Method and information processing device

Abstract

The present disclosure enables an external device to use a trained machine learning model. A first computer receives a first request for a first service including provision of a first machine learning model that has been trained by a plurality of devices or provision of an inference result by the first machine learning model, and provides the first service in response to the first request. The first computer or a second computer distributes an incentive for the first machine learning model by the first service to the plurality of devices that participated in learning of the first machine learning model.

Description

Method, and Information Processing Apparatus

[0001] The present disclosure relates to a communication network system.

[0002] In a 5G core network, it is disclosed that federated learning of a machine learning model is performed by a plurality of NWDAFs (for example, Patent Document 1).

[0003] 3GPP TS 23.288, “Architecture enhancements for 5G System (5GS) to support network data analytics services”, V19.0.0, September 2024, 6.2C Federated Learning among Multiple NWDAFs

[0004] An object of the present disclosure is to provide a method for making a learned machine learning model available to an external device and an information processing apparatus.

[0005] One aspect of the present disclosure includes a first computer receiving a first request for a first service including providing a first machine learning model learned by a plurality of devices or providing an inference result by the first machine learning model, the first computer providing the first service in response to the first request, and the first computer or a second computer distributing an incentive for the first machine learning model by the first service to the plurality of devices that participated in the learning of the first machine learning model.

[0006] Another aspect of the present disclosure is an information processing apparatus including a control unit that executes distributing an incentive for the first machine learning model by the first service to the plurality of devices that participated in the learning of the first machine learning model when a first service including providing a first machine learning model learned by a plurality of devices or providing an inference result by the first machine learning model is provided.

[0007] One aspect of the present disclosure is an information processing device comprising: a control unit that performs: transmitting a first request for a first service which includes providing a first machine learning model that has been trained by a plurality of devices or providing inference results by the first machine learning model; receiving the first service in response to the first request; and providing an incentive for the first machine learning model by the first service to the plurality of devices that have participated in training the first machine learning model.

[0008] According to this disclosure, trained machine learning models can be provided to external devices.

[0009] Figure 1 shows an example of a machine learning model provision service by the system. Figure 2 shows an example of a machine learning model provision service by the system. Figure 3 shows an example of a procedure between the system and an entity related to the ML model provision service. Figure 4 shows an example of the architecture of a fifth-generation mobile communication system. Figure 5 shows an example of the hardware configuration of an information processing device that can operate as each NF within 5GC. Figure 6 shows an example of the functional configuration of an NEF. Figure 7 shows an example of the functional configuration of an NWDAF. Figure 8 shows an example of the functional configuration of an ADRF. Figure 9 shows an example of ML model publication information held in the ML model publication information database. Figure 10 is an example of a processing sequence in the ML model publication setting procedure. Figure 11 is an example of a flowchart of the FL server processing in the ML model publication setting procedure. Figure 12 is an example of a processing sequence in the ML model publication information acquisition procedure. Figure 13 is a flowchart of the NEF processing according to the first embodiment. Figure 14 is an example of a flowchart of the NWDAF processing in the ML model publication information acquisition procedure. Figure 15 is an example of a sequence in the ML model provision procedure when the target is an ML model. Figure 16 is an example of a flowchart of the NWDAF processing in the ML model provision procedure when the target is an ML model. Figure 17 is an example of a sequence in the ML model provision procedure when the target is the prediction results of an ML model. Figure 18 is an example of a flowchart of the NWDAF processing in the ML model provision procedure when the target is the prediction results from an ML model. Figure 19 is an example of a flowchart of the UE processing in the ML model provision procedure when the target is the prediction results of an ML model. Figure 20 is an example of a sequence in the incentive distribution procedure. Figure 21 is an example of a flowchart of the FL server processing in the incentive distribution procedure. Figure 22 is a diagram showing an example of an ML model provision service.

[0010] In one aspect of this disclosure, when a machine learning model trained by multiple devices is provided outside the system that built the machine learning model, incentives are distributed to the multiple devices that trained the machine learning model. More specifically, one aspect of this disclosure is a method comprising: a first computer receiving a first request for a first service which includes providing a first machine learning model trained by multiple devices or providing prediction results by the first machine learning model; the first computer providing the first service in response to the first request; and the first or second computer distributing incentives for the first machine learning model from the first service to the multiple devices that participated in training the first machine learning model.

[0011] The first computer and the second computer may be computers included in the same system or computers included in different systems. The system including the first computer and the second computer may, for example, be a core network system of a 5G (5th Generation) or later mobile communication system. However, the system including at least one of the first computer and the second computer is not limited to a core network system of a 5G or later mobile communication system. The system including at least one of the first computer and the second computer may, for example, be a system for training machine learning models.

[0012] When the first and second computers are included in a 5G core network system, the first and second computers are computers that implement instances of Network Functions (NFs) that perform predetermined functions within the core network of a mobile communication system. These NF instances are realized, for example, by running virtualized computing such as containers on the computers. The first and second computers are, for example, Network Data Analytics Functions (NWDAFs), Network Exposure Functions (NEFs), or other NFs.

[0013] The multiple devices used to train the first machine learning model include, for example, servers and terminals. If the first computer and the second computer are included in the core network system of a 5G or later generation mobile communication system, the multiple devices used to train the first machine learning model include NWDAF, AF (Application Function), and UE (User Equipment). However, the multiple devices used to train the first machine learning model are not limited to these. UE includes, for example, in-vehicle devices mounted on mobile bodies such as vehicles, mobile terminals with wireless communication functions such as smartphones and tablet terminals, and stationary computers with wireless communication functions. UE may also include the vehicle itself equipped with wireless communication functions. In addition to vehicles, UE may also include mobile bodies such as ships, aircraft, and train cars, and devices mounted on these mobile bodies. Furthermore, the multiple devices used to train the first machine learning model may be included in the same system as the first computer and the second computer, or in different systems.

[0014] Methods for training machine learning models using multiple devices include, for example, horizontal and vertical federated learning (FL), distributed machine learning (DML), split learning (SL), and privacy-preserving machine learning (PPML). The first service is not limited to providing the first machine learning model itself or providing the prediction results from the first machine learning model. Incentives provided to the multiple devices that trained the first machine learning model include, for example, points, money, coupons, and priority.

[0015] According to one aspect of this disclosure, the first machine learning model or the prediction results obtained by the first machine learning model can be provided upon request. Furthermore, since incentives are distributed to multiple devices that have trained the first machine learning model, these multiple devices can offset the costs incurred in training the first machine learning model, or they can be provided with benefits exceeding those costs.

[0016] In one aspect of this disclosure, the first computer may receive, along with the first request, information indicating the subject matter of the provision relating to the first machine learning model. In this case, the first or second computer may determine an incentive for the first machine learning model depending on the subject matter of the provision relating to the first machine learning model. For example, the incentive may be higher when the first machine learning model itself is provided than when the prediction results from the first machine learning model are provided, or vice versa. For example, if a fee for the first service is charged to the user for the first machine learning model, the user can choose the subject matter of the provision from the first machine learning model itself and the prediction results, depending on the fee.

[0017] In one aspect of this disclosure, the first computer may accept the first request if the sender of the first request satisfies the first condition, and return a rejection response to the first request if the sender of the first request does not satisfy the first condition. The first condition is a condition relating to users who are permitted to receive the first service for the first machine learning model. For example, the first condition is that the user is a contractor who enters into a contract with the administrator of the first machine learning model regarding the provision of the first machine learning model, or a user under the control of said contractor. However, the first condition is not limited to this. This makes it possible to limit the recipients of the first service for the first machine learning model to devices that satisfy the first condition.

[0018] In one aspect of this disclosure, if the first computer receives information indicating that the subject of the provision relating to the first machine learning model is the first machine learning model itself, along with the first request, the first computer may transmit some or all of the data of the first machine learning model as information relating to the first machine learning model. Some or all of the data relating to the first machine learning model may be some or all of the data used to reproduce the first machine learning model, such as weights learned on training data, parameters, model structure, and preprocessing and / or postprocessing details. A user who receives the first machine learning model itself can, for example, use the first machine learning model to make inferences or use the first machine learning model for purposes other than its original purpose. Furthermore, the user can build a new machine learning model by further training the first machine learning model with training data they have prepared themselves (transfer learning), in which case the user can reduce time, resources, and costs compared to building a machine learning model from scratch.

[0019] In one aspect of this disclosure, if the first computer receives information along with the first request indicating that the subject of the provision relating to the first machine learning model is the prediction result by the first machine learning model, the first computer may transmit the prediction result by the first machine learning model as information relating to the first machine learning model. This minimizes the disclosure of the first machine learning model to external parties.

[0020] Furthermore, the first computer may accept a request for the first computer to specify a recipient for the prediction results of the first machine learning model, and may send the prediction results of the first machine learning model to the specified recipient in response to the first request. This makes it possible to provide the prediction results of the first machine learning model to parties other than the sender of the first request, within the scope specified by the sender of the first request.

[0021] In this case, the first computer may accept a request for the first computer to forward the prediction results of the first machine learning model from a designated recipient, and may send the prediction results of the first machine learning model and an instruction to forward the prediction results of the first machine learning model to the designated recipient in response to the first request. For example, if the designated recipient is a UE, the UE may forward the results to surrounding UEs via vehicle-to-vehicle communication, Wi-Fi communication, BLE (Bluetooth® Low Energy) communication, or 5G side-link communication, etc. This allows the prediction results of the first machine learning model to be distributed more widely.

[0022] Furthermore, the first computer may send the prediction results of the first machine learning model to a designated recipient if the designated recipient satisfies a second condition indicating which recipients are permitted to receive the prediction results of the first machine learning model, and may return a rejection response to the first request if the designated recipient does not satisfy the second condition. This makes it possible to limit the destinations to which the prediction results of the first machine learning model specified for the first request are forwarded to destinations that satisfy the second condition.

