Information processing apparatus and information processing method

Abstract

To provide a technology for cooperatively detecting abnormality without disclosing localization data by a client who participates in federated learning.SOLUTION: A learning step of acquiring a first intermediate result based on learning data from each of the plurality of client devices and obtaining a probability distribution of a value of the learning data from the first intermediate result by federated learning, and an anomaly detection step of acquiring a second intermediate result based on detection target data from each of the plurality of client devices and determining whether or not an anomaly has occurred in the detection target data using the second intermediate result and the probability distribution.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device and an information processing method. [Background technology]

[0002] Federated Learning (FL) is a task in which multiple entities keep data localized without sharing it, and use that localized data to train a machine learning model. In federated learning, each piece of localized data has an owner (called a client) with permission to view it, and no entity can view all of the data. Therefore, even if an abnormality could be detected if the entire data were viewable, there are cases in which the abnormality of the localized data cannot be detected from the perspective of each client that can only view the localized data.

[0003] In federated learning, if an anomaly occurs in the data held by the client, the machine learning model cannot be trained properly, so it is desirable to detect anomalies in the data.

[0004] Patent Document 1 discloses a method for detecting anomalies in federated learning without the client disclosing data. Specifically, Patent Document 1 discloses a method for detecting attacks in federated learning, in which a malicious participant (client) inserts an exit (subtree) that leaks input data into a decision tree, in which a decision tree-based model is passed between clients. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2024-17375 Summary of the Invention [Problem to be solved by the invention]

[0006] Patent Document 1 targets a specific attack, namely the insertion of a subtree, in the associative learning of a learning model using a decision tree. Patent Document 1 also assumes supervised associative learning. However, it is desirable to be able to detect data anomalies in a more general way.

[0007] The present disclosure aims to provide a technology that enables clients participating in federated learning to cooperatively detect anomalies without disclosing localized data. [Means for solving the problem]

[0008] One aspect of the present disclosure is an information processing device for detecting anomalies in a federated learning system including a plurality of client devices, the information processing device comprising a control unit that executes a learning step of acquiring a first intermediate result based on learning data from each of the plurality of client devices and determining a probability distribution of values ​​of the learning data from the first intermediate result by federated learning, and an anomaly detection step of acquiring a second intermediate result based on detection target data from each of the plurality of client devices and determining whether an anomaly has occurred in the detection target data using the second intermediate result and the probability distribution.

[0009] Another aspect of the present disclosure is an information processing method performed by an information processing device for detecting an anomaly in a federated learning system including a plurality of client devices, the method including: a learning step of acquiring a first intermediate result based on learning data from each of the plurality of client devices, and determining a probability distribution of values ​​of the learning data from the first intermediate result by federated learning; and an anomaly detection step of acquiring a second intermediate result based on the detection target data from each of the client devices, and determining whether or not an anomaly has occurred in the detection target data using the second intermediate result and the probability distribution.

[0010] Other aspects include an information processing system composed of 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. [Effects of the Invention]

[0011] According to the present disclosure, clients participating in federated learning can cooperatively detect anomalies without disclosing localized data. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating an overview of an information processing system according to the present disclosure. [Figure 2] FIG. 2 is a diagram illustrating an overview of the anomaly detection method according to the present disclosure. [Figure 3] FIG. 3 is a diagram illustrating a configuration of an information processing device according to the present disclosure. [Figure 4] FIG. 4 is a flowchart showing the flow of the pre-learning step in the anomaly detection process according to the present disclosure. [Figure 5] FIG. 5 is a flowchart showing the flow of the anomaly detection step in the anomaly detection process according to the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating data stored in each client device according to the first embodiment. [Figure 7] FIG. 7 is a diagram illustrating the pre-learning step in the first embodiment. [Figure 8] FIG. 8 is a diagram illustrating the abnormality detection step in the first embodiment. [Figure 9] FIG. 9 is a diagram illustrating the abnormality detection process according to the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. The configurations of the following embodiments are examples, and the present disclosure is not limited to the configurations of the embodiments.

[0014] <Embodiment 1> (overview) The present disclosure relates to a data anomaly detection method in an information processing system that performs federated learning. FIG. 1 is a diagram illustrating a configuration of an information processing system according to the present disclosure. As illustrated in the diagram, the information processing system according to this embodiment includes at least one server device 100 and at least one client device 200 connected to N networks. Each of the client devices 200 participates in federated learning. The localized data held by each of the client devices 200 is not shared with other devices. In this situation, the server device 100 detects an anomaly in the data held by any of the client devices 200.

