Information processing apparatus, information processing system, information processing method, and program

Abstract

To appropriately select a partner of associative learning.SOLUTION: An information processing apparatus capable of communicating with an information processing apparatus participating in federated learning via a network includes an acquisition unit that acquires client-related information based on information registered by a client from the information processing apparatus participating in the federated learning, an evaluation unit that evaluates a similarity between the client-related information of the information processing apparatus participating in the federated learning and client-related information of a partner candidate of the federated learning, and a screen output unit that outputs screen data for selecting a partner of the federated learning based on the similarity.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an information processing device, an information processing system, an information processing method, and a program. [Background technology]

[0002] There is a well-known machine learning technique called federated learning. In federated learning, machine learning is performed while the data held by each business is distributed, without consolidating it. Federated learning makes it possible to build models that utilize data from multiple businesses, as if the data were linked, while ensuring privacy and security.

[0003] For example, Patent Document 1 discloses a data provision method for a learning model based on distributed learning. In the data provision method disclosed in Patent Document 1, each node learns a local model using local data processed by the multiple nodes, global model data is updated based on multiple local model data acquired from each of the multiple nodes, the local model data and global model data are stored in association with version information, and model data selected from the global model and local models associated with different version information is distributed to multiple nodes or other devices. Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the conventional technology, it is difficult to appropriately select a partner for federated learning. For example, in the conventional technology, a user selects a partner for federated learning based on client-related information, but it is difficult to appropriately select a partner for federated learning based on client-related information alone.

[0005] One embodiment of the present invention aims to appropriately select partners for federated learning. [Means for solving the problem]

[0006] An information processing device that is one embodiment of the present invention is an information processing device that can communicate via a network with information processing devices participating in federated learning, and is equipped with an acquisition unit that acquires client-related information based on information registered by clients from the information processing devices participating in the federated learning, an evaluation unit that evaluates the similarity between the client-related information of the information processing devices participating in the federated learning and the client-related information of potential partners for the federated learning, and a screen output unit that outputs screen data for selecting a partner for the federated learning based on the similarity. [Effects of the Invention]

[0007] According to one embodiment of the present invention, partners for federated learning can be appropriately selected. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a block diagram showing an example of the overall configuration of a federated learning system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram showing an example of a hardware configuration of a computer according to an embodiment of the present invention. [Figure 3] 1 is a block diagram showing an example of the functional configuration of a federated learning system according to an embodiment of the present invention. [Figure 4] 1 is a flowchart illustrating an example of a federated learning method according to an embodiment of the present invention. [Figure 5] 10 is a flowchart illustrating an example of a general-purpose model registration process according to an embodiment of the present invention. [Figure 6] FIG. 10 is a diagram showing an example of a general-purpose model addition application screen according to an embodiment of the present invention. [Figure 7] 10 is a flowchart illustrating an example of a federation partner selection process according to an embodiment of the present invention. [Figure 8] FIG. 2 is a diagram showing a first example of a federation model selection screen according to one embodiment of the present invention. [Figure 9]FIG. 10 is a diagram showing a second example of a federation model selection screen according to one embodiment of the present invention. [Figure 10] FIG. 10 is a diagram showing a third example of a federation model selection screen according to one embodiment of the present invention. [Figure 11] FIG. 10 is a diagram showing a fourth example of a federation model selection screen according to one embodiment of the present invention. [Figure 12] 10 is a flowchart illustrating an example of a federated learning process according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, components having the same functions are designated by the same reference numerals, and redundant description will be omitted.

[0010] [Embodiment] One embodiment of the present invention is an information processing system that builds a machine learning model through federated learning. Hereinafter, the information processing system according to this embodiment will be referred to as a "federated learning system." The federated learning system has a function for selecting a partner (hereinafter also referred to as a "federated partner") to participate in federated learning when starting federated learning. A user of the federated learning system can arbitrarily select a federated partner from multiple clients participating in federated learning.

[0011] Conventionally, federated learning was based on the premise that all participating clients would be working on the same task. Therefore, when starting federated learning, there was no need to consider differences in tasks, and federated partners were selected based on, for example, industry, task type, or amount of data. However, in reality, each client wanted to work on different tasks, and federated learning was required between clients working on different tasks.

[0012] In recent years, techniques such as distillation learning have made it possible to perform federated learning across different tasks. However, in federated learning with federated partners working on different tasks, it is unclear whether model accuracy can be improved based solely on factors such as industry, task type, or amount of data, making it difficult to select an appropriate federated partner.

[0013] An object of one embodiment of the present invention is to appropriately select a federation partner. To this end, in this embodiment, client-related information based on information registered by each client is acquired from each client, and screen data for selecting a federation partner from multiple clients is output based on the similarity of the client-related information.

[0014] In one aspect, according to this embodiment, a screen for selecting a federation partner based on the similarity of client-related information is displayed, allowing an appropriate federation partner to be selected. In another aspect, according to this embodiment, federated learning is performed with an appropriate federation partner, allowing the accuracy of the machine learning model to be improved even in federated learning between different tasks.

[0015] <Overall structure of the federated learning system> The overall configuration of the federated learning system will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the overall configuration of a federated learning system according to one embodiment of the present invention.

[0016] 1, the federated learning system 1000 includes a central device 10, a central storage device 15, a plurality of client devices 20 (20-1 to 20-N), a plurality of client storage devices 25 (25-1 to 25-N), a client device 30, and a client storage device 35, where N is an integer equal to or greater than 2.

[0017] The central device 10 and the client device 20 are connected to a communication network N1. The central device 10 and the client device 30 are connected to a communication network N2. The communication networks N1 and N2 are configured so that the connected devices can communicate with each other. The communication networks N1 and N2 may be integrated into a single communication network.

[0018] The communication networks N1 and N2 are constructed by a network using wired communication, such as the Internet, a local area network (LAN), or a wide area network (WAN). The communication network N1 may include not only wired communication but also wireless communication such as a wireless LAN or short-range wireless communication, or a network using mobile communication such as WiMAX (Worldwide Interoperability for Microwave Access), LTE (Long Term Evolution), or 5G (5th Generation).

