Learning device, learning system, learning method, and program

Abstract

Provided are a learning device, a learning system, a learning method, and a program with which a global model can be constructed when networks of multiple organizations are not always connected. A learning device (1) comprises a communication establishment unit (11) that establishes secure communication with information terminals (2) placed in networks (N) of organizations, an acquisition unit (12) that acquires a local model trained on a data set for each organization from the information terminals (2) using secure communication, and an integration unit (13) that integrates the multiple acquired local models.

Description

Learning device, learning system, learning method, and computer-readable medium

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

[0002] Patent Literature 1 discloses a technique for performing machine learning to build an AI (Artificial Intelligence) model (also called a local model) personalized for a user.

[0003] Special Publication No. 2020-531999

[0004] It is known that integrating multiple local AI models can build a better AI model (also called a global model). The server collects user data, allowing the server to build the local model and the global model.

[0005] When the user is an organization, it is desirable to build a network that connects multiple organizations because it is necessary to collect data owned by each organization. However, there is a problem in that it is difficult to build a network that connects multiple organizations with different security concepts.

[0006] Therefore, one of the objectives that the embodiments disclosed in this specification aim to achieve is to provide a learning device, a learning system, a learning method, and a computer-readable medium that can build a global model when the networks of multiple organizations are not always connected to each other.

[0007] A learning device according to a first aspect of the present disclosure includes a communication establishment means for establishing secure communication with an information terminal located on a network of each organization; an acquisition means for acquiring, from the information terminal using the secure communication, a local model trained on a data set for each organization; and an integration means for integrating the acquired local models.

[0008] A computing system according to a second aspect of the present disclosure is a learning system comprising: an information terminal disposed on a network of each organization; and a learning device, wherein the learning device establishes secure communication with the information terminal, uses the secure communication to obtain from the information terminal a local model trained on a data set for each organization, and integrates the obtained local models.

[0009] In a calculation method according to a third aspect of the present disclosure, a computer establishes secure communication with an information terminal located on a network of each organization, acquires a local model trained on a data set for each organization from the information terminal using the secure communication, and integrates the acquired multiple local models.

[0010] A non-transitory computer-readable medium according to a fourth aspect of the present disclosure stores a program for causing a computer to execute the following processes: establishing secure communication with an information terminal located on each organization's network; acquiring, using the secure communication, from the information terminal, a local model trained on a data set for each organization; and integrating the acquired multiple local models.

[0011] According to the present disclosure, it is possible to provide a learning device, a learning system, a learning method, and a computer-readable medium that are capable of building a global model when the networks of multiple organizations are not always connected to each other.

[0012] Fig. 1 is a block diagram showing the configuration of a learning device according to embodiment 1. Fig. 2 is a block diagram showing the configuration of a learning system according to embodiment 2. Fig. 3 is a block diagram showing the configuration of a learning device according to embodiment 2. Fig. 4 is a flowchart showing the flow of an operation for generating a local model. Fig. 5 is a block diagram showing the configuration of a learning system according to embodiment 2.

[0013] First Embodiment Fig. 1 is a block diagram showing the configuration of a learning device 1 according to a first embodiment. The learning device 1 includes a communication establishment unit 11, an acquisition unit 12, and an integration unit 13. The learning device 1 is connected to a public network (not shown). The public network (not shown) is connected to the network of each organization. An information terminal (not shown) is disposed in the network of each organization. The information terminal constructs a local model that has learned a dataset for each organization. The information terminal may also be a repository in which datasets owned by each organization are stored.

[0014] The communication establishment unit 11 establishes secure communication with an information terminal located on the network of each organization. The communication establishment unit 11 may establish secure communication at a predetermined timing.

[0015] For example, the communication establishment unit 11 connects the learning device 1 to each organization's network via a VPN (Virtual Private Network). In this case, confidentiality of communication between the learning device 1 and the information terminal is maintained by encryption and encapsulation. In other words, secure communication is established between the learning device 1 and the information terminal.

[0016] The communication establishment unit 11 may establish secure communication using a technology other than VPN. The communication establishment unit 11 may control communication using a protocol including encryption (e.g., SSL / TLS, SSH (Secure Shell), FTPS (File Transfer Protocol over SSL / TLS)).

[0017] The acquisition unit 12 acquires, via secure communication, a local model trained on a data set for each organization from an information terminal.

[0018] The integration unit 13 integrates the acquired local models.

