Learning-model processing device, remote learning system and learning-model processing program

Abstract

To provide a learning-model processing device, remote learning system and learning-model processing program that perform private machine learning on a learning model transferred from a different device without bringing out data for learning model.SOLUTION: A learning-model processing device of the present invention comprises a processor, and the processor accepts a transfer of an encrypted learning model from a learning-model generating device, acquires a group of data for learning model, makes the learning model learn the group of data for learning model by secure computation and derives an encrypted learnt model, and transfers the learnt model to the learning-model generating device.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to a learning model processing device, a remote learning system, and a learning model processing program.

Background Art

[0002] When performing machine learning using the health information of a plurality of individuals on a learning model generated by data users such as pharmaceutical companies and medical device manufacturers, it is necessary to take out the health information across borders from the data management bases such as health examination facilities and medical institutions that manage the health information to the data utilization bases managed by the data users, or to move the learning model for performing machine learning to the data management bases. Health information includes personal information, and a large number of health information inputs are required for machine learning. On the other hand, when moving the learning model, the learning model, which is confidential information, will be disclosed to the data management base side.

[0003] Patent Document 1 describes a distributed machine learning system that provides an encrypted learned model between a client device and a server device using a secure communication path established by authenticating each other's startup correctness. From each client device, the server device is provided with an encrypted client model, and from the server device to each client device, an encrypted global model is provided. It is also described that the server device aggregates a plurality of client models provided from each client device in a state of homomorphic encryption to obtain a homomorphic encrypted global model.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] In Patent Document 1, there is a description of transferring a learned learning model between different devices and aggregating a plurality of learned models into one learned model while keeping them in homomorphic encryption. However, there is no description of inputting learning model data at the transfer destination for the encrypted model information to perform machine learning, and there is a risk that the learning model in the encrypted state may not be available for machine learning.

[0006] An object of the present invention is to provide a learning model processing apparatus, a remote learning system, and a learning model processing program that do not take out learning model data and perform machine learning on a learning model transferred from different devices in a non-disclosed manner.

Means for Solving the Problems

[0007] The learning model processing apparatus of the present invention includes a processor. The processor receives the transfer of an encrypted learning model from a learning model generation device, acquires a group of learning model data, learns the group of learning model data with the learning model by confidential computing to derive an encrypted learned model, and transfers the learned model to the learning model generation device.

[0008] Preferably, the processor separates the group of learning model data acquired in advance into a training data group and an evaluation data group, inputs the evaluation data group into the learned model to obtain an evaluation result, and transfers quality data based on the evaluation result to the learning model generation device.

[0009] Preferably, the processor performs anonymization processing on the evaluation result to generate quality data.

[0010] Preferably, the processor determines the price of the learned model to be paid by the user from the evaluation result and transfers the claim amount based on the price to the learning model generation device.

[0011] Preferably, the processor determines the price according to the type of data included in the group of learning model data.

[0012] In deriving the learned model, the processor preferably obtains the training evaluation result representing the accuracy during training, calculates the degree of overfitting from the deviation amount between the training evaluation result and the evaluation result, and reduces the claim amount according to the degree of overfitting.

[0013] The processor stores the provider of the data constituting the data group for the learning model, receives, from the learning model generation device, the target variable information specifying the conditions of the data constituting the data group for the learning model together with the learning model, and preferably determines the amount of remuneration to be paid for each provider based on the price, the target variable information, and the total number of records in the data group for the learning model.

[0014] The processor preferably groups the providers according to the type of data in the target variable information, and sets the amount of remuneration paid to the provider of the data that matches the type to be higher than the amount of remuneration paid to the provider of the data that does not match the type.

[0015] The processor preferably presents the amount of remuneration and the target variable information to the provider.

[0016] The data group for the learning model is preferably a data group related to health, and the learned model is preferably a learned model obtained by learning about data related to health.

[0017] In the remote learning system of the present invention, the data group for the learning model used for learning the learning model is distributed and held by a plurality of learning model processing devices as different data groups for distributed learning models. The learning model processing devices perform distributed learning on the data groups for distributed learning models they each hold using confidential computing means on the learning model generated and encrypted by the learning model generation device. The learning model generation device receives the input of the learned model derived by the distributed learning.

[0018] As a plurality of learning model processing devices that perform distributed learning, there are a first learning model processing device that holds a first distributed learning model data group, which is a data group for a distributed learning model, and a second learning model processing device that holds a second distributed learning model data group, which is a data group for a distributed learning model. The first learning model processing device inputs the first distributed learning model data group to the learning model transferred from the learning model generation device and derives an in-learning model in which learning has been stopped in the middle of learning. The second learning model processing device preferably acquires the in-learning model transferred from the first learning model processing device, inputs the second distributed learning model data group, and resumes learning.

[0019] A plurality of learning model processing devices that perform distributed learning input a data group for a distributed learning model to the learning model acquired from the learning model generation device and respectively derive a distributed learning completed model. It is preferable that any one of the learning model processing devices integrates the plurality of distributed learning completed models and derives a learned model.

[0020] The learning model processing program of the present invention causes a computer to function as a learning model transfer means that receives the transfer of the encrypted learning model from the learning model generation device and transfers the derived learned model to the learning model generation device, a data acquisition means that acquires a data group for a learning model, and a learning model training means that causes the learning model to learn the data group for a learning model by confidential calculation and derives a learned model.

Effects of the Invention

[0021] According to the present invention, it is possible to perform machine learning on a learning model transferred from different devices without taking out the data for the learning model and without disclosing it.

Brief Description of the Drawings

[0022]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Embodiments for Carrying Out the Invention

[0023] [First Embodiment] As shown in FIG. 1, the remote learning system 10 includes a learning model generation device 11 that generates a learning model, a learning model processing device 12 that performs machine learning on the learning model using a data group for the learning model, and a provider terminal 13 that receives a notification from the learning model processing device 12.

[0024] The learning model generation device 11 is a device having a function of generating, transmitting, and receiving a learning model owned by a data user such as a pharmaceutical company or a medical device manufacturer. Based on the instructions and operations of the data user, the generated learning model is encrypted and sent to the learning model processing device 12.

[0025] The learning model processing device 12 is a device that enables a data administrator, such as a health examination facility or a medical institution, to perform machine learning on a learning model to derive a learned model. It acquires the learning model through transfer from the learning model generation device 11 and transfers the derived learned model to the learning model generation device 11. The machine learning on the learning model is executed without disclosing it to the users or administrators of the learning model processing device 12 by using confidential computing means realized by hardware encryption. In addition, it calculates the price of the learned model and notifies a plurality of provider terminals 13 of the amount of remuneration for the data provider based on the price.