[0019] The central device 10 and the central storage device 15 are electrically connected. The client device 20 and the client storage device 25 are electrically connected. The client device 30 and the client storage device 35 are electrically connected. The central storage device 15, the client storage device 25, and the client storage device 35 may be connected to the communication network N1 or the communication network N2.

[0020] The central device 10 is an example of an information processing device that generates a central model. An example of the central device 10 is a computer such as a personal computer, a workstation, or a server. The central device 10 is a server that participates in federated learning.

[0021] The central model is a machine learning model that integrates multiple client models generated by each client device 20-1, . . . , 20-N. Examples of machine learning models include neural networks, deep neural networks, support vector machines, and random forests.

[0022] The central storage device 15 is an example of a storage device that stores the central model. An example of the central storage device 15 is an external storage device such as a disk array or a NAS (Network Attached Storage). The central storage device 15 may be built into the central device 10.

[0023] The client device 20 is an example of an information processing device that generates a client model. An example of the client device 20 is a computer such as a personal computer, a workstation, or a server. The client device 20 is a node that participates in federated learning. The client device 20 may execute a predetermined task using a trained central model or trained client models (hereinafter collectively referred to as "trained models").

[0024] The client model is a machine learning model generated using client data. The client data is a data set for learning the client model. The client data may be prepared in advance by the user of the client device 20. The content of the client data varies depending on the task to be executed using the client model.

[0025] The client data may include a general-purpose data set (hereinafter referred to as a "general-purpose data set") that is open to the general public on the web, etc. The general-purpose data set may be prepared based on a request from the client device 20.

[0026] The client storage device 25 is a storage device that stores the client model and client data. An example of the client storage device 25 is an external storage device such as a disk array or NAS. The client storage device 25 may be built into the client device 20.

[0027] The client device 30 is an example of an information processing device that uses a trained model. An example of the client device 30 is a computer such as a personal computer, a workstation, or a server. The client device 30 does not participate in federated learning, but performs a predetermined task using the trained model generated by the federated learning. The client device 30 does not need to be included in the federated learning system 1000.

[0028] The client storage device 35 is an example of a storage device that stores a trained model. An example of the client storage device 35 is an external storage device such as a disk array or a NAS. The client storage device 35 may be built into the client device 30. If the client device 30 is not included in the federated learning system 1000, the client storage device 35 may not be included in the federated learning system 1000.

[0029] The central device 10, the client device 20, or the client device 30 is not limited to a computer as long as it is a device equipped with a communication function. The central device 10, the client device 20, or the client device 30 may be, for example, an image forming device (printer, fax machine, MFP (Multifunction Peripheral / Product / Printer: digital multifunction machine), scanner device), PJ (Projector: projector), IWB (Interactive White Board: white board with electronic blackboard function capable of intercommunication), output device such as digital signage, HUD (Head Up Display) device, industrial machinery, imaging device, sound collection device, medical equipment, network home appliance, automobile (Connected Car), notebook PC (Personal Computer), mobile phone, smartphone, tablet terminal, game console, PDA (Personal Digital Assistant), digital camera, wearable PC, desktop PC, or the like.

[0030] The configuration of the federated learning system 1000 shown in Figure 1 is an example, and other configurations are also possible. For example, the central device 10, the client devices 20, and the client devices 30 may be realized by a single information processing device, or may be a system realized by multiple information processing devices. For example, the central storage device 15 may be built into the central device 10, the client storage device 25 may be built into the client device 20, and the client storage device 35 may be built into the client device 30. The federated learning system 1000 includes various devices that input and / or output electronic data, and these devices may use various services provided by the federated learning system 1000.

[0031] <Hardware configuration of the federated learning system> The hardware configuration of each device included in the federated learning system 1000 will be described with reference to Fig. 2. The central device 10, client device 20, and client device 30 included in the federated learning system 1000 may be realized by a computer, for example. Fig. 2 is a block diagram showing an example of the hardware configuration of a computer according to an embodiment of the present invention.

[0032] As shown in FIG. 2, the computer 500 includes a CPU 501, a ROM 502, a RAM 503, a HD 504, a HDD (Hard Disk Drive) controller 505, a display 506, an external device connection I / F (Interface) 508, a network I / F 509, a bus line 510, a keyboard 511, a pointing device 512, a DVD-RW (Digital Versatile Disk Rewritable) drive 514, and a media I / F 516.

[0033] Of these, the CPU 501 controls the operation of the entire computer. The ROM 502 stores programs used to drive the CPU 501, such as the IPL. The RAM 503 is used as a work area for the CPU 501. The HD 504 stores various data such as programs. The HDD controller 505 controls the reading and writing of various data from and to the HD 504 under the control of the CPU 501.

[0034] The display 506 displays various types of information such as a cursor, menus, windows, characters, or images. The external device connection I / F 508 is an interface for connecting various types of external devices. In this case, the external devices are, for example, USB (Universal Serial Bus) memories or printers. The network I / F 509 is an interface for data communication using the communication network N1. The bus line 510 is an address bus, a data bus, or the like for electrically connecting the components such as the CPU 501 shown in FIG. 2.

[0035] The keyboard 511 is a type of input means having multiple keys for inputting characters, numbers, various instructions, etc. The pointing device 512 is a type of input means for selecting and executing various instructions, selecting a processing target, moving a cursor, etc. The DVD-RW drive 514 controls reading and writing of various data from a DVD-RW 513, which is an example of a removable recording medium. Note that this is not limited to a DVD-RW, and may be a DVD-R, etc. The media I / F 516 controls reading and writing (storing) of data from a recording medium 515, such as a flash memory.

[0036] <Functional configuration of the federated learning system> The functional configuration of the federated learning system 1000 will be described with reference to Fig. 3. Fig. 3 is a block diagram showing an example of the functional configuration of the federated learning system according to one embodiment of the present invention.

[0037] Central Storage Device As shown in FIG. 3, the central storage device 15 includes a central data storage unit 151, a central model storage unit 152, and a client model storage unit 153.

[0038] Central data is stored in the central data storage unit 151. The central data is data used for various processes by the central device 10. In this embodiment, the central data includes client-related information related to the client model.

