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

Abstract

An information processing system according to the present invention is made up of a plurality of participating nodes and performs federated learning on the basis of sequentially obtained local data. The information processing system includes a detecting means that detects a data abnormality in at least one of the plurality of participating nodes, and a control means that determines which of a plurality of countermeasures is to be implemented when a data abnormality occurs.

Description

Information Processing System, Information Processing Apparatus, Information Processing Method, and Program

[0001] The present disclosure relates to an information processing system, an information processing apparatus, an information processing method, and a program.

[0002] Federated Learning (FL) is a learning method in which multiple entities train a machine learning model by utilizing their local data while keeping the data localized without sharing it. Non-Patent Document 1 describes the procedure of federated learning by multiple NWDAFs (Network Data Analytics Function) in a 5G system.

[0003] In federated learning, as shown in Non-Patent Document 2, it has also been considered to perform sequential learning using dynamic sampling in an environment where data is given sequentially.

[0004] 3GPP TS 23.288 V18.3.0 (2023-09) Section 6.2COthmane Marfoq, et al., Federated Learning for Data Stream, arXiv:2301.01542, <URL: https: / / arxiv.org / pdf / 2301.01542>

[0005] When performing dynamic sampling, it is assumed that some data cannot be acquired after data alignment is completed and federated learning is started. Since there is a risk that model learning cannot be performed correctly if a data error occurs, it is desirable to perform appropriate recovery processing. However, in the prior art, no consideration has been given to what recovery processing should be performed when what kind of error occurs.

[0006] An object of the present disclosure is to provide a method that enables appropriate recovery processing when a data error occurs during the execution of federated learning.

[0007] One aspect of this disclosure is an information processing system comprising a plurality of participating nodes that perform federated learning based on locally obtained data sequentially, the system comprising: detection means for detecting data anomalies in at least one of the plurality of participating nodes; and control means for determining which of a plurality of response methods to implement when a data anomaly occurs.

[0008] Another aspect of the present disclosure is an information processing device in an information processing system which consists of a plurality of participating nodes and performs federative learning based on locally data obtained sequentially, the information processing device having a processor that performs a receiving step of receiving data indicating that a data anomaly has occurred from at least one of the plurality of participating nodes, a control step of determining which of a plurality of response methods to implement when the data anomaly occurs, and a notification step of notifying the plurality of participating nodes to implement the determined response method.

[0009] Other embodiments include an information processing system comprising the above-mentioned information processing device, an information processing method corresponding to the above-mentioned information processing device, a program for causing a computer to execute the information processing method, or a computer-readable storage medium that non-temporarily stores the program.

[0010] According to this disclosure, appropriate recovery processing will be possible if a data error occurs during the execution of associative learning.

[0011] Figure 1 is a diagram illustrating an overview of the information processing system related to this disclosure. Figure 2 is a diagram illustrating an overview of vertical federated learning. Figure 3 is a diagram illustrating the flow of federated learning processing related to this disclosure. Figure 4 is a diagram illustrating the flow of federated learning processing related to this disclosure. Figure 5 is a diagram illustrating the flow of error recovery processing related to this disclosure. Figure 6 is a diagram illustrating the types of data errors related to this disclosure. Figure 7 is a diagram illustrating the flow of error recovery processing related to this disclosure. Figure 8 is a diagram illustrating an example of the information processing system related to this disclosure. Figure 9 is a diagram illustrating an example of the information processing device related to this disclosure.

[0012] 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.

[0013] <Embodiment 1> (Summary) This disclosure relates to a recovery method when a data error occurs in an information processing system 10 that performs federated learning. Figure 1 is a diagram showing the configuration of the information processing system 10 according to this disclosure. As shown in the figure, the information processing system 10 according to this embodiment includes at least one server device 100 and at least one client device 200 connected to a network N. Each of the server device 100 and the client device 200 participates in federated learning. Localized data held by each of the client devices 200 is not shared with other devices.

[0014] The information processing system 10 performs federated learning in cooperation with the server device 100 and the client device 200. The federated learning performed by the information processing system 10 is assumed to be vertical federated learning (VFL) in this case, but it may also be horizontal federated learning (HFL).

[0015] Figure 2 is an explanatory diagram of Vertical Flour Learning (VFL). VFL is a federated learning method in which learning is performed without exchanging or sharing the local data held by each entity, and is a learning technique used when the local data held by each entity has different features of the same sample.