[0023] In one aspect of this disclosure, a first computer, a second computer, or a third computer may send a second request to a plurality of devices participating in the training of a first machine learning model, inquiring whether they are willing to provide a first service for the first machine learning model, and receive information from the plurality of devices as a response to the second request indicating permission or denial of provision of the first service for the first machine learning model. The first computer may accept the first request if permission to provide the first service for the first machine learning model has been obtained from all of the plurality of devices participating in the training of the first machine learning model, and may send a denial response to the first request if permission to provide the first service for the first machine learning model has not been obtained from at least one of the plurality of devices participating in the training of the first machine learning model. The third computer may be a computer that is part of the same system as either the first computer or the second computer, or a computer that is part of a different system from either of them.

[0024] Any of the first to fourth computers may determine that the first service for the first machine learning model cannot be provided if a refusal to provide the first service for the first machine learning model has been obtained from at least one of the multiple devices that participated in training the first machine learning model, and determine that the first service for the first machine learning model can be provided if permission to provide the first service for the first machine learning model has been obtained from all of the multiple devices that participated in training the first machine learning model. The fourth computer may be a computer that is part of the same system as any of the first to third computers, or a computer that is part of a different system. Based on the approval or disapproval of each of the multiple devices that participated in training the first machine learning model, a determination is made as to whether or not to accept the first request and whether or not to provide the first service for the first machine learning model. This makes it possible to provide the first service for the first machine learning model while taking into account the intentions of the administrators of each of the multiple devices that participated in training the first machine learning model.

[0025] In one aspect of this disclosure, a first computer may receive a third request from a third computer and a fifth computer requesting to obtain first information relating to a first machine learning model, which includes at least information indicating whether or not a first service relating to the first machine learning model is available, and transmit the first information as a response to the third request. The fifth computer may be a computer that is part of the same system as any of the first to fourth computers, or a computer that is part of a different system. This makes it possible to provide the sender of the third request with information indicating whether or not a first service relating to the first machine learning model is available.

[0026] This disclosure can also be identified as, in one other aspect, a system comprising at least one of the first to fifth computers that perform the above method. Furthermore, this disclosure can also be identified as, in one other aspect, an information processing device corresponding to one of the first to fifth computers that perform the above method. Furthermore, this disclosure can also be identified as, in one other aspect, a program for causing a computer to perform the processing of one of the first to fifth computers that perform the above method, and a non-temporary, computer-readable recording medium on which said program is recorded.

[0027] Another aspect of the present disclosure is an information processing device comprising a control unit that, when a first service is provided, including the provision of a first machine learning model trained by a plurality of devices or the provision of inference results by the first machine learning model, distributes an incentive for the first machine learning model provided by the first service to the plurality of devices that participated in training the first machine learning model. The information processing device may, for example, be a computer implementing an instance of an NF that performs a predetermined function within a core network system of a mobile communication system of a 5G or later generation. However, it is not limited to this, and the information processing device may, for example, be a computer such as a server included in a system that trains a machine learning model. The control unit included in the information processing device may be, for example, a processor such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The present disclosure can also be identified as another aspect of the present disclosure as a method by which a computer performs the processing performed by the above information processing device. The present disclosure can also be identified as another aspect of the present disclosure as a program for causing a computer to perform the processing of the above information processing device, and a non-temporary, computer-readable recording medium on which the program is recorded.

[0028] Another aspect of the present disclosure is an information processing device comprising: a control unit that performs the following actions: sending a first request for a first service which includes providing a first machine learning model that has been trained by a plurality of devices or providing inference results by the first machine learning model; receiving the first service in response to the first request; and providing incentives to the first machine learning model by the first service to the plurality of devices that have participated in training the first machine learning model. The information processing device may be, for example, a device within a system that manages the first machine learning model, or a device outside of such a system. The information processing device may be a server or a terminal. Another aspect of the present disclosure may be identified as a method by which a computer performs the processing performed by the information processing device. Another aspect of the present disclosure may be identified as a program for causing a computer to perform the processing performed by the information processing device, and a non-temporary, computer-readable recording medium on which the program is recorded.

[0029] Embodiments of this disclosure will be described below with reference to the drawings. The configurations of the following embodiments are illustrative, and this disclosure is not limited to the configurations of these embodiments.

[0030] <First Embodiment> Figure 1 shows an example of a machine learning model provision service by system 100. In the first embodiment, system 100 is a 5G core system. In the first embodiment, system 100 holds a machine learning model that has undergone federated learning by multiple devices under its management. Hereinafter, the machine learning model will simply be referred to as the ML model. The federated learning performed in system 100 may be either horizontal or vertical.

[0031] Furthermore, the device that controls the federated learning of the ML model within system 100 is referred to as the FL server. Devices that participate in the federated learning of the ML model are referred to as FL clients. The FL server and FL clients may be devices under the management of system 100 or devices outside the management of system 100. In the first embodiment, the FL server is assumed to be the NWDAF within system 100. Also in the first embodiment, the FL clients are the NF (Network Function) within system 100, the UE which is a subscriber to system 100, and external computers etc. which have a contract with system 100. Note that, outside of the 5G core system, the FL server may be referred to as the coordinator and the FL client as the participant.

[0032] In the first embodiment, system 100 is a system that provides an ML model provision service, which provides an ML model in response to a request. The provision includes two things: the ML model itself and the prediction results from the ML model. Figure 1 shows an example of providing the ML model itself. The ML model provision service is an example of the "first service".

[0033] System 100 receives an ML model provision request from Entity 500. Entity 500 may be, for example, a device under the management of System 100, or a device outside the management of System 100. More specifically, Entity 500 may be an NF within System 100, an AF (Application Function) trusted by System 100, another 5G core system, or a UE (User Equivalent) joining System 100. UEs include, for example, smartphones, tablet devices, in-vehicle devices, vehicles with communication capabilities, wearable devices such as smartwatches, and PCs. However, UEs are not limited to these. Entity 500 may also be a device that participated in learning the ML model, i.e., an FL client. Entity 500 also transmits information indicating the subject of the provision along with the ML model provision request. In the example shown in Figure 1, the information indicating the subject of the provision indicates the ML model itself.

[0034] When system 100 receives an ML model provision request that targets the ML model itself, it sends some or all of the ML model data to entity 500. Some of the ML model data is, for example, some or all of the data used to reproduce the ML model, such as weights learned from training data, parameters, model structure, and preprocessing and / or postprocessing details. Whether some or all of the ML model data is provided may be specified, for example, in the ML model provision request, or it may be set by the administrator of system 100 depending on the type or use of the ML model.

[0035] If entity 500 receives a portion of the ML model data from system 100, it will reconstruct the ML model from that portion of data. Entity 500 may then use the ML model to perform inferences in its own system, or it may further train the ML model with new training data for another task (transfer learning).

[0036] Meanwhile, system 100 distributes incentives for the target ML model provided by the ML model provision service to FL clients who participated in learning the target ML model. Incentives may include, for example, points, money, coupons, and priority. Incentives may also be distributed to the FL server.

[0037] Figure 2 shows an example of a machine learning model provision service provided by System 100. Figure 2 shows an example of providing prediction results using an ML model. In the example shown in Figure 2, Entity 500 sends information indicating the prediction results as information indicating the target of the provision, along with an ML model provision request. Furthermore, if the target of the provision is prediction results, Entity 500 can specify the recipient of the prediction results from the ML model. The specification of the recipient of the prediction results from the ML model is sent to System 100 along with the ML model provision request.

[0038] When system 100 receives an ML model provision request that specifies the provision of ML model prediction results, it sends the ML model prediction results to the specified recipient. For example, if no recipient for the prediction results is specified, the prediction results may be sent to the source of the ML model provision request. Recipients of the ML model prediction results can include, for example, a server in another system, a UE subscribed to another mobile communication system, another 5G core system, and entity 500 itself, the source of the ML model provision request.

[0039] Furthermore, even when the provided item is the prediction result of an ML model, an incentive is provided from system 100 to the FL client that participated in training the ML model. The incentive when the provided item is the prediction result of an ML model may be less than or the same as when the provided item is the ML model itself.

[0040] Figure 3 is a diagram illustrating an example of the procedure between System 100 and Entity 500 related to the ML model provision service. Hereafter, when simply referred to as "service," it will refer to the ML model provision service. In S1, the ML model publication setting procedure is executed, and publication information regarding the ML model is set. The publication information regarding the ML model will be hereinafter referred to as ML model publication information. The ML model publication information includes, for example, information such as the intended use of the ML model, whether or not the ML model can be provided to the ML model provision service, the target users who are permitted to provide the ML model to the ML model provision service for the ML model, and the scope of users who are permitted to enjoy the ML model provision service for the ML model.

[0041] In S2, the procedure for obtaining publicly available ML model information is executed, and entity 500 obtains publicly available ML model information of the ML model held by system 100 from system 100. In S3, the procedure for providing ML models is executed, and entity 500 is provided with the ML model or the prediction results based on the ML model. In S4, the procedure for distributing incentives is executed, and incentives are distributed in system 100 to the FL clients that participated in learning the ML model. Details of each procedure from S1 to S4 will be described later.

[0042] Figure 4 shows an example of the architecture of a fifth-generation mobile communication system. The fifth-generation mobile communication network will be referred to as the 5G network below. The 5G network has a 5G core network (5GC) and an access network ((RAN)). User Equipment (UE) 50, Data Network (DN), and Application Function (AF) are connected to the 5G network. The UE 50 is the user's (subscriber's) terminal. The RAN (Radio Access Network) is the access network to the 5GC. The RAN includes a base station (gNB) as one of its nodes. Hereafter, when simply referred to as RAN, it refers to a node within the RAN.

[0043] Figure 4 shows some of the components included in 5GC. Also, in Figure 4, components according to the first embodiment are denoted by reference numerals. In 5G, the software that implements network functions and the hardware on which that software is executed are separated using hardware abstraction technology. This allows various network function software to operate on common hardware resources, regardless of the configuration of each hardware product. Figure 4 shows the network functions (NFs) included in 5GC. Each of the multiple NFs included in 5GC is implemented by one or more computers (information processing devices) executing a program. However, a single computer may implement any two or more NFs.

[0044] The UPF (User Plane Function) performs routing, forwarding, packet inspection, and QoS processing of user packets. User packets are packets in the user plane that the UE transmits and receives.

[0045] The AMF (Access and Mobility Management Function) accommodates the RAN and performs registration management, connection management, and mobility management of the UE in the 5GC. In addition, the AMF also relays messages between the SMF and the UE.

