Data generation device

Abstract

This data generation device (10) according to the present disclosure is provided with: a generation unit (11) that, on the basis of a first parameter, generates false data by entering noise data into a generation model that generates false data imitating original data which is time-series data; a conversion unit (12) that generates data to be determined by converting data to be processed, which is one of the false data and the original data, on the basis of an arbitrary time-series feature quantity; and a detection unit (13) that, on the basis of a second parameter, determines whether the data to be determined is the original data or false data by entering the data to be determined into a determination model that determines whether input data is the original data or false data. The generation unit (11) updates the first parameter on the basis of the determination result, and the detection unit (13) updates the second parameter on the basis of the determination result.

Description

Data Generator

[0001] The present disclosure relates to a data generating device.

[0002] Non-Patent Document 1 describes a technology called a Generative Adversarial Network (GAN), which generates data (fake data) that imitates original data. Fig. 10 is a diagram showing an example of the configuration of a data generation device 1 that generates fake data using a GAN.

[0003] As shown in FIG. 10 , the data generating device 1 includes a generating unit 11 and a detecting unit 13 .

[0004] Noise data (random noise) is input to the generation unit 11. The generation unit 11 generates the fake data by inputting the input noise data to a machine learning model (generative model) that generates fake data that imitates original data from input data based on parameters (first parameters).

[0005] The detection unit 13 randomly inputs either the fake data or the original data generated by the generation unit 11 as data to be determined. The detection unit 13 inputs the input data to be determined into a machine learning model (determination model) that determines whether the input data is original data or fake data based on a parameter (second parameter), and determines whether the data to be determined is original data or fake data.

[0006] The generation unit 11 updates the parameters of the generative model (first parameters) based on the result of the judgment by the detection unit 13. Specifically, the generation unit 11 evaluates how much the detection unit 13 made an incorrect judgment, and updates the parameters of the generative model based on the result. The detection unit 13 also updates the parameters of the judgment model (second parameters) based on the result of the judgment. Specifically, the detection unit 13 evaluates the result of the judgment using a loss function, and updates the parameters of the judgment model based on the result. The updating of the parameters of the generative model and the judgment model is repeated (iterative learning) until the error rate of the detection unit 13 is minimized.

[0007] When data is generated after the generative model and discriminative model have been trained, false data close to the original data can be generated by inputting noise data into the generative model after adjusting the parameters.

[0008] GANs are generally used for image generation and audio generation. This is because images have hundreds of thousands of pixels, making it possible to efficiently capture spatial features. Audio generally has a high sampling rate, such as 1 kHz, making it easy to capture temporal features.

[0009] Non-Patent Document 2 describes a conditional GAN ​​that uses additional information to generate fake data and determine whether the data is original or fake.

[0010] “Generative Adversarial Networks”, Ian J. Goodfellow et al. (2014) “Conditional Generative Adversarial Nets”, Mehdi Mirza and Simon Osindero (2014)

[0011] As mentioned above, GANs are generally applied to large data such as image or audio generation. However, when GANs are applied to small data such as traffic data in a communication network, it becomes difficult to maintain a balance between the generator and the detector. For example, if the generator can fool the detector with the generated false data once, it will only generate similar false data until the detector changes the parameters of the decision model (mode collapse). Furthermore, if the generator cannot fool the detector with the generated false data even once, the parameters of the generative model will not converge.

[0012] Furthermore, for example, traffic in a communication network fluctuates over time depending on the time of day (day of the week, day or night, etc.) or events (for example, game app updates, online live performances, etc.) Conventional GANs have the problem of being unable to generate data that takes into account time-series features (time of day, events, etc.).

[0013] In view of the above circumstances, an object of the present disclosure is to provide a data generation device that can generate more diverse fake data by taking into account time-series feature quantities.

[0014] In order to solve the above problem, the data generation device according to the present disclosure includes a generation unit that inputs noise data into a generation model that generates fake data that imitates original data, which is time-series data, based on a first parameter, to generate the fake data; a conversion unit that generates data to be determined by converting processing target data, which is either the fake data or the original data, based on an arbitrary time-series feature; and a detection unit that inputs the data to be determined into a determination model that determines whether input data is the original data or the fake data based on a second parameter, and determines whether the data to be determined is the original data or the fake data, wherein the generation unit updates the first parameter based on a result of the determination, and the detection unit updates the second parameter based on a result of the determination.