[0026] The provider terminal 13 is an information terminal such as a PC (Personal Computer), a smartphone, or a wearable device that can be accessed by each provider group that directly or indirectly provides each piece of learning model data constituting the learning model data group to the learning model processing device 12. The provider terminal 13 may be a personally owned information terminal, or may be a medical information management device that is a server for managing the health data of multiple people provided in each medical institution or medical-related company, or a medical device that is a wearable device that measures and transmits health data lent by a medical institution or the like to a provider.

[0027] In the remote learning system 10, the learning model generation device 11 does not disclose the learning model to the learning model processing device 12, and the learning model processing device 12 performs machine learning without taking out the learning model data group to the outside.

[0028] As shown in FIG. 2, in the remote learning system 10, the learning model processing device 12 receives the transfer of the decryption key, the target variable information, and the encrypted pre-learning model image from the learning model generation device 11, and performs machine learning to learn the learning model data group acquired in advance. The learning model data group used for machine learning is separated into a training data group and an evaluation data group. Machine learning is performed by learning the training data group on the learning model through confidential computing, and an encrypted learned model is derived. The pre-learning model image has an unlearned learning model.

[0029] After the learned model is derived, the learned model generation device 11 receives the transfer of quality data, billing information, and the encrypted learned model image from the learned model processing device 12. The quality data is generated by performing anonymization processing on the evaluation results obtained by inputting the evaluation data group into the learned model by the learned model processing device 12. The billing information is information on the amount to be billed based on the price of the learned model paid by the user, and the price of the learned model is determined from the evaluation results.

[0030] The remote learning system 10 connects the learned model generation device 11 and the learned model processing device 12 in accordance with REST (Representational State Transfer). That is, the transfer of the learned model is performed on the web with a unified interface in which the data formats are common, addressability in which all information has a unique identifier, connectivity in which the communicated information includes hyperlinks, and statelessness in which the communication is completed each time.

[0031] The data group for the learned model is data provided by a plurality of providers and is used for machine learning for the learned model. For example, the provided data is medical data including personal information regarding the health of the provider, such as PHR (Personal Health Record) data. In that case, the learned model is a learned model that has learned about health-related data. The PHR data is data including at least any one of the results of hospital examinations and tests, prescriptions, the results of regular health checkups, past medical history, allergies, the progress of pregnancy and childbirth, and vital data such as blood pressure and pulse measured at home.

[0032] As shown in FIG. 3, in the learning model generation device 11, a central control unit (not shown) constituted by a data control processor causes a program in the program memory to operate, thereby realizing the functions of a learning model generation unit 20, an encryption processing unit 21, a decryption key setting unit 22, a target variable information determination unit 23, a data transmission / reception unit 24, an input reception unit (not shown), and an output control unit (not shown). The input reception unit is connected to a user interface and receives operations from the user.

[0033] The learning model generation device 11 is electrically connected to a display (not shown) and a user interface (UI) (not shown). The display displays the generation status of the learning model and information on the connected learning model processing device 12. The user interface is an input device through which a user of the learning model generation device 11 performs setting input for learning model generation, input of target variable information, etc., and includes a keyboard, a mouse, and the like.

[0034] The learning model generation unit 20 generates a learning model that executes preprocessing, learning model training processing (training processing), and learning model evaluation processing (evaluation processing). The learning model is expressed in a format such as a VM (Virtual Machine) image or a Docker image, and has a configuration for storing data input and output in each process. Further, the learning model has an interface through which preprocessing, training processing, and evaluation processing can be started from the outside.

[0035] The preprocessing reads in a data group for the learning model and performs regularization to exclude outliers. The regularized data group for the learning model is stored as a regularized training data group and a regularized evaluation data group. The training processing reads in the regularized training data group and performs training. The learned model obtained by the training is stored. The evaluation processing reads in the regularized evaluation data group and evaluates the learned model. The evaluation result for the learned model is output and stored.

[0036] The encryption processing unit 21 encrypts the learning model image having the generated learning model. For encryption, for example, the BitLocker method is used. Also, in the Linux (registered trademark) environment, the LUKS (Linux Unified Key Setup on disk format) method may be used.

[0037] The decryption key setting unit 22 sets a decryption key in an unencrypted state for the encrypted learning model image. Also, instead of transferring the decryption key together with the learning model image, the decryption key may be obtained from a URL set and managed by the data user, and the learning model may be decrypted with stronger security.

[0038] The target variable information determination unit 23 realizes the target variable input means, receives the input of the user via the user interface, and determines the target variable information to be sent to the learning model processing device 12. The target variable information is information that specifies the types of health information included in the learning model data group used for generating the learned model. The data user inputs, for example, "diabetes" which is a medical history, or "blood pressure range or threshold" which is health information. Instead of the input of the user via the user interface, previously registered past target variable information or the like may be used. By inputting the target variable information, learning using the learning model data group having case tendency values or medical images equal to or higher than a threshold value set in advance for a specific disease is requested.

[0039] The data transmission / reception unit 24 transfers the encrypted learning model image, the decryption key, and the target variable information to the learning model processing device 12. Also, it acquires the learned learning model image, quality data, and charge information from the learning model processing device 12. The data transmission / reception unit 24 realizes the learning model image transfer means in the learning model generation device 11.

[0040] As shown in FIG. 4, in the learning model processing apparatus 12, a central control unit (not shown) constituted by a data control processor causes a program in the program memory to operate, thereby providing a learning model data storage unit 30, a confidential calculation unit 31, a price determination unit 40, a quality data creation unit 42, a reward calculation unit 43, and a data transmission / reception unit 44. The confidential calculation unit 31 includes an evaluation data extraction unit 32, a preprocessing unit 33, a training processing unit 34, and an evaluation processing unit 35. Further, the price determination unit 40 includes an overfitting determination unit 41.

[0041] In addition, the learning model processing apparatus 12 has an output control unit (not shown) that outputs data to a display which is an external device, and an input reception unit (not shown) that receives an input such as a unit price of data in price determination from a user interface (UI) which is an input device. The output control unit has a program related to processing such as image processing stored in a program memory (not shown), and displays the learning status of the learning model on the electrically connected display as needed.

[0042] The learning model data storage unit 30 includes a storage memory (not shown) and stores a learning model data group used by the confidential calculation unit 31. The learning model data group is obtained by pre-collecting learning model data from a plurality of providers. The collection may be obtained from the provider terminals 13 respectively, or may be collected via a medical institution or the like.