[0015] An overview of an anomaly detection method in an information processing system according to the present disclosure will be described with reference to FIG.

[0016] In this embodiment, each of the client devices 200 transmits to the server device 100 at least values ​​that occur during the learning, analysis, and inference calculations or values ​​that are finally obtained by those calculations. In this embodiment, these transmitted and received values ​​are called intermediate results. The intermediate results may be, for example, the output of a layer in the middle of the forward propagation process in a neural network, or in the case of a matrix decomposition model, all or part of either or both of the matrices A and B where D≒AB is obtained by decomposing the original table data D. The intermediate results may be gradient data used for model updating or the model itself, or the encrypted results of these may be used as the intermediate results.

[0017] The server device 100 performs pre-learning of the probability distribution of localized data based on the intermediate results of the existing data transmitted from the client device 200.

[0018] The client device 200 calculates the same intermediate result as above based on the data added thereafter and transmits it to the server device 100.

[0019] The server device 100 detects anomalies in the data based on the transmitted intermediate results and the probability distribution learned by pre-learning. For example, if the likelihood of the localized data calculated based on the probability distribution is equal to or less than a threshold, it determines that an anomaly has occurred.

[0020] (composition) 3 schematically illustrates an example of the hardware configuration of an information processing device (computer) 300 that operates as the server device 100 or the client device 200 in this embodiment. As illustrated in Fig. 3, the information processing device 300 according to this embodiment is a computer in which a control unit 301, a storage unit 302, an input device 303, an output device 304, and a communication interface 305 are electrically connected.

[0021] The control unit 301 is a hardware processor, such as a CPU (Central Processing Unit). The control unit 301 includes a RAM (Random Access Memory), a ROM (Read Only Memory), etc., and is configured to execute information processing based on programs and various data. The control unit 301 (CPU) is an example of a processor resource.

[0022] The storage unit 302 is an example of a memory resource, and is configured, for example, by a hard disk drive, a solid state drive, etc. In this embodiment, the storage unit 302 stores various information such as a program describing information processing operations and data 4 that is the target of information processing.

[0023] The input device 303 is a device for inputting information, such as a mouse, keyboard, keypad, or touch panel device. The output device 304 is a device for outputting information, such as a display or speaker. A user can operate the information processing device 300 by using the input device 303 and the output device 304. The input device 303 and the output device 304 may be integrated into one device, such as a touch panel display.

[0024] The communication interface 305 is, for example, a wired LAN (Local Area Network) module, a wireless LAN module, etc., and is an interface for performing wired or wireless communication via a network. The information processing device 300 can perform data communication with other computers via the communication interface 305.

[0025] It should be noted that, with regard to the specific hardware configuration of the information processing device 300, components may be omitted, replaced, or added as appropriate depending on the embodiment. For example, the control unit 301 may include multiple hardware processors. The hardware processors may be a microprocessor, a field-programmable gate array (FPGA), a digital signal processor (DSP), etc. The storage unit 302 may be configured by RAM and ROM included in the control unit 301. At least one of the input device 303, the output device 304, and the communication interface 305 may be omitted. The server device 100 and the client device 200 may be configured by a plurality of computers. In this case, the hardware configurations of the computers may or may not be the same. In addition, the server device 100 and the client device 200 may be configured by a plurality of computers. The client device 200 may be an information processing device designed specifically for the service to be provided, a general-purpose server device, a general-purpose PC (Personal Computer), an industrial PC, or the like.

[0026] The server device 100 and the client device 200 cooperate to detect anomalies in data held by the client device 200 by the control unit 301 executing a program.

[0027] (process) The anomaly detection process according to the present disclosure has two main steps. The first step is a pre-learning step in which a model for estimating the probability distribution of data is learned through pre-learning based on existing data or learning data. The second step is an anomaly detection step in which it is determined whether an anomaly has occurred in new data or data subject to anomaly detection.

[0028] FIG. 4 is a flowchart showing the flow of the pre-learning steps.