[0015] According to the data generation device according to the present disclosure, it is possible to generate more diverse fake data that takes into account time-series characteristics.

[0016] FIG. 1 is a diagram illustrating an example configuration of a data generating device according to a first embodiment of the present disclosure. FIG. 2 is a diagram illustrating an example of a feature quantity shown in FIG. 1. FIG. 3 is a diagram illustrating an example configuration of a conversion unit shown in FIG. 1. FIG. 4 is a flowchart illustrating an example of an operation of the data generating device shown in FIG. 1. FIG. 5 is a diagram illustrating an example configuration of a data generating device according to a second embodiment of the present disclosure. FIG. 6 is a diagram illustrating an example configuration of a data generating device according to a third embodiment of the present disclosure. FIG. 7 is a diagram illustrating an example configuration of a data generating device according to a fourth embodiment of the present disclosure. FIG. 8 is a diagram illustrating an example configuration of a data generating device according to a fifth embodiment of the present disclosure. FIG. 9 is a diagram illustrating an example hardware configuration of a data generating device according to the present disclosure. FIG. 10 is a diagram illustrating an example configuration of a conventional data generating device.

[0017] (First Embodiment) FIG. 1 is a diagram illustrating an example configuration of a data generating device 10 according to a first embodiment of the present disclosure. The data generating device 10 according to the present disclosure generates fake data that imitates original data. The original data is, for example, one-dimensional time series data of an arbitrary length seq (1×seq length). The original data is, for example, traffic data in a communication network, but the present disclosure is not limited to this and can be applied to various time series data. In FIG. 1, components similar to those in FIG. 10 are denoted by the same reference numerals, and descriptions thereof will be omitted as appropriate.

[0018] 1, a data generating device 10 according to this embodiment includes a generating unit 11, a converting unit 12, and a detecting unit 13. The data generating device 10 according to this embodiment differs from the data generating device 1 shown in FIG. 10 in that a converting unit 12 is added.

[0019] The generator 11 receives noise data, which is random noise according to a probability distribution or the like. The noise data is Dim-dimensional data (N x Dim) with a length of N. The generator 11 generates the false data by inputting the noise data into a generative model that generates false data that imitates original data, which is time-series data, from the input data based on a parameter (first parameter). That is, the generator 11 generates, from the noise data, false data that is one-dimensional time-series data (1 x seq length) of an arbitrary length seq, the same as the original data. The generative model is, for example, a machine learning model such as a CNN (Convolutional Neural Network) or a Transformer, but is not limited to these.

[0020] The conversion unit 12 receives as input the original data and the fake data generated by the generation unit 11. The conversion unit 12 also receives as input any time-series feature. The time-series feature is, for example, a feature representing a time period such as a month, a week, or a date and time, or a feature representing an event, but is not limited to these, and may be any feature corresponding to the time-series feature required for the data generated by the data generation device 10.

[0021] The converter 12 randomly selects either the pseudo data or the original data as the processing target data, and generates data to be judged by transforming the processing target data based on any input time-series feature. Specifically, the converter 12 generates the data to be judged by embedding a feature vector obtained by vector-transforming the input feature into the processing target data, or by adding the feature vector to the processing target data. The converter 12 may generate the data to be judged by transforming the processing target data based on the feature vector using, for example, a process called embedding or encoding. The converter 12 may also generate the data to be judged by transforming the processing target data based on the feature vector using, for example, a predetermined function (e.g., a trigonometric function, a Fourier transform representing frequency feature values, a wavelet transform representing the relationship between time and frequency, etc.). The converter 12 may also generate the data to be judged by transforming the processing target data based on the feature vector using, for example, statistical processing such as average or variance. The converter 12 may also modify the processing target data itself using, for example, a multilayer perceptron, a machine learning model. In this case, the conversion unit 12 may change the parameters of the machine learning model based on the loss of the detection unit 13.

[0022] The type (n×m× . . . ) of the determination target data generated by the conversion unit 12 may be any type as long as it can be recognized by the detection unit 13. The conversion unit 12 outputs the generated determination target data to the detection unit 13.