[0039] The client-related information is information about a client model or a generic model for federated learning. The generic model is a client model generated based on a generic dataset. The client-related information may be generated based on information registered by a user of the client device 20. The client-related information may include data to be provided to the client device 20 for selecting a federation partner. The client-related information may include, for example, industry type, task type, number of data, dataset, and characteristic data of the dataset.

[0040] The business type may be a business type targeted for a task performed by the client device 20. The client-related information may include at least one of a business or a purpose of use instead of or in addition to the business type. The business may be a business that utilizes a dataset. The purpose of use may be a purpose for using a trained model.

[0041] Examples of industries include retail, wholesale, food, communications, agriculture, forestry, mining, construction, transportation, finance, insurance, medical care, welfare, and services. Examples of business operations include research, development, production, sales, inquiries, surveys, and daily reports. Examples of usage purposes include stock price prediction, energy demand prediction, and object detection.

[0042] The task type is a classification of the task executed by the trained model. For example, the task type may include regression, classification, clustering, etc., or may be a classification that is a more detailed classification of these. For example, a classification task may be further subdivided into negative / positive classification or topic classification. Note that negative / positive classification is a task of classifying whether data to be classified is positive information or negative information. Topic classification is a task of classifying the type of topic that is being discussed in the data to be classified.

[0043] The number of data items is the number of data items to be learned that are included in the dataset. The dataset is a collection of data items in which the data items to be learned are associated with labels that are the correct values ​​for the data items. The labels may be predetermined depending on the task.

[0044] The feature data is data indicating the features of a dataset. The feature data may include one or more features related to the dataset. The feature data may include a feature indicating the features of the entire dataset. The feature indicating the features of the entire dataset may include, for example, a feature indicating the distribution of data in the dataset. The feature data may include a feature indicating the features of labels included in the dataset. The feature indicating the features of the labels may include a feature for each label. The feature for each label may be, for example, an embedding vector for the label or an embedding vector for data associated with the label.

[0045] A central model is stored in the central model storage unit 152. The central model stored in the central model storage unit 152 is a central model generated by the learning control unit 107 of the central device 10.

[0046] A client model is stored in the client model storage unit 153. The client model stored in the client model storage unit 153 is a client model generated by the learning unit 206 of the client device 20.

[0047] <Central Device> As shown in FIG. 3, the central device 10 includes a communication control unit 101, a storage control unit 102, an acquisition unit 103, a registration unit 104, an evaluation unit 105, a screen output unit 106, and a learning control unit 107.

[0048] The communication control unit 101, the memory control unit 102, the acquisition unit 103, the registration unit 104, the evaluation unit 105, the screen output unit 106, and the learning control unit 107 are realized, for example, by processing that is executed by the CPU 501 of a program expanded from the HD 504 shown in Figure 2 onto the RAM 503.

[0049] The communication control unit 101 controls communication with the client device 20. The communication control unit 101 may receive client-related information from the client device 20. The communication control unit 101 may transmit screen data to the client device 20. The communication control unit 101 may transmit a central model to the client device 20. The communication control unit 101 may receive a client model from the client device 20.

[0050] The storage control unit 102 controls the storage of data in the central storage device 15. The storage control unit 102 may write a trained client model or client-related information received from the client device 20 to the central storage device 15. The storage control unit 102 may write a trained central model generated by the learning control unit 107 to the central storage device 15. The storage control unit 102 may read a trained model to be sent to the client device 20 from the central storage device 15.

[0051] The acquisition unit 103 acquires various types of data from the client device 20. The acquisition unit 103 may acquire a client model or client-related information from the client device 20. The acquisition unit 103 may acquire the client-related information including one or more feature amounts selected by the client device 20. The feature amounts included in the client-related information may be selected based on a security policy set by the client device 20.

[0052] The registration unit 104 registers a generic model. The registration unit 104 may accept a generic model addition request from the client device 20. The registration unit 104 may generate client-related information related to the generic model based on the generic model addition request. The feature data of the client-related information related to the generic model may include all available feature amounts.

[0053] The evaluation unit 105 evaluates the similarity of the client-related information. The evaluation unit 105 may evaluate the similarity of the client-related information based on feature data included in the client-related information. As an example, the evaluation unit 105 may evaluate the similarity of the client-related information by calculating the cosine similarity between the feature data. The evaluation unit 105 may evaluate the similarity of the client-related information based on other information included in the client-related information (e.g., industry, business, purpose of use, task type, number of data, etc.).

[0054] The evaluation unit 105 may evaluate the similarity of the client-related information in response to a request from the client device 20. The evaluation unit 105 may evaluate the similarity of the client-related information between the client-related information of the requesting client device 20 and the client-related information of another client device 20. The client-related information of the other client device 20 may include client-related information related to a generic model. When evaluating the similarity between the client-related information of the client devices 20, the evaluation unit 105 may evaluate the similarity using a feature commonly included in each piece of client-related information.

[0055] The screen output unit 106 outputs screen data for displaying various screens on the client device 20. The screen data may be, for example, screen data written in HTML (Hyper Text Markup Language) or the like. The screen data may include, for example, an application written in JavaScript (registered trademark) or the like. The screen data output by the screen output unit 106 is transmitted to the client device 20 by the communication control unit 101.

[0056] The screen output unit 106 may output screen data for displaying a general-purpose model addition application screen. The general-purpose model addition application screen is a screen for requesting addition of a general-purpose model from the client device 20 to the central device 10.

[0057] Screen output unit 106 may output screen data for displaying a federated model selection screen. The federated model selection screen is a screen for selecting a model to be a partner in federated learning from among federated learning partner candidates. The federated learning partner candidates may include client models (federated partner models) or general-purpose models generated by client devices 20-1 to 20-N included in federated learning system 1000. The federated model selection screen may display either the federated partner models or the general-purpose models in a selectable manner, or may display both in a selectable manner.