[0043] The confidential computing unit 31 implements confidential computing means and is implemented on a CVM (Confidential Virtual Machine) with memory protected using hardware encryption processing so that data cannot be viewed from the outside. Since the data on the confidential computing unit 31 is also protected from the users and administrators of the learning model processing device 12, machine learning can be performed with the configuration of the learning model decrypted using a decryption key in a non-disclosed state. The confidential computing unit 31 is compatible with the encryption used by the learning model generation device 11 and performs the same encryption when transmitting the learning model image to the learning model generation device 11. Also, it implements data acquisition means and acquires the learning model data group stored in the learning model data storage unit 30.

[0044] The evaluation data extraction unit 32 implements evaluation data extraction means and performs an extraction process of separating the acquired learning model data group into a training data group used for training at a specific ratio and an evaluation data group used for evaluating the learned model, and extracting the evaluation data group not used for training. The extraction process is preferably at the timing when the learning model and the target variable information are acquired from the learning model generation device 11. Also, a plurality of training data groups and evaluation data groups may be generated by K-fold cross-validation for cross-validation.

[0045] The preprocessing unit 33 implements preprocessing means and has a function of executing preprocessing. In the preprocessing, preprocessing including regularization is performed on the training data group and the evaluation data group obtained in the evaluation data extraction unit 32. Outlier data in the training data group and the evaluation data group can be excluded by regularization. The preprocessing unit obtains a preprocessed training data group and a preprocessed evaluation data group.

[0046] The training processing unit 34 implements learning model training means and has a function of executing a training process on the learning model. In the training process, learning is performed on the learning model included in the pre-learning learning model image using the preprocessed training data group, and a learned model is generated. The learned model has a function of discriminating the items input to the target variable information by machine learning.

[0047] The evaluation processing unit 35 implements learning model evaluation means, and performs evaluation processing for inputting the preprocessed evaluation data group into the learned model and outputting the evaluation result. The evaluation result is represented by an AUC (Area Under the Curve) value that takes a value from 0 to 1.0. The AUC value is a value indicating the discrimination ability to discriminate the target variable information of the learned model. The closer the value is to 1.0, the higher the discrimination ability. When the AUC value is around 0.5, for example, in the range of 0.4 to 0.6, there is no discrimination ability. When the discrimination ability is random, it indicates 0.5. Therefore, when the AUC value is at least greater than 0.6, it can be determined that the learned model has significant discrimination ability. Even when the AUC value is lower than 0.4, for example, if it is 0.1 or less, the determination result may be reversed, and it can be determined that the discrimination ability is high.

[0048] In addition, the preprocessed evaluation data group may be input to the learning model during training for which the training process has been temporarily stopped to obtain the evaluation result during training. The evaluation result during training represents the accuracy during training.

[0049] The evaluation process is also affected by the quality of the learning model image created by the data user in addition to the data group for the learning model. Therefore, even when it is evaluated that there is no discrimination ability in the evaluation process and training is performed again with a different data group for the learning model, an upper limit on the number of times of repeated execution of training is provided. For example, the number of times training can be repeated is set to 2 or less, and it is not repeated after the third evaluation process. The degree of variation of the AUC values obtained multiple times can also be evaluated as the quality of the data.

[0050] The price determination unit 40 implements price determination means, obtains the evaluation result from the evaluation processing unit 35, and determines the price of the learned model based on the evaluation result. Based on the determined price, the billing amount, which is the total amount to be paid by the data user such as the operator of the learning model generation device 11, is calculated. The price is determined based on the configuration of the data group for the learning model in addition to the evaluation result. For example, it is determined according to the evaluation result, the unit price based on the types of data included in the data group for the learning model, and the total number of data in the data group for the learning model.

[0051] The overfitting determination unit 41 implements overfitting determination means, temporarily stops the training process in the training processing unit 34, and calculates the degree of overfitting when the training evaluation result indicating the accuracy during training is obtained in the evaluation processing unit 35. The degree of overfitting is a value calculated based on the deviation amount between the training evaluation result and the evaluation result, and the price determination unit 40 reduces the claimed amount according to the degree of overfitting. Overfitting may occur, for example, when the input training data is biased, when training data is input further after sufficient learning, and when a large number of inappropriate training data are included.

[0052] In the determination of overfitting, for example, information on the evaluation value and the number of input evaluation data is obtained from the training evaluation result and the evaluation result respectively, the relationship between the number of input training data with respect to the evaluation value at the time of temporary stop and the time of training processing is compared, and the increase rate of the evaluation value with respect to the increase rate of the number of training data is obtained. The greater the increase rate of the evaluation value with respect to the increase rate of the number of training data deviates from a specific appropriate range, the higher the degree of overfitting, and it can be determined that the learning by inputting evaluation data is not properly performed.

[0053] The quality data creation unit 42 implements evaluation data providing means, processes the regularized evaluation data group to create quality data. Since the evaluation data group has personal information etc. and cannot be transmitted from the learning model processing device 12 to the learning model generation device 11, the part corresponding to personal information is processed into a transmissible state by replacement processing or masking processing. On the other hand, the quality data obtained by the processing can recognize the presence or absence of items related to the target variable information included in the evaluation data group. Thereby, the data user can confirm the quality of the learned model by referring to the quality data.

[0054] Also, the quality data creation unit 42 may include the evaluation result. In that case, it is preferable to add an explanation to the data user regarding the discriminative ability indicated by the AUC value which is the evaluation result. For example, the discriminative ability according to the AUC value is evaluated in stages.

[0055] The reward calculation unit 43 implements the individual reward amount calculation means, and calculates the reward amount to be paid to the provider of each data constituting the learning model data group from the total reward amount based on the price of the learned model. It uses the information of the provider stored in association with the learning model data corresponding to each provider in the learning model data storage unit 30. It performs grouping and ratio calculation to determine the individual reward, which is the reward amount to be paid to each provider. Since there is one piece of data per person for the learning model data, the total number of records of the learning model data constituting the learning model data group is the number of people to whom the reward is paid.

[0056] Grouping groups the providers of the data used for learning according to the target variable information that specifies the type of the learning model data or the type of data that the learning model data has, which is obtained from the learning model generation device 11. When there is one item of target variable information, it divides into a group having the data corresponding to the target variable information and a group not having the corresponding data. Also, when there are multiple items of target variable information, groups are set for each corresponding number.

[0057] In the ratio calculation, in combination with grouping, the reward amount to be paid to the provider of the data that matches the type of data indicated by the target variable information is made higher than the reward amount to be paid to the provider of the data that does not match the type of data. For example, the total reward amount is equally divided by the number of groups to set the reward amount for each group equally, and in each group, the value obtained by equally dividing the reward amount for each group by the number of members is used as the individual reward. As a result, the provider belonging to a group with a small number of people receives a higher individual reward amount than other groups.