[0019] The learning device 1 includes a processor, a memory, and a 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 communication establishment unit 11, an acquisition unit 12, and an integration unit 13.

[0020] Alternatively, the communication establishment unit 11, the acquisition unit 12, and the integration unit 13 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 programs. Furthermore, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an FPGA (Field-Programmable Gate Array), etc. may be used as the processor.

[0021] Furthermore, when some or all of the components of the learning device 1 are implemented using 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 implemented as a client-server system, cloud computing system, or other system connected via a communication network. Furthermore, the functions of the learning device 1 may be provided in the form of SaaS (Software as a Service).

[0022] The learning device according to the first embodiment establishes secure communication with an information terminal connected to the network of each organization and acquires a local model using the secure communication. Therefore, according to the first embodiment, a global model can be constructed even when the networks of multiple organizations are not always connected to each other.

[0023] <Embodiment 2> Embodiment 2 is a specific example of embodiment 1. Fig. 2 is a schematic diagram showing the configuration of a learning system 100 according to embodiment 2. Learning system 100 includes information terminal 2a, information terminal 2b, information terminal 2c, VPN device 3a, VPN device 3b, VPN device 3c, and learning device 4. Learning device 4 is a specific example of the learning device 1 described above.

[0024] An information terminal 2a and a VPN device 3a are arranged in a network Na of organization A. An information terminal 2b and a VPN device 3b are arranged in a network Nb of organization B. An information terminal 2c and a VPN device 3c are arranged in a network Nc of organization C.

[0025] A data set owned by organization A is stored in information terminal 2a. A data set owned by organization B is stored in information terminal 2b. A data set owned by organization C is stored in information terminal 2c.

[0026] Furthermore, information terminal 2a constructs a local model La that has learned from a dataset owned by organization A. Information terminal 2b constructs a local model Lb that has learned from a dataset owned by organization B. Information terminal 2c constructs a local model Lc that has learned from a dataset owned by organization C. Information terminals 2a, 2b, and 2c update local models La, Lb, and Lc in accordance with the accumulation of datasets. When local models La, Lb, and Lc are not to be distinguished from one another, they may be simply referred to as local model L.

[0027] The number of organizations is not limited to three. It may be two, four, or more. Each organization is, for example, a pharmaceutical manufacturer or a chemical manufacturer. In this case, the dataset is a dataset of compounds. Each record included in the dataset of compounds lists structural information and property information of the compound. The structure of the compound is expressed as 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). Property values ​​(e.g., tensile strength) may be values ​​obtained through experiments, simulations, or theoretical calculations. For example, data generated daily in research and development activities of organization A is stored in information terminal 2a. Of course, the dataset is not limited to a dataset of compounds, but may be a dataset of any object.

[0028] When the information terminals 2a, 2b, and 2c are not distinguished from one another, they may be simply referred to as information terminals 2. When the networks Na, Nb, and Nc are not distinguished from one another, they may be simply referred to as network N. The network N may be a LAN (Local Area Network) or a network in which multiple LANs are connected. The network N is connected to a public network PN such as the Internet.

[0029] The VPN devices 3a, 3b, and 3c are VPN servers or VPN-compatible routers. When the VPN devices 3a, 3b, and 3c are not distinguished from one another, they may be simply referred to as VPN devices 3. The IP (Internet Protocol) address of the learning device 4 may be set in advance in the VPN device 3. The VPN may be an Internet VPN, an IP-VPN, or a wide area Ethernet.

[0030] 3 is a block diagram illustrating the configuration of the learning device 4. The learning device 4 is connected to the network PN. The learning device 4 includes a communication establishment unit 41, an acquisition unit 42, and an integration unit 43.

[0031] The communication establishment unit 41 is a specific example of the communication establishment unit 11 described above. The communication establishment unit 41 establishes secure communication with the information terminal 2. Specifically, the communication establishment unit 41 connects to a VPN device 3, such as a VPN server, via a public network PN and requests a VPN connection from the VPN device 3. First, a TCP / IP connection is established between the learning device 4 and the VPN device 3. Then, the learning device 4 is authenticated, and a VPN session is established between the learning device 4 and the VPN device 3. After the acquisition unit 42 acquires the local model L, the communication establishment unit 41 terminates the VPN session. The learning device 4 may be connected to the network N via a remote access VPN.