[0046] The SMF (Session Management Function) manages the PDU (Protocol Data Unit) session, assigns and manages IP addresses to the UE, and selects and controls the UPF. The PDU session is a virtual communication path for data exchange between the UE and the DN. The DN is an external data network (cloud, Internet, etc.) of the 5GC.

[0047] The PCF (Policy Control Function) manages, for example, access and mobility management policies, session management policies, and charging policies, and provides policy-related information to the AMF and the SMF. Policies related to access and mobility include, for example, routing, etc. Policies related to session management include QoS, filtering, etc.

[0048] UDM (Unified Data Management) manages subscriber information, authentication information, etc. regarding subscribers who subscribe to the operator. Subscriber information includes, for example, access and mobility subscription data used for UE registration and mobility management, slice selection subscription data used for network slice selection, SMF selection subscription data used for SMF selection, and session management subscription data used for PDU session establishment. A network slice is a virtual network with specifications according to the application.

[0049] UDR (Unified Data Repository) stores data used by UDM, PCF, and NEF, and provides search for these data. More specifically, the data held by UDR includes, for example, subscriber information, authentication information, and policy data, etc.

[0050] NEF 1 provides a function to expose the services of each NF. More specifically, NEF1 provides a function to securely expose the capabilities and event information disclosed by network functions in the 5G system to external applications such as AF. Also, NEF 1 provides a function to receive information from permitted external applications into the network. That is, NEF 1 plays a role as an interface with the outside. AF is an application server (external server) that provides auxiliary services other than 5GC specifications.

[0051] The Network Repository Function (NRF) 2 provides the function of registering services for each Network Facility (NF) within the 5GC (e.g., AMF, SMF, UPF, etc.). NRF 2 stores and manages information about each NF in which a service is registered. In response to an inquiry regarding a desired service, NRF 2 can return to the inquirer a list of candidate NFs from among the registered NFs that provide the said service.

[0052] NWDAF 3 provides, for example, analytical and statistical information within the network, and analytical and predictive information regarding the movement of UEs. NWDAF 3 also has functions such as training ML models and providing trained ML models or prediction results from ML models. In the first embodiment, NWDAF 3 performs federated learning of ML models.

[0053] ADRF (Analytics Data Repository Function) 4 provides the function of collecting and storing data from each NF, and the function of holding ML models learned by NWDAF 3. ADRF 4 performs registration, reading, updating, and deletion of ML models in response to access from NWDAF 3. CHF (Charging Function) 5 provides billing functionality in the 5G network.

[0054] In 5GC, multiple NFs of the same type may exist. For example, a certain type of NF may be provided for each data center. Alternatively, one NF may be shared among data centers. The correspondence between NFs and data centers can be set as appropriate.

[0055] Figure 5 shows an example of the hardware configuration of an information processing device that can operate as an NF within 5GC. The information processing device 1000 can be configured using an information processing device (computer) such as a personal computer (PC), workstation (WS), or server machine. The information processing device 1000 may also be a collection of one or more computers (cloud). In addition, an NF within the 5G core network may be a device equipped with an electrical circuit such as a dedicated FPGA (Field-Programmable Gate Array) or ASIC (Application Specific Integrated Circuit) to execute the relevant processing.

[0056] The information processing device 1000 comprises a CPU 1001, a main memory 1002, an auxiliary storage device 1003, and a communication device 1004 as its hardware configuration. The main memory 1002 and the auxiliary storage device 1003 are recording media readable by a computer. The CPU 1001, the auxiliary storage device 1003, and the communication device 1004 are electrically connected by a bus.

[0057] The auxiliary storage device 1003 stores programs used to operate as one of the NFs within the 5G core network, and data used by the CPU 1001 when executing each program. The auxiliary storage device 1003 is, for example, an EPROM (Erasable Programmable ROM), a hard disk drive, or an SSD (Solid State Drive). Programs held in the auxiliary storage device 1003 include, for example, an operating system (OS) and the control program for the corresponding NF.

[0058] The main memory 1002 is a memory device that provides the CPU 1001 with a memory area and a work area for loading programs stored in the auxiliary memory device 1003, and is also used as a buffer. The main memory 1002 includes, for example, semiconductor memory such as ROM (Read Only Memory) and RAM (Random Access Memory).

[0059] The CPU 1001 executes the processing corresponding to each NF in the 5G core network by loading the OS held in the auxiliary storage device 1003 and a program related to one of the NFs in the 5G core network into the main memory device 1002 and executing them. The CPU 1001 is an example of a processor. A GPU or a DSP (Digital Signal Processor) may also be provided as the processor. There is not limited to one CPU 1001, but there may be multiple CPUs 1001.

[0060] The communication device 1004 is, for example, a NIC (Network Interface Card), an optical line interface, etc. The communication device 1004 may also be, for example, a wireless communication circuit connected to a wireless network such as a wireless LAN. The hardware configuration of the information processing device 1000 that realizes the functions of each NF within the 5G core network is not limited to that shown in Figure 5.

[0061] Furthermore, entities 500 such as AF, NF, and UE also have a hardware configuration similar to that of the information processing device 1000, including a CPU, main memory, auxiliary memory, and communication device. When entity 500 is a UE, the communication device is a device that performs wireless communication compatible with the 5G mobile communication system.

[0062] Figure 6 is a diagram showing an example of the functional configuration of NEF 1. In Figure 6, the functional components of the first embodiment are extracted and shown from the functional configuration of NEF 1. The same applies to the functional configuration of each subsequent NF. NEF 1 includes a control unit 11 as a functional component. The processing performed by the control unit 11 is achieved by the CPU 1001 of the information processing device 1000 operating as NEF 1 executing a predetermined program.

[0063] In the first embodiment, the control unit 11 performs processing as an interface with the entity 500 related to the ML model provision service. The control unit 11 also performs processing as a consumer of the Nnwdaf_AnalyticsSubscription service, Nnwdaf_AnalyticsInfo service, Nnwdaf_MLModelProvision service, and Nnwdaf_MLModelInfo service. More specifically, when the control unit 11 receives a request from the entity 500 related to the ML model provision service, it forwards the request to the NWDAF 3 corresponding to the request. Furthermore, when the control unit 11 receives a response from the NWDAF 3, the destination of the forwarded request, it forwards the response to the entity 500. Details of the processing of the control unit 11 will be described later. Note that the functional configuration of NEF 1 is not limited to the example shown in Figure 6.

[0064] Figure 7 shows an example of the functional configuration of NWDAF 3. NWDAF 3 comprises an analysis control unit 31, an ML model control unit 32, and an ML model learning information storage unit 33. The processing performed by these functional components is achieved by the CPU 1001 of the information processing device 1000, which operates as NWDAF 3, executing a predetermined program.

[0065] The analysis control unit 31 corresponds to, for example, an AnLF (Analytics Logical Function). The analysis control unit 31 performs processing related to, for example, the Nnwdaf_AnalyticsSubscription service and the Nnwdaf_AnalyticsInfo service. More specifically, the analysis control unit 31 performs statistical analysis and prediction of 5G network information to acquire and publish analytical information. In the first embodiment, in the ML model provision procedure (S3 in Figure 2), the analysis control unit 31 performs estimation using the ML model in response to an ML model provision request for which the estimation result of the ML model from entity 500 is to be provided, and transmits the estimation result to entity 500 via NEF 1. The analysis control unit 31 may acquire the ML model to be requested from ADRF 4 via ML model control unit 32.

[0066] The ML model control unit 32 corresponds to, for example, the MTLF (Model Training Logical Function). The ML model control unit 32 performs processing related to, for example, the Nnwdaf_MLModelProvision service, the Nnwdaf_MLModelInfo service, the Nnwdaf_MLModelTraining service, and the Nnwdaf_MLModelTraining service. The ML model control unit 32 also performs processing as a consumer of, for example, the Nadrf_MLModelManagement service. More specifically, the ML model control unit 32 trains the ML model and makes the ML model public. The ML model control unit 32 operates as either an FL server or an FL client in federated learning. Whether it operates as an FL server or an FL client is, for example, specified by the administrator of system 100 for each ML model, or determined in the discovery procedure for finding an FL client. The ML model control unit 32 is also an interface for accessing ADRF 4. The ML model control unit 32 stores trained ML models in ADRF 4, updates trained models stored in ADRF 4, and retrieves ML models from ADRF 4.

[0067] In the first embodiment, when the ML model control unit 32 operates as an FL server or an FL client, it performs the ML model publication setting procedure (S1 in Figure 2). In the ML model publication setting procedure, if the ML model control unit 32 is an FL client, it responds, for example, regarding the ML model in question, whether or not it can be provided to the ML model provision service, who is permitted to receive the provision, and the scope of users who can use it. If the ML model control unit 32 is an FL server, it compiles the responses from the FL clients, generates ML model publication information for the ML model, and registers it with ADRF 4.

[0068] In the first embodiment, in the ML model public information acquisition procedure (S2 in Figure 2), when the ML model control unit 32 receives an ML model public information request from the NEF 1, it accesses the ADRF 4, acquires the ML model public information of the corresponding ML model, and transmits it to the NEF 1. In the first embodiment, in the ML model provision procedure (S3 in Figure 2), in response to an ML model provision request for which the ML model itself is to be provided, the ML model control unit 32 acquires all or part of the data of the corresponding ML model from the ADRF 4 and transmits it to the entity 500 via the NEF 1. In the first embodiment, in the incentive distribution procedure (S4 in Figure 2), the ML model control unit 32 decides to distribute the incentives generated by the ML model provision service for an ML model that it has trained as an FL server to the FL clients that participated in training the ML model. The recipients of the incentives may also include FL servers.

[0069] The ML model learning information storage unit 33 is created, for example, in the storage area of ​​the auxiliary storage device 1003 of the information processing device 1000 operating as NWDAF 3. The ML model learning information storage unit 33 holds ML model learning information about ML models in which the ML model control unit 32 participated in learning as an FL server or FL client. More specifically, the ML model learning information includes identification information of the ML model, information indicating whether it is an FL server or an FL client, and identification information of the NWDAF 3 which is an FL client if it is an FL server, or identification information of the NWDAF 3 which is an FL server if it is an FL client. However, the information included in the ML model learning information held in the ML model learning information storage unit 33 is not limited to these. Note that the functional configuration of NWDAF 3 is not limited to the example shown in Figure 7.