[0058] The determined reward amount and the target variable information are presented to the provider. The reward paid to an individual varies depending on the content of the data provided. In grouping, by grouping valuable data etc. that does not correspond to the target variable information for which the user specifies the data desired for learning, the provider of the valuable data can also obtain a high individual reward. Valuable data includes data of providers with diseases, data of providers with rare blood types, etc.

[0059] The data transmission / reception unit 44 realizes the learning model transfer means and transfers the learned learning model image, quality data, and invoice information to the learning model generation device 11. Further, it may acquire the encrypted learning model image, decryption key, and target variable information from the learning model processing device 12. The data transmission / reception unit 44 realizes the learning model image transfer means in the learning model processing device 12. In addition, the target variable information is presented to the provider terminal 13 as the target variable disclosure means together with the calculated reward amount.

[0060] An example of the remote learning system 10 that uses LUKS for the method of encrypting the learning model and performs machine learning on PHR data where the target variable information is "hypertension" will be described.

[0061] The generation of the learning model image is performed at the data utilization site, which is a pharmaceutical company or a medical device manufacturer having the learning model generation device 11, and the machine learning of the PHR data is performed at the PHR data site, which is a medical institution having the learning model processing device 12 and holding the PHR data collected from a plurality of providers. Further, the learning model processing device 12 realizes the function of the confidential calculation unit 31 using the second-generation AMD EPYC processor (manufactured by AMD) equipped with the "Secure Encrypted Virtualization" function.

[0062] The learning model generation device 11 encrypts the partition and below by applying LUKS to the learning model image, which is the VM image generated by the learning model generation unit 20, in the encryption processing unit 21. The learning model image has a root partition where each process of machine learning can be executed in the learning model processing device 12. The decryption key setting unit 22 sets a non-encrypted partition for starting the machine learning of the learning model processing device 12 in the VM image. The learning model is provided with NFS (Network File System) mount points for reading PHR data for the training data group and the evaluation data group.

[0063] As shown in FIG. 5, the learning model image generated by the learning model generation unit 20 includes a decryption key storage point 51, a training data group storage point 52, an evaluation data group storage point 53, a preprocessed training data group storage point 54, a preprocessed evaluation data group storage point 55, a learned model storage point 56, and an evaluation result storage point 57 as folders used for input or output in each process of machine learning executed by the learning model processing device 12. It is preferable to set each folder, which is a storage point, to be read-only or read-write according to its role.

[0064] The decryption key storage point 51 is a folder assigned to “ / boot” and stores the decryption key set by the decryption key setting unit 22. The decryption key storage point 51 is an unencrypted partition, and causes the confidential calculation unit 31 to decrypt the learning model image encrypted with the decryption key and start machine learning. Alternatively, instead of the decryption key, a decryption key acquisition program may be set, and the decryption key may be acquired from a URL (Uniform Resource Locator) for decryption managed by the data user by executing the decryption key acquisition program.

[0065] The training data group storage point 52 is a folder assigned to “ / mnt / data / train” and stores the training data group separated from the data group for the learning model. The training data group is used for preprocessing.

[0066] The evaluation data group storage point 53 is a folder assigned to “ / mnt / data / test” and stores the evaluation data group separated from the data group for the learning model. The evaluation data group is used for preprocessing.

[0067] The preprocessed training data group storage point 54 is a folder assigned to “ / mnt / work / train” and stores the preprocessed training data group, which is the training data group output by preprocessing. The preprocessed training data group is used for training processing.

[0068] The preprocessed evaluation data group storage point 55 is a folder assigned to " / mnt / work / test" and stores the preprocessed evaluation data group, which is the evaluation data group output by the preprocessing. The preprocessed training data group is used for the evaluation process.

[0069] The learned model storage point 56 is a folder assigned to " / result / model" and stores the learned model output by the training process. The learned model is used for the evaluation process and is transferred to the learning model generation device 11 after the evaluation process.

[0070] The evaluation result storage point 57 is a folder assigned to "result / eval" and stores the evaluation results output by the evaluation process. The evaluation results are represented by the AUC value.

[0071] Also, the learning model image is a REST-API (Application Programing Interface) corresponding to REST, which is an interface for executing preprocessing, training processing, and evaluation processing in response to requests in the confidential computing unit 31. The REST-API has a URL path for starting each process in response to an HTTP operation by the confidential computing unit 31. The URL paths are, for example, " / preproc" for preprocessing, " / train" for training processing, and " / eval" for evaluation processing. Each URL path is specified in a GET request, which is an HTTP verb by the confidential computing unit 31, and each process starts in response to the request.

[0072] The encryption processing unit 21 performs encryption processing on the generated learning model image and protects the data of the learning model image until it is decrypted by confidential computing in the confidential computing unit 31 of the learning model processing device 12. The decryption key setting unit 22 sets the decryption key in the non-encrypted state.

[0073] Since the learned model trained for hypertension is the objective, the target variable information determination unit 23 inputs blood pressure information as the target variable information via the user interface. Systolic blood pressure of 130 mmHg or higher and diastolic blood pressure of 80 mmHg or higher are input as the criteria for discriminating hypertension.

[0074] The transmission data having at least the re-encrypted learned model image, the decryption key, and the target variable information is transferred via the data transmission / reception unit 44. The transmission and reception of the learned model image between the learning model generation device 11 and the learning model processing device 12 are implemented as a general web application. The data user transfers the URL to the learning model processing device 12 to perform data transfer while maintaining security.

[0075] When transferring the pre-learning learned model image from the learning model generation device 11 to the learning model processing device 12, for example, the URL set as "https: / / data center host / learning" is transmitted to the learning model processing device 12, and the learning model processing device 12 accesses this URL to obtain the encrypted pre-learning learned model image. Note that the data handled by the data center host is PHR data or the like.

[0076] In the transfer of the learned model from the learning model processing device 12 to the learning model generation device 11, for example, the URL set as "https: / / data center host / learned" is transmitted to the learning model generation device 11, and the data user accesses this URL to obtain the learned model image after learning.

[0077] The confidential computing unit 31 that starts the CVM and obtains the pre-training model image stores the PHR data as an NFS server. The confidential computing unit 31 that has stored a sufficient number of data groups for the learning model separates the data groups for the learning model collected from the provider terminal 13 in the evaluation data extraction unit 32. For example, PHR data for a thousand people is used as a sufficient number of data groups for the learning model. In the separation, to avoid overfitting to specific items in the training data group and the evaluation data group and perform training and evaluation with random PHR data, the data is shuffled. The shuffled PHR data is divided into a training data group and an evaluation data group at a ratio of 7:3. The evaluation data group extracted by the division is stored in the evaluation data group storage point 53.