[0058] The screen output unit 106 may generate screen data for displaying an association model selection screen based on the similarity of the client-related information evaluated by the evaluation unit 105. For example, the screen output unit 106 may generate screen data of an association model selection screen that displays, in a selectable manner, client devices 20 whose similarity of client-related information is equal to or greater than a predetermined threshold. For example, the screen output unit 106 may generate screen data of an association model selection screen that displays information about the client devices 20 in association with the similarity of the client-related information. For example, the screen output unit 106 may generate screen data of an association model selection screen that displays information about the client devices 20 in descending order of the similarity of the client-related information.

[0059] The learning control unit 107 controls the federated learning. The learning control unit 107 may distribute an initial model to the client device 20. The initial model is a central model in an initial state. The learning control unit 107 may acquire a trained client model from the client device 20. The learning control unit 107 may generate a trained central model based on the trained client model. The learning control unit 107 may distribute the trained central model to the client device 20. The learning control unit 107 may repeatedly perform the federated learning until an end condition for ending the federated learning is met.

[0060] Client Storage Device As shown in FIG. 3, the client storage device 25 includes a client data storage unit 251 and a client model storage unit 252.

[0061] Client data is stored in the client data storage unit 251. The client data is data used for various processes by the client device 20. In this embodiment, the client data includes a data set for learning a client model.

[0062] The client model is stored in the client model storage unit 252. The client model is a trained client model generated by the training unit 206 of the client device 20.

[0063] <Client device> As shown in FIG. 3, the client device 20 includes a communication control unit 201, a storage control unit 202, a display control unit 203, a setting unit 204, a registration unit 205, a learning unit 206, and an application unit 207.

[0064] The communication control unit 201, the memory control unit 202, the display control unit 203, the setting unit 204, the registration unit 205, the learning unit 206 and the application unit 207 are realized, for example, by processing that is executed by the CPU 501 of a program expanded from the HD 504 shown in Figure 2 onto the RAM 503.

[0065] The communication control unit 201 controls communication with the central device 10. The communication control unit 201 may transmit client-related information to the central device 10. The communication control unit 201 may receive screen data from the central device 10. The communication control unit 201 may receive a central model from the central device 10. The communication control unit 201 may transmit a client model to the central device 10.

[0066] The storage control unit 202 controls the storage of data in the client storage device 25. The storage control unit 202 may read a dataset used for training a client model from the client storage device 25. The storage control unit 202 may write a trained client model generated by the training unit 206 to the client storage device 25. The storage control unit 202 may read a client model to be transmitted to the client device 20 from the client storage device 25.

[0067] The display control unit 203 displays various screens on the display 506 based on the screen data received from the central device 10. The display control unit 203 may display a general-purpose model addition application screen on the display 506 based on the screen data of the general-purpose model addition application screen. The display control unit 203 may display a federation model selection screen on the display 506 based on the screen data of the federation model selection screen.

[0068] The setting unit 204 sets a security policy for data to be registered in the central device 10. The security policy may include a sharing range or an access level of a client model. The security policy may also include information indicating the type of feature to be included in the feature data of the client-related information.

[0069] The registration unit 205 registers the client model in the central device 10. The registration unit 205 may generate client-related information regarding the client model and transmit it to the central device 10. The registration unit 205 may calculate feature data of the dataset based on the client data read from the client storage device 25. The registration unit 205 may select feature amounts to be included in the feature data based on the security policy set by the setting unit 204.

[0070] The registration unit 205 may, for example, calculate the distribution of data included in the dataset and generate feature quantities that indicate characteristics of the distribution. The registration unit 205 may, for example, generate embedding vectors of labels or data included in the dataset based on a language model such as TF-IDF (Term Frequency - Inverse Document Frequency), Word2Vec (Word to Vector), or BERT (Bidirectional Encoder Representations from Transformers).

[0071] The learning unit 206 generates a trained client model based on the client data read from the client storage device 25. The learning unit 206 may generate the trained client model by learning the client data with respect to the central model received from the central device 10 according to a predetermined learning algorithm.

[0072] The application unit 207 deploys the trained model. The application unit 207 may select a trained model to deploy from a trained central model or a trained client model. The application unit 207 may execute a predetermined task using the deployed trained model.

[0073] <Processing procedure of the federated learning method> The federated learning method executed by federated learning system 1000 will be described with reference to Figures 4 to 12. Figure 4 is a flowchart showing an example of the federated learning method according to one embodiment of the present invention.

[0074] In step S1, the registration unit 205 of the client device 20 accepts input of a dataset. The registration unit 205 may also accept input of a portion of client-related information along with the dataset. For example, the registration unit 205 may accept input of industry, business, purpose of use, task type, etc. The registration unit 205 sends the accepted dataset and a portion of the client-related information to the storage control unit 202. The storage control unit 202 stores the dataset and a portion of the client-related information in the client data storage unit 251 of the client storage device 25.

[0075] In step S2, the registration unit 104 of the central device 10 registers the generic model in response to a request from the client device 20. Specifically, the registration unit 104 accepts a generic model addition application from the client device 20 and acquires the generic data set indicated in the application. The registration unit 104 generates client-related information related to the generic model based on the acquired generic data set. The registration unit 104 sends the generated client-related information to the storage control unit 102. The storage control unit 102 stores the client-related information related to the generic model in the central data storage unit 151 of the central storage device 15.

[0076] <<General model registration process>> The general-purpose model registration process (step S2 in Fig. 4) will be described in more detail with reference to Fig. 5 and Fig. 6. Fig. 5 is a flowchart showing an example of the general-purpose model registration process according to one embodiment of the present invention.

[0077] In step S2-1, the display control unit 203 of the client device 20 requests a general-purpose model addition application screen from the central device 10. In response to the request from the client device 20, the screen output unit 106 of the central device 10 generates screen data for displaying the general-purpose model addition application screen. The screen output unit 106 sends the generated screen data to the communication control unit 101. The communication control unit 101 transmits the screen data for displaying the general-purpose model addition application screen to the client device 20.

[0078] In the client device 20, the communication control unit 201 receives screen data for displaying a general-purpose model addition application screen from the central device 10. The display control unit 203 displays the general-purpose model addition application screen on the display 506 based on the screen data received by the communication control unit 201. The display control unit 203 transmits a general-purpose model addition application to the central device 10 in response to an operation on the general-purpose model addition application screen.