[0023] The detection unit 13 receives the data to be determined from the conversion unit 12. The detection unit 13 inputs the data to be determined to a determination model that determines whether the input data is original data or false data based on a parameter (second parameter), and determines whether the data to be determined is original data or false data.

[0024] The generation unit 11 updates the parameters of the generative model (first parameters) based on the results of the determination by the detection unit 13. Specifically, the generation unit 11 evaluates how much the detection unit 13 made an incorrect determination, and updates the parameters of the generative model based on the result. Furthermore, the detection unit 13 updates the parameters of the determination model (second parameters) based on the results of the determination. Specifically, the detection unit 13 evaluates the results of the determination using a loss function, and updates the parameters of the determination model based on the result. By repeating this process, the generation unit 11 sets the parameters of the generative model (first parameters) that maximize the error rate of the determination results by the detection unit 13, and the detection unit 13 sets the parameters of the determination model (second parameters) that minimize the error rate of the determination results by the detection unit 13.

[0025] When the learning of the parameters of the generative model and the judgment model is completed and data is generated by the data generating device 10, noise data is input to the generating unit 11, and false data generated by the generating unit 11 based on the input noise data is output as generated data by the data generating device 10.

[0026] FIG. 2 is a diagram illustrating an example of a feature. In FIG. 2, the original data is traffic data of a communication network. In this case, the feature may include, for example, an "event," a "time period," and "time-series data," as shown in FIG. 2. An "event" is an event that affects traffic on a communication network (e.g., a game app update, an online live performance by a popular artist, etc.), quantified according to the degree of its impact. A "time period" is a time period at a predetermined time interval (every hour in the example shown in FIG. 2). When the data generating device 10 generates data after a certain time period, the "time-series data" is a quantified version of traffic data observed for each time period before that time period. Considering "event" as a feature enables data generation that takes burst traffic into account. Considering "time period" and "time-series data" enables data generation that takes into account periodic traffic fluctuations, such as day of the week or day and night.

[0027] Fig. 3 is a diagram showing an example of the configuration of the conversion unit 12. Fig. 3 illustrates an example in which a plurality of feature amounts are input as shown in Fig. 2.

[0028] As shown in FIG. 3, the conversion unit 12 includes feature quantity conversion units 121a, 121b, and 121c, and encoding units 122a, 122b, and 122c.

[0029] The feature transforming unit 121a transforms, from among the input features, features with relatively long periods, such as "month" and "day of the week," into feature vectors, and outputs the vectors to the encoding unit 122a.

[0030] The encoding unit 122a receives data to be processed. The encoding unit 122a converts the input data to be processed by encoding the data to be processed based on the feature vector output from the feature conversion unit 121a. The encoding unit 122a outputs the converted data to the encoding unit 122b.

[0031] The feature conversion unit 121b converts, from among the multiple input features, a feature with a relatively short period, such as "time," which is different from the feature used by the feature conversion unit 121a, into a feature vector and outputs it to the encoding unit 122b.

[0032] The encoding unit 122b receives the processing target data converted by the encoding unit 122a. The encoding unit 122b converts the processing target data by encoding the input processing target data based on the feature vector output from the feature conversion unit 121b. The encoding unit 122b outputs the converted processing target data to the encoding unit 122c.

[0033] The feature conversion unit 121c converts, from among the multiple input features, burst features such as "events," which are different from the features used by the feature conversion units 121a and 121b, into feature vectors and outputs them to the encoding unit 122c.

[0034] The encoding unit 122c receives the processing target data converted by the encoding unit 122b. The encoding unit 122c converts the processing target data by encoding the input processing target data based on the feature vector output from the feature conversion unit 121c. The encoding unit 122c outputs the converted processing target data as determination target data.

[0035] Note that the configuration of the conversion unit 12 shown in Figure 3 is merely an example, and the number of feature conversion units 121 and encoding units 122 may be changed as appropriate depending on the number of features used to convert the data to be processed.

[0036] Next, the operation of the data generating device 10 according to this embodiment will be described. Fig. 4 is a flowchart showing an example of the operation of the data generating device 10 according to this embodiment, and is a diagram for explaining a data generation method executed by the data generating device 10. Note that Fig. 4 explains the operation of the data generating device 10 when learning a generative model and a determination model.

[0037] The generator 11 generates fake data by inputting noise data into a generative model that generates fake data that imitates original data based on parameters (first parameters) (step S11).