[0078] After the extraction process, the pre-training model image is decrypted using the decryption key within the CVM whose memory is protected by hardware encryption. By decrypting within the CVM, training processing and evaluation processing on the learning model image can be executed in a non-public state for the operator and administrator of the learning model processing device 12.

[0079] After decryption, the training processing unit 34 makes an HTTP GET request for preprocessing to the learning model image. The learning model image returns an HTTP200, which is a status code indicating "OK" to the training processing unit 34 in response to the GET request, and executes preprocessing on the training data group stored in the training data group storage point 52 and the evaluation data group stored in the evaluation data group storage point 53. The preprocessed training data group and the preprocessed evaluation data group obtained by the preprocessing are stored in the preprocessed training data group storage point 54 and the preprocessed evaluation data group storage point 55. In the regularization in the preprocessing, outliers and the like are excluded.

[0080] After preprocessing, the training processing unit 34 makes an HTTP GET request for training processing to the learning model image. The learning model image returns HTTP200 to the training processing unit 34 in response to the GET request, and executes training processing using the preprocessed training data group stored in the preprocessed training data group storage point 54 to generate a learned model. The learned model is stored in the learned model storage point 56.

[0081] After training processing, the evaluation processing unit 35 makes an HTTP GET request for evaluation processing to the learning model image. The learning model image returns HTTP200 to the evaluation processing unit 35 in response to the GET request, and executes evaluation processing using the preprocessed evaluation data group for the learned model stored in the learned model storage point 56, and outputs an evaluation result. The AUC value obtained as the evaluation result is also output as an HTTP response. Also, the AUC value is stored in the evaluation result storage point 57.

[0082] After evaluation processing, re-encryption processing and decryption key setting are performed on the learning model image storing the learned model, and it is transmitted from the confidential calculation unit 31 to the data transmission / reception unit 44.

[0083] After the completion of each process in the confidential calculation unit 31, the price determination unit 40 determines the price of the learned model image using the total number of PHR data records, which is the total number of data groups for the learning model, and the evaluation result. The determined price is the amount to be billed to the data user. The price is determined based on the discrimination ability of the learned model, the unit price of the PHR data, and the total number of PHR records. For example, the price is obtained by calculating (AUC value - 0.5) × unit price of PHR data × total number of PHR data records.

[0084] After the price is determined, the quality data creation unit 42 creates quality data from the evaluation results and the evaluation data group. For the evaluation data group, processing such as data deletion, anonymization, and pseudonymization is performed so that it can be output from the learning model processing device 12 to the learning model generation device 11. For example, it may be processed by deleting information that does not affect the evaluation, anonymization or pseudonymization such as noise processing of some information, or converting the blood pressure information, which is the target variable information, into an approximate number represented by 5 mmHg each. An explanation regarding the quality level indicated by the AUC value is added to the evaluation result.

[0085] After the quality data is created, the quality data, the claim amount based on the price, and the re-encrypted learning model image are sent to the learning model generation device 11 via the data transmission / reception unit 44.

[0086] After the price is determined, the reward calculation unit 43 calculates the reward amount to be paid for each provider of the learning model data based on the price. The individual reward amount is the total reward amount divided by the number of groups and then equally divided by the number of people in the group to which the individual belongs. Therefore, the calculation of the reward amount involves grouping and ratio calculation. The grouping determines two groups based on whether the PHR data corresponds to hypertension, which is the target variable information, and groups each PHR data. The ratio calculation assigns an equal reward amount to each of the group with hypertension and the group without hypertension, and obtains the value divided by the number of people in each group. For example, if there are 1000 providers of PHR record data and 200 of them have hypertension, the individual reward for those with hypertension is (total reward amount÷2)÷200. Also, the individual reward for those without hypertension is (total reward amount÷2)÷800.

[0087] After the reward is calculated, the calculated individual reward amount and the target variable information are notified to the provider terminal 13 that each provider has via the data transmission / reception unit 44. As a result, each provider can specifically know for what purpose or reason their PHR data has been utilized.

[0088] As shown in FIG. 6, a flowchart showing a series of processes in which the learning model processing device 12 performs machine learning on a learning model transferred from a different device without taking out the learning model data in a non-disclosed manner will be described. The learning model processing device 12 acquires an encrypted pre-learning model image, a decryption key, and target variable information from the learning model generation device 11 (step ST110). In the confidential calculation unit 31 that has acquired the pre-learning model image, the CVM, which is a confidential computer, is activated by hardware encryption processing (step ST120). The confidential calculation unit 31 acquires the decryption key, decrypts the encrypted pre-learning model image, and acquires a pre-learning model image capable of executing preprocessing, training processing, and evaluation processing (step ST130).

[0089] The evaluation data extraction unit 32 acquires the learning model data group stored in the learning model data storage unit 30 in advance, and divides the learning model data group into a training data group and an evaluation data group (step ST140). Using the training data group regularized by preprocessing in the preprocessing unit 33, the training processing unit 34 performs training processing to generate a learned model (step ST150). For the generated learned model, using the evaluation data group regularized by preprocessing in the preprocessing unit 33, the evaluation processing unit 35 performs evaluation processing to acquire the evaluation result of the learned model (step ST160). When the evaluation result evaluated by the AUC value is not sufficient for the discrimination ability of the learned model, such as a value of 0.4 to 0.6, and the number of repeatable times of training is 1 or more (N in step ST170), the learning model data group is acquired again from the learning model data storage unit 30, divided into a training data group and an evaluation data group, and a learned model is generated (step ST140).

[0090] When the discriminative ability of the learned model is sufficient, such as when the evaluation result is greater than 0.6, or when the number of repeatable training times is 0 (Y in step ST170), the learned model is stored in the learning model image (step ST180). After storage, in the price determination unit 40, based on the evaluation result, the unit price of the learning model data, and the number of the learning model data groups, the billing amount of the learned model for billing the data user is calculated (step ST190). In the quality data creation unit 42, editing for the evaluation result and anonymization such as noise immunity for the evaluation data group are performed to create quality data to be sent to the data user (step ST200). The learned model, the billing amount, and the quality data generated in the learning model processing device 12 are sent to the learning model generation device 11 (step ST210).

[0091] According to the above content, in the generation of the learned model, the learning model generation device 11 can perform training without disclosing the learning model image to the learning model processing device 12. Also, the learning model processing device 12 can perform training on the learning model image generated by another device without outputting the learning model data group having personal information.

[0092] Also, in the first embodiment, a learned model is generated by machine learning executed by one learning model processing device 12 for the learning model image acquired from the learning model generation device 11, but distributed learning may be executed in which machine learning for the learning model image is distributed and performed by a plurality of learning model processing devices to generate a learned model.

