the embodiment of the present invention, before step S33, it further includes:

[0055] Step S35: Obtain the current load rate of the third subsystem;

[0056] Step S36: When the load rate of the third subsystem is lower than the second set value, then enter step S33.

[0057] That is, only when the load rate of the third subsystem is low can it be borrowed.

[0058] Further, in the embodiment of the present invention, the third subsystem is set to be multiple;

[0059] After step S35, it further includes:

[0060] Step S36: When the load rate of the third subsystem is greater than the second set value and less than the third set value, the scheduling and management system splits a part of the specified data resources of the party with the higher load rate among the first subsystem and the second subsystem according to the proportion of the load rates of each third subsystem, transfers the split multiple parts of the specified data resources to the corresponding multiple third subsystems through MPC respectively, and transfers the artificial intelligence model to each third subsystem through MPC respectively;

[0061] Step S37: Each third subsystem trains the artificial intelligence model according to the corresponding split specified data resources to obtain local model parameters, and transfers the local model parameters to a subsystem through MPC.

[0062] In an embodiment of the present invention, for the case where there are multiple third subsystems, and when the load rate of the third subsystems is relatively high, it can be shared by multiple subsystems. Specifically, when the third subsystem with a high load rate is allocated less resource data for training, and when the third subsystem with a low load rate is allocated more resource data for training.

[0063] In an embodiment of the present invention, the artificial intelligence model includes one of a machine learning model, a deep learning model, and a reinforcement learning model.

[0064] In an embodiment of the present invention, the data resources include medical images or medical text information.

[0065] The present invention also provides a multi-party secure computing system based on medical artificial intelligence. The multi-party secure computing system based on medical artificial intelligence includes multiple subsystems and a scheduling and management system. The multiple subsystems are correspondingly arranged at multiple hospital terminals. The multiple subsystems include a first subsystem and a second subsystem. Each of the subsystems is communicatively connected through an MPC protocol. The scheduling and management system is communicatively connected to each of the subsystems through an MPC protocol.

[0066] Specifically, in an embodiment of the present invention, as Figure 2 shown, the multi-party secure computing system based on medical artificial intelligence includes a basic layer, a data transmission layer, a data calculation layer, a service layer, a display layer, and a front-end UI layer, as Figure 1 shown. Specifically:

[0067] Basic layer: It includes various server clusters, specifically the Linux server clusters owned by the hospitals participating in the AI project;

[0068] Data transmission layer: It adopts a multi-party secure computing (MPC) data transmission protocol, and the encryption algorithm is homomorphic encryption, which is a type of encryption method with special natural properties and can perform data operations in the ciphertext domain.

[0069] Data calculation layer: It includes traditional machine learning classification and regression algorithms, and also includes deep learning algorithms such as CNN (Convolutional Neural Network), YOLO (Object Detection Algorithm), and GCN (Graph Neural Network Algorithm); it also includes the calculation of joint modeling, specifically the joint parameter collaborative calculation between servers;

[0070] Service layer: It mainly includes common service requirements of medical artificial intelligence, such as medical image classification, lesion target detection, and lesion image segmentation.

[0071] Display layer: It includes template engine rendering and Ajax interaction technologies (including POST, GET, PUT requests, etc.);

[0072] Front - end UI: including Html, CCS, JQuery, pictures, texts, etc., mainly for user front - end interaction layout and rendering technology.

[0073] The specific process of modeling a multi - party secure computing system based on a medical artificial intelligence is as Figure 3 shown:

[0074] S21, taking Hospital A as an example, algorithm engineer a in Hospital A builds models, including machine learning models, deep learning models, etc.;

[0075] S22, algorithm engineer a initiates a model training request to the multi - party secure computing system through the multi - party secure computing subsystem. The request includes specified data resources and artificial intelligence models in each multi - party secure computing subsystem. This data resource is the data resource marked by each hospital in the system, including but not limited to medical images, medical text information, etc.;

[0076] S23, the scheduling system transmits the architecture of the model to the subsystems of each hospital. The scheduling system involved is the TensorFlow TF - Encrypted multi - party secure computing distributed system, and the homomorphic encryption algorithm is used in the transmission process;

[0077] S24, other subsystems perform training based on local pictures and the model. The model is the encrypted model established by algorithm engineer a;

[0078] S25, other subsystems transmit the trained model parameters to the subsystem of Hospital A through the MPC protocol. The model parameters involved are the weight parameters of the machine learning and deep learning algorithm network structures. The MPC protocols involved are homomorphic encryption, secret sharing, zero - knowledge proof, etc.;

[0079] S26, the subsystem of Hospital A completes the model training and saves it locally;

[0080] S27, model deployment: algorithm engineer a in Hospital A deploys the trained model for service. The service deployment uses the TensorFlow Serving model deployment service, which conducts interface data interaction with the service to be predicted through an interface form;

[0081] S28, model prediction. The external service makes predictions by calling the TensorFlow Serving prediction interface.

[0082] The multi - party secure computing mechanism between each hospital system is as Figure 4 shown, Figure 4 shown. Hospitals A, B, and C are only for illustrating the interaction between hospital systems, and there is no clear limit on the number of hospitals. Figure 4The sub-system performs the MPC calculation as shown in the figure. The scheduling management system is simply referred to as scheduling management. The medical data resources of Hospitals A, B, and C exist in the server cluster systems of their respective hospitals.