[0016] As shown in Figure 2, the information processing system 10 that performs vertical federative learning consists of one server device 100 and multiple client devices 200A to 200C (three in Figure 2). Among the multiple participants in VFL, the participant (also called a party) that has the correct labels in machine learning and controls the overall learning is called the server, and the other participants (participants that do not have correct labels) are called clients. Note that the device that controls the overall learning and the device that has the correct labels may be different; in this case, the former is called the coordinator and the latter is called the active participant. Clients are also called passive participants.

[0017] Server device 100 holds the local model 111 and the correct labels 112 for each sample. Client devices 200A to 200C each hold the local models 221A to 221C and the training sample data 222A to 222C for each sample. In the following description, client devices 200A to 200C may be referred to as client device 200, local models 221A to 221C as local model 221, and training sample data 222A to 222C as training sample data 222.

[0018] In this embodiment, the learning sample data 222A to 222C held by the client devices 200A to 200C are acquired sequentially after the start of learning and used for learning. Sample data acquired before the start of learning may also be used for learning.

[0019] Local model 111 and local models 221A-221C are each neural networks, and there is a one-to-one correspondence between each node in the downstream layer of local model 111 and each node in the upstream layer of local models 221A-221C. Therefore, local model 111 and local models 221A-221C constitute a single learning model or neural network. Local model 111 is also called the top model because it is the upstream model, and local models 221A-221C are also called the bottom models because they are the downstream models.

[0020] The correct label 112 and the training sample data 222A to 222C contain the correct label and features (attributes) for common samples. Furthermore, each of the training sample data 222A to 222C contains different features (attributes). Note that the training sample data 222A to 222C may contain data from different (non-common) samples, or they may contain overlapping features (attributes). However, when performing vertical learning, each client device 200A to 200C learns using the different features of the common samples.

[0021] Figure 3 is a flowchart illustrating the flow of federated learning processing by the information processing system 10 according to this embodiment.

[0022] In step S10, the server device 100 determines the learning participants (client devices 200) that will participate in the federated learning. The learning participants are selected from devices that possess the data necessary for learning and that intend to participate in the learning.

[0023] In step S20, data alignment is performed between the server device 100 and client devices 200 participating in the learning process. During the data alignment process, each client device 200 determines which feature data to use for learning, and processes such as unifying the data format and normalizing the data are performed.

[0024] In step S30, federated learning based on the learning data is performed between the server device 100 and the client device 200 participating in the learning process. Since the client device 200 acquires learning data sequentially even after data alignment, the information processing system 10 periodically performs additional learning based on the newly acquired learning data.

[0025] Figure 4 shows a detailed flow of the associative learning process in step S30.

[0026] In step S31, the client device 200 collects training sample data. In step S32, the client device performs local training. Local training is the process of inputting the training sample data C held by the client device 200 into a local model and obtaining the output of the top layer of the local model. In step S33, the client device 200 transmits the local training result, i.e., the output of the top layer of the local model, to the server device 100.

[0027] In step S34, the server device 100 receives local learning results from each client device 200. In step S35, it inputs these results into the local model 111 to obtain the final output and calculates the error (loss function) by comparing the final result with the correct label. This completes the forward propagation process in the learning process. Based on the calculated error, the server device 100 finds the gradient, and sequentially finds the gradient at each layer of the local model through backpropagation. It also updates the local model 111 (updates the weight parameters) based on the gradient. Once the server device 100 has found the gradient at the furthest downstream layer, in step S36, it sends that gradient to each client device 200 as a model update parameter.

[0028] In step S37, the client device 200 receives the model update parameters (gradients), and in step S38, it sequentially determines the gradients at each layer of the local model by backpropagation and updates the local model (updates the weight parameters) based on the gradients.

[0029] The processing from step S32 to step S38 completes one epoch of learning. In Figure 4, it is shown that only one epoch of learning is performed each time training data is acquired, but multiple epochs of learning may also be performed.

[0030] If new training data is obtained, return to step S31 and repeat the above process.

[0031] (Data Error Recovery Process - Example 1) In this embodiment, the recovery process when a data error occurs after the start of associative learning will be described. Figure 5 is a flowchart showing the flow of the recovery process in this embodiment, and Figure 6 is a diagram illustrating the types of data errors.

[0032] Here, the process shown in Figure 5 is described as being performed by the server device 100. In step S51, the server device 100 detects the occurrence of a data error in one of the client devices 200, and in step S52, it determines the type of error. The entity that performs the processes in steps S51 and S52 corresponds to the detection means of the present invention.

