Learning system, learning method, and program

Abstract

Provided are a learning system and a learning method that can improve the accuracy of a machine learning model while preventing the leakage of information regarding datasets owned by organizations. This learning system (10) comprises: a first learning unit (11) that generates a local model by training a global model generated by federated learning on first data included in a dataset; a second learning unit (12) that trains a first machine learning model by machine learning using second data, which differs from the first data among the data included in the dataset; a integration unit (13) that integrates the local model or the global model with the first machine learning model; and a generation unit (14) that uses the local model to generate a new global model.

Description

[Technical Field]

[0001] The present disclosure relates to a learning system, a learning method, and a computer-readable medium. [Background technology]

[0002] Patent Document 1 discloses an information processing system that uses federated learning. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2021 / 205959 Summary of the Invention [Problem to be solved by the invention]

[0004] Federated learning technology has been proposed, in which local models are trained using datasets owned by each organization and a global model that integrates the local models is distributed. Federated learning technology can conceal the datasets used for training. However, if there is an organization that records the time series changes in the global model, there is a risk that the learning data used to train the most recent local model can be inferred by reverse engineering the global model.

[0005] Therefore, one of the objectives that the embodiments disclosed in this specification aim to achieve is to provide a learning system, a learning method, and a program that improve the accuracy of machine learning models while preventing the leakage of data owned by each organization. [Means for solving the problem]

[0006] A learning system according to a first aspect of the present disclosure includes: a first learning means for learning a first machine learning model by machine learning using first data that is highly confidential among data included in the dataset; a second learning means for generating a local model by having a global model generated by federated learning learn second data, which is different from the first data, among data included in the dataset; and an integration means for integrating the local model or the global model with the first machine learning model; a generating means for generating a new global model using the local model; It is equipped with:

[0007] In the learning method according to the second aspect of the present disclosure, training a first machine learning model through machine learning using first data that is highly confidential among data included in the dataset; generating a local model by training a global model generated by federated learning on second data, which is different from the first data, among data included in the dataset; Integrating the local model or the global model with the first machine learning model; The local model is used to generate a new global model.

[0008] A non-transitory computer-readable medium according to a third aspect of the present disclosure, comprising: On the computer, a process of training a first machine learning model by machine learning using first data that is highly confidential among data included in the dataset; generating a local model by training a global model generated by federated learning on second data, which is different from the first data, among data included in the dataset; integrating the local model or the global model with the first machine learning model; generating a new global model using the local model; The program for executing the above is stored. [Effects of the Invention]

[0009] According to the present disclosure, it is possible to provide a learning system, a learning method, and a program that improve the accuracy of machine learning models while preventing the leakage of highly confidential data. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a block diagram showing the configuration of a related learning system. [Figure 2] 1 is a block diagram showing the configuration of a learning system according to a first embodiment. [Figure 3] FIG. 10 is a block diagram showing the configuration of a learning system according to a second embodiment. [Figure 4] FIG. 10 is a block diagram showing the configuration of a client terminal according to the second embodiment. [Figure 5] 10 is a flowchart showing the flow of operations of a classification unit. [Figure 6] FIG. 10 is a block diagram showing the configuration of a learning system according to a third embodiment. [Figure 7] FIG. 10 is a block diagram showing the configuration of a learning system according to a fourth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] <Background to the embodiment> First, an overview of federated learning will be described. First, a related learning system 1 includes a client terminal 2x, a client terminal 2y, a client terminal 2z, and a server 3.

[0012] The client terminal 2x generates a machine learning model (called a local model 4x) from a dataset owned by the organization X. The client terminal 2x transmits the local model 4x to the server 3.

[0013] The client terminal 2y generates a machine learning model (called a local model 4y) from a dataset owned by organization Y. The client terminal 2y transmits the local model 4y to the server 3.

[0014] The client terminal 2z generates a machine learning model (called a local model 4z) from a dataset owned by the organization Z. The client terminal 2z transmits the local model 4z to the server 3.

[0015] The server 3 generates a global model by integrating the local model 4x, the local model 4y, and the local model 4z. The server 3 may generate the global model by, for example, taking the arithmetic mean of the model parameters. Note that the method of integrating the model parameters is not limited to the arithmetic mean. The server 3 transmits the global model to the client terminals 2x, 2y, and 2z.