[0070] Figure 8 shows an example of the functional configuration of ADRF 4. ADRF 4 comprises a control unit 41, an ML model public information DB 42, and an ML model DB 43. The processing performed by these functional components is achieved by the CPU 1001 of the information processing device 1000, which operates as ADRF 4, executing a predetermined program.

[0071] The control unit 41 controls access to the ML model public information DB 42 and the ML model DB 43. The control unit 41 performs processing related to the Nadrf_MLModelManagement service, for example. More specifically, the control unit 41 performs registration, update, deletion, and retrieval processing of information in the ML model public information DB 42 and the ML model DB 43 in response to requests.

[0072] The ML Model Public Information DB 42 and the ML Model DB 43 are created in the storage area of ​​the auxiliary storage device 1003 of the information processing device 1000, which operates as ADRF 4. The ML Model Public Information DB 42 holds the ML Model Public Information. The ML Model Public Information is information about ML models that is permitted to be made public outside of the system 100. Details of the ML Model Public Information will be described later.

[0073] ML Model DB 43 stores ML models. ML models are stored in file format, for example, as ML model files. ML model files contain information such as the architecture of the ML model, trained parameters, and optimization algorithms. The architecture of the ML model, for example in the case of a neural network, includes information such as the number and type of layers, the number of units in each layer, and the type of activation function. Trained parameters are parameters that the ML model has adjusted based on the training data, such as weights and biases. Trained parameters include, for example, the weights corresponding to each connection in the neural network. Optimization algorithm information includes, for example, the algorithm used during training, and settings for hyperparameters such as the learning rate and batch size. Note that the functional configuration of ADRF 4 is not limited to the example shown in Figure 8.

[0074] Figure 9 shows an example of ML model public information held in the ML model public information DB 42. The ML model public information includes, for example, fields for ML model identification information (ID), purpose, availability, permitted recipients, and scope of users. The purpose field stores information indicating the purpose of the ML model. The availability field stores information indicating permission or denial for the provision of the ML model to the ML model provision service.

[0075] The field for which provision is permitted stores information indicating the subjects to whom provision of the ML model is permitted, provided that the field for permission to provide is stored with permission. In the first embodiment, the information indicating the subjects to whom provision is permitted includes information indicating the ML model and information indicating the prediction results. If the field for permission to provide is stored with rejection, the field for which provision is permitted will be empty. If the field for permission to provide is stored with permission and the field for which provision is permitted is empty, for example, the subjects to whom provision is permitted may include both the ML model and the prediction results. However, this is not limited to this.

[0076] The "Scope of Available Users" field stores information indicating users who are permitted to receive the ML model or the prediction results generated by the ML model, provided that the "Availability" field contains information indicating permission. Users permitted to receive the ML model include, for example, users subscribed to system 100, groups of MNOs with capital ties to system 100, and systems that have contracts with system 100 for the use of ML model provision services. For example, as shown in Figure 9, if the "Scope of Available Users" field contains information indicating "Contracting MNO," it indicates that ML model provision requests from MNOs (Mobile Network Operators) that have usage contracts with system 100 for ML model provision services are permitted. If the "Availability" field contains information indicating denial, the "Scope of Available Users" field will be empty. If the "Availability" field contains information indicating permission and the "Scope of Available Users" field is empty, it may indicate, for example, that all users are permitted. The recipients of the prediction results generated by the ML model are also called the recipients of the prediction results, although this is not limited to this definition. Furthermore, the information included in the ML model public information held in ADRF 4's ML model public information DB 42 is not limited to the information shown in Figure 9.

[0077] (Details of the ML Model Publication Setup Procedure) Figure 10 shows an example of the processing sequence in the ML model publication setup procedure. The ML model publication setup procedure is performed by NWDAF 3 acting as an FL server, NWDAF 3 acting as an FL client, and ADRF 4. Hereinafter, the NWDAF acting as an FL server will be simply referred to as the FL server. The NWDAF acting as an FL client will be simply referred to as the FL client. In Figure 10, the ML model that FL server 3A, FL client 3B, and FL client 3B have participated in learning will be referred to as the target ML model. In Figure 10, it is assumed that the learning of the target ML model has been completed by FL server 3A, FL client 3B, and FL client 3B. The ML model publication setup procedure is started, for example, after the discovery procedure for finding FL clients is completed, or after the learning of the target ML model is completed.

[0078] In S101, the FL server 3A sends a publishing configuration request to the FL clients 3B and 3C. The publishing configuration request is a request to inquire about the information to be published for the target ML model. Along with the publishing configuration request, for example, identification information of the target ML model is also sent.

[0079] In S102, FL clients 3B and 3C each send a response to the public access request. Along with the response to the public access request, information is also sent indicating whether the ML model in question is permitted or denied for service provision, information indicating who is permitted to provide the service, and information indicating the scope of users who can access the service. This information is, for example, pre-configured by the administrator of each FL client.

[0080] In S103, FL Server 3A creates ML model publication information for the target ML model based on the responses to the publication setting requests received from FL Clients 3B and 3C, and the information set in FL Server 3A itself. Similar to FL Clients 3B and 3C, FL Server 3A is also pre-configured with information indicating permission or denial of provision of the target ML model to the service, information indicating those who are permitted to provide the service, and information indicating the scope of those who can use the service. The information indicating those who are permitted to provide the service and the information indicating the scope of those who can use the service are optional and may not be sent from each FL Client. For example, if even one of FL Server 3A, FL Clients 3B and 3C refuses to provide the target ML model to the service, FL Server 3A will determine that provision of the target ML model to the service is "denied".

[0081] In S104, FL Server 3A sends a public information registration request to ADRF 4. The public information registration request is a request to register the public information of the ML model. Along with the public information registration request, the public information of the target ML model is also sent. For example, the Nadrf_MLModelManagement_StorageRequest request message may be used for the public information registration request. However, it is not limited to this.

[0082] In S105, ADRF 4 receives a public information registration request and the ML model public information for the target ML model from FL server 3A, registers the ML model public information in ML model public information DB 42, and sends a response to FL server 3A. Note that if the administrator of the FL server and all FL clients are the same, the ML model public information for the target ML model may be registered in ADRF 4 by that administrator, and the ML model public setting procedure shown in Figure 10 may not be performed. FL server 3A in Figure 10 is an example of a "third computer". The public setting request is an example of a "second request".

[0083] Figure 11 is an example of a flowchart of the FL server's processing during the ML model publishing setup procedure. The processing shown in Figure 11 is initiated when a predetermined event occurs, such as the completion of the discovery procedure or the completion of training for the target ML model. The entity executing the processing shown in Figure 11 is the CPU 1001 of the information processing device 1000, which operates as an FL server, but for convenience, the explanation will focus on the functional components. The same applies to the flowcharts from Figure 11 onward.

[0084] In OP101, the ML model control unit 32 of NWDAF 3, which operates as an FL server, sends a public setting request to all FL clients that participated in the training of the target ML model. The processing of OP101 corresponds to the processing of S101 in Figure 10. In the following description of Figure 11, the ML model control unit 32 refers to the ML model control unit 32 of NWDAF 3, which operates as an FL server.

[0085] In OP102, the ML model control unit 32 determines whether or not it has received a response from all FL clients that participated in training the target ML model. The ML model control unit 32 remains in a standby state until it receives a response from all FL clients that participated in training the target ML model (OP102: NO). If it receives a response from all FL clients that participated in training the target ML model (OP102: YES), the process proceeds to OP103. The process in OP102 corresponds to the process in S102 of Figure 10.

[0086] In OP103, the ML model control unit 32 determines whether there are any devices among the FL server and all FL clients that refuse to provide the target ML model to the service. If there are devices that refuse to provide the target ML model to the service (OP103: YES), the process proceeds to OP104. In OP104, the ML model control unit 32 generates ML model publication information for the target ML model, which includes information indicating the refusal to provide the ML model to the service.

[0087] If there are no devices that refuse to provide the target ML model to the service (OP103: NO), the process proceeds to OP105. In OP105, the ML model control unit 32 determines whether there are any devices among the FL server and all FL clients for which the permitted provision is only for prediction results. If there are devices for which the permitted provision is only for prediction results (OP105: YES), the process proceeds to OP106. In OP106, the ML model control unit 32 generates ML model publication information for the target ML model, including information indicating permission to provide the ML model to the service and that the permitted provision is for prediction results. The ML model control unit 32 also identifies overlapping ranges in the range of users specified for the FL server and all FL clients, and sets the overlapping range as the range of users for the target ML model. However, the method of setting the range of users for the target ML model is not limited to this. For example, the ML model control unit 32 may define the range of available users as the range specified by more devices than the range specified by the FL server and all FL clients, or it may define the range of available users as all of the ranges specified by the FL server and all FL clients.

[0088] If there are no devices for which the authorized provision target is only the prediction result (OP105: NO), the process proceeds to OP107. In OP107, the ML model control unit 32 generates ML model disclosure information for the target ML model, which includes information indicating that the provision of the ML model to the service is authorized and that the authorized provision target is the ML model and the prediction result. The method for setting the scope of users is the same as in OP106. The processing from OP103 to OP107 corresponds to the processing in S103 of Figure 10.

[0089] In OP108, the ML model control unit 32 sends a public information registration request and the ML model public information of the target ML model to ADRF 4. The processing in OP108 corresponds to the processing in S104 of Figure 10. In OP109, the ML model control unit 32 receives a public information registration response from ADRF 4. The processing in OP109 corresponds to the processing in S105 of Figure 10. After that, the processing shown in Figure 11 is completed.

[0090] In the process shown in Figure 11, the ML model control unit 32 assumes that devices whose permitted provision targets are only estimation results intend to minimize disclosure regarding the ML model, and if there is at least one device whose permitted provision targets are only estimation results, it sets the permitted provision target of the ML model to "estimation results". On the other hand, for devices whose permitted provision targets include ML models, the ML model control unit 32 assumes that there is no intention to restrict disclosure regarding the ML model, and if the permitted provision targets for the FL server and all FL clients include ML models, it sets the permitted provision targets of the ML model to "ML model" and "estimation results". However, the method of setting the permitted provision targets of the ML model is not limited to this. For example, if the permitted provision targets for the FL server and all FL clients are only ML models, the ML model control unit 32 may set the permitted provision target of the ML model to "ML model".