[0079] <<Application screen for adding a generic model>> 6 is a diagram showing an example of a general-purpose model addition application screen according to one embodiment of the present invention. As shown in FIG. 6, a general-purpose model addition application screen 600 has a location information input section 601, a task type input section 602, a usage conditions input section 603, a remarks input section 604, and an application button 605.

[0080] The location information input field 601 is an area for inputting location information where a generic dataset can be obtained. The location information may be, for example, a URL (Uniform Resource Locator) where the generic dataset is publicly available. The task type input field 602 is an area for inputting the task type to which the generic dataset can be applied. The usage conditions input field 603 is an area for inputting usage conditions such as license information for using the generic dataset. The remarks input field 604 is an area for inputting additional information regarding the generic dataset or the usage application. The application button 605 is a button for sending a generic model addition application to the central device 10.

[0081] Returning to Figure 5, the explanation will be given below. In step S2-2, the registration unit 104 of the central device 10 accepts a generic model addition request from the client device 20. The administrator of the federated learning system 1000 determines whether or not to add a generic model based on the generic model addition request. If the administrator determines to add a generic model (addition possible), the registration unit 104 proceeds to step S2-3. On the other hand, if the administrator determines not to add a generic model (addition not possible), the registration unit 104 skips step S2-3 and terminates the generic data registration process.

[0082] The administrator of the federated learning system 1000 may determine whether to add a general-purpose model according to the following criteria, for example: However, these criteria are merely examples, and the administrator may determine whether to add a general-purpose model according to other criteria. Check commercial use: Check the license information for the general-purpose dataset to see if commercial use is permitted and if the license-based restrictions on use (such as whether the data can be redistributed or modified) are met. Security and privacy checks: Ensure generic datasets do not contain personal information and are appropriately anonymized. Check data quality: Check whether there are missing or outliers in the general-purpose dataset. Evaluate the performance when training only on the general-purpose dataset that performs additional judgments, and check whether appropriate classification labels are assigned. Task generality: Check whether the labels assigned to the general dataset are general-purpose and whether they can be expected to be effective when combined with the model.

[0083] In step S2-3, the registration unit 104 of the central device 10 acquires a general-purpose data set based on the general-purpose model addition application. The registration unit 104 generates client-related information related to the general-purpose model based on the acquired general-purpose data set. The registration unit 104 sends the generated client-related information to the storage control unit 102. The storage control unit 102 stores the client-related information related to the general-purpose model in the central data storage unit 151 of the central storage device 15.

[0084] In addition, the administrator of federated learning system 1000 constructs a client device 20 and a client storage device 25 for training a client model using the added general-purpose dataset, and adds them to federated learning system 1000. The administrator of federated learning system 1000 acquires the general-purpose dataset and stores it in the added client storage device 25.

[0085] 4, in step S3, the setting unit 204 of the client device 20 sets a security policy. For example, the setting unit 204 may display a screen for setting a security policy on the display 506, and set the security policy in response to an operation on the screen by the user of the client device 20.

[0086] In step S4, the registration unit 104 of the central device 10 registers a client model in response to a request from the client device 20. Specifically, the storage control unit 202 of the client device 20 reads out a part of the data set and client-related information registered in step S1 from the client data storage unit 251 of the client storage device 25. The storage control unit 202 sends the read data set and part of the client-related information to the registration unit 205. The registration unit 205 generates feature data of the data set. At this time, the registration unit 205 selects feature amounts to be included in the feature data based on the type of feature amount indicated in the security policy set in step S3.

[0087] The registration unit 205 generates client-related information related to the client model by including the feature data of the dataset in part of the client-related information. The registration unit 205 sends the generated client-related information to the communication control unit 201. The communication control unit 201 transmits the client-related information to the central device 10.

[0088] In the central device 10, the communication control unit 201 receives the client-related information from the client device 20. The communication control unit 201 sends the received client-related information to the storage control unit 202. The storage control unit 202 stores the client-related information in the central data storage unit 151 of the central storage device 15.

[0089] In step S5, the display control unit 203 of the client device 20 accepts the selection of a federation partner. Specifically, the display control unit 203 of the client device 20 displays an federation model selection screen on the display 506, and accepts the selection of a client device 20 to be a federation partner in response to an operation on the federation model selection screen.

[0090] <Alliance Partner Selection Process> The federation partner selection process (step S5 in FIG. 4) will be described in more detail with reference to Figures 7 to 11. Figure 7 is a flowchart showing an example of the federation partner selection process according to one embodiment of the present invention.

[0091] In step S5-1, the display control unit 203 of the client device 20 requests a federation model selection screen from the central device 10. In the central device 10, the storage control unit 202 reads out client-related information from the central data storage unit 151 of the central storage device 15 in response to the request from the client device 20. The storage control unit 202 sends the read out client-related information to the evaluation unit 105.

[0092] The evaluation unit 105 evaluates the similarity of the client-related information between the client-related information of the requesting client device 20 and the client-related information of another client device 20. The evaluation unit 105 sends the evaluated similarity of the client-related information to the screen output unit 106.

[0093] The screen output unit 106 generates screen data for displaying an association model selection screen based on the similarity of the client-related information evaluated by the evaluation unit 105. The screen output unit 106 associates the client-related information with the similarity of the client-related information and embeds the association model selection screen in screen data for displaying the association model selection screen. The screen output unit 106 sends the generated screen data to the communication control unit 101. The communication control unit 101 transmits the screen data for displaying the association model selection screen to the client device 20.

[0094] In the client device 20, the communication control unit 201 receives screen data for displaying the federation model selection screen from the central device 10. The display control unit 203 displays the federation model selection screen on the display 506 based on the screen data received by the communication control unit 201.

[0095] <Alliance Model Selection Screen> 8 is a diagram showing a first example of a federation model selection screen according to an embodiment of the present invention. As shown in FIG. 8, the federation model selection screen 700 includes a client model display section 701, a generic model display section 702, and a selection button 703.

[0096] The client model display unit 701 displays client-related information related to the client model. As an example, the client model display unit 701 may display the business type, task type, number of data, and update date. The business type, task type, and number of data are included in the client-related information. The update date is the date and time when the client model was last updated.