[0016] Here, a dataset owned by an organization may contain data (e.g., data on a compound being developed) that needs to be kept confidential from other organizations. For example, if an organization starts developing a compound that exhibits a specific effect, it may want to keep this a secret. However, since the dataset owned by the organization contains a large amount of data on compounds that exhibit a specific effect, there is a risk that the organization's start of development of such a compound may be inferred by reverse engineering the global model. The inventor of the present application conceived the invention of embodiment 1 based on the above considerations.

[0017] <Embodiment 1> 2 is a block diagram showing the configuration of a learning system 10 according to embodiment 1. The learning system 10 includes a first learning unit 11, a second learning unit 12, an integration unit 13, and a generation unit 14.

[0018] The first learning unit 11 learns a first machine learning model through machine learning using first data that is highly confidential among data included in the dataset.

[0019] The second learning unit 12 generates a local model by having the global model generated by associative learning learn second data, which is different from the first data, among the data included in the dataset.

[0020] The integration unit 13 integrates the first machine learning model with the local model or the global model, and the integrated model is called a second machine learning model.

[0021] The generator 14 generates a new global model using the local model.

[0022] The learning system 10 according to the first embodiment can generate a highly accurate second machine learning model while preventing leakage of highly confidential data.

[0023] The learning system 10 includes a processor, memory, and storage device (not shown). The storage device stores a computer program that implements the processing of the learning method according to this embodiment. The processor then loads the computer program from the storage device into the memory and executes the computer program. This allows the processor to realize the functions of a first learning unit 11, a second learning unit 12, an integration unit 13, and a generation unit 14.

[0024] Alternatively, the first learning unit 11, the second learning unit 12, the integration unit 13, and the generation unit 14 may each be realized by dedicated hardware. Furthermore, some or all of the components of each device may be realized by general-purpose or dedicated circuits, processors, etc., or a combination of these. These may be configured by a single chip, or by multiple chips connected via a bus. Some or all of the components of each device may be realized by a combination of the above-mentioned circuits, etc., and a program. Furthermore, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field-Programmable Gate Array), etc., may be used as the processor.

[0025] Furthermore, when some or all of the components of learning system 10 are realized by multiple information processing devices, circuits, etc., the multiple information processing devices, circuits, etc. may be centrally or decentralized. For example, the information processing devices, circuits, etc. may be realized as a client-server system, cloud computing system, etc., connected via a communication network. Furthermore, the functions of learning system 10 may be provided in the form of SaaS (Software as a Service).

[0026] <Embodiment 2> Figure 3 is a block diagram showing the configuration of a learning system 100 according to embodiment 2. Learning system 100 is a specific example of learning system 10 according to embodiment 1. Learning system 100 includes client terminals 20x, 20y, and 20z, and a server 30. Each client terminal is a terminal of an organization (e.g., a pharmaceutical manufacturer or a chemical manufacturer) that uses learning system 1.

[0027] The client terminals 20x, 20y, and 20z are communicably connected to the server 30 via a network N. The network N may be wired or wireless. The network N may be, for example, a VPN (Virtual Private Network).

[0028] Hereinafter, when there is no need to distinguish between the client terminals 20x, 20y, and 20z, they may be simply referred to as client terminals 20. Note that the number of client terminals 20 is not limited to three, and may be two, or four or more.

[0029] Next, the client terminal 20 will be described with reference to Fig. 4. The client terminal 20 includes a storage unit 21, a classification unit 22, a first learning unit 23, a second learning unit 24, an integration unit 25, and a setting unit 26. The first learning unit 23 is a specific example of the first learning unit 11, the second learning unit 24 is a specific example of the second learning unit 12, and the integration unit 25 is a specific example of the integration unit 13.

[0030] The memory unit 21 is a storage that stores a dataset 211, a global model 212, a local model 213, a first machine learning model 214, and a second machine learning model 215 owned by each organization.

[0031] The dataset 211 includes multiple records. A record is also called data. The dataset 211 is, for example, a dataset of chemical compounds. In this case, the dataset includes multiple data (records), and each data item lists values ​​such as the structure and properties of the chemical compound. The structure of the chemical compound is expressed by a fixed-length bit string, and each bit in the bit string indicates the presence or absence of a specific structure (e.g., a benzene ring). The property value (e.g., tensile strength value) may be a value obtained by experiment, or may be a value obtained by simulation or theoretical calculation. The properties include, for example, strength, elastic modulus, transition temperature, optical properties, mechanical properties, and thermal properties. The data may include the name and composition of the chemical compound in addition to or instead of the chemical compound structure.