[0091] (ML Model Public Information Acquisition Procedure) Figure 12 shows an example of the processing sequence in the ML model public information acquisition procedure. The ML model public information acquisition procedure is performed by entity 500, NEF 1, NWDAF 3, and ADRF 4.

[0092] In S201, entity 500 sends an ML model publication information request to NEF 1. The ML model publication information request is a request for ML model publication information of an ML model held by system 100. In S202, NEF 1 forwards the ML model publication information request received from entity 500 to NWDAF 3. NEF 1 may, for example, use an Nnwdaf_MLModelInfo request to forward the ML model publication information request to NWDAF 3. In this case, the Nnwdaf_MLModelInfo request may include information indicating that it is an ML model publication information request. In S202, NEF 1 may also query NRF 2 to identify NWDAF 3.

[0093] In S203, the NWDAF 3 (ML Model Control Unit 32) receives an ML Model Publication Information Request from the NEF 1 and sends a request to the ADRF 4, for example, a Nadrf_MLModelManagement_RetrievalRequest request, in order to obtain the ML Model Publication Information. The Nadrf_MLModelManagement_RetrievalRequest request may include information indicating that it is an ML Model Publication Information Request.

[0094] In S204, ADRF 4 receives a Nadrf_MLModelManagement_RetrievalRequest request from NWDAF 3, reads the ML model publication information from the ML model publication information DB 42, and sends it to NWDAF 3 along with the Nadrf_MLModelManagement_RetrievalRequest response.

[0095] In S205, NWDAF 3 sends the response received from ADRF 4 and the ML model publication information to NEF 1. For example, if an ML model publication information request is received in S202 using the Nnwdaf_MLModelInfo request, NWDAF 3 sends the ML model publication information response using the Nnwdaf_MLModelInfo response.

[0096] In S206, NEF 1 forwards the response received from NWDAF 3 and the ML model publication information to entity 500. Note that the messages used in the ML model publication information request and response in processing S202 and S205 are examples only and are not limited thereto. NEF 1 or NWDAF 3 in Figure 12 is an example of the "fifth computer". The ML model publication information request is an example of the "third request". The ML model publication information is an example of the "first information".

[0097] Figure 13 is a flowchart of the processing of NEF 1 according to the first embodiment. The processing shown in Figure 13 is repeatedly executed at a predetermined cycle while NEF 1 is running, for example.

[0098] In OP201, the control unit 11 determines whether or not it has received a request from entity 500 related to the ML model provision service. Requests related to the ML model provision service include the ML model publication information request in the ML model publication information acquisition procedure and the ML model provision request in the ML model provision procedure described later. If a request related to the ML model provision service is received from entity 500 (OP201: YES), the process proceeds to OP202. If a request related to the ML model provision service is not received from entity 500 (OP201: NO), the process shown in Figure 13 ends. A process that results in a positive determination in OP201 corresponds, for example, to receiving an ML model publication information request from entity 500 in S201 of Figure 12.

[0099] In OP202, the control unit 11 forwards the request for the ML model provision service received in OP201 to NWDAF 3. In OP202, the control unit 11 may convert the request for the ML model provision service into the format of another request, depending on the request. The processing in OP202 corresponds, for example, to forwarding the ML model publication information request in S202 of Figure 12.

[0100] In OP203, the control unit 11 determines whether or not it has received a response from NWDAF 3 to the request sent in OP202. Until a response is received from NWDAF 3 (OP203: NO), NWDAF 3 remains in a waiting state. If a response is received from NWDAF 3 (OP203: YES), the process proceeds to OP204. The process in OP203 corresponds, for example, to receiving a response from NWDAF 3 in S205 of Figure 12.

[0101] In OP204, the control unit 11 sends the response received from NWDAF 3 to the appropriate destination as a response to the request received in OP201. If the control unit 11 has not received a destination specification along with the response received from NWDAF 3, it sends the response to entity 500, the source of the request received in OP201. If the control unit 11 has received a destination specification along with the response received from NWDAF 3, it sends the received response to the specified destination. The processing in OP203 corresponds, for example, to sending a response to entity 500 in S206 of Figure 12. After that, the processing shown in Figure 13 is completed.

[0102] Figure 14 is an example of a flowchart of the NWDAF 3 processing in the ML model public information acquisition procedure. The processing shown in Figure 14 is executed repeatedly, for example, at a predetermined interval.

[0103] In OP251, the ML model control unit 32 determines whether or not it has received an ML model publication information request from NEF 1. If an ML model publication information request has been received from NEF 1 (OP251: YES), the process proceeds to OP252. If an ML model publication information request has not been received from NEF 1 (OP251: NO), the process shown in Figure 14 ends. The process that results in a positive determination in OP251 corresponds, for example, to receiving an ML model publication information request from NEF 1 in S202 of Figure 12.

[0104] In OP252, the ML model control unit 32 sends a Nadrf_MLModelManagement_RetrievalRequest request to the ADRF 4 to obtain ML model publication information. The Nadrf_MLModelManagement_RetrievalRequest request may include information indicating that it is a request for ML model publication information. The processing in OP252 corresponds, for example, to the processing in S203 in Figure 12.

[0105] In OP253, the ML model control unit 32 determines whether or not it has received a response from ADRF 4. Until a response is received from ADRF 4 (OP253: NO), the ML model control unit 32 remains in a standby state. If a response is received from ADRF 4 (OP253: YES), the process proceeds to OP254. Along with the response, ADRF 4 also receives ML model publication information for the ML model held in the system 100. The process that results in a positive determination in OP253 corresponds, for example, to receiving a response from ADRF 4 in S204 of Figure 12.

[0106] In OP254, the ML model control unit 32 sends a response to NEF 1 for the ML model publication information request received in OP251. Along with the response, the ML model publication information of the ML model held in system 100 is also sent. After that, the process shown in Figure 14 is completed. The process in OP254 corresponds, for example, to the process in S205 of Figure 12.

[0107] In Figures 12 to 14, the procedure for obtaining ML model publication information was described as sending ML model publication information for all ML models held by system 100 to entity 500, but this is not limited to this. For example, if system 100 recognizes the identification information of the target ML model for which ML model publication information is desired, the identification information of that target ML model may also be sent along with the ML model publication information request, and only the ML model publication information for that target ML model may be sent to entity 500.

[0108] (ML model provision procedure when providing ML models) Figure 15 shows an example of the sequence in the ML model provision procedure when providing ML models. The ML model provision procedure when providing ML models is carried out by entity 500, NEF 1, NWDAF 3, and ADRF 4.

[0109] In S301, entity 500 sends an ML model provision request to NEF 1. An ML model provision request is a request for the provision of an ML model or prediction results. Along with the ML model provision request, the identification information of the target ML model, information indicating that the item to be provided is an ML model (as shown in Figure 15), and the user ID of the requester are also sent.

[0110] In S302, NEF 1 forwards the ML model provision request received from entity 500 to NWDAF 3. NEF 1 may, for example, forward the ML model provision request to NWDAF 3 using an Nnwdaf_MLModelInfo request. In this case, the Nnwdaf_MLModelInfo request may include information indicating that it is an ML model provision request, identification information of the target ML model, information indicating that the item being provided is an ML model, and the user ID of the requester. In addition, in S302, NEF 1 may query NRF 2 to identify NWDAF 3.

[0111] In S303, NWDAF 3 receives an ML model provision request from NEF 1 and determines whether to accept the request. For example, the decision to accept the request is made based on whether the recipient matches the permitted recipients in the ML model publication information of the target ML model, and whether the requesting user ID indicates a user included in the scope of users in the ML model publication information of the target ML model.

[0112] In S304, the NWDAF 3 (ML Model Control Unit 32) sends a Nadrf_MLModelManagement_RetrievalRequest request to the ADRF 4 in order to retrieve the target ML model. The Nadrf_MLModelManagement_RetrievalRequest request sent in S304 includes identification information of the target ML model.

[0113] In S305, ADRF 4 receives a Nadrf_MLModelManagement_RetrievalRequest request from NWDAF 3, reads the data of the target ML model from ML model DB 43, and sends it to NWDAF 3 along with the Nadrf_MLModelManagement_RetrievalRequest response. The data of the target ML model may be the ML model file itself, or it may be some information used to reproduce the ML model, such as the parameters of the target ML model. Note that if NWDAF 3 already has the target ML model, the processing in S304 and S305 does not need to be executed.

[0114] In S306, NWDAF 3 sends the response received from ADRF 4 and the data of the target ML model to NEF 1. For example, in S302, if an ML model provision request is received using an Nnwdaf_MLModelInfo request, NWDAF 3 sends an ML model provision response using an Nnwdaf_MLModelInfo response.

[0115] In S307, NEF 1 forwards the response received from NWDAF 3 and the data of the target ML model to entity 500. Note that the request and response messages used in the ML model provision request and response in some of the processing in Figure 15 are examples and are not limited thereto. NEF 1 or NWDAF 3 in Figure 15 is an example of the "first computer". NWDAF 3 in Figure 15 is an example of the "fourth computer". The target ML model is an example of the "first machine learning model". The ML model provision request is an example of the "first request".

[0116] Figure 16 is an example of a flowchart of the NWDAF 3 processing in the ML model provision procedure for which ML models are provided. The processing shown in Figure 16 is executed repeatedly, for example, at a predetermined cycle.

[0117] In OP301, the ML model control unit 32 determines whether or not it has received an ML model provision request from NEF 1 that includes an ML model as the target. If an ML model provision request that includes an ML model as the target is received from NEF 1 (OP301: YES), the process proceeds to OP302. If an ML model provision request that includes an ML model as the target is not received from NEF 1 (OP301: NO), the process shown in Figure 16 ends. The process that results in a positive determination in OP301 corresponds, for example, to receiving an ML model provision request from NEF 1 in S302 of Figure 15.