[0032] The timing at which the communication establishment unit 41 establishes secure communication, i.e., the timing at which the learning device 4 connects to the network N via VPN, will be described later. This is because the timing may be related to the progress of processing in the integration unit 43, which will be described later. The timing at which secure communication is established with the information terminal 2a, the timing at which secure communication is established with the information terminal 2b, and the timing at which secure communication is established with the information terminal 2c may be different from each other.

[0033] The acquisition unit 42 is a specific example of the above-mentioned acquisition unit 12. The acquisition unit 42 acquires the local model L from the information terminal 2 after the learning device 4 is connected to the network N via VPN.

[0034] The integrating unit 43 is a specific example of the integrating unit 13 described above. The integrating unit 43 integrates the local models La, Lb, and Lc acquired by the acquiring unit 42. The integrated model is called a global model. The integrating unit 43 may integrate the local models La, Lb, and Lc at a predetermined timing (e.g., once a day, once every few months). The global model has higher performance than the local models La, Lb, and Lc. The integrating unit 43 may perform a process of integrating the local models La, Lb, and Lc when the local models La, Lb, and Lc are updated.

[0035] The integrating unit 43 may generate the global model by, for example, taking the arithmetic mean of the model parameters of the local model La, the model parameters of the local model Lb, and the model parameters of the local model Lc. Note that the method for integrating the model parameters is not limited to the arithmetic mean.

[0036] After the integrator 43 generates the global model, the learning device 4 distributes the global model to the information terminals 2 a, 2 b, and 2 c. For example, after the process of generating the global model is completed, the learning device 4 may connect to the networks Na, Nb, and Nc in order via VPN and transmit the global model to the information terminals 2 a, 2 b, and 2 c.

[0037] Furthermore, the learning device 4 may connect to the network N via VPN in response to a request from each information terminal 2 and transmit the global model to the information terminal 2. Each information terminal 2 can import the global model at any time. Organizations A, B, and C can utilize a high-performance global model that combines data sets owned by multiple organizations.

[0038] Constructing multiple local models L and integrating the multiple local models L is also called federated learning. It can also be said that the learning device 4 performs federated learning.

[0039] The learning device 4 sequentially repeats the process of establishing secure communication and the process of acquiring the local model L. This allows the performance of the global model to be improved based on the data set accumulated daily in each information terminal 2. Note that the process of integrating multiple local models may be performed at any timing.

[0040] Next, the timing at which the communication establishment unit 41 establishes secure communication will be described. The communication establishment unit 41 may establish secure communication at a predetermined timing. The predetermined timing may be once every few months or once every few days.

[0041] Furthermore, the communication establishment unit 41 may establish secure communication in response to receiving a request from each information terminal 2. For example, the information terminal 2 transmits the local model after having the local model L newly learn a data set exceeding a predetermined amount. The information terminal 2 may transmit the request when the model parameters of the local model L converge in learning the data set exceeding the predetermined amount.

[0042] When training one data set on the local model L, the data set is divided into multiple batches, and the local model L is trained on the multiple batches in order. The process of dividing the data set into batches and training the multiple batches is repeated a predetermined number of times. The predetermined number of times is set so that the model parameters of the local model L converge. However, the predetermined number of times needs to be set to a small number that does not cause overtraining.

[0043] If the dataset is divided into five batches and the training is repeated 10 times, the request may be sent when the training is complete, i.e., when the 10th training run is completed. The request may also be sent when the training is nearing completion, for example, when the fourth batch of the 10th training run is completed.

[0044] The communication establishment unit 41 may establish the next secure communication based on the progress of the process of integrating multiple local models L. If the process in the integrating unit 43 is not a simple arithmetic average or if there are a large number of organizations, the process by the integrating unit 43 may take a long time. It is efficient if the learning device 4 can start the next process after the process in the integrating unit 43 is completed.

[0045] Furthermore, when a secure computation technique is applied, the processing by the integrating unit 43 may take a long time. It is known that the dataset used for learning can be inferred by reverse engineering the local model L. Therefore, in order to improve the confidentiality of the local model L, it is desirable to perform secure computation when integrating the local models L. Secure computation is a technique for performing computational processing on data while keeping it encrypted, and known secure computation techniques include multi-party computation (MPC) and homomorphic encryption.

[0046] 4 is a flowchart showing the flow of processing for updating the local model L. It is assumed that the learning device 4 stores the initial local model L (step S101).