[0029] In step S401, each of the client devices 200 calculates an intermediate result from the learning data. The intermediate result is a value generated in the calculation of the learning process, and is a value from which the original data cannot be obtained. In step S402, each of the client devices 200 transmits the intermediate result to the server device 100.

[0030] In step S403, the server device 100 receives intermediate results from each of the client devices 200. In step S403, the server device 100 updates a model for estimating a probability distribution of data based on the obtained intermediate results. Any model may be used, and for example, a neural network, a matrix decomposition model, a decision tree model, or the like may be adopted. In step S405, the server device 100 transmits parameters for updating the model in the client device 200 to the client device 200.

[0031] In step S406, client device 200 receives the update parameters from server device 100. In step S407, client device 200 updates the local model based on the update parameters. After step S407, the processes from step S401 onwards may be repeated. Note that in the above process, steps S405 to S407 may be omitted, and the entire process need not be repeated multiple times.

[0032] FIG. 5 is a flowchart showing the flow of the anomaly detection step.

[0033] In step S501, each of the client devices 200 calculates an intermediate result from the detection target data. In step S502, each of the client devices 200 transmits the intermediate result to the server device 100.

[0034] In step S503, the server device 100 receives the intermediate results from each of the client devices 200. In step S504, the server device 100 determines whether the data is within a possible normal range based on the acquired intermediate results. For example, the likelihood of the detection target data is calculated using the received intermediate results and a model obtained by pre-learning, and if the likelihood is equal to or less than a threshold, it can be determined that the data is outside the normal range. Alternatively, the possible range of the intermediate results can be calculated using a model obtained by pre-learning, and if the acquired intermediate results are outside this range, it can be determined that the data is outside the normal range. Note that the determination of whether the intermediate results are within the normal range is performed for each piece of data held by the client device 200. If the intermediate results are outside the normal range (S505-YES), in step S506 The server device 100 can determine that an abnormality has occurred in the detection target data held by the client device 200 that sent this intermediate result, and notifies the client device 200 or an administrator to that effect.

[0035] Example 1 This embodiment is an unsupervised anomaly detection method using vertical federated learning.

[0036] First, the data held by each client will be explained with reference to FIG. 6. Here, it is assumed that there are K client devices (K is an integer equal to or greater than 1). Each client device i (i is an integer equal to or greater than 1 and equal to or less than K) holds Mi different feature quantities for a common sample. When expressing the features held by client i collectively, d i := (d i,1 , …, di, Mi ) to represent the entire data held by all clients, D := (d1, d2, …, d M ) is expressed as follows. The data held is stored as preprocessed values ​​such as categorical variables and real numbers so that they can be input into machine learning models or statistical models.

[0037] Here, it is assumed that a sample ID (uid) is assigned to each sample. In other words, when each client is queried for the sample ID in the nth row, it is guaranteed that it is the feature of the same sample. Sample IDs can be assigned using a known data alignment method. In the following, when referring to data with a specific uid, we will emphasize that it is data in the nth row and use d n , d n i , d n i,Mi Write the UID on the right side like this: Posted.

[0038] In this embodiment, since unsupervised anomaly detection is assumed, none of the clients possesses the truth labels or the objective variables. However, even if the clients possess the truth labels or the objective variables, this does not prevent the application of the technology of this embodiment.

[0039] The anomaly detection method of this embodiment roughly consists of two steps. The first step is to perform principal component analysis with a vertical association learning problem setting through prior learning based on existing data. The second step is to perform anomaly detection and anomaly analysis on the target data using the weight matrix obtained by the principal component analysis.

[0040] [First step: Privacy-Preserving Principal Component Analysis (PPPCA)] The first step will be described with reference to Figure 7. Learning data (existing data) D∈R N×M The data column is M=M1+...+MK The system is divided vertically into K clients, and managed so that they cannot access each other. D=[d1,…,d N ]d n ∈R M corresponds to the nth sample, and consists of a total of N samples.

[0041] Consider a data matrix D, which is subjected to principal component analysis so that D ≒ WP. T = I, W∈R N ×r , P∈R r×M , I∈R r×r where matrix P is an orthogonal matrix and M-dimensional basis vectors It is composed of r pieces of W=[w1,…, w N ] vector w n ∈R r is a set of r basis vectors The weight of the torque is determined, and the data d n It is used to restore d n =w n P by d n can be restored.