[0038] The conversion unit 12 converts the processing target data, which is either the false data or the original data, based on any time-series feature amount to generate determination target data (step S12).

[0039] The detection unit 13 inputs the data to be judged into a judgment model that judges whether the input data is original data or fake data based on the parameter (second parameter), and judges whether the data to be judged is original data or fake data (step S13).

[0040] The generator 11 updates the parameters (first parameters) of the generative model based on the result of the determination by the detector 13. The detector 13 also updates the parameters (second parameters) of the determination model based on the result of the determination by the detector 13 (step S14). Here, the generator 11 updates the parameters (first parameters) of the generative model, for example, so that the error rate of the determination result by the detector 13 is maximized. The detector 13 also updates the parameters (second parameters) of the determination model, for example, so that the error rate of the determination result by the detector 13 is minimized (step S15).

[0041] As described above, the data generating device 10 according to this embodiment includes a generating unit 11, a converting unit 12, and a detecting unit 13. The generating unit 11 generates the fake data by inputting noise data into a generating model that generates fake data that imitates original data, which is time-series data, based on parameters (first parameters). The converting unit 12 generates data to be determined by converting processing target data, which is either the fake data or the original data, based on an arbitrary time-series feature. The detecting unit 13 inputs the data to be determined into a determining model that determines whether the input data is original data or fake data based on parameters (second parameters), and determines whether the data to be determined is original data or fake data. The generating unit 11 also updates the first parameters based on the determination results. The detecting unit 13 updates the second parameters based on the determination results.

[0042] In this manner, in this embodiment, the fake data and original data are converted based on arbitrary time-series features. This generates a variety of data, making it possible to maintain a balance between the generator 11 and the detector 13 (preventing mode collapse). Therefore, the data generator 10 according to this embodiment can generate a greater variety of fake data that takes time-series features into account.

[0043] Second Embodiment Fig. 5 is a diagram illustrating an example configuration of a data generating device 10A according to a second embodiment of the present disclosure. As illustrated in Fig. 5, the data generating device 10A according to this embodiment includes a generating unit 11, a converting unit 12A, and a detecting unit 13. The data generating device 10A according to this embodiment differs from the data generating device 10 according to the first embodiment in that the converting unit 12 is replaced with a converting unit 12A.

[0044] Similar to the converter 12, the converter 12A generates data to be determined by converting processing target data, which is either false data or original data, based on any time-series feature amount. Here, the converter 12A cuts out the processing target data based on the feature amount, and converts the cut processing target data to generate data to be determined.

[0045] For example, suppose that the data generating device 10A generates data for a part of a predetermined period (for example, a period from Friday to Sunday in a week). In this case, the feature quantities for that part of the period are input to the conversion unit 12A.

[0046] When generating data for a portion of a predetermined period (e.g., data for some days of a week), generating data for the entire predetermined period may result in higher learning accuracy. Therefore, the generation unit 11 generates fake data for the entire predetermined period. The conversion unit 12A cuts out the processing target data, which is either the fake data or the original data generated by the generation unit 11, for a portion of the period for which a feature is given, and converts the cut processing target data based on the feature to generate data to be determined.

[0047] As in the first embodiment, once learning of the parameters of the generative model and the judgment model is completed, when data generation device 10A generates data, generator 11 generates fake data from input noise data. Converter 12A cuts out the generated fake data for a period for which a feature is given. Converter 12A then converts the cut-out fake data based on the feature and outputs it as generated data of data generation device 10A.

[0048] This makes it possible to improve the learning accuracy and generate data of a desired size.

[0049] Third Embodiment Fig. 6 is a diagram illustrating an example configuration of a data generating device 10B according to a third embodiment of the present disclosure. As illustrated in Fig. 6, the data generating device 10B according to this embodiment includes a generating unit 11B, a converting unit 12, and a detecting unit 13B. The data generating device 10B according to this embodiment differs from the data generating device 10 according to the first embodiment in that the generating unit 11 is changed to a generating unit 11B and the detecting unit 13 is changed to a detecting unit 13B.

[0050] The noise data and the feature amount are input to the generating unit 11B. The generating unit 11B inputs the feature amount as additional information to a generative model, and generates false data from the noise data.