[0097] The client model display unit 701 has a selection unit 704 and a filtering unit 705. The selection unit 704 is a screen component for selecting the client device 20 that generates the client model as a federation partner. The filtering unit 705 is a screen component for switching whether or not to filter the client model displayed in the client model display unit 701.

[0098] Filtering is a function that displays only client models that are expected to improve the accuracy of the model by selecting them as federation partners. Specifically, it is a function that displays only client models whose similarity in client-related information is equal to or greater than a predetermined threshold, and hides client models whose similarity in client-related information is less than the threshold.

[0099] The filtering criteria may be set by the user of the client device 20. The filtering criteria may include, for example, the type of similarity used for filtering and a threshold value for each similarity. The type of similarity may specify a feature amount included in the feature data. For example, the type of similarity may include the similarity of the entire data set, the similarity of labels, the similarity of data for each label, etc. The type of similarity may be a combination of multiple similarities.

[0100] The generic model display unit 702 displays client-related information related to the generic model. For example, the generic model display unit 702 may display the industry type, task type, number of data, and update date, similar to the client model display unit 701. Because generic datasets are datasets that do not need to be kept confidential, the generic model display unit 702 may display details of labels or data included in the generic dataset.

[0101] The generic model display unit 702 has a selection unit 706 and a filtering unit 707. The selection unit 706 is a screen component for selecting the client device 20 that generates the generic model as a federation partner. The filtering unit 707 is a screen component for switching whether or not to filter the generic model displayed in the generic model display unit 702. The function of the filtering unit 707 may be the same as the function of the filtering unit 705.

[0102] The select button 703 is a button for confirming the selection of a federation partner. When the user of the client device 20 presses the select button 703, the client device 20 that generates the client model selected in the selection section 704 and the client device 20 that generates the generic model selected in the selection section 706 are selected as federation partners.

[0103] FIG. 9 is a diagram showing a second example of a federation model selection screen according to an embodiment of the present invention. As shown in FIG. 9, the federation model selection screen 700 may display a label similarity 711 in the client model display section 701. The federation model selection screen 700 may also display a label similarity 712 in the generic model display section 702. That is, the federation model selection screen 700 may display client-related information and the similarity of the client-related information in association with each other. While FIG. 9 shows an example in which label similarity is displayed as an example of the similarity of the client-related information, other similarities may be displayed, or multiple similarities may be displayed.

[0104] Fig. 10 is a diagram showing a third example of a federation model selection screen according to an embodiment of the present invention. As shown in Fig. 10, the federation model selection screen 700 may display client-related information sorted in descending order of label similarity 711 in the client model display section 701. The federation model selection screen 700 may also display client-related information sorted in descending order of label similarity 712 in the generic model display section 702. While Fig. 10 shows an example in which the client-related information is sorted and displayed in descending order of similarity, the client-related information may also be sorted and displayed in ascending order of similarity, or the descending order or ascending order may be switchable.

[0105] 11 is a diagram showing a fourth example of a federation model selection screen according to an embodiment of the present invention. As shown in FIG. 11, the federation model selection screen 700 may display label similarity 711 in the client model display section 701, and label similarity 712 and dataset similarity 713 in the generic model display section 702. That is, the federation model selection screen 700 may display different similarities of client-related information between the client model and the generic model.

[0106] Returning to Figure 7, the explanation will be given below. In step S5-3, the learning unit 206 of the client device 20 accepts the selection of a federation partner in response to an operation on the federation model selection screen 700. Specifically, when the selection button 703 on the federation model selection screen 700 is pressed, the learning unit 206 selects, as federation partners, the client device 20 that generates the client model selected in the selection unit 704 and the client device 20 that generates the generic model selected in the selection unit 706 on the federation model selection screen 700. The learning unit 206 transmits information indicating the selected federation partners to the central device 10.

[0107] Returning to Fig. 4, in step S6, the learning control unit 107 of the central device 10 executes federated learning. The federated learning is executed among the central device 10, the client devices 20 that selected federation partners in step S5, and the client devices 20 that were selected as federation partners in step S5.

[0108] In federated learning, the learning unit 206 of the client device 20 generates a trained client model, and the learning control unit 107 of the central device 10 generates a trained central model repeatedly until a termination condition is met. When the federated learning is completed, the application unit 207 of the client device 20 deploys the trained model. The application unit 207 executes a predetermined task using the deployed trained model.

[0109] <Associative Learning Processing> The federated learning process (step S6 in FIG. 4) will be described in more detail with reference to FIG. 12. FIG. 12 is a flowchart showing an example of the federated learning process according to one embodiment of the present invention.

[0110] In step S6-1, the learning control unit 107 of the central device 10 generates an initial model. The learning control unit 107 sends the generated initial model to the communication control unit 101. The communication control unit 101 distributes the initial model to each of the client devices 20 participating in the federated learning.

[0111] In the client device 20, the communication control unit 201 receives the initial model from the central device 10. The storage control unit 202 stores the initial model received by the communication control unit 201 in the client model storage unit 252 of the client storage device 25.

[0112] In step S6-2, the storage control unit 202 of the client device 20 reads the initial model from the client model storage unit 252 of the client storage device 25. The storage control unit 202 also reads the client data from the client data storage unit 251 of the client storage device 25.

[0113] The learning unit 206 of the client device 20 learns the client data with respect to the initial model according to a predetermined learning algorithm. As a result, a trained client model is generated. The learning unit 206 sends the trained client model to the communication control unit 201 and the storage control unit 202. The learning unit 206 may evaluate the trained client model using the client data (or a part of it), and send the trained client model together with the evaluation result to the communication control unit 201 and the storage control unit 202.

[0114] In step S6-3, the communication control unit 201 of the client device 20 receives the trained client model from the learning unit 206. The communication control unit 201 transmits the trained client model to the central device 10. The communication control unit 201 may transmit the evaluation result together with the trained client model to the central device 10. Furthermore, the memory control unit 202 stores the trained client model received from the learning unit 206 in the client model memory unit 252 of the client storage device 25. The memory control unit 202 may store the evaluation result together with the trained client model in the client model memory unit 252 of the client storage device 25.