[0042] If matrices W and P are managed in the same place, D can be restored from W and P. To avoid this, W is managed by the server device (platformer), and P is managed by dividing it vertically for each client. By managing it in this way, the server device cannot restore the data matrix D. Also, P:=[P1, ..., P K ], K clients store matrix P in vertical partitions. By doing so, each client k =WP k This type of privacy-preserving principal component analysis can be achieved by the method in the following document, for example, but other methods are also available. The method may be adopted.

[0043] Jaberi, Mehrad, and Hamid Mala. "Privacy-preserving multi-party PCA computation on horizontally and vertically partitioned data based on outsourced QR decomposition." The Journal of Supercomputing 79.13 (2023): 14358-14387.

[0044] In privacy-preserving principal component analysis, clients send calculation results (intermediate results) based on localized data to a server device, which then aggregates the calculation results from each client and feeds back parameters for matrix calculation. Principal component analysis can be performed by repeating this process between the client device and the server device. In this example, the probability distribution p of the decomposition matrices P, W, and w corresponds to a model for determining the range or distribution of possible values ​​of the intermediate results.

[0045] Data point d n is a point w on the r-dimensional subspace spanned by the basis vectors of P n The server device is expressed as w1, ..., w N The probability distribution p(w) of w is estimated from the following equation. For example, this can be done by assuming a statistical model such as a mixed normal distribution and estimating its parameters using an optimization algorithm such as maximum likelihood estimation, Bayesian estimation, or the EM algorithm.

[0046] [Second step: Anomaly detection and analysis using the weight matrix obtained from principal component analysis] When new data D' is obtained, using the principal component matrix P calculated in the first step, D' can be written as D'=W'P+Δ. In other words, the weight matrix is ​​calculated from the data D' and the principal component matrix P. Column W':=[w'1, …, w' N ] and residual Δ:=[δ1, …, δ N ] can be calculated. Projection of D' to W' The order is p which each client manages separately. K ∈R r×MK Using d' k P T k ∈R J×r This can be done by calculating and adding the results. k P T k If you do not want to disclose it to a third party In this calculation of W' and Δ, secrecy can be maintained by using additive homomorphic encryption. The client device sends the calculation result (intermediate result) for the data D' to the server device. The client device and the server device can work together to find W' and Δ.

[0047] The server device calculates w' using the probability distribution p(w) obtained in the first step. i Evaluate each likelihood. If the probability distribution p(w) is normal, this evaluation is Hotelling t 2 This corresponds to anomaly detection. Generally, p(w) is not necessarily normally distributed, so anomaly detection is performed according to the likelihood. w' i If the likelihood of is less than a threshold (e.g., 5%), then this w' i And the corresponding d' i It can be determined that an abnormality has occurred.

[0048] The client calculates Δ=D'-W'P, which is the leaked subspace when D' is projected onto W'. Among the residuals δi, the d' corresponding to the column that is greater than the threshold is i It is determined that an abnormality has occurred in The client may make the abnormality determination and notify the server of the determination result, or the client may send Δ to the server, which may then make the abnormality determination.

[0049] By detecting anomalies using the probability distribution p(w), it becomes possible to detect anomalies that cannot be detected based on residuals. The fact that the data d' has a low likelihood in the subspace means that it is an unusual basis vector. Even with such an anomaly, the combination of basis vectors As long as it is constructed, there are no residuals and anomaly detection cannot be performed using only residuals. Anomaly detection using the probability distribution p(w) is a method for detecting the use of unknown basis vectors, and is complementary to anomaly detection based on residuals.

[0050] Example 2 Similar to the first embodiment, this embodiment is also an unsupervised anomaly detection method using vertical federated learning, and includes two steps: a pre-learning step and an anomaly detection step.

[0051] In this embodiment, some of the data (vector) X is considered as feature quantities and objective variables, and the remaining Anomaly detection is performed by regarding all or some of the variables as explanatory variables. In other words, the data set has an input-output relationship (X -m , x m ) and the probability distribution p(x m |X -m ) and perform unsupervised anomaly analysis. -m is the vector X to x m represents a vector excluding

[0052] In the first step, the pre-training step, as shown in Figure 9(A), the input-output data set (X -m , x m ) to the probability distribution p(x m |X -m ) is learned through federated learning. A regression tree-based method is known as such a learning method. In other words, the function f is derived from a regression tree or a decision tree. In addition to regression tree-based methods, there is also a neural network-based method called NGBoost. Network-based methods may be used to train the regression function.