[0047] Next, the communication establishment unit 41 of the learning device 4 determines whether it is time to establish secure communication (step S102). If it is not time to establish secure communication (NO in step S102), the process returns to step S102.

[0048] If it is time to establish secure communication (YES in step S102), the communication establishment unit 41 establishes secure communication between the information terminal 2 and the learning device 4, and the acquisition unit 42 acquires the local model L from the information terminal 2 (step S103). This updates the local model L that serves as the basis for constructing the global model. Thereafter, the communication establishment unit 41 terminates the secure communication.

[0049] In step S103, multiple local models L may be acquired. First, secure communication is established between the information terminal 2a and the learning device 4, the acquisition unit 42 acquires the local model La from the information terminal 2a, and the communication establishment unit 41 terminates the secure communication. Then, secure communication is established between the information terminal 2b and the learning device 4, the acquisition unit 42 acquires the local model Lb from the information terminal 2b, and the communication establishment unit 41 terminates the secure communication. Then, secure communication is established between the information terminal 2c and the learning device 4, the acquisition unit 42 acquires the local model Lc from the information terminal 2c, and the communication establishment unit 41 terminates the secure communication. Of course, in step S103, the local model L may be acquired from any of the information terminals 2a, 2b, and 2c. After acquiring (updating) the local model L, the process returns to step S102. The process of integrating multiple local models L may be performed at any timing.

[0050] The learning device according to the second embodiment connects to the network of each organization via VPN at appropriate communication timing to acquire the local model, thereby enabling the local model to be received securely and updated at appropriate timing.

[0051] Note that secure communication is not limited to communication via a VPN. Secure communication may be communication using any secure communication protocol (e.g., an encryption protocol). The local model may be sent from the information terminal 2 to the learning device 4 by email using a secure communication protocol (e.g., S / MIME).

[0052] <Modification of Second Embodiment> A repository that stores data sets owned by each organization may be provided in a location other than the information terminal 2 that builds the local model L. In this case, the information terminal 2 may establish secure communication (e.g., SSL) with the repository as needed to acquire the data sets required for learning. This not only makes it possible to secure communication between the local model L and the global model, but also between the local model L and the repository.

[0053] <Third Embodiment> The third embodiment is a specific example of the second embodiment. The learning device according to the third embodiment integrates model parameters of local models using secure computation. Fig. 5 is a block diagram showing the configuration of a learning system 100a according to the third embodiment. Comparing Fig. 2 with Fig. 5, a server group 5 has been added.

[0054] The server group 5 includes a plurality of secure computation servers 51. The number of secure computation servers 51 is not limited to three. However, in order to perform secure computation, it is preferable that the number of secure computation servers 51 is three or more.

[0055] The server group 5 integrates the local model La, the local model Lb, and the local model Lc, and transmits the result of the secure computation to the learning device 4.

[0056] The integrating unit 43 of the learning device 4 divides the model parameters of the local model La into multiple shares (e.g., three) and transmits the multiple shares to the multiple secure computation servers 51. The integrating unit 43 divides the model parameters of the local model Lb into multiple shares and transmits the multiple shares to the multiple secure computation servers 51. The integrating unit 43 divides the model parameters of the local model Lc into multiple shares and transmits the multiple shares to the multiple secure computation servers 51.

[0057] Each secure computation server 51 uses the received shares to perform secure computation to calculate the global model. The local model cannot be discovered from the shares, and computation using the shares can be considered secure computation. Multiple secure computation servers 51 may cooperate to perform multi-party computation (MPC). Since the amount of computation required to integrate the local model L is sufficiently small, it is believed that the server group 5 can perform secure computation in a realistic amount of time.

[0058] Furthermore, all or part of the functions of the learning device 4 may be provided in the server group 5. Secure communication may be established between the server group 5 and the information terminal 2 by connecting multiple secure computation servers 51 to the network N via VPN. Model parameters of the local model L may be acquired by multiple secure computation servers 51 receiving multiple shares. Multiple secure computation servers 51 may integrate model parameters of multiple local models L by performing secure computation.

[0059] The third embodiment also provides the same effects as the second embodiment. Furthermore, according to the third embodiment, the calculation for integrating the global model can be kept confidential.

[0060] 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.

[0061] 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.