[0093] A remote learning system that performs distributed learning has a learning model generation device 11 and a plurality of learning model processing devices that perform distributed learning on the pre-learning model image generated and encrypted by the learning model generation device 11. In that case, the learning model data group is distributed and held by a plurality of learning model processing devices as different distributed learning model data groups. Each learning model processing device executes distributed learning using the distributed learning model data group it holds, with the confidential computing means for the learning model image generated and encrypted by the learning model generation device 11. The learning model generation device 11 receives the input of a learning model image having a learned model derived by distributed learning. The second and third embodiments are forms in which a learned model is derived by distributed learning.

[0094] [Second Embodiment] In the remote learning system according to the second embodiment, confidential computing of distributed learning is executed by sequentially transferring one learning model image to a plurality of learning model processing devices and repeating the learning. In each learning model processing device, the in-learning model derived using the distributed learning model data group, which is the respective learning model data group, is transferred to the next site, and the learning for the transferred in-learning model is resumed. Note that descriptions of the same content as in the first embodiment are omitted.

[0095] As shown in FIG. 7, a remote learning system 60 that realizes distributed learning means includes a learning model generation device 11 and, as a plurality of learning model processing devices, a first learning model processing device 61 that executes the first distributed learning to derive an in-learning model, a second learning model processing device 62 that performs distributed learning on the in-learning model, and a third learning model processing device 63 that executes the last distributed learning to derive a learned model. The learning model image storing the in-learning model is encrypted and transferred as an in-learning learning model image.

[0096] The second learning model processing device 62 is included in the remote learning system 60 in a number corresponding to the number of times of distributed learning performed until a learned model is derived. That is, in distributed learning executed in a distributed manner at n locations (n ≥ 2), distributed learning is performed in n - 2 second learning model processing devices 62. In distributed learning using a data group for a learning model distributed at two locations, distributed learning is performed by the first learning model processing device 61 and the third learning model processing device 63.

[0097] The first learning model processing device 61, the second learning model processing device 62, and the third learning model processing device 63 have, in addition to each function of the learning model processing device 12 (see FIG. 4), a function of acquiring input / output information, which is information on the input destination and output destination of the learning model image, and a function of controlling distributed learning. The input / output information is information for recognizing whether it is either the learning model generation device 11 or the learning model processing device.

[0098] The input / output information in the first learning model processing device 61 becomes a learning model processing device whose input source is the learning model generation device 11 and whose transmission destination is the second learning model processing device 62 or the third learning model processing device 63. The first learning model processing device 61 executes a stop process for stopping the training process by the training processing unit 34 for the pre-learning model image in a learning-in-progress state where learning can be resumed, and a transfer process for transferring the learning model image in the learning-in-progress state to another learning model processing device where machine learning has not been performed.

[0099] The input / output information in the second learning model processing device 62 is such that the input source is a learning model processing device that is the first learning model processing device 61 or the second learning model processing device 62 that has already performed distributed learning, and the destination is a learning model processing device that is the second learning model processing device 62 that has not performed distributed learning or the third learning model processing device 63. The second learning model processing device 62 receives the transfer of the learning model image storing the in-learning model, and executes a resumption process for resuming the training process for the in-learning model. Further, it executes a stop process for stopping the training process by the training processing unit 34 in a learning-in-progress state where learning can be resumed, and a transfer process for transferring the learning model image in the learning-in-progress state to another learning model processing device where machine learning has not been performed.

[0100] The input / output information in the third learning model processing device 63 is such that the input source is a learning model processing device that is the first learning model processing device 61 or the second learning model processing device 62, and the destination is the learning model generation device 11. The third learning model processing device 63 receives the transfer of the learning model image storing the in-learning model, and executes a resumption process for resuming the training process for the in-learning model. The training process resumed in the training processing unit is completed to obtain a learned model.

[0101] For calculating the price of the learned model derived in the third learning model processing device 63, since information such as the total number of records and unit price information of the data group for the learning model is used, information necessary for price determination in each distributed learning model data group is obtained from the first learning model processing device 61 and the second learning model processing device 62 that have performed distributed learning. For example, the first learning model processing device 61 and the second learning model processing device output the price calculation information generated by performing anonymization processing on personal information included in the distributed learning model data group and deleting unnecessary information, together with the learning model image having the in-learning model, to the output destination learning model processing device. Further, the generated price calculation information may be directly transmitted to the third learning model processing device 63.

[0102] Also, the reward amount calculated by the third learning model processing device 63 is transmitted to the first learning model processing device 61 and the second learning model processing device 62 that have performed distributed learning based on the acquired price calculation information. As a result, each learning model processing device can notify the provider terminal of the information on the reward amount.

[0103] In the first learning model processing device 61 and another second learning model processing device 62, the training process may be performed using all of the learning model data groups as training data groups without performing the extraction process and the evaluation process. In that case, the evaluation process is executed only by the third learning model processing device 63.

[0104] As an example in the second embodiment, a remote learning system 60 that performs distributed learning with three learning model processing devices will be described. For example, each device included in the remote learning system 60 belongs to a different base, and includes a learning model usage base having a learning model generation device 11, a first training base to which the first learning model processing device 61 that holds the first distributed learning model data group belongs, a second training base to which the second learning model processing device 62 that holds the second distributed learning model data group belongs, and a third training base having a third learning model processing device 63. The transfer of the learning model image is performed, and a learned model is derived.

[0105] The first learning model processing device 61 inputs a first distributed training data group separated from the first distributed learning model data group to the pre-learning model image acquired from the learning model generation device 11, and derives a learning-in-progress model in which learning has been stopped during the learning process. The learning-in-progress model is stored in the learning model image. The learning-in-progress learning model image, which is the learning model image storing the learning-in-progress model, is encrypted and transferred to the second learning model processing device 62 together with the price calculation information.

[0106] The second learning model processing device 62 acquires the learning model image having the learning-in-progress model transferred from the first learning model processing device 61, inputs the second distributed training data group separated from the second distributed learning model data group, and resumes learning.

[0107] In distributed processing, the learning-in-progress model is stored in a learning-in-progress model storage point in the learning model image that is different from the trained model storage point 56. The learning-in-progress model storage point is a folder assigned to " / middle / model" and stores the learning-in-progress model output by the training process. The learning model image that stores the learning-in-progress model that was stopped during training is re-encrypted and transferred in a state where the training process can be resumed in the second learning model processing device 62 or the third learning model processing device 63.

[0108] When the training process is resumed, a training process is executed in which a group of learning model data stored in each learning model processing device is input to the in-progress model stored in the in-progress model storage point. The second learning model processing device 62 again derives the in-progress model that was stopped during learning, and the third learning model processing device 63 derives a trained model.