[0033] Data errors can be classified into four types, as shown in Figures 6(A) to 6(D). Figures 6(A) to 6(D) show combinations of samples and features used in the training data. Here, shading indicates missing data. For example, in Figure 6(A), "Feature 1" of sample "BBB" is missing.

[0034] Figure 6(A) shows a data unit error. A data unit error is an error that prevents the acquisition of a specific feature for a particular sample. Although Figure 6(A) shows a case where only one feature for one sample becomes unavailable, a data unit error can also be considered when one or more features for one or more samples become unavailable (except in the cases described below).

[0035] Figure 6(B) shows a sample-level error. A sample-level error is an error in which all features for a particular sample cannot be obtained. Although Figure 6(B) shows the case where only the features of one sample become unavailable, cases where the features of multiple samples become unavailable (except in the case described below) can also be considered sample-level errors. A sample-level error can also be called a row-level error.

[0036] Figure 6(C) shows a feature-level error. A feature-level error is an error that prevents a particular feature from being obtained from all samples. Although Figure 6(C) shows the case where only one feature becomes unavailable, cases where multiple features become unavailable (except as described below) can also be considered feature-level errors. A feature-level error can also be called a column-level error.

[0037] Figure 6(D) shows a client-specific error. A client-specific error is an error in which no data can be acquired at all from a particular client device 200.

[0038] Data-level errors, sample-level errors, feature-level errors, and client-level errors correspond to the first, second, third, and fourth types of data anomalies of the present invention, respectively.

[0039] The client device 200 can first detect data errors at the data unit, sample unit, and feature unit levels. Therefore, the client device 200 can notify the server device 100 of the occurrence of any of these data errors, and the server device 100 can detect the occurrence and type of data error upon receiving this notification. The method of notifying the server device 100 of the occurrence of a data error from the client device 200 is not particularly limited, but for example, it can be notified by an error message. For example, if the client device 200 is operating, it can notify the server device 100 of a client-level data error. However, if the client device 200 fails or becomes unable to communicate with the server device 100, the server device 100 can determine that a client-level data error has occurred when it becomes unable to communicate with the client device 200.

[0040] The server device 100 determines whether to recover from a data error by implementing one of the response methods in steps S53 to S55, depending on the type of data error that occurred. The entity that makes this determination corresponds to the control means of the present invention.

[0041] If the data error is a data-unit error, the process proceeds to step S53, where the server device 100 decides to correct or delete the anomalous data and continue learning. Correcting anomalous data involves, for example, estimating the missing data based on past values ​​of the same sample and / or current values ​​of other samples. The estimation can be obtained, for example, by the mean, median, or mode.

[0042] If the data error is an error at the sample unit or the feature quantity unit, the process proceeds to step S54, and the server device 100 re-executes the data alignment process (step S20 in FIG. 3) and determines to continue learning after the data alignment is completed.

[0043] If the data error is an error at the client unit, the process proceeds to step S55, and the server device 100 determines to re-execute the federated learning from the beginning (step S10 in FIG. 3).

[0044] According to the present embodiment, appropriate recovery processing can be performed according to the type of the generated data error, and the learning results obtained so far can be effectively utilized.

[0045] (Data Error Recovery Process - Example 2) FIG. 7 is a flowchart showing another example of the flow of the recovery process in the present embodiment. In this example, among a plurality of corresponding methods, recovery is performed so that the accuracy of the model becomes equal to or higher than a preset threshold value.

[0046] In step S71, the server device 100 detects the occurrence of a data error in any of the client devices 200. Since the detection method is the same as in the above example, the description thereof is omitted. In this example, although the determination of the type of the data error is not essential, the type of the generated data error may be determined.

[0047] In step S73, the server device 100 estimates the learning accuracy of the learning model when the error data is corrected or deleted and learning is continued, and determines whether the estimated accuracy is equal to or higher than the accuracy threshold value. The accuracy threshold value is set from the outside at the start of learning. If this estimated accuracy is equal to or higher than the threshold value (S73 - YES), the process proceeds to step S74, and it is determined to correct or delete the abnormal data and continue learning. If the condition is not satisfied, the process proceeds to step S75.