[0083] The process of the multi-party secure computing system is described as follows (taking Hospital A as an example):

[0084] Algorithm engineers in Hospital A build an artificial intelligence model, which includes but is not limited to machine learning, deep learning, reinforcement learning, etc.

[0085] The data required for the model comes from Hospitals A, B, and C. The model framework is TF-Encrypted secure multi-party computing (MPC). The server clusters of Hospitals A, B, and C as a whole form a distributed cluster system. Data transmission and interaction between sub-systems are carried out through MPC communication.

[0086] Hospital A loads and trains the model with local data through the local cluster. At the same time, TF-Encrypted secure multi-party computing, through scheduling management, trains the model with the local data of the clusters of Hospitals B and C. Among them, the model encrypts the model in the sub-system of Hospital A and transmits it to the sub-systems of Hospitals B and C through MPC communication and system scheduling management.

[0087] The model parameters trained by the sub-systems of Hospitals A, B, and C are encrypted and transmitted to Hospital A through MPC communication for further parameter reduction operations. The reduction operations involved are function operations related to the model structure, such as addition, multiplication, AES, set intersection, etc.; it is also applicable to all general algorithms that can be represented as a calculation process.

[0088] After distributed MPC calculation, the sub-system of Hospital A obtains the final parameters of the model.

[0089] The above is only the preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structural transformation made under the inventive concept of the present invention, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A modeling method for a multi-party secure computing system based on medical artificial intelligence, characterized in that, The multi-party secure computing system based on medical artificial intelligence includes multiple subsystems and a scheduling and management system. The multiple subsystems are correspondingly arranged in multiple hospital terminals. The multiple subsystems include a first subsystem and a second subsystem. Each of the subsystems is communicatively connected through an MPC protocol. The scheduling and management system is communicatively connected to each of the subsystems through an MPC protocol. The modeling method of the multi-party secure computing system based on medical artificial intelligence includes the following steps: Step S10: The first subsystem receives a model training request initiated by a staff member from the hospital terminal where it is located. The model training request includes the specified data resources of the multiple subsystems and an artificial intelligence model. The specified data resources are correspondingly located in the first subsystem and the second subsystem. The artificial intelligence model is trained using the resource data of the first subsystem to obtain local model parameters, and the model training request is forwarded to the scheduling and management system; Step S20: The scheduling and management system forwards the artificial intelligence model in the model training request to the second subsystem; Step S30: The second subsystem trains the artificial intelligence model using its specified data resources to obtain off-site model parameters, and transmits the off-site model parameters to one of the subsystems through MPC; The multiple subsystems further include a third subsystem, and the third subsystem does not have the specified data resources; After step S30, it includes: Step S31: The scheduling and management system obtains the current load rate of the first subsystem and the current load rate of the second subsystem; Step S32: Compare the current load rate of the first subsystem with the current load rate of the second subsystem to obtain a difference; Step S33: When the difference is greater than a first set value, the scheduling and management system transmits part of the specified data resources of the party with the higher load rate among the first subsystem and the second subsystem to the third subsystem with a load rate lower than a second set value among the multiple subsystems through MPC, and sends the artificial intelligence model to the third subsystem; Step S34: The third subsystem trains the artificial intelligence model using the part of the specified data resources to obtain local model parameters, and transmits the local model parameters to one of the subsystems through MPC.

2. The modeling method for a multi-party secure computing system based on medical artificial intelligence according to claim 1, characterized in that, Before step S33, it further includes: Step S35: Obtain the current load rate of the third subsystem; Step S36: When the load rate of the third subsystem is lower than the second set value, then enter step S33.

3. The modeling method for a multi-party secure computing system based on medical artificial intelligence according to claim 2, characterized in that, The third subsystem is set to be multiple; After step S35, it further includes: Step S36: When the load rate of the third subsystem is greater than the second set value and less than the third set value, the scheduling and management system splits a part of the designated data resources of the party with the higher load rate among the first subsystem and the second subsystem according to the proportion of the load rates of the respective third subsystems, and transmits the split multiple parts of the designated data resources to the corresponding multiple third subsystems through MPC respectively, and transmits the artificial intelligence model to each of the third subsystems through MPC respectively; Step S37: Each of the third subsystems trains the artificial intelligence model according to the corresponding split designated data resources to obtain local model parameters, and transmits the local model parameters to one of the subsystems through MPC.

4. The modeling method for a multi-party secure computing system based on medical artificial intelligence according to claim 1, characterized in that, The artificial intelligence model includes one of a machine learning model, a deep learning model, and a reinforcement learning model.

5. The modeling method for a multi-party secure computing system based on medical artificial intelligence according to claim 1, characterized in that, The data resources include medical images or medical text information.

6. A multi-party secure computing system based on medical artificial intelligence, characterized in that, The multi-party secure computing system based on medical artificial intelligence includes multiple subsystems and a scheduling and management system. The multiple subsystems are correspondingly arranged at multiple hospital ends. The multiple subsystems include a first subsystem and a second subsystem. Each of the subsystems is communicatively connected through the MPC protocol. The scheduling and management system is communicatively connected to each of the subsystems through the MPC protocol. The multi-party secure computing system based on medical artificial intelligence is used to implement the modeling method of the multi-party secure computing system based on medical artificial intelligence according to any one of claims 1-5.

Images