[0118] In OP302, the ML model control unit 32 obtains ML model publication information for the ML model that is the target of the ML model provision request from ADRF 4. The processing of OP302 is performed, for example, by the ML model control unit 32 sending a Nadrf_MLModelManagement_RetrievalRequest request to ADRF 4 and receiving a Nadrf_MLModelManagement_RetrievalRequest response from ADRF 4.

[0119] In OP303, the ML model control unit 32 determines whether or not to accept the ML model provision request received in OP301. For example, if the provision availability field in the ML model publication information of the target ML model is "permitted," "ML model" is included in the permitted provision targets, and the requesting user ID matches the user indicated in the range of available users field, then OP303 makes a positive determination. If the provision availability field is "denied," "ML model" is not included in the permitted provision targets, or the requesting user ID does not match the user indicated in the range of available users field, then OP303 makes a negative determination.

[0120] If it is determined in OP301 to accept the ML model provision request received (OP303: YES), the process proceeds to OP305. If it is determined in OP301 to not accept the ML model provision request received (OP303: NO), the process proceeds to OP304. In OP304, the ML model control unit 32 sends a rejection response to NEF 1 as an ML model provision response. After that, the process shown in Figure 16 ends. The process that results in a positive determination in OP303 corresponds, for example, to the process in S303 of Figure 15.

[0121] In OP305, the ML model control unit 32 determines whether or not it holds the ML model that is the target of the ML model provision request. If it holds the target ML model (OP305: YES), the process proceeds to OP308. If it does not hold the target ML model (OP305: NO), the process proceeds to OP306.

[0122] In OP306, the ML model control unit 32 sends a Nadrf_MLModelManagement_RetrievalRequest request to ADRF 4. The Nadrf_MLModelManagement_RetrievalRequest request may include identification information of the target ML model. The processing in OP306 corresponds, for example, to the processing in S304 of Figure 15.

[0123] In OP307, the ML model control unit 32 determines whether or not it has received a response from ADRF 4 to the ML model provision request sent in OP306. Until a response is received from ADRF 4 (OP307: NO), the ML model control unit 32 remains in a standby state. If a response is received from ADRF 4 (OP307: YES), the process proceeds to OP308. Along with the response, the data for the target ML model is also received from ADRF 4. The process that results in a positive determination in OP307 corresponds, for example, to receiving a response from ADRF 4 in S305 of Figure 15.

[0124] In OP308, the ML model control unit 32 sends a response to NEF 1 for the ML model provision request received in OP301. Along with the response, the data of the target ML model is also sent. The transmitted ML model data may be, for example, the entire ML model file, or it may be a part of the information contained in the ML model file. After that, the process shown in Figure 16 is completed. The processing in OP308 corresponds, for example, to the processing in S306 in Figure 15.

[0125] Note that the processing of NWDAF 3 shown in Figure 16 is just one example, and the processing of NWDAF 3 in an ML model provision procedure that provides ML models is not limited to the processing shown in Figure 16. In an ML model provision procedure that provides ML models, the processing of NEF 1 can be the processing shown in Figure 13. Also, the determination of whether or not to accept an ML model provision request may be performed by NEF 1.

[0126] (ML model provision procedure when providing ML model prediction results) Figure 17 shows an example of the sequence in the ML model provision procedure when providing ML model prediction results. The ML model provision procedure when providing ML model prediction results is carried out by entity 500, NEF 1, NWDAF 3, ADRF 4, and designated notification recipient 600.

[0127] In S351, entity 500 sends an ML model provision request to NEF 1. Along with the ML model provision request, the identification information of the target ML model, information indicating that the provision target is an estimation result (as shown in Figure 17), and the user ID of the requester are also sent. If the provision target is an estimation result, optionally, along with the ML model provision request, the identification information of the device or user to whom the estimation result will be notified, and information indicating whether or not the estimation result will be expanded may also be sent.

[0128] Recipients of the prediction results may include, for example, UEs, other 5GCs, AFs, and servers that are members of System 100 and are different from the requester. Entity 500, the requester of the ML model provision request, may also be included as a recipient of the prediction results. If no recipients for the prediction results are specified, for example, the ML model may be notified only to Entity 500.

[0129] The deployment of prediction results refers to the process where, if the recipient of the prediction results is a User Engine (UE), that UE notifies surrounding UEs of the prediction results via, for example, Wi-Fi, BLE, vehicle-to-vehicle communication, or side-link communication. "Deployment of prediction results enabled" can be specified, for example, when the recipient of the prediction results is a UE. If there is no specification regarding the presence or absence of deployment of prediction results, it may be treated as, for example, no deployment of prediction results.

[0130] In S352, NEF 1 forwards the ML model provision request received from entity 500 to NWDAF 3. NEF 1 may also forward the ML model provision request to NWDAF 3 using, for example, an Nnwdaf_AnalyticsInfo_Request or an Nnwdaf_AnalyticsSubscription request. In this case, the Nnwdaf_AnalyticsInfo_Request or Nnwdaf_AnalyticsSubscription request may include information indicating that it is an ML model provision request, identification information of the target ML model, information indicating that the provision target is an estimation result, and the user ID of the requester. In addition, identification information of the device or user to which the estimation result will be notified, and information indicating whether or not the estimation result will be expanded may also be transmitted. Note that in S352, NEF 1 may query NRF 2 to identify NWDAF 3.

[0131] Whether to use an Nnwdaf_AnalyticsInfo_Request or an Nnwdaf_AnalyticsSubscription request may be specified by entity 500, for example, by sending information indicating whether it is a request or a subscription along with the ML model provision request as a method for notifying the prediction results. Alternatively, for example, NEF 1 may obtain information about the target ML model from ADRF 4 via NWDAF 3 and determine whether to use an Nnwdaf_AnalyticsInfo_Request or an Nnwdaf_AnalyticsSubscription request from the information about the ML model. In the case of a subscription, the prediction results are notified each time predictions are made by the target ML model, or when the prediction results meet the notification conditions. The notification conditions may be specified by entity 500 along with the transmission of the ML model provision request.

[0132] In S353, NWDAF 3 receives an ML model provision request from NEF 1 and determines whether to accept the request. For example, the decision to accept the request is made based on whether the recipient matches the permitted recipients in the ML model publication information of the target ML model, and whether the requesting user ID indicates a user included in the scope of users in the ML model publication information of the target ML model.

[0133] In S354, NWDAF 3 (ML Model Control Unit 32) sends a Nadrf_MLModelManagement_RetrievalRequest request to ADRF 4 in order to retrieve the target ML model. The Nadrf_MLModelManagement_RetrievalRequest request includes identification information of the target ML model. In S355, ADRF 4 receives the Nadrf_MLModelManagement_RetrievalRequest request from NWDAF 3, reads the data of the target ML model from ML Model DB 43, and sends it to NWDAF 3 along with the Nadrf_MLModelManagement_RetrievalRequest response.

[0134] In S361, NWDAF 3 performs inference using the target ML model received from ADRF 4. In S362, NWDAF 3 sends the inference result to NEF 1. For example, if an ML model provision request is received in S352 using Nnwdaf_AnalyticsInfo_Request or Nnwdaf_AnalyticsSubscription request, NWDAF 3 sends the inference result using Nnwdaf_AnalyticsInfo_Request response or Nnwdaf_AnalyticsSubscription_Notify. Also, if a notification destination is specified along with the ML model provision request in S352, the specified notification destination is sent as the destination along with the inference result. Also, if a UE is specified as the notification destination and information indicating expansion is available is received along with the ML model provision request in S352, the specified notification destination and expansion instructions are sent as the destination along with the inference result.

[0135] In S363, NEF 1 forwards the inference result received from NWDAF 3 to the designated notification destination 600. If the designated notification destination 600 is a UE and has received an expansion instruction along with the inference result, the expansion instruction is also sent to the designated notification destination 600.

[0136] In S364, if both the prediction result and the deployment instruction are received, the designated notification recipient UE 600 will, in accordance with the deployment instruction, broadcast or multicast the prediction result of the target ML model to surrounding UEs via, for example, Wi-Fi, BLE, vehicle-to-vehicle communication, or side-link communication.

[0137] If the ML model provision request in S351 or S352 is a subscription request, the processing from S361 to S364 is performed each time an inference is made or each time a notification condition is met. The determination of whether the notification condition has been met may be made by NWDAF 3 after the inference. The processing performed by the subscription is repeated until a stop instruction such as Nnwdaf_AnalyticsSubscription_Unsubscribe is received. The messages used in the ML model provision request and response in S352 and S362 are examples only and are not limited thereto. NWDAF 3 in Figure 17 is an example of a "first computer". The target ML model is an example of a "first machine learning model". The ML model provision request is an example of a "first request". The notification destination for the ML model inference result is an example of a "destination for providing the inference result of the first machine learning model". The deployment instruction is an example of a "transfer instruction".

[0138] Figure 18 is an example of a flowchart of the NWDAF 3 processing in an ML model provision procedure that provides inference results from an ML model. The processing shown in Figure 18 is executed repeatedly, for example, at a predetermined cycle.

[0139] In OP351, the analysis control unit 31 determines whether or not it has received an ML model provision request from NEF 1 that provides the prediction results of the ML model. If it has received an ML model provision request from NEF 1 that provides the prediction results (OP351: YES), the process proceeds to OP352. If it has not received an ML model provision request from NEF 1 that provides the prediction results (OP351: NO), the process shown in Figure 18 ends. The process that results in a positive determination in OP351 corresponds, for example, to receiving an ML model provision request from NEF 1 in S352 of Figure 17.

[0140] In OP352, the analysis control unit 31 obtains ML model publication information for the ML model that is the target of the ML model provision request from ADRF 4 via the ML model control unit 32. The processing of OP352 is performed, for example, by the analysis control unit 31 making a request to the ML model control unit 32, the ML model control unit 32 sending a Nadrf_MLModelManagement_RetrievalRequest request to ADRF 4, and receiving a Nadrf_MLModelManagement_RetrievalRequest response from ADRF 4.