[0109] This allows machine learning to be performed on a learning model generated by another device, even if the data sets for the learning model are held at multiple locations or multiple learning model processing devices, without outputting the data sets for the learning model held by each location or device. Furthermore, by performing distributed learning, machine learning can be performed using a sufficient amount of data, even if the number of data sets for the learning model held by each learning model processing device is small.

[0110] [Third embodiment] In the remote learning system of the third embodiment, distributed learning is performed by collecting multiple individually derived trained models in one location and integrating them into a single trained model using a stacking method. Each learning model processing device holds an ensemble learning data set, which is a data set for each different distributed learning model, and derives a trained model using the ensemble learning data set. Note that a description of other aspects that are the same as those of the first or second embodiment will be omitted.

[0111] As shown in FIG. 8, a remote learning system 70 that implements distributed learning means includes a learning model generation device 11, a plurality of fourth learning model processing devices 71 that derive a distributed learning completed model by machine learning using a pre-learning model image, and a fifth learning model processing device 72 that integrates each of the distributed learning completed models transferred from the plurality of fourth learning model processing devices 71 into one learning completed model.

[0112] In addition to the same functions as the fourth learning model processing device 71, the fifth learning model processing device 72 has a function of performing integration processing by confidential computing, integrates the learning completed models stored in each distributed learning completed learning model image into one, stores the integrated learning completed model in the learning model image, and transfers the encrypted learning completed learning model image to the learning model generation device 11.

[0113] Also, the fifth learning model processing device 72 may also acquire the pre-learning model image transferred from the learning model generation device 11 and derive a distributed learning completed model. In that case, the fourth learning model processing device 71 performs distributed learning at at least one location.

[0114] As an example of the remote learning system 70 in the third embodiment, distributed learning in which an ensemble learning data group, which is a data group for a distributed learning model, is input to a pre-learning model image acquired from the learning model generation device 11 at four different bases will be described. In each learning model processing device that inputs the ensemble learning data group, a distributed learning completed model is derived, and the distributed learning completed models are integrated in a learning model processing device at one of the bases to derive one learning completed model. Each base is equipped with a learning model processing device. One of the bases that integrates the learning completed models has the fifth learning model processing device 72, and the other bases have the fourth learning model processing device 71.

[0115] For example, the remote learning system 70 transfers a learning model image among a learning model utilization site that generates a learning model, a first ensemble learning site that performs machine learning using a first ensemble learning data group, a second ensemble learning site that performs machine learning using a second ensemble learning data group, a third ensemble learning site that performs machine learning using a third ensemble learning data group, and a learned model integration site that integrates the learned models derived at the first to third ensemble learning sites, and performs distributed learning. Note that for each transfer, the encryption process and the setting of the decryption key are configured to be different each time.

[0116] At the learning model utilization site, a pre-learning model image is generated by the learning model generation device 11, and target variable information, a decryption key, and the encrypted pre-learning model image are transferred to the fourth learning model processing device 71 at the first to third ensemble learning sites.

[0117] At each ensemble learning site, extraction processing, preprocessing, training processing, and evaluation processing are performed on the pre-learning model image acquired by the fourth learning model processing device 71, and a learned model and an evaluation result are derived. The learned model is stored at the learned model storage point 56 in the learning model image, encrypted as a distributed learning-completed learning model image, and transferred to the learned model integration site.

[0118] At the learned model integration site, the integration process of the generated learned model and the learned models respectively extracted from the transferred distributed learning-completed learning model images is performed in the confidential calculation unit of the fifth learning model processing device 72. Evaluation processing is also performed on the learned model derived by the integration process, the learned model for which the evaluation result is obtained is stored in the learning model image, and transferred to the learning model generation device 11.

[0119] As described above, by generating a two-stage learned model, a highly accurate learned model can be derived. In addition, by using data sufficient for generating a learned model collected by each site from its respective provider, a highly versatile learned model can be derived.

[0120] In the above embodiment, in the machine learning at each ensemble learning site, in the extraction process, the data group for the ensemble learning model is separated into an ensemble training data group and an ensemble evaluation data group, and after performing preprocessing in the same manner as in the first embodiment, training processing and evaluation processing are performed. On the other hand, without performing the extraction process and the evaluation process in the same manner as in the second embodiment, each data group for the ensemble learning model may be used as an ensemble training data group for the training process.

[0121] In each embodiment, the hardware structure of the processing unit that executes various processes such as the central control unit, input reception unit, output control unit, learning model generation unit 20, encryption processing unit 21, decryption key setting unit 22, target variable information determination unit 23, data transmission / reception unit 24 in the learning model generation device, and the central control unit, input reception unit, output control unit, confidential calculation unit 31, price determination unit 40, quality data creation unit 42, reward calculation unit 43, data transmission / reception unit 44 in the learning model processing device is various processors as shown below. The various processors include a CPU (Central Processing Unit), which is a general-purpose processor that executes software (program) and functions as various processing units, a programmable logic device (PLD), which is a processor such as an FPGA (Field Programmable Gate Array) whose circuit configuration can be changed after manufacture, and an application-specific electric circuit, which is a processor having a circuit configuration specifically designed to execute various processes.

[0122] One processing unit may be composed of one of these various processors, or may be composed of a combination of two or more processors of the same or different types (for example, a plurality of FPGAs, or a combination of a CPU and an FPGA). Also, a plurality of processing units may be composed of one processor. As an example of configuring a plurality of processing units with one processor, firstly, as represented by a computer such as a client or a server, one processor is configured by a combination of one or more CPUs and software, and this processor functions as a plurality of processing units. Secondly, as represented by a System On Chip (SoC), etc., there is a form in which a processor that realizes the functions of the entire system including a plurality of processing units with one IC (Integrated Circuit) chip is used.

[0123] Thus, as a hardware structure, the various processing units are configured using one or more of the above various processors. Furthermore, the hardware structure of these various processors is, more specifically, an electrical circuit (Circuitry) in the form of a combination of circuit elements such as semiconductor elements.

[0124] The storage memory stores a data group for a learning model collected from a number of providers. It is preferable to use a built-in SSD (Solid State Drive) for storage. Instead of the SSD, a recording medium such as a USB (Universal Serial Bus) memory or an HDD (Hard Disc Drive) may be used.

[0125] Also, from the above description, the learning model processing device described in the following Supplementary Notes 1 to 10 can be understood.