[0032] The dataset 211 includes first data 2111 and second data 2112. The first data 2111 is used to train the first machine learning model 214. The second data 2112 is used to train the local model 213. The first data 2111 is not used to train the local model 213. The first data 2111 and the second data 2112 are classified by the classification unit 22.

[0033] The first data 2111 and the second data 2112 may be identified using a flag. Since there is a risk that the learning data may be guessed in associative learning, associative learning is performed using only the second data 2112, which is a part of the data set. The first data 2111 is more confidential data than the second data 2112 (e.g., data on a compound under development).

[0034] The global model 212 is a global model that integrates local models learned by a plurality of client terminals 20.

[0035] The local model 213 is a machine learning model that is learned by a second learning unit 24, which will be described later. The local model 213 is a machine learning model that is obtained by having the global model 212 learn second data 2112. The local model 213 is used by the server 30 to generate a new global model.

[0036] The first machine learning model 214 is a machine learning model that is trained by the first training unit 23, which will be described later.

[0037] The second machine learning model 215 is a machine learning model generated by the integration unit 25, which will be described later.

[0038] The global model 212, the local model 213, the first machine learning model 214, and the second machine learning model 215 may be, for example, models that infer properties from the structure of a compound.

[0039] The classification unit 22 classifies each piece of data included in the data set 211 into first data 2111 or second data 2112. Note that the data set 211 may include data that is not classified into either the first data 2111 or the second data 2112.

[0040] 5 is a flowchart showing an example of the operation of the classification unit 22. The processes of steps S101 to S105 are performed for each piece of data included in the data set 211.

[0041] First, the classification unit 22 determines whether the data is to be used for associative learning (step S101). If the confidentiality of the data is low, the classification unit 22 may determine that the data is to be used for associative learning. If the data is to be used for associative learning (YES in step S101), the data is classified as second data 2112 (step S102).

[0042] If the data will not be used in the federated learning (NO in step S101), the classification unit 22 determines whether to use the data for training the first machine learning model 214 (step S102). The first machine learning model 214 is a machine learning model trained using data not used in the federated learning. If the reliability of the data is high, the classification unit 22 may determine to use the data for training the first machine learning model 214. Furthermore, if the data has already been used in training the first machine learning model 214, the classification unit 22 may determine not to use the data for training the first machine learning model 214.

[0043] If the data is to be used for training the first machine learning model 214 (YES in step S103), the data is classified as first data 2111 (step S104). If the data is not to be used for training the first machine learning model 214 (NO in step S102), the classification unit 22 does not classify the data as either the first data 2111 or the second data 2112 (step S105).

[0044] A period during which data included in the data set 211 is not used for learning the local model 213 may be set for each data. In this case, in step S101, the classification unit 22 classifies data for which the period has elapsed as second data 2112. The period may be set longer for data with higher confidentiality.

[0045] The first learning unit 23 generates a first machine learning model 214 by machine learning using the first data 2111. The first learning unit 23 is a specific example of the above-described first learning unit 11. The first learning unit 23 may generate the first machine learning model 214 by machine learning using both the first data 2111 and the second data 2112. Specifically, the first machine learning model 214 is a machine learning model trained only on a dataset 211 owned by a single organization.

[0046] The first learning unit 23 may perform machine learning after converting the format of the first data into a data format (referred to as a predetermined format) used for learning the local model 213. The dataset 211 is typically data collected by each client terminal 20 during research and development of compounds, etc., and may be saved in a format different from the predetermined format.

[0047] The second learning unit 24 generates the local model 213 by having the global model 212 learn the second data 2112. The second learning unit 24 is a specific example of the above-mentioned second learning unit 12. The local model 213 is transmitted to the server 30 and used to generate a new global model. Note that the second learning unit 24 may convert the data format of the second data 2112 into a predetermined format and have the global model 212 learn the second data 2112 converted into the predetermined format.

[0048] A weight may be set for each second data 2112 when training the local model 213. In this case, the second training unit 24 trains the local model 213 based on the set weight. For example, a higher weight may be set for the second data 2112 with higher reliability. Note that the model parameters of the machine learning model may also be referred to as weights, but care must be taken not to confuse them with the weights of the training data.