[0141] In OP353, the analysis control unit 31 determines whether or not to accept the ML model provision request received in OP351. For example, if the provision availability field in the ML model publication information of the target ML model is "allowed," the permitted provision items include "prediction results," and the requesting user ID matches the user indicated in the range of available users field, then OP353's determination is affirmative. In addition, if a notification destination for the ML model prediction results is specified, the notification destination must also match the user indicated in the range of available users field, which is also a condition for OP353's affirmative determination. If the provision availability field is "rejected," the permitted provision items do not include "prediction results," the requesting user ID does not match the user indicated in the range of available users field, or the specified notification destination for the ML model prediction results does not match the user indicated in the range of available users field, then OP353's determination is negative. The scope of users who can access the service for the requesting user ID and the designated recipient of the ML model's prediction results may be the same, or different scopes may be set for each. The scope of users who can access the service for the requesting user ID is an example of "Condition 1". The scope of users who can access the service for the designated recipient of the ML model's prediction results is an example of "Condition 2".

[0142] If OP351 determines that the received ML model provision request should be accepted (OP353: YES), the process proceeds to OP355. If OP351 determines that the received ML model provision request should not be accepted (OP353: NO), the process proceeds to OP354. In OP354, the analysis control unit 31 sends a rejection response to NEF 1 as an ML model provision response. After that, the process shown in Figure 18 ends. The process that results in a positive determination in OP353 corresponds, for example, to the process in S353 of Figure 17.

[0143] In OP355, the analysis control unit 31 determines whether the ML model targeted by the ML model provision request is held in NWDAF 3. If the target ML model is held in NWDAF 3 (OP355: YES), the process proceeds to OP358. If the target ML model is not held in NWDAF 3 (OP355: NO), the process proceeds to OP356.

[0144] In OP356, the analysis control unit 31 requests the ML model control unit 32 to acquire the target ML model, and the ML model control unit 32 sends a Nadrf_MLModelManagement_RetrievalRequest request to ADRF 4 in order to acquire the target ML model. The processing in OP356 corresponds, for example, to the processing in S304 of Figure 15.

[0145] In OP357, the analysis control unit 31 determines, via the ML model control unit 32, whether or not it has received a Nadrf_MLModelManagement_RetrievalRequest response from ADRF 4. Until a response is received from ADRF 4 (OP357: NO), the analysis control unit 31 remains in a standby state. If a response is received from ADRF 4 (OP357: YES), the process proceeds to OP358. Along with the response, the data of the target ML model is also received from ADRF 4. The process that results in a positive determination in OP357 corresponds, for example, to receiving a response from ADRF 4 in S355 of Figure 17.

[0146] In OP358, the analysis control unit 31 performs estimation using the target ML model. More specifically, the analysis control unit 31 inputs predetermined data into the target ML model and obtains estimation results from the output results. The processing in OP358 corresponds, for example, to the processing in S361 in Figure 17.

[0147] In OP359, the analysis control unit 31 sends the prediction result of the target ML model to NEF 1 along with the response. Along with the prediction result of the target ML model, information about the designated notification destination and, if the destination includes a UE and expansion is required, an expansion instruction is also sent. After that, the process shown in Figure 18 is completed. The process in OP359 corresponds, for example, to the process in S362 of Figure 17.

[0148] Note that the processing of NWDAF 3 shown in Figure 18 is just one example, and the processing of NWDAF 3 in an ML model provision procedure that provides the prediction results of an ML model is not limited to the processing shown in Figure 18. In an ML model provision procedure that provides the prediction results of an ML model, the processing of NEF 1 can be the processing shown in Figure 13.

[0149] Figure 19 is an example of a flowchart of the UE's processing in an ML model provision procedure that provides the prediction results of an ML model. The UE is the UE designated as the recipient of the notification of the ML model's prediction results. The processing shown in Figure 19 is executed repeatedly, for example, at a predetermined interval. The entity executing the processing shown in Figure 19 is the UE's CPU, but for convenience, Figure 19 is explained with the UE as the main entity.

[0150] In OP381, the UE determines whether or not it has received the ML model prediction result. If the ML model prediction result has been received (OP381: YES), the process proceeds to OP382. If the ML model prediction result has not been received (OP381: NO), the process shown in Figure 19 ends. The process that results in a positive determination in OP381 corresponds, for example, to the process of receiving the ML model prediction result from 5GC in S362 of Figure 17.

[0151] In OP382, the UE determines whether or not it has received an expansion instruction along with the ML model prediction result. If an expansion instruction is received (OP382: YES), the process proceeds to OP383. If an expansion instruction is not received (OP382: NO), the process shown in Figure 19 ends.

[0152] In OP383, the UE notifies surrounding UEs of the ML model prediction results via broadcast or multicast, for example, via Wi-Fi, BLE, vehicle-to-vehicle communication, or sidelink communication. After that, the process shown in Figure 19 is completed. The process in OP383 corresponds, for example, to the process in S364 of Figure 17.

[0153] (Incentive Distribution Procedure) Figure 20 shows an example of the sequence in the incentive distribution procedure. The incentive distribution procedure is performed by the FL server and the FL client. In Figure 20, the incentive distribution procedure is from S411 to S413.

[0154] S401 to S405 is an example of a payment sequence for Entity 500's ML model provision service. Entity 500's payment for the ML model provision service is one of the triggers for initiating the incentive distribution procedure described later.

[0155] In S401, CHF 5 transmits billing information for entity 500, including the fee for the ML model provision service, to the billing support system. For example, NWDAF 3 may report to CHF 5 each time it receives an ML model provision request, so that CHF 5 maintains historical information of entity 500's ML model provision requests. CHF 5 may, for example, determine the fee for entity 500's ML model provision service based on the historical information of entity 500's ML model provision requests. The billing support system is a system that manages billing and payment of fees to subscribers of system 100.

[0156] In S402, the billing support system issues a bill to entity 500. In S403, entity 500 makes the payment. In S404, the billing support system confirms the payment by entity 500 and sends a payment completion notification for entity 500 to CHF 5. In S405, CHF 5 notifies FL server 3A of the payment completion notification for entity 500. Entity 500's payment for the ML model provision service is an example of providing incentives to FL servers and FL clients that participated in learning the ML model.

[0157] In S411, the FL server 3A obtains information regarding incentives for the target ML model. The target ML model in the incentive distribution procedure is the ML model that the FL server 3A was responsible for training.

[0158] Information regarding incentives for the target ML model may include, for example, the usage fee for one use of the target ML model by a single user, the number of times the target ML model or its prediction results were provided during a specified period, and the total incentives generated by the services provided by the target ML model during a specified period. Information regarding incentives for the target ML model may be obtained, for example, by notifications from a billing support system that manages billing, notifications from a device that manages statistics on ML model service provision, or by acquisition by FL server 3A. The device that manages statistics on ML model service provision may be, for example, NWDAF 3, CHF 5, and ADRF 4. The device that manages statistics on ML model service provision may, for example, obtain statistics on ML model service provision by receiving reports from NWDAF 3 regarding the acceptance of ML model provision requests.

[0159] In S412, FL Server 3A obtains incentives for the target ML model from the incentive information and distributes them to FL Client 3B and FL Client 3C. FL Server 3A may also be included as a recipient. If the incentive information is the number of times the target ML model or its prediction results have been provided during a predetermined period, the incentive for the target ML model during that predetermined period is obtained by multiplying the number of times by the unit price of the incentive for providing the target ML model or its prediction results.

[0160] The unit price of the incentive may differ depending on the target recipient. The unit price of the incentive may be higher for those who receive ML models than for those who receive prediction results, and vice versa.

[0161] Furthermore, the incentive unit price may be uniform across ML models, or it may vary depending on the type and application of the ML model. For example, the incentive unit price may be lower or higher for ML models used in high-demand applications. For example, the incentive unit price may be higher for ML models with high training costs.

[0162] Furthermore, the unit price of the incentive may differ depending on the type of entity 500. For example, the unit price of the incentive may be higher for UE, AF, and the other 5GC in that order. The unit price of the incentive may be set by the system 100.

[0163] The distribution of incentives may be, for example, equally divided between FL servers and FL clients that participated in training the target ML model, or the FL servers may receive a higher share than the FL clients. Furthermore, among FL clients, the higher the contribution to the target ML model, the higher the incentive distribution. The degree of contribution to the target ML model may be determined based on, for example, the number of training data points, the number of training iterations, etc.

[0164] In S413, FL Server 3A notifies FL Clients 3B and 3C of information regarding the incentives for the target ML model. FL Clients 3B and 3C save the notified information regarding the incentives. In Figure 20, FL Server 3A is an example of a "second computer".

[0165] Figure 21 is an example of a flowchart of the processing performed by the FL server 3A in the incentive distribution procedure. The processing shown in Figure 21 is executed repeatedly, for example, at a predetermined interval. The processing shown in Figure 20 is executed for each ML model that the FL server 3A has been involved in training.

[0166] In OP401, the ML model control unit 32 determines whether or not it has obtained information regarding incentives for the target ML model. If information regarding incentives has been obtained for the target ML model (OP401: YES), the process proceeds to OP402. If information regarding incentives has not been obtained for the target ML model (OP401: NO), the process shown in Figure 21 ends.

[0167] In OP402, the ML model control unit 32 distributes incentives to FL clients that participated in training the target ML model. The FL server 3A may also be included as a recipient of the incentives. In OP403, the ML model control unit 32 notifies the FL clients of the incentives for the target ML model that have been distributed to each FL client. After that, the process shown in Figure 21 is completed.

[0168] In the examples shown in Figures 20 and 21, the FL server 3A distributes incentives for the target ML model, but this is not limited to it. The distribution of incentives to devices that participated in learning the target ML model may be performed by a device other than the FL server 3A. A device capable of distributing incentives to devices that participated in learning the target ML model may be, for example, an NWDAF other than the NWDAF operating as the FL server 3A, a NEF, a CHF, or a billing support system. A device other than the FL server 3A may, for example, obtain ML model learning information from the FL server 3A or ADRF 4 and identify the devices that participated in learning the target ML model. Alternatively, the device that distributes incentives may be made to hold the ML model learning information for each ML model.

[0169] <Specific Example> Figure 22 shows an example of an ML model provision service. In Figure 22, the ML model is a model that performs congestion prediction. AF#1 corresponds to an entity and sends an ML model provision request to 5GC#1, specifying that the prediction results will be provided. In the example in Figure 22, the prediction result of the ML model is a congestion prediction, which is a prediction that congestion will occur. In addition, AF#1 sends the ML model provision request along with the specification of AF#1 as the recipient of the congestion prediction notification, the UE that is subscribed to 5GC#1, the specification of 5GC#2, and the specification that the congestion prediction should be expanded.