[0053] In this embodiment, the data x m The device that holds the data acts as a server and mIn federated learning, data x m The device with data X is the active participant, and the other devices are the passive participants. -m From x m Learning a regression function to estimate During this learning process, passive participants -m Actual intermediate results based on The active participants update the model based on the intermediate results and send the updated parameters of the model to the passive participants, who then update their local models based on the updated parameters.

[0054] In the second step, the anomaly detection step, as shown in Figure 9(B), the passive participant applies new data to its local model and sends the intermediate results to the active participant. The active participant then applies the intermediate results to its own local model to calculate the likelihood of the new data. If the likelihood is smaller than a threshold (e.g., 5%), the active participant rejects the data x m It is determined that an abnormality has occurred.

[0055] In this example, the probability distribution p(x m |X -m ), so x m This allows for likelihood evaluation of This makes likelihood-based anomaly detection possible. Anomaly detection methods that utilize probability distributions simply use decision trees or regression trees to find the regression function x m =f(X -m ), rather than anomaly detection methods that learn x m Many Anomaly detection is possible effectively in cases where the distribution follows a peak. One of the features of this embodiment is that the probability distribution p(x m |X -m ) rather than using decision tree or regression tree based methods to learn The point is to learn a probability distribution, and the learning method is not limited to decision tree or regression tree based methods, and any method can be used.

[0056] The above probability distribution p(x m |X -m ) and anomaly detection using the learning for all m, it becomes possible to detect anomalies occurring in any feature of any data. That is, each information processing device functions as a server in learning a probability distribution for estimating its own feature, and also functions as a client in learning a probability distribution for estimating feature quantities held by other devices. More specifically, when functioning as a server, in the learning step, a probability distribution for at least one feature of the data held by the information processing device when other features are given is calculated through federated learning with the other devices. In addition, in the anomaly detection step, a second interim result based on the detection target data is obtained from the other devices, and a likelihood of the at least one feature of the detection target data is calculated based on the second interim result and the probability distribution. If the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the at least one feature of the detection target data. On the other hand, when functioning as a client, in the learning step, a first interim result is calculated based on the training data and transmitted to the other devices in order to perform federated learning of a probability distribution for at least one feature of the data held by the other devices when other features are given. In addition, in the anomaly detection step, a second intermediate result based on the detection target data is calculated and transmitted to the other device in order to calculate the likelihood of at least one feature of the data held by the other device.

[0057] <Other embodiments> The above-described embodiment and modifications are merely examples, and the present disclosure may be modified as appropriate within the scope of the present disclosure. Furthermore, the processes and means described in the present disclosure may be freely combined and implemented as long as no technical contradiction occurs.

[0058] The technique of the present disclosure is applicable not only to vertical associative learning but also to horizontal associative learning.

[0059] The information processing system of the present disclosure may be any system. As an example, the information processing system of the present disclosure may be a 5G system, which includes various NFs (Network Functions) such as NWDAF (Network Data Analytics Function), AMF (Access and Mobility Management Function), SMF (Session Management Function), UDM (Unified Data Management), PCF (Policy Control Function), NRF (Network Repository Function), NEF (Network Exposure Function), and AF (Application Function) and OAM (Operation, Administration, and Maintenance) devices. The client device and server device of the present disclosure may be any of these NFs and OAM terminals.

[0060] The present disclosure can also be realized by providing a computer program implementing the functions described in the above embodiments 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 via a non-transitory computer-readable storage medium connectable to the computer's system bus or via a network. Non-transitory computer-readable storage media include any type of medium suitable for storing electronic instructions, such as any type of disk, including magnetic disks (e.g., floppy disks, hard disk drives (HDDs), etc.), optical disks (e.g., CD-ROMs, DVDs, Blu-ray disks), read-only memory (ROM), random access memory (RAM), EPROM, EEPROM, magnetic cards, flash memory, or optical cards. [Explanation of symbols]

[0061] 100: Server device 200: Client device 300: Information processing device (server device, client device) 301: Control unit

Claims

1. An information processing device for detecting anomalies in a federated learning system including a plurality of client devices, a learning step of acquiring a first intermediate result based on training data from each of the plurality of client devices and determining a probability distribution of values ​​of the training data from the first intermediate result by federated learning; an anomaly detection step of acquiring a second interim result based on the detection target data from each of the plurality of client devices, and determining whether or not an anomaly has occurred in the detection target data using the second interim result and the probability distribution; An information processing device comprising a control unit that executes the above.