[0062] Some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes. (Supplementary Note 1) A learning device comprising: communication establishment means for establishing secure communication with an information terminal located on a network of each organization; acquisition means for acquiring, from the information terminal using the secure communication, a local model trained on a dataset for each organization; and integration means for integrating the acquired multiple local models. (Supplementary Note 2) The learning device according to Supplementary Note 1, wherein the communication establishment means establishes the secure communication in response to a request from each information terminal, the request being transmitted after the local model has trained on a dataset exceeding a predetermined amount. (Supplementary Note 3) The learning device according to Supplementary Note 2, wherein the request is transmitted when model parameters of the local model have converged in new training on a dataset exceeding the predetermined amount. (Supplementary Note 4) The learning device according to Supplementary Note 1, wherein the communication establishment means establishes the secure communication at a predetermined timing. (Supplementary Note 5) The learning device according to Supplementary Note 1, wherein the communication establishment means establishes the next secure communication based on the progress of a process of integrating the multiple local models. (Supplementary Note 6) The learning device according to Supplementary Note 5, wherein the integrating means integrates the multiple local models using secure computing technology. (Supplementary Note 7) The learning device according to any one of Supplements 1 to 6, wherein the communication establishing means establishes the secure communication by connecting the learning device to the network via a VPN (Virtual Private Network). (Supplementary Note 8) A learning system comprising: an information terminal disposed on a network of each organization; and a learning device, wherein the learning device establishes secure communication with the information terminal, uses the secure communication to obtain from the information terminal a local model trained on a dataset for each organization, and integrates the obtained multiple local models. (Supplementary Note 9) The learning system according to Supplementary Note 8, wherein the learning device establishes the secure communication in response to receiving a request from each information terminal, and the request is transmitted after the local model has newly learned a dataset exceeding a predetermined amount.(Supplementary Note 10) A learning method in which a computer establishes secure communication with an information terminal located in a network of each organization, obtains, from the information terminal using the secure communication, a local model trained with a dataset for each organization, and integrates the multiple acquired local models. (Supplementary Note 11) A non-transitory computer-readable medium storing a program for causing a computer to execute the following processes: establishing secure communication with an information terminal located in a network of each organization, obtaining, from the information terminal using the secure communication, a local model trained with a dataset for each organization, and integrating the multiple acquired local models.

[0063] 1, 4 Learning device 11, 41 Communication establishment unit 12, 42 Acquisition unit 13, 43 Integration unit 2, 2a, 2b, 2c Information terminal 3, 3a, 3b, 3c VPN device 100, 100a Learning system 5 Server group 51 Secure computation server N, Na, Nb, Nc Network PN Public network

Claims

1. Communication establishment means for establishing secure communication with information terminals arranged in the network of each organization, Acquisition means for acquiring a local model obtained by training a data set for each organization from the information terminal using the secure communication, Integration means for integrating the plurality of acquired local models A learning device comprising.

2. The communication establishment means establishes the secure communication in response to receiving a request from each information terminal, The request is transmitted after newly training a data set exceeding a predetermined amount in the local model. The learning device according to claim 1.

3. The request is transmitted when the model parameters of the local model converge in the new training. The learning device according to claim 2.

4. The communication establishment means establishes the secure communication at a predetermined timing. The learning device according to claim 1.

5. The communication establishment means establishes the next secure communication based on the progress of the process of integrating the plurality of local models. The learning device according to claim 1.

6. The integration means integrates the plurality of local models using secret computing technology. The learning device according to claim 5.

7. The communication establishment means establishes the secure communication by connecting the learning device to the network via a VPN (Virtual Private Network). The learning device according to any one of claims 1 to 6.

8. An information terminal arranged in the network of each organization, A learning device, A learning system comprising, The learning device, Establishes secure communication with the information terminal, Acquires a local model obtained by training a data set for each organization from the information terminal using the secure communication, Integrates the plurality of acquired local models. Learning system.

9. The learning device establishes the secure communication in response to receiving a request from each information terminal, The request is transmitted after newly training a data set exceeding a predetermined amount in the local model. The learning system according to claim 8.

10. A computer, Establishes secure communication with information terminals arranged in the network of each organization, Acquires a local model obtained by training a data set for each organization from the information terminal using the secure communication, Integrates the plurality of acquired local models. Learning method.

11. A program for causing a computer to perform: a process of establishing secure communication with information terminals arranged in the networks of respective organizations; a process of obtaining local models learned with dataset for each organization from the information terminals by using the secure communication; and a process of integrating the obtained plurality of local models. ​