[0170] Therefore, in the example shown in Figure 22, for example, when congestion is predicted in area #A by the ML model in 5GC#1, the congestion prediction for area #A is notified from 5GC#1 to AF#1, the UE of 5GC#1, and 5GC#2. The UE of 5GC#1 is notified of the congestion prediction along with a deployment instruction, so the UE of 5GC#1 notifies the surrounding UEs of the congestion prediction for area #A by means of Wi-Fi, BLE, vehicle-to-vehicle communication, or side-link communication, for example.

[0171] For example, AF#1 may utilize the congestion forecast for Area #A by notifying UEs subscribed to AF#1's service of the congestion forecast for Area #A. For example, 5GC#2 may utilize the congestion forecast for Area #A by notifying UEs subscribed to 5GC#2 of the congestion forecast for Area #A.

[0172] <Effects of the First Embodiment> In the first embodiment, the ML model that the system 100 has trained can also be provided to parties outside the system 100. In the first embodiment, incentives for providing the ML model are distributed to the FL server and FL client that participated in training the ML model. This makes it possible to recover the costs incurred in training the ML model. Furthermore, it is possible to encourage the participation of FL clients when training other ML models.

[0173] Entity 500, which receives the ML model provision service, no longer needs to prepare the ML model itself, thus reducing costs. Furthermore, Entity 500, which receives the ML model as a service provider, may train the ML model using training data suitable for other purposes (repurposed learning), which can save time and resources compared to training an ML model from scratch.

[0174] <Other Embodiments> The embodiments described above are merely examples, and this disclosure may be modified as appropriate without departing from its essence.

[0175] In the first embodiment, the ML model publication setting procedure, the ML model publication information acquisition procedure, and the ML model provision procedure were described as separate procedures, but the embodiment is not limited thereto. For example, when an ML model publication information request or an ML model provision request is received from entity 500, the NWDAF 3 that received the request may perform the ML model publication setting procedure for the target ML model on the FL server and FL client to obtain information indicating whether or not the target ML model can be provided to the service.

[0176] In the first embodiment, the ML model is assumed to be trained by associative learning. However, it is not limited to this, and any learning method that involves multiple devices may be used to train the ML model. Examples of learning methods that involve multiple devices include Distributed Machine Learning (DML), Split Learning (SL), and Privacy-Preserving Machine Learning (PPML).

[0177] In the first embodiment, information regarding the ML model is held in ADRF 4, but the device that holds information regarding the ML model is not limited to ADRF 4. For example, public information about the ML model may be held in NRF 2 or NWDAF 3. Data about the ML model may be held in NWDAF 3, which operates as an FL server that trained the ML model.

[0178] In the first embodiment, the explanation was based on the premise that ML model training is performed under the management of 5GC as system 100. However, the ML model provision service described in the first embodiment can also be applied when ML model training is performed in a system other than 5GC. In systems other than 5GC, the procedures performed by the ML model provision service are the same as in the first embodiment: ML model publication setting procedure, ML model publication information acquisition procedure, ML model provision procedure, and incentive distribution procedure. The devices that execute each procedure will be changed as appropriate depending on the system configuration.

[0179] In systems other than 5GC, the processing of NEF 1 and NWDAF 3 may be performed by a device operating as an FL server. The ML model publication information and ML models held by ADFR 4 may be held by any device within the system.

[0180] The processes and methods described in this disclosure can be freely combined and implemented, provided that no technical inconsistencies arise.

[0181] Furthermore, a process described as being performed by a single device may be divided and executed by multiple devices. Conversely, a process described as being performed by different devices may be executed by a single device. In a computer system, the hardware configuration (server configuration) by which each function is implemented can be flexibly changed.

[0182] The present disclosure can also be realized by supplying a computer program implementing the functions described in the embodiments above to a computer, and having one or more processors in the computer read and execute the program. Such a computer program may be provided to the computer by a non-temporary computer-readable storage medium that can be connected to the computer's system bus, or it may be provided to the computer via a network. Non-temporary computer-readable storage mediums include, for example, any type of disk such as magnetic disks (floppy disks, hard disk drives (HDDs), etc.), optical disks (CD-ROMs, DVDs, Blu-ray discs, etc.), read-only memory (ROM), random access memory (RAM), EPROM, EEPROM, magnetic cards, flash memory, optical cards, and any type of medium suitable for storing electronic instructions.

[0183] 1. NEF 2. NRF 3. NWDAF 4. ADRF 5. CHF 11. Control Unit 31. Analysis Control Unit 32. ML Model Control Unit 33. ML Model Learning Information Storage Unit 41. Control Unit 42. ML Model Public Information DB 43. ML Model DB 100. System 500. Entity 1000. Information Processing Device 1001. CPU 1002. Main Memory 1003. Auxiliary Memory 1004. Communication Device

Claims

1. A method comprising: a first computer receiving a first request for a first service which includes providing a first machine learning model that has been trained by a plurality of devices or providing inference results by the first machine learning model; the first computer providing the first service in response to the first request; and the first or second computer distributing an incentive for the first machine learning model provided for the first service to the plurality of devices that participated in training the first machine learning model.

2. The method according to claim 1, wherein the first computer receives, along with the first request, information indicating the subject of the provision relating to the first machine learning model, and the first computer or the second computer determines an incentive for the first machine learning model according to the subject of the provision relating to the first machine learning model.

3. The method according to claim 1, further comprising the first computer accepting the first request if the sender of the first request satisfies a first condition which is a condition relating to a user permitted to receive the first service with respect to the first machine learning model, and returning a rejection response to the first request if the sender of the first request does not satisfy the first condition.

4. The method according to claim 1, wherein, upon receiving information indicating that the subject of the provision relating to the first machine learning model is the first machine learning model itself, the first computer transmits some or all of the data of the first machine learning model as information relating to the first machine learning model.

5. The method according to claim 1, wherein, upon receiving information indicating that the subject of the provision relating to the first machine learning model is the prediction result by the first machine learning model, along with the first request, the first computer transmits the prediction result by the first machine learning model as information relating to the first machine learning model.

6. The method according to claim 5, wherein the first computer accepts a designation of a recipient for the first request and transmits the first machine learning model's prediction results to the designated recipient in response to the first request.

7. The method according to claim 6, wherein the first computer receives a designation that the prediction results of the first machine learning model be transferred from the designated recipient in response to the first request, and transmits the prediction results of the first machine learning model and an instruction to transfer the prediction results of the first machine learning model to the designated recipient in response to the first request.

8. The method according to claim 6, wherein the first computer transmits the prediction results of the first machine learning model to the designated recipient if the designated recipient satisfies a second condition indicating a recipient who is permitted to receive the prediction results of the first machine learning model, and replies with a rejection response to the first request if the designated recipient does not satisfy the second condition.

9. The method according to claim 1, further comprising: the first computer, the second computer, or the third computer sending a second request to the plurality of devices participating in the training of the first machine learning model to inquire whether they are willing to provide the first service with respect to the first machine learning model; and receiving information from the plurality of devices as a response to the second request indicating permission or denial of the provision of the first service with respect to the first machine learning model.

10. The method according to claim 9, wherein the first computer accepts the first request if it has obtained permission from all of the plurality of devices that have participated in training the first machine learning model to provide the first service for the first machine learning model, and sends a rejection response to the first request if it has not obtained permission from at least one of the plurality of devices that have participated in training the first machine learning model to provide the first service for the first machine learning model.

11. The method according to claim 9, further comprising: determining that the first service for the first machine learning model is unavailable if any of the third computers or a fourth computer from the first computer has obtained a refusal to provide the first service for the first machine learning model from at least one of the plurality of devices that participated in training the first machine learning model; and determining that the first service for the first machine learning model is available if permission to provide the first service for the first machine learning model has been obtained from all of the plurality of devices that participated in training the first machine learning model.

12. The method according to claim 9, further comprising: the first computer receiving a third request from any of the third computers or a fifth computer requesting to obtain first information relating to the first machine learning model, which includes at least information indicating whether or not the first service relating to the first machine learning model is available; and transmitting the first information as a response to the third request.

13. An information processing device comprising: a control unit that, when a first service is provided which includes the provision of a first machine learning model that has been trained by multiple devices or the provision of inference results by the first machine learning model, distributes the incentive for the first machine learning model provided by the first service to the multiple devices that participated in the training of the first machine learning model.

14. The information processing apparatus according to claim 13, wherein the control unit determines an incentive for the first machine learning model provided by the first service, depending on the subject of the provision relating to the first machine learning model.

15. An information processing device comprising: a control unit that performs: sending a first request for a first service which includes providing a first machine learning model that has been trained by multiple devices or providing inference results by the first machine learning model; receiving the first service in response to the first request; and providing an incentive for the first machine learning model by the first service to the multiple devices that participated in training the first machine learning model.

16. The information processing apparatus according to claim 15, wherein the control unit also transmits, along with the first request, information indicating the subject of the provision relating to the first machine learning model, and provides an incentive for the first machine learning model according to the subject of the provision relating to the first machine learning model.

17. The information processing apparatus according to claim 15, wherein the control unit transmits, along with the first request, information indicating that the subject of the provision relating to the first machine learning model is the prediction result by the first machine learning model, and further performs the operation of specifying the recipient of the prediction result by the first machine learning model for the first request.

18. The information processing apparatus according to claim 17, wherein the control unit further transmits, along with the first request, an instruction for the transfer of the prediction results by the first machine learning model to the designated recipient.

19. The information processing apparatus according to claim 18, which, upon receiving the prediction result from the first machine learning model and a transfer instruction for the prediction result from the first machine learning model, transmits the prediction result from the first machine learning model to another device via a wireless communication method that does not involve a relay device.

20. The information processing apparatus according to claim 15, further comprising: the control unit transmitting a third request for obtaining first information relating to the first machine learning model, which includes at least information indicating whether or not the first service relating to the first machine learning model can be provided; and receiving the first information as a response to the third request.

Images