2. In the anomaly detection step, a likelihood of a value of the detection target data is calculated based on the second intermediate result and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the detection target data.

2. The information processing apparatus according to claim 1, wherein:

3. the learning step estimates the probability distribution from a weight matrix obtained by performing privacy-preserving principal component analysis on an intermediate result based on the learning data; In the anomaly detection step, a likelihood of a value of the detection target data is calculated based on the second intermediate result and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the detection target data.

3. The information processing apparatus according to claim 2, wherein:

4. In the anomaly detection step, the detection target data is determined to be anomaly based on a residual when the detection target data is projected using a projection matrix obtained by performing privacy-preserving principal component analysis on an intermediate result based on the learning data.

4. The information processing apparatus according to claim 3,

5. the information processing device is one of the plurality of client devices, In the learning step, a probability distribution for at least one feature of data held by the information processing device when a feature other than the at least one feature is given is obtained by federated learning with the plurality of client devices other than the information processing device; In the anomaly detection step, the second interim results based on the detection target data are acquired from the plurality of client devices other than the information processing device, a likelihood of the at least one feature of the detection target data is calculated based on the second interim results and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the at least one feature of the detection target data.

3. The information processing apparatus according to claim 2, wherein:

6. In the learning step, in order to perform associative learning of a probability distribution for at least one feature of data held by another client device when a feature other than the one feature is given, a first intermediate result is calculated based on the learning data and transmitted to the other client device; In the anomaly detection step, a second intermediate result is calculated based on the detection target data to calculate a likelihood of at least one feature of the data held by the other client device, and the second intermediate result is transmitted to the other client device.

6. The information processing apparatus according to claim 5,

7. Information for detecting anomalies in a federated learning system including multiple client devices An information processing method performed by a processing device, a learning step of acquiring a first intermediate result based on training data from each of the plurality of client devices and determining a probability distribution of values ​​of the training data from the first intermediate result by federated learning; an anomaly detection step of acquiring a second interim result based on the detection target data from each of the plurality of client devices, and determining whether or not an anomaly has occurred in the detection target data using the second interim result and the probability distribution; An information processing method including:

8. In the anomaly detection step, a likelihood of a value of the detection target data is calculated based on the second intermediate result and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the detection target data.

8. The information processing method according to claim 7,

9. the learning step estimates the probability distribution from a weight matrix obtained by performing privacy-preserving principal component analysis on an intermediate result based on the learning data; In the anomaly detection step, a likelihood of a value of the detection target data is calculated based on the second intermediate result and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the detection target data.

9. The information processing method according to claim 8.

10. In the anomaly detection step, the detection target data is determined to be anomaly based on a residual when the detection target data is projected using a projection matrix obtained by performing privacy-preserving principal component analysis on an intermediate result based on the learning data.

10. The information processing method according to claim 9.

11. the information processing device is one of the plurality of client devices, In the learning step, a probability distribution for at least one feature of data held by the information processing device when a feature other than the at least one feature is given is obtained by federated learning with the plurality of client devices other than the information processing device; In the anomaly detection step, the second interim results based on the detection target data are acquired from the plurality of client devices other than the information processing device, a likelihood of the at least one feature of the detection target data is calculated based on the second interim results and the probability distribution, and if the likelihood is equal to or less than a threshold, it is determined that an anomaly has occurred in the at least one feature of the detection target data.

9. The information processing method according to claim 8.

12. In the learning step, in order to perform associative learning of a probability distribution for at least one feature of data held by another client device when a feature other than the one feature is given, a first intermediate result is calculated based on the learning data and transmitted to the other client device; In the anomaly detection step, a second intermediate result is calculated based on the detection target data to calculate a likelihood of at least one feature of the data held by the other client device, and the second intermediate result is transmitted to the other client device.

12. The information processing method according to claim 11.

13. A program for causing an information processing device to execute each step of the information processing method according to any one of claims 7 to 12.

Images