DATA GENERATION PROGRAM, INFORMATION PROCESSING APPARATUS, AND DATA GENERATION METHOD

By evaluating and regenerating data within a generative adversarial network to exclude specific features, the method accelerates the production of high-quality training data, addressing the inefficiency of existing GAN-based data generation methods.

JP7583256B2Active Publication Date: 2024-11-14FUJITSU LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
JP2020211460
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2020-12-21
Publication Date
2024-11-14
Estimated Expiration
2040-12-21

AI Technical Summary

Technical Problem

Generative adversarial networks (GANs) require a long time to generate fake data that can be used as training data due to the use of randomly generated noise.

Method used

A data generation program and method that utilizes a generative adversarial network to generate multiple data points, evaluates their authenticity, identifies common features among low-evaluation data, and regenerates data without those features, thereby shortening the time required to produce high-quality training data.

Benefits of technology

This approach significantly reduces the time needed to generate teacher data by controlling the generation and evaluation process, ensuring higher-quality output more efficiently.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007583256000001
    Figure 0007583256000001
  • Figure 0007583256000002
    Figure 0007583256000002
  • Figure 0007583256000003
    Figure 0007583256000003
Patent Text Reader

Abstract

To provide a data generation program, an information processing device, and a data generation method capable of reducing a time period required for generation of teacher data.SOLUTION: A data generation program generates a plurality of pieces of first data from noises generated at random, calculates evaluation values regarding authenticity for each of the generated plurality of pieces of first data, identifies a characteristic common for a plurality of pieces of second data having evaluation values lower than reference data among the generated plurality of pieces of first data, and re-generates a plurality of pieces of first data not having the identified characteristic.SELECTED DRAWING: Figure 6
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

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

[0002] For example, a method using generative adversarial networks (GANs) is known as a method for automatically generating training data used to generate a machine learning model.

[0003] This generative adversarial network consists of a generator that generates data (hereinafter also referred to as fake data) by using noise, and a discriminator that determines whether the fake data generated by the generator is true or false.

[0004] Specifically, the generative adversarial network improves the accuracy of the generator and the discriminator by repeating the generation of fake data by the generator and the determination of the authenticity of the fake data by the discriminator. After sufficiently improving the accuracy of the generator and the discriminator, the generative adversarial network generates training data to be used in generating a machine learning model by using, as training data, fake data that the discriminator determines to be real data among the fake data generated by the generator (for example, see Patent Documents 1 to 3). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2018-163554 A [Patent Document 2] International Publication No. 2020 / 008710 [Patent Document 3] Patent No. 6687962 Summary of the Invention [Problem to be solved by the invention]

[0006] However, in the above-mentioned generative adversarial network, for example, fake data is generated by using randomly generated noise. Therefore, in the generative adversarial network, it may take a long time to generate fake data that can be used as training data (fake data that the discriminator judges to be real data).

[0007] Therefore, in one aspect, the present invention aims to provide a data generation program, an information processing device, and a data generation method that enable the time required to generate teacher data to be reduced. [Means for solving the problem]

[0008] In one aspect of the embodiment, a data generation program causes a computer to execute a process for generating data using a generative adversarial network, which generates a plurality of first data from randomly generated noise, calculates an evaluation value regarding the truth or falsity of each of the generated plurality of first data, identifies common features among a plurality of second data among the generated plurality of first data, the evaluation value of which is lower than that of a reference data, and re-generates the plurality of first data that do not have the identified features. Effect of the Invention

[0009] According to one aspect, it is possible to reduce the time required to generate training data. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating the configuration of an information processing system 10. As shown in FIG. [Diagram 2] FIG. 2 is a diagram illustrating a generative adversarial network. [Diagram 3] FIG. 3 is a diagram illustrating the hardware configuration of the information processing device 1. As shown in FIG. [Figure 4] FIG. 4 is a functional block diagram of the information processing device 1. As shown in FIG. [Diagram 5] FIG. 5 is a flowchart outlining the data generation process in the first embodiment. [Figure 6] FIG. 6 is a flowchart outlining the data generation process in the first embodiment. [Figure 7] FIG. 7 is a flowchart illustrating details of the data generation process in the first embodiment. [Figure 8] FIG. 8 is a flowchart illustrating details of the data generation process in the first embodiment. [Figure 9] FIG. 9 is a flowchart illustrating details of the data generation process in the first embodiment. [Figure 10] FIG. 10 is a flowchart illustrating details of the data generation process in the first embodiment. [Figure 11] FIG. 11 is a flowchart illustrating details of the data generation process in the first embodiment. [Figure 12] FIG. 12 is a diagram for explaining a specific example of the process from S41 to S45. [Figure 13] FIG. 13 is a diagram for explaining a specific example of the process from S41 to S45. [Figure 14] FIG. 14 is a diagram for explaining a specific example of the process from S41 to S45. [Figure 15] FIG. 15 is a diagram illustrating another specific example of the process from S41 to S45. [Figure 16] FIG. 16 is a diagram illustrating another specific example of the process from S41 to S45. [Figure 17] FIG. 17 is a diagram illustrating another specific example of the process from S41 to S45. [Figure 18] FIG. 18 is a flowchart illustrating the data generation process according to the second embodiment. [Figure 19]FIG. 19 is a flowchart illustrating the data generation process according to the third embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] [Information processing system configuration] First, a description will be given of the configuration of the information processing system 10. FIG 1 is a diagram for explaining the configuration of the information processing system 10.

[0012] 1 includes, for example, an information processing device 1 consisting of one or more physical machines, and an operation terminal 2 into which an operator (hereinafter also simply referred to as an operator) who generates teacher data inputs necessary information. The operation terminal 2 may be, for example, a personal computer (PC). The information processing device 1 and the operation terminal 2 are connected to each other via a network NW such as the Internet.

[0013] The information processing device 1 executes, for example, a process of generating training data using a generative adversarial network (hereinafter also referred to as a data generation process). The generative adversarial network will be described below.

[0014] [Generative Adversarial Networks] FIG. 2 is a diagram illustrating a generative adversarial network.

[0015] 2 includes a noise generating unit 113, a data generating unit 114, and a data discriminator 115. In the following description, the data generating unit 114 corresponds to a generator, and the data discriminator 115 corresponds to a discriminator.

[0016] The data generation unit 114 generates multiple pieces of fake data DT2, for example, by adding noise randomly generated by the noise generation unit 113 to multiple pieces of pre-prepared teacher data DT1 (hereinafter also referred to as real data DT1).

[0017] Moreover, the data identification unit 115, for example, learns about a plurality of pieces of genuine data DT1, and then judges whether a plurality of pieces of fake data DT2 generated by the data generation unit 114 are genuine or false.

[0018] The data generating unit 114 and the data identifying unit 115 alternately repeat the generation of fake data DT2 and the determination of whether the data DT2 is fake, thereby improving the accuracy of each unit.

[0019] Specifically, the data generation unit 114 performs repeated learning so as to be able to generate fake data DT2 that the data discrimination unit 115 determines not to be fake data (fake data DT2 that the data discrimination unit 115 determines to be real data) by using