[0048] In step S75, the server device 100 estimates the learning accuracy of the learning model when data alignment is re-executed, and determines whether the estimated accuracy is equal to or higher than an accuracy threshold. If this estimated accuracy is equal to or higher than the threshold (S75 - YES), the process proceeds to step S76, where it is determined to re-execute the data alignment and continue learning after the data alignment is completed. If the estimated accuracy due to re-executing the data alignment is lower than the threshold (S75 - NO), the process proceeds to step S77, where it is determined to re-execute the learning from the beginning.

[0049] Note that the prediction of the learning accuracy in steps S73 and S75 can be executed based on prior learning. For example, by performing learning after modifying or deleting some of the data in the learning dataset of federated learning and obtaining the learning accuracy thereof, it becomes possible to grasp how much learning accuracy can be obtained when a defect occurs in which data. The same applies when data errors occur in units of samples and feature quantities. It is only necessary to perform prior learning on how the learning accuracy changes due to a decrease in the number of samples or the number of feature quantities. In the case of a data error in units of feature quantities, it may be determined whether it is possible to obtain the feature quantities in which an error has occurred from other client devices 200 currently participating in federated learning, and if it is possible to obtain them, it may be determined that learning can be performed with the same accuracy as the current time.

[0050] (Data error recovery process - modified example) In the above description, when a data error occurs, the server device 100 performs detection of the error, determination of the error type, and determination of what recovery process to execute. However, other devices may perform these processes. For example, if a coordinator is included in the information processing system 10, the coordinator may execute the above processes. Also, the client device 200 may determine the recovery method when a data error occurs. When a data unit error occurs, the client device 200 may continue local learning using the corrected data without reporting it to the server device 100 or the coordinator. Also, when a data error occurs in units of samples or feature quantities, the client device 200 may request the server device 100 or the coordinator to re-execute from data alignment.

[0051] (Example of an Information Processing System) Figure 8 shows an example configuration of a 5G system, which is an example of an information processing system that performs federated learning. In Figure 8, UE (User Equipment) 2 is the user's (subscriber's) terminal. RAN (Radio Access Network) 3 is the access network to the 5G core network (5GC). RAN3 is composed of base stations (gNB). The 5G network has a 5G core network (5GC) and an access network ((RAN)), and UE2, DN5, and AF12 are connected to the 5G network. Each of NF11a to 11k is a function realized by one or more computers (information processing devices) executing a program. However, a single computer may realize two or more of NF11a to 11k.

[0052] 5GC is composed of a set of components having predetermined functions called NF (Network Function). Figure 8 illustrates the following as NF11 that constitute 5GC. In Figure 8, they are shown as rectangles with thick lines.

[0053] UPF (User Plane Function) 11a AMF (Access and Mobility Management Function) 11b SMF (Session Management Function) 11c PCF (Policy Control Function) 11d NEF (Network Exposure Function) 11e NRF (Network Repository Function) 11g NSSF (Network Slice Selection Function) 11h AUSF (Authentication Server) Function) 11i UDM (Unified Data Management) 11j NWDAF (Network Data Analytics Function) 11k

[0054] Active and passive participants in vertical associative learning can be any NF, for example, NWDAF11k or AF12. Note that multiple instances of each NF may exist within the system, and all active and passive participants in associative learning may be NWDAF11k.

[0055] As an example, it is assumed that multiple NWDAF11k devices will perform federated learning as client devices. For example, in an example of federated learning using data about UEs, the sample is the UE, and the features are data obtained about the UE. Examples of data about the UE include the location and communication volume of the UE, the age and occupation of the UE owner, etc. A data-level data error would occur if some features could not be obtained from the UE. A sample-level data error would be due to an abnormality or failure of the UE, termination of the UE's service usage contract, or a change in the type of data provided (discontinuation of the provision of the features). A feature-level data error would be due to a change in the type of data provided (discontinuation of the provision of the features).

[0056] (Example of Information Processing Device Configuration) Figure 9 shows an example of an information processing device configuration that can operate as an NF11a to 11k, an OAM terminal, and an external server, respectively. In Figure 9, the information processing device 90 can be configured using a dedicated or general-purpose information processing device (computer) such as a personal computer (PC), workstation (WS), or server machine. However, the information processing device 90 may also be a collection of one or more computers (cloud).

[0057] The information processing device 90 includes a control unit 91, a storage device 92, an input device 93, an output device 94, and a communication interface 95 (communication IF 95), all of which are interconnected via a bus.