[0051] The detection unit 13B receives the data to be determined and the feature amount. The detection unit 13B inputs the feature amount as additional information to a determination model, and determines whether the data to be determined is original data or fake data.

[0052] As in the first embodiment, once the learning of the parameters of the generative model and the judgment model is completed, when data is generated by the data generating device 10B, the generator 11 inputs the features to the generative model as additional information and also inputs noise data to the generative model, and outputs the data generated by the generative model as generated data of the data generating device 10B.

[0053] By inputting the features as additional information into a generative model to generate fake data, it becomes possible to generate fake data that takes the features into consideration. Also, by inputting the features as additional information into a judgment model to determine whether the data to be judged is original data or fake data, it becomes possible to make judgments that take the features into consideration. As a result, it becomes possible to generate data that conforms to the desired features and make judgments that conform to the desired features.

[0054] The above-described method of inputting features as additional information into a generative model and a decision model is known as a conditional GAN. Conditional GAN ​​is described in Non-Patent Document 2, and therefore will not be described here.

[0055] 7 is a diagram illustrating an example configuration of a data generating device 10C according to a fourth embodiment of the present disclosure. As illustrated in FIG. 7, the data generating device 10C according to this embodiment includes a plurality of generating units 11B, a converting unit 12, a detecting unit 13B, and a combining unit 14. The data generating device 10C according to this embodiment differs from the data generating device 10B according to the third embodiment in that it includes a plurality of generating units 11B and in that it additionally includes a combining unit 14.

[0056] The plurality of generators 11B each receive noise data and feature quantities as input. The plurality of generators 11B each inputs the same or different feature quantities as additional information to a corresponding generative model, thereby generating fake data from the noise data.

[0057] The combining unit 14 combines the false data generated by the multiple generating units 11 and outputs the combined false data to the detecting unit 13B.

[0058] When the multiple generators 11B each input the same feature amount as additional information to the corresponding generative models, the combiner 14 may combine the fake data generated by each of the multiple generators 11B into one piece of data with the same dimensions. The combiner 14 may also combine the fake data generated by each of the multiple generators 11B into one piece of data with a larger number of dimensions than the original fake data. When the multiple generators 11B each input the same feature amount as additional information to the corresponding generative models, different parameters (first parameters) may be set for the generative models corresponding to each of the multiple generators 11B. This allows for the generation of a variety of data.

[0059] When the multiple generators 11B input different feature quantities as additional information to the corresponding generative models, for example, one generator 11B may use a feature quantity indicating a periodic change as additional information, and another generator 11B may use a feature quantity indicating a longer-term trend as additional information. The combining unit 14 may combine the fake data generated by the multiple generators 11B into one data, or may output the data to the detection unit 13B as is without combining them.

[0060] Each of the multiple generators 11B updates the parameters (first parameters) of the corresponding generative model based on the determination result of the detector 13B.

[0061] As in the first embodiment, once the learning of the parameters of the generative model and the judgment model is completed, when the data generating device 10C generates data, one or more generators 11B input desired features as additional information to the generative model and also input noise data to the generative model to generate fake data. The combiner 14 combines the fake data generated by the one or more generators 11B and outputs the combined data as generated data of the data generating device 10C.

[0062] By providing multiple generators 11B, it is possible to separate multiple features of the original data, such as periodicity and trends, and generate data based on each feature, thereby enabling the generation of data that is more similar to the original data.

[0063] Fifth Embodiment Fig. 8 is a diagram illustrating an example configuration of a data generating device 10D according to a fifth embodiment of the present disclosure. As illustrated in Fig. 8, the data generating device 10D according to this embodiment includes multiple generating units 11B, a converting unit 12, multiple detecting units 13B, and a combining unit 14. The data generating device 10D according to this embodiment differs from the data generating device 10C according to the fourth embodiment in that it includes multiple detecting units 13B.

[0064] The converter 12 converts the processing target data based on the feature and generates determination target data, and outputs the generated data to each of the multiple detectors 13B. The converter 12 may output the same determination target data to each of the multiple detectors 13, or may output different determination target data. When outputting different determination target data to the multiple detectors 13, the converter 12 may convert the processing target data based on, for example, a periodic feature and a burst feature to generate determination target data, and output each determination target data to a different detector 13B.