[0049] The integrating unit 25 integrates the global model 212 or the local model 213 with the first machine learning model 214 to generate the second machine learning model 215. The integrating unit 25 is a specific example of the integrating unit 13 described above. The integrating unit 25 may generate the second machine learning model 215, for example, by taking the arithmetic mean of the model parameters of the global model 212 or the local model 213 and the model parameters of the first machine learning model 214. The model parameters of the global model 212 or the local model 213 are referred to as first model parameters. The model parameters of the first machine learning model 214 are referred to as second model parameters. Note that the model integration method is not limited to the arithmetic mean. The integrating unit 25 may also calculate a weighted average of the first model parameters and the second model parameters based on integration weights described below.

[0050] The second machine learning model 215 is a machine learning model that integrates the global model 212 and the first machine learning model 214 that has been trained using second data 2112 that has not been used in the federated learning. Therefore, the second machine learning model 215 has higher accuracy than the global model 212. The second machine learning model 215 may be used, for example, by users belonging to each organization to infer the structure and properties of compounds.

[0051] The setting unit 26 sets an integration weight, which is a weight of the first machine learning model 214 when integrating the global model 212 or the local model 213 with the first machine learning model 214. The setting unit 26 may set the integration weight in accordance with an input to the client terminal 20.

[0052] Specifically, the higher the reliability of the first machine learning model 214 compared to the reliability of the global model 212, etc., the higher the integrated weight is set. For example, the greater the amount of first data 2111 compared to the amount of second data 2112, the higher the integrated weight may be set. Also, the smaller the number of client terminals 20 participating in the federated learning, the higher the integrated weight may be set. This is because when the number of client terminals 20 is small, the accuracy of the global model 212 is low.

[0053] Next, the server 30 will be described with reference to Fig. 3. The server 30 includes a generating unit 31. The generating unit 31 is a specific example of the generating unit 14 described above.

[0054] The generation unit 31 integrates the local model 213 trained in the client terminal 20x, the local model 213 trained in the client terminal 20y, and the local model 213 trained in the client terminal 20z to generate a new global model. The generation unit 31 distributes the new integrated global model to the client terminal 20x, the client terminal 20y, and the client terminal 20z. Information about the first data 2111 is not leaked from the new global model.

[0055] When the server 30 manages the data sets of each organization, the first learning unit 23, the second learning unit 24, and the integrating unit 25 may be provided in the server 30.

[0056] The learning system according to the second embodiment can generate a highly accurate second machine learning model while preventing the first data from being leaked.

[0057] <Embodiment 3> Fig. 6 is a block diagram showing the configuration of a learning system 101 according to the third embodiment. Learning system 101 is a modified example of the learning system 100 described above. Comparing Fig. 3 with Fig. 6, client terminals 20x, 20y, and 20z have been replaced with client terminals 200x, 200y, and 200z, and information terminals 5x, 5y, and 5z have been added. Elements whose functions overlap with those of the second embodiment are given the same reference numerals, and their description will be omitted.

[0058] The client terminals 200x, 200y, and 200z are connected to an external network N of each organization. When the client terminals 200x, 200y, and 200z are not to be distinguished from one another, they are simply referred to as client terminals 200.

[0059] Information terminal 5x is a terminal that manages a data set owned by organization X, information terminal 5y is a terminal that manages a data set owned by organization Y, and information terminal 5z is a terminal that manages a data set owned by organization Z. When there is no need to distinguish between information terminals 5x, 5y, and 5z, they may be simply referred to as information terminals 5.

[0060] Data acquired by the research and development department of each organization is newly registered in the information terminal 5. Unlike the client terminal 20, the information terminal 5 is not connected to an external network N. This makes it possible to prevent the data managed by the information terminal 5 from being leaked.

[0061] 4, the information terminal 5 includes a storage unit 21, a classification unit 22, a first learning unit 23, a second learning unit 24, an integration unit 25, and a setting unit 26. Note that the client terminal 200 does not include these functions.

[0062] Because information terminal 5 is not connected to an external network N, global model 212 is transferred from client terminal 200 to information terminal 5 using a storage medium such as a USB (Universal Serial Bus) memory. Because information terminal 5 is not connected to network N, it is possible to prevent first machine learning model 214 and second machine learning model 215 from being leaked. Therefore, it is possible to prevent first data 2111 from being deduced by reverse engineering.