[0058] The control unit 91 is composed of a processor, RAM, and ROM, etc. The processor is typically a CPU, but it may be a processor other than a CPU, such as a DSP or GPU, or a combination of a CPU and a processor. Furthermore, the processing and operation by the processor (the processing and operation as each NF performed by the execution of the processor's program) may be performed using semiconductor devices (hardware) such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit), or using a combination of a processor and hardware such as a SoC (System On a Chip). The processor (controller), FPGA, ASIC, and SoC are each examples of "circuitry".

[0059] The storage device 92 includes main memory and auxiliary storage. The main memory is used as at least one of the following: a program and data storage area, a program deployment area, a program work area, and a communication data buffer area. The main memory consists of RAM (Random Access Memory), or a combination of RAM and ROM (Read Only Memory). The auxiliary storage is used as a data and program storage area. Non-volatile storage media are used for the auxiliary storage. Examples of non-volatile storage media include hard disks, solid state drives (SSDs), flash memory, or EEPROMs (Electrically Erasable Programmable Read-Only Memory). The storage device 92 may also include a drive device for a disk recording medium.

[0060] Information processing device 90 participating in vertical federative learning as an active participant stores the local model (top model) and the correct labels for each sample in the storage device 92. Information processing device 90 participating in vertical federative learning as a passive participant stores the local model (bottom model) and training sample data in the storage device 92.

[0061] The input device 93 includes keys, buttons, pointing devices, and touch panels, and is used for inputting information. The output device 94 is, for example, a liquid crystal display, and displays information and data.

[0062] Communication IF95 is a circuit that performs communication processing. For example, communication IF95 is a network interface card (NIC). Alternatively, communication IF95 may be a wireless communication circuit that performs wireless communication (5G, wireless LAN (Wi-Fi®), BLE, etc.). Furthermore, communication IF95 may be a combination of a circuit that performs wired communication processing and a wireless communication circuit.

[0063] The control unit 91 performs various processes by executing various programs stored in the storage device 92. By the control unit 91 executing the programs stored in the storage device 92, the information processing device 90 can operate as NF11a to 11k, OAM terminals, and external servers, respectively.

[0064] <Other Embodiments> The embodiments and modifications described above are merely examples, and this disclosure may be modified as appropriate without departing from its essence. Furthermore, the processes and means described in this disclosure can be freely combined and implemented as long as no technical inconsistencies arise.

[0065] While four examples of data error types have been given—data-level errors, sample-level errors, feature-level errors, and client-level errors—the types of data errors do not necessarily have to include all four, and may also include types other than these four. Furthermore, the recovery method when a specific type of data error occurs is not limited to the methods shown above. This disclosure can be understood as an invention that, when a data error occurs, determines which of the multiple types of data error has occurred and takes a recovery method corresponding to the error type.

[0066] The above explanation used an example where associative learning is vertical associative learning, but similar data error recovery processing can be applied to horizontal associative learning as well. Furthermore, while the example shown uses a neural network as the learning model, the learning model may also be constructed based on other algorithms such as decision trees.

[0067] 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.

[0068] The client device in this disclosure may be any device. For example, when applying the technology of this disclosure to a mobile communication system, the client device may be a user terminal (UE), a base station, or any network function / network entity that constitutes the core system. The same applies to the server device.

[0069] Furthermore, the client device may notify the core network in advance of the possible solutions from among several possible solutions as Capability information. The Capability information may be stored in the core network's known or new NF. The server device may refer to the Capability information and have the client device execute the possible solution.

[0070] This disclosure can be understood as an information processing system as a whole, or as the server and client devices that constitute said information processing system.

[0071] The information processing system according to this disclosure is an information processing system comprising a plurality of participating nodes that perform federated learning based on locally data obtained sequentially, and includes: detection means for detecting a data anomaly in at least one of the plurality of participating nodes; and control means for determining which of a plurality of response methods to implement when the data anomaly occurs.

[0072] The server device relating to this disclosure is an information processing device that participates as a server device in an information processing system which consists of a plurality of participating nodes and performs federated learning based on locally data obtained sequentially, and has a processor that performs: a receiving step of receiving data from a client device from at least one of the plurality of participating nodes indicating that a data anomaly has occurred; a control step of determining which of a plurality of response methods to implement when the data anomaly occurs; and a notification step of notifying the plurality of client devices to implement the determined response method.