[0115] In the central device 10, the communication control unit 101 receives client models from each client device 20. The storage control unit 102 stores the multiple client models received by the communication control unit 101 in the client model storage unit 153 of the central storage device 15. When the evaluation results are received together with the client models, the storage control unit 102 may store only the client models with good evaluation results in the client model storage unit 153 of the central storage device 15.

[0116] In step S6-4, the learning control unit 107 of the central device 10 reads out the multiple client models generated by each client device 20 from the client model storage unit 153 of the central storage device 15. The learning control unit 107 integrates the multiple read client models. As a result, a new central model is generated. The learning control unit 107 sends the new central model obtained by the integration to the storage control unit 102. The storage control unit 102 stores the new central model in the central model storage unit 152 of the central storage device 15.

[0117] In step S6-5, the central device 10 determines whether or not a termination condition has been met. The termination condition is a condition that must be met in order to terminate the update of the central model. For example, the termination condition may be that the difference before and after the update has converged, or that a predetermined number of updates has been reached. If it is determined that the termination condition has not been met (NO), the central device 10 proceeds to step S6-6. On the other hand, if it is determined that the termination condition has been met (YES), the central device 10 proceeds to step S6-7.

[0118] In step S6-6, the memory control unit 102 of the central device 10 reads the central model from the central model memory unit 152 of the central storage device 15. The communication control unit 101 distributes the central model read by the memory control unit 102 to each of the client devices 20 participating in the federated learning.

[0119] Thereafter, the federated learning system 1000 repeats the processes from step S6-2 to step S6-4 for the central model distributed in step S6-6. In this way, the federated learning system 1000 repeatedly distributes the central model, learns the client models, and updates the central model until it is determined in step S6-5 that the termination condition is met.

[0120] In step S6-7, the application unit 207 of the client device 20 selects a trained model to deploy from the trained central model and the trained client model. The application unit 207 transmits a request signal to the central device 10 to request transmission of the selected trained model.

[0121] In the central device 10, the communication control unit 101 receives a request signal from the client device 20. The memory control unit 102 reads out the trained model indicated in the request signal received by the communication control unit 101 from the central model memory unit 152 or the client model memory unit 153 of the central storage device 15. The communication control unit 101 transmits the trained model read out by the memory control unit 102 to the client device 20.

[0122] In step S6-8, the communication control unit 201 of the client device 20 receives the trained model from the central device 10. The communication control unit 201 sends the received trained model to the application unit 207. The application unit 207 stores the trained model in the client model storage unit 252 of the client storage device 25. The application unit 207 deploys the trained model. Thereafter, the application unit 207 uses the deployed trained model to execute a predetermined task or business.

[0123] <Effects of the embodiment> The central device 10 in this embodiment acquires client-related information based on information registered by the client from the client device 20, and outputs screen data for selecting a federation partner based on the similarity of the client-related information. In one aspect, according to this embodiment, a screen for selecting a federation partner is displayed based on the similarity of the client-related information, allowing an appropriate federation partner to be selected.

[0124] In particular, in federated learning between different tasks, it is difficult to select an appropriate federated partner based only on, for example, the industry, task type, or number of data, etc. According to this embodiment, a screen for selecting a federated partner is displayed based on the similarity of client-related information, making it possible to appropriately select a federated partner for performing federated learning between different tasks.

[0125] The central device 10 may output screen data that displays selectable client devices 20 whose client-related information has a similarity equal to or greater than a threshold. According to one aspect, the present embodiment displays only client devices 20 whose client-related information is similar, making it easy to select a client device 20 that is suitable as a federation partner.

[0126] The central device 10 may output screen data that displays information about the client devices 20 in association with the similarity of the client-related information. In one aspect, according to this embodiment, the similarity of the client-related information for each client device 20 is displayed, making it easy to select an appropriate client device 20 as a federation partner.

[0127] The central device 10 may output screen data that displays information about the client devices 20 in descending order of similarity of the client-related information. According to one aspect, according to this embodiment, the client devices 20 with higher similarity of the client-related information are displayed in positions that are easier to see, making it easy to select a client device 20 that is appropriate as a federation partner.

[0128] The central device 10 may acquire one or more feature data selected by the client device 20. In one aspect, according to this embodiment, the similarity of the client-related information can be evaluated using only feature data that matches the security policy of the client device 20.

[0129] The feature data may include a feature that indicates the characteristics of the entire dataset. The feature that indicates the characteristics of the entire dataset may include a feature that indicates the distribution of data in the dataset. In one aspect, according to this embodiment, client devices 20 that have similar characteristics of the entire dataset can be selected as federation partners.

[0130] The feature data may include features that indicate characteristics of labels included in the dataset. The features that indicate characteristics of the labels may include an embedding vector of the label or an embedding vector of data associated with the label. In one aspect, according to this embodiment, client devices 20 with similar label characteristics can be selected as federation partners.

[0131] [supplement] Each function of the above-described embodiments can be realized by one or more processing circuits. Here, the term "processing circuit" in this specification includes a processor programmed to perform each function by software, such as a processor implemented by an electronic circuit, as well as devices such as an ASIC (Application Specific Integrated Circuit), a DSP (Digital Signal Processor), an FPGA (Field Programmable Gate Array), and conventional circuit modules designed to perform each of the above-described functions.

[0132] The devices described in the examples are merely one of several computing environments for implementing the embodiments disclosed herein. In one embodiment, the central device 10 or the client device 20 includes multiple computing devices, such as a server cluster, configured to communicate with each other via any type of communication link, including a network, shared memory, etc., and to perform the processes disclosed herein.

[0133] For example, aspects of the present invention are as follows.

[0134] (Appendix 1) An information processing device capable of communicating with information processing devices participating in federated learning via a network, an acquisition unit that acquires client-related information based on information registered by the client from the information processing devices participating in the federated learning; an evaluation unit that evaluates the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a partner candidate for the federated learning; a screen output unit that outputs screen data for selecting a partner for associative learning based on the similarity; An information processing device comprising:

[0135] (Appendix 2) the screen output unit outputs the screen data that displays the partner candidates for the federated learning whose similarity is equal to or greater than a threshold in a selectable manner. 10. The information processing device according to claim 1.