[0063] 6, the client terminal 200 transmits the local model 213 generated by the information terminal 5 to the server 30. The client terminal 200 also receives the global model 212 generated by the server 3.

[0064] The learning system according to the third embodiment also achieves the same effects as those of the second embodiment. Since the information terminal 5 is isolated from the external network, the risk of the first data being leaked can be further reduced.

[0065] <Embodiment 4> FIG. 7 is a block diagram showing the configuration of a learning system 102 according to the fourth embodiment. The learning system 102 is a modified example of the learning system 100. Comparing FIG. 3 with FIG. 7, the server 30 is replaced with a server group 300. The server group 300 includes multiple servers 32. The number of servers 32 is not limited to three. However, in order to perform secure computation, it is preferable that the number of servers 32 be three or more.

[0066] The server group 300 securely integrates the global model 212 and transmits the results of the secure computation to the client terminals 20x, 20y, and 20z.

[0067] As in the second embodiment, each client terminal 20 learns a local model 213 by machine learning using the second data 2112. Then, each client terminal 20 divides each parameter of the local model 213 into multiple (e.g., three) shares and transmits the multiple shares to multiple servers 32.

[0068] Each server 32 uses the received shares to perform secure computation to calculate the global model 212. Each server 32 may generate the global model 212 at a predetermined time. The local model is not revealed from the shares, and computation using the shares can be considered secure computation. Multiple servers 32 may cooperate to perform multi-party computation (MPC). Since the amount of computation required to integrate the local models 213 is sufficiently small, it is believed that the server group 300 can perform secure computation in a reasonable amount of time.

[0069] The fourth embodiment also provides the same effects as the second embodiment. Furthermore, according to the fourth embodiment, the calculation for integrating the global model can be kept confidential.

[0070] The above-described program includes a set of instructions (or software code) that, when loaded into a computer, causes the computer to perform one or more functions described in the embodiments. The program may be stored in a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, CD-ROM, digital versatile disc (DVD), Blu-ray (registered trademark) disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.

[0071] Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the invention. [Explanation of symbols]

[0072] 1, 10, 100, 101, 102 Learning Systems 2, 2x, 2y, 2z, 20, 20x, 20y, 20z, 200, 200x, 200y, 200z Client terminal 21 Memory section 211 datasets 2111 First Data 2112 Second Data 212 Global Model 4, 4x, 4y, 4z, 213 local models 214 First Machine Learning Model 215 Second Machine Learning Model 22 Classification Department 11, 23 First Study Section 12, 24 Second Study Section 13, 25 Integrated Department 26 Setting section 3, 30, 32 servers 14, 31 Generation part 300 servers 5, 5x, 5y, 5z information terminal N Network

Claims

1. a first learning means for learning a first machine learning model by machine learning using first data that is highly confidential among data included in the dataset; a second learning means for generating a local model by having a global model generated by federated learning learn second data, which is different from the first data, among data included in the dataset; and an integration means for integrating the local model or the global model with the first machine learning model; a generating means for generating a new global model using the local model; A learning system that includes:

2. Equipped with an information terminal that is not connected to an external network, The integration means is provided in the information terminal. The learning system of claim 1 .

3. a first setting means for setting an integration weight, which is a weight of the first machine learning model when integrating the local model or the global model with the first machine learning model; The learning system according to claim 1 or 2, comprising:

4. The first data and the second data are identified using a flag. The learning system according to claim 1 or 2.

5. A period in which data included in the data set is not used for training the local model is set, Classifying means for classifying data for which the period has elapsed into the second data The learning system according to claim 1 or 2, comprising:

6. A weight for each second data when training the local model is set. The learning system according to claim 1 or 2.

7. The first learning means The format of the first data is converted into a format of data used for training the local model, and then the machine learning is performed. The learning system according to claim 1 or 2.

8. training a first machine learning model by machine learning using first data that is highly confidential among data included in the dataset; generating a local model by training a global model generated by federated learning on second data, which is different from the first data, among data included in the dataset; Integrating the local model or the global model with the first machine learning model; Generate a new global model using the local model How to learn.

9. On the computer, a process of training a first machine learning model by machine learning using first data that is highly confidential among data included in the dataset; generating a local model by training a global model generated by federated learning on second data, which is different from the first data, among data included in the dataset; Integrating the local model or the global model with the first machine learning model; generating a new global model using the local model; A program to execute.

Images