[0065] The plurality of detection units 13B each receive the target data and feature values ​​as input. The plurality of detection units 13B each inputs the same or different feature values ​​as additional information to a corresponding determination model, and determines whether the target data is original data or fake data.

[0066] When the multiple detection units 13B each input the same feature amount as additional information to the corresponding determination model, the multiple detection units 13B may each update the parameters of the corresponding determination model based on the determination results of the multiple detection units 13. For example, the multiple detection units 13B may each update the parameters of the corresponding determination model based on the sum or average of the evaluations of the determination results of all the detection units 13 using a loss function. Furthermore, the multiple detection units 13B may each update the parameters of the corresponding determination model based on the best value or worst value of the evaluations of the determination results of all the detection units 13 using a loss function.

[0067] When multiple detection units 13B input different feature quantities as additional information to corresponding models, for example, one detection unit 13B inputs a periodic feature quantity as additional information to a corresponding judgment model, and another detection unit 13B inputs a burst feature quantity as additional information to the corresponding judgment model. In this case, the conversion unit 12 converts the processing target data based on the feature quantity input by the detection unit 13B to the judgment model as additional information, and outputs the generated judgment target data to the detection unit 13B. Furthermore, the generation unit 11B and the detection unit 13B that input the same feature quantity to models as additional information may simultaneously update the parameters of the corresponding models.

[0068] As in the fourth embodiment, once the learning of the parameters of the generative model and the judgment model is completed, when the data generating device 10D generates data, one or more generators 11B input desired features as additional information to the generative model and also input noise data to the generative model to generate false data. The combiner 14 combines the false data generated by the one or more generators 11B and outputs the combined data as generated data.

[0069] By providing multiple detectors 13B, it is possible to separate multiple features of the original data, such as periodicity and trends, and make judgments based on each feature. This improves the accuracy of judgment, and as a result, it becomes possible to generate data that is more similar to the original data.

[0070] The data generating devices 10, 10A, 10B, 10C, and 10D described above can be realized by a computer 20 shown in FIG. 9 . A program for causing the computer 20 to function as the data generating devices 10, 10A, 10B, 10C, and 10D may be provided. The program may be stored on a storage medium or provided via a network. FIG. 9 is a block diagram showing a schematic configuration of a computer 20 functioning as the data generating devices 10, 10A, 10B, 10C, and 10D. The computer 20 may be a general-purpose computer, a dedicated computer, a workstation, a personal computer (PC), an electronic notepad, or the like. The program instructions may be program code, code segments, or the like for performing necessary tasks.

[0071] 9, the computer 20 includes a processor 21, a ROM (Read Only Memory) 22, a RAM (Random Access Memory) 23, a storage 24, an input unit 25, a display unit 26, and a communication interface (I / F) 27. Each component is communicably connected to one another via a bus 29. The processor 21 is specifically a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), a SoC (System on a Chip), or the like, and may be configured with multiple processors of the same or different types.

[0072] The processor 21 is a control unit that controls each component and performs various arithmetic operations. That is, the processor 21 reads a program from the ROM 22 or the storage 24 and executes the program using the RAM 23 as a work area. The processor 21 controls each component and performs various arithmetic operations in accordance with the program stored in the ROM 22 or the storage 24. In this embodiment, the ROM 22 or the storage 24 stores a program for causing the computer 20 to operate as the data generation devices 10, 10A, 10B, 10C, and 10D according to the present disclosure. The program is read and executed by the processor 21 to realize each component of the data generation devices 10, 10A, 10B, 10C, and 10D, i.e., the generation units 11 and 11B, the conversion units 12 and 12A, and the detection units 13 and 13B.

[0073] The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), a USB (Universal Serial Bus) memory, etc. The program may also be provided in a form downloaded from an external device via a network.

[0074] The ROM 22 stores various programs and various data. The RAM 23 temporarily stores programs or data as a working area. The storage 24 is configured with an HDD (Hard Disk Drive) or an SSD (Solid State Drive) and stores various programs including the operating system and various data.

[0075] The input unit 25 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0076] The display unit 26 is, for example, a liquid crystal display, and displays various information. The display unit 26 may be a touch panel type and function as the input unit 25.

[0077] The communication interface 27 is an interface for communicating with other devices, for example, an interface for a LAN.

[0078] The following additional notes are provided regarding the above-described embodiments.