[0136] (Appendix 3) the screen output unit outputs the screen data that displays information about the partner candidate for the federated learning in association with the similarity. 3. The information processing device according to claim 1 or 2.

[0137] (Appendix 4) the screen output unit outputs the screen data that displays information about the partner candidates for the federated learning arranged in descending order of the similarity. 4. The information processing device according to claim 3.

[0138] (Appendix 5) the evaluation unit evaluates the degree of similarity between feature data relating to the datasets. 5. An information processing device according to any one of appendices 1 to 4.

[0139] (Appendix 6) the acquisition unit acquires one or more pieces of feature data selected by the information processing devices participating in the federated learning. 6. The information processing device according to claim 5.

[0140] (Appendix 7) the feature data includes a feature amount indicating a feature of the entire data set; 7. The information processing device according to claim 5 or 6.

[0141] (Appendix 8) the feature quantity indicating the characteristics of the entire dataset includes a feature quantity indicating the distribution of data in the dataset; 8. The information processing device according to claim 7.

[0142] (Appendix 9) the feature data includes a feature amount indicating a feature of a label included in the dataset; 9. An information processing device according to any one of appendices 5 to 8.

[0143] (Appendix 10) The feature indicating the feature of the label includes an embedding vector of the label or an embedding vector of data associated with the label. 10. The information processing device according to claim 9.

[0144] (Appendix 11) An information processing system in which a first information processing device and a second information processing device participating in federated learning can communicate with each other via a network, The first information processing device a communication control unit that transmits client-related information based on information registered by the client to the second information processing device; The second information processing device an acquisition unit that acquires the client-related information from the first information processing device; an evaluation unit that evaluates the degree of similarity between the client-related information of the first information processing device and the client-related information of a partner candidate for federated learning; a screen output unit that outputs screen data for selecting a partner for the associative learning based on the similarity; An information processing system comprising:

[0145] (Appendix 12) A computer that can communicate with information processing devices participating in the federated learning via a network, a step of acquiring client-related information based on information registered by the client from the information processing devices participating in the federated learning; a step of evaluating the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a candidate partner of the federated learning; a step of outputting screen data for selecting a partner for federated learning based on the similarity; An information processing method that performs the above.

[0146] (Appendix 13) A computer that can communicate with information processing devices participating in federated learning via a network, a step of acquiring client-related information based on information registered by the client from the information processing devices participating in the federated learning; a step of evaluating the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a candidate partner of the federated learning; a step of outputting screen data for selecting a partner for federated learning based on the similarity; A program to execute.

[0147] Although the embodiments of the present invention have been described in detail above, the present invention is not limited to these embodiments, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0148] 10: Central device 15: Central storage device 20: Client device 25: Client storage device 30: Client device 35: Client storage device 101: Communication control unit 102: Memory control unit 103: Acquisition Department 104: Registration Department 105: Evaluation Department 106: Screen output section 107: Learning control unit 201: Communication control unit 202: Memory control unit 203: Display control unit 204: Setting section 205: Registration Department 206: Learning Department 207: Application section 1000: Federated Learning System [Prior art documents] [Patent documents]

[0149] [Patent Document 1] Japanese Patent Publication No. 2022-76277

Claims

1. An information processing device capable of communicating with information processing devices participating in federated learning via a network, an acquisition unit that acquires client-related information based on information registered by the client from the information processing devices participating in the federated learning; an evaluation unit that evaluates the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a partner candidate for the federated learning; a screen output unit that outputs screen data for selecting a partner for associative learning based on the similarity; An information processing device comprising:

2. the screen output unit outputs the screen data that displays the partner candidates for the federated learning whose similarity is equal to or greater than a threshold in a selectable manner. The information processing device according to claim 1 .

3. the screen output unit outputs the screen data that displays information about the partner candidate for the federated learning in association with the similarity. The information processing device according to claim 1 .

4. the screen output unit outputs the screen data that displays information about the partner candidates for the federated learning arranged in descending order of the similarity. The information processing device according to claim 3 .

5. the evaluation unit evaluates the degree of similarity between feature data relating to the datasets.

5. The information processing device according to claim 1.

6. the acquisition unit acquires one or more pieces of feature data selected by the information processing devices participating in the federated learning. The information processing device according to claim 5 .

7. the feature data includes a feature amount indicating a feature of the entire data set; The information processing device according to claim 5 .

8. the feature quantity indicating the characteristics of the entire dataset includes a feature quantity indicating the distribution of data in the dataset; The information processing device according to claim 7 .

9. the feature data includes a feature amount indicating a feature of a label included in the dataset; The information processing device according to claim 5 .

10. The feature indicating the feature of the label includes an embedding vector of the label or an embedding vector of data associated with the label. The information processing device according to claim 9 .

11. An information processing system in which a first information processing device and a second information processing device participating in federated learning can communicate with each other via a network, The first information processing device a communication control unit that transmits client-related information based on information registered by the client to the second information processing device; The second information processing device an acquisition unit that acquires the client-related information from the first information processing device; an evaluation unit that evaluates the degree of similarity between the client-related information of the first information processing device and the client-related information of a partner candidate for federated learning; a screen output unit that outputs screen data for selecting a partner for the associative learning based on the similarity; An information processing system comprising:

12. A computer that can communicate with information processing devices participating in the federated learning via a network, a step of acquiring client-related information based on information registered by the client from the information processing devices participating in the federated learning; a step of evaluating the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a candidate partner of the federated learning; a step of outputting screen data for selecting a partner for federated learning based on the similarity; An information processing method that performs the above.

13. A computer that can communicate with information processing devices participating in federated learning via a network, a step of acquiring client-related information based on information registered by the client from the information processing devices participating in the federated learning; a step of evaluating the degree of similarity between client-related information of the information processing device participating in the federated learning and client-related information of a candidate partner of the federated learning; a step of outputting screen data for selecting a partner for federated learning based on the similarity; A program to execute.

Images