[0073] The client device relating to this disclosure is an information processing device that participates as a client device in an information processing system which consists of multiple participating nodes and performs federated learning based on locally data obtained sequentially, and has a processor that performs: a transmission step of sending data indicating that a data anomaly has occurred to a server device; a reception step of receiving a notification from the server device indicating which of several response methods to the data anomaly should be implemented; and an anomaly recovery step of implementing the response method notified by the server device.

[0074] 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. The non-temporary computer-readable storage medium includes 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, or optical cards, and any type of medium suitable for storing electronic instructions.

[0075] 100: Server device 200: Client device

Claims

1. An information processing system comprising multiple participating nodes that performs federative learning based on locally obtained data sequentially, comprising: detection means for detecting data anomalies in at least one of the multiple participating nodes; and control means for determining which of a plurality of response methods to implement when a data anomaly occurs.

2. The information processing system according to claim 1, wherein the plurality of response methods include at least two of the following: correcting the data in which an anomaly occurred and continuing learning, re-executing data alignment and continuing learning, and retraining the model.

3. The information processing system according to claim 1, characterized in that the detection means also determines the type of the detected data anomaly, and the control means determines which control to implement, taking into consideration the type of data anomaly.

4. The information processing system according to claim 3, wherein the information processing system includes a plurality of participating nodes, each having local data with different feature quantities for the same sample, and performs vertical federated learning, and the types of data anomalies include at least two of the following: a first type of data anomaly in which some feature quantities are missing for some samples; a second type of data anomaly in which data is missing on a sample-by-sample basis; a third type of data anomaly in which data is missing on a feature-by-feature basis; and a fourth type of data anomaly in which all data is missing or the participating node itself malfunctions.

5. The information processing system according to claim 4, characterized in that, when a data anomaly of the first type occurs, the control means controls the system to correct the anomalyed data and continue learning.

6. The information processing system according to claim 4, characterized in that, when a data anomaly of the second or third type occurs, the control means controls the system to re-execute data alignment and continue learning.

7. The information processing system according to claim 4, characterized in that, when a data anomaly of the fourth type occurs, the control means controls the model to retrain.

8. The information processing system according to claim 1, characterized in that the control means estimates the learning accuracy of the learning model when the plurality of corresponding methods are adopted and determines which corresponding method to adopt by comparing it with a threshold.

9. An information processing device in an information processing system composed of multiple participating nodes that performs federative learning based on locally obtained data sequentially, the information processing device having a processor that performs: a receiving step of receiving data indicating that a data anomaly has occurred from at least one of the multiple participating nodes; a control step of determining which of a plurality of response methods to implement when the data anomaly occurs; and a notification step of notifying the multiple participating nodes to implement the determined response method.

10. The information processing apparatus according to claim 9, wherein the plurality of response methods include at least two of the following: correcting the data in which an anomaly occurred and continuing learning, re-executing data alignment and continuing learning, and retraining the model.

11. The information processing apparatus according to claim 9, characterized in that, in the control step, the type of data anomaly is also determined from the data indicating that the data anomaly has occurred, and a decision is made on which control to implement, taking into consideration the type of data anomaly.

12. The information processing apparatus according to claim 9, characterized in that the data indicating the occurrence of the data anomaly includes the type of the data anomaly, and in the control step, a decision is made on which control to implement, taking into consideration the type of data anomaly.

13. The information processing system according to claim 11 or 12, wherein the information processing system includes a plurality of participating nodes, each having local data with different feature quantities for the same sample, and is a system that performs vertical federated learning using the plurality of participating nodes, and the types of data anomalies include at least two of the following: a first type of data anomaly in which some feature quantities are missing for some samples; a second type of data anomaly in which data is missing on a sample-by-sample basis; a third type of data anomaly in which data is missing on a feature-by-feature basis; and a fourth type of data anomaly in which all data is missing or the participating node itself malfunctions.

14. An information processing method in an information processing system that consists of multiple participating nodes and performs federative learning based on locally obtained data sequentially, comprising: a detection step of detecting a data anomaly in at least one of the multiple participating nodes; and a control step of determining which of a plurality of response methods to implement when the data anomaly occurs.

15. An information processing method performed by a management node in an information processing system composed of multiple participating nodes that performs federative learning based on locally obtained data sequentially, the method comprising: a receiving step of receiving data indicating that a data anomaly has occurred from at least one of the multiple participating nodes; a control step of determining which of a plurality of response methods to implement when the data anomaly occurs; and a notification step of notifying the multiple participating nodes to implement the determined response method.

16. A program for causing a computer to perform each step of the information processing method described in claim 14 or 15.

Images