[0126] [Supplementary Note 1] Equipped with a processor, The processor receives the transfer of the encrypted learning model from the learning model generation device, acquires the data group for the learning model, Derive a trained model that has been encrypted by training the data group for the learning model using confidential computing, A learning model processing device that transfers the trained model to the learning model generation device. [Appendix 2] The processor Separate the data group for the learning model into a training data group and an evaluation data group, Input the evaluation data group into the trained model to obtain an evaluation result, The learning model processing device according to Appendix 1, which transfers quality data based on the evaluation result to the learning model generation device. [Appendix 3] The processor The learning model processing device according to Appendix 2, which performs anonymization processing on the evaluation result to generate the quality data. [Appendix 4] The processor Determine the price of the trained model to be paid by the user from the evaluation result, The learning model processing device according to Appendix 2 or 3, which transfers a claim amount based on the price to the learning model generation device. [Appendix 5] The processor The learning model processing device according to Appendix 4, which determines the price according to the type of data included in the data group for the learning model. [Appendix 6] The processor In the derivation of the trained model, obtain an evaluation result during training that represents the accuracy during training, Calculate the degree of overfitting from the deviation amount between the evaluation result during training and the evaluation result, The learning model processing device according to Appendix 4 or 5, which reduces the claim amount according to the degree of overfitting. [Appendix 7] The processor Store the provider of the data constituting the data group for the learning model, Receive, from the learning model generation device, target variable information that specifies the conditions of the data constituting the data group for the learning model together with the learning model, A learning model processing apparatus according to any one of Appendices 4 to 6 that determines the amount of remuneration to be paid for each provider based on the price, the target variable information, and the total number of records in the data group for the learning model. [Appendix 8] The processor groups the providers according to the type of data in the target variable information, and a learning model processing apparatus according to Appendix 7 that sets the amount of remuneration to be paid to the provider of data that matches the type to be higher than the amount of remuneration to be paid to the provider of data that does not match the type. [Appendix 9] The processor is a learning model processing apparatus according to Appendix 7 or 8 that presents the amount of remuneration and the target variable information to the provider. [Appendix 10] The data group for the learning model is a data group related to health, and the learned model is a learned model obtained by learning data related to health. A learning model processing apparatus according to any one of Appendices 1 to 9.

Explanation of Signs

[0127] 10 Remote learning system 11 Learning model generation device 12 Learning model processing device 13 Provider terminal 20 Learning model generation unit 21 Encryption processing unit 22 Decryption key setting unit 23 Target variable information determination unit 24 Data transmission / reception unit 30 Data storage unit for learning model 31 Confidential calculation unit 32 Evaluation data extraction unit 33 Pretreatment unit 34 Training processing unit 35 Evaluation processing unit 40 Price determination unit 41 Overfitting determination unit 42 Quality data creation unit 43 Reward calculation unit 44 Data transmission / reception unit 51 Decryption key storage point 52 Training data storage point 53 Evaluation data storage point 54 Preprocessed training data storage point 55 Preprocessed evaluation data storage point 56 Trained model storage point 57 Evaluation result storage point 60 Remote learning system 61 First learning model processing device 62 Second learning model processing device 63 Third learning model processing device 70 Remote learning system 71 Fourth learning model processing device 72 Fifth learning model processing device ST110~ST210 Steps

Claims

1. A learning model processing apparatus comprising a processor, wherein the processor receives the transfer of an encrypted learning model from a learning model generation apparatus, obtains a data group for the learning model, performs confidential computing to train the data group for the learning model on the learning model to derive an encrypted learned model, and transfers the learned model to the learning model generation apparatus.

2. The processor separates the data group for the learning model into a training data group and an evaluation data group, inputs the evaluation data group into the learned model to obtain an evaluation result, and transfers quality data based on the evaluation result to the learning model generation apparatus. The learning model processing apparatus according to claim 1.

3. The processor performs anonymization processing on the evaluation result to generate the quality data. The learning model processing apparatus according to claim 2.

4. The processor determines the price of the learned model to be paid by the user from the evaluation result, and transfers a claim amount based on the price to the learning model generation apparatus. The learning model processing apparatus according to claim 2.

5. The processor determines the price according to the type of data included in the data group for the learning model. The learning model processing apparatus according to claim 4.

6. The processor obtains an evaluation result during training representing the accuracy during training in the derivation of the learned model, calculates the degree of overfitting from the deviation amount between the evaluation result during training and the evaluation result, and reduces the claim amount according to the degree of overfitting. The learning model processing apparatus according to claim 4.

7. The processor stores the providers of the data constituting the data group for the learning model, receives, from the learning model generation apparatus, target variable information specifying the conditions of the data constituting the data group for the learning model together with the learning model, and determines the amount of remuneration to be paid for each provider based on the price, the target variable information, and the total number of records in the data group for the learning model. The learning model processing apparatus according to claim 4.

8. The processor groups the providers according to the type of data in the target variable information, and the amount of remuneration paid to the providers of the data matching the type is higher than the amount of remuneration paid to the providers of the data not matching the type. The learning model processing apparatus according to claim 7.

9. The processor The learning model processing apparatus according to claim 7, which presents the reward amount and the target variable information to the provider.

10. The data group for the learning model is a data group related to health, The learning model processing apparatus according to any one of claims 1 to 9, wherein the learned model is a learned model obtained by learning data related to health.

11. The data group for the learning model used for learning the learning model is distributed and held by a plurality of learning model processing apparatuses as mutually different data groups for distributed learning, The learning model processing apparatus performs distributed learning on the data group for the distributed learning model held by each of them using a confidential computing means on the learning model generated and encrypted by the learning model generation apparatus, The learning model generation apparatus is a remote learning system that receives an input of the learned model derived by the distributed learning.

12. As the plurality of learning model processing apparatuses that perform the distributed learning, there are a first learning model processing apparatus that holds a first data group for distributed learning model which is the data group for the distributed learning model, and a second learning model processing apparatus that holds a second data group for distributed learning model which is the data group for the distributed learning model, The first learning model processing apparatus inputs the first data group for distributed learning model to the learning model transferred from the learning model generation apparatus, and derives a learning model in progress in which learning has been stopped in the middle of learning, The second learning model processing apparatus acquires the learning model in progress transferred from the first learning model processing apparatus, inputs the second data group for distributed learning model, and resumes learning. The remote learning system according to claim 11.

13. The plurality of learning model processing apparatuses that perform the distributed learning input the data group for the distributed learning model to the learning model acquired from the learning model generation apparatus, and respectively derive a distributed learning completed model, Any one of the learning model processing apparatuses integrates the plurality of distributed learning completed models and derives the learned model. The remote learning system according to claim 11.

14. Learning model transfer means for receiving the transfer of the encrypted learning model from the learning model generation apparatus and transferring the derived learned model to the learning model generation apparatus, Data acquisition means for acquiring the data group for the learning model, A learning model processing program that causes a computer to function as a learning model training means for training the learning model with the data group for the learning model by confidential calculation to derive a learned model.

Images