[0079] [Supplementary Item 1] A data generating device comprising a control unit, wherein the control unit generates fake data that imitates original data, which is time-series data, by inputting noise data into a generation model that generates the fake data based on a first parameter; generates data to be judged by converting processing target data, which is one of the fake data and the original data, based on an arbitrary time-series feature; inputs the data to be judged into a judgment model that judges whether input data is the original data or the fake data based on a second parameter, and judges whether the data to be judged is the original data or the fake data; the generation unit updates the first parameter based on a result of the judgment; and the detection unit is configured to update the second parameter based on a result of the judgment.

[0080] [Supplementary Item 2] In the data generating device according to Supplementary Item 1, the control unit cuts out the processing target data based on the feature amount, and converts the cut processing target data to generate the judgment target data.

[0081] [Supplementary Item 3] In the data generating device described in Supplementary Item 1, the control unit inputs the feature amount as additional information to the generation model to generate the false data from the noise data, and inputs the feature amount as additional information to the judgment model to perform the judgment.

[0082] [Supplementary Item 4] In the data generating device described in Supplementary Item 3, the control unit inputs the same or different feature quantities as the additional information to the corresponding generative models, respectively, to generate the false data from the noise data, and updates the first parameter of the corresponding generative model based on the result of the determination.

[0083] [Supplementary Item 5] In the data generating device described in Supplementary Item 3 or 4, the control unit inputs the same or different features as the additional information into the corresponding judgment models, respectively, to perform the judgment, and updates the second parameter of the corresponding judgment model based on the result of its own judgment or the results of the multiple judgments.

[0084] [Supplementary Item 6] A data generation method executed by a data generation device, comprising: inputting noise data into a generation model that generates fake data that imitates original data, which is time-series data, based on a first parameter, to generate the fake data; generating data to be judged by converting processing target data, which is one of the fake data and the original data, based on an arbitrary time-series feature; inputting the data to be judged into a judgment model that judges whether input data is the original data or the fake data based on a second parameter, and judging whether the data to be judged is the original data or the fake data; updating the first parameter based on the result of the judgment; and updating the second parameter based on the result of the judgment.

[0085] [Supplementary Item 7] A non-transitory storage medium storing a program executable by a computer, the non-transitory storage medium storing the program causing the computer to operate as the data generation device described in any one of Supplementary Items 1 to 5.

[0086] All publications, patent applications, and technologies mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, and technology was specifically and individually indicated to be incorporated by reference.

[0087] Although the above-described embodiments have been described as typical examples, it will be apparent to those skilled in the art that many modifications and substitutions can be made within the spirit and scope of the present disclosure. Therefore, the present invention should not be construed as being limited by the above-described embodiments, and various modifications or alterations are possible without departing from the scope of the claims.

[0088] 1, 10, 10A, 10B, 10C, 10D Data generating device 11, 11B Generating unit 12, 12A Converting unit 13, 13B Detecting unit 121A, 121B, 121C Feature converting unit 122A, 122B, 122C Encoding unit 20 Computer 21 Processor 22 ROM 23 RAM 24 Storage 25 Input unit 26 Output unit 27 Communication interface 28 Bus

Claims

1. A data generation device comprising: a generation unit that inputs noise data into a generation model that generates fake data that imitates original data, which is time-series data, based on first parameters, to generate the fake data; a conversion unit that generates data to be judged by converting processing target data, which is either the fake data or the original data, based on any time-series feature; and a detection unit that inputs the data to be judged into a judgment model that judges whether input data is the original data or the fake data based on second parameters, and judges whether the data to be judged is the original data or the fake data, wherein the generation unit updates the first parameter based on the result of the judgment, and the detection unit updates the second parameter based on the result of the judgment.

2. A data generating device according to claim 1, wherein the conversion unit cuts out the processing target data based on the feature amount, and converts the cut out processing target data to generate the judgment target data.

3. A data generating device according to claim 1, wherein the generating unit inputs the feature amount as additional information into the generating model to generate the false data from the noise data, and the detecting unit inputs the feature amount as additional information into the judgment model to make the judgment.

4. A data generating device according to claim 3, comprising a plurality of generating units that input the same or different features as the additional information into the corresponding generative models to generate the false data from the noise data, and each of the plurality of generating units updates the first parameter of the corresponding generative model based on the result of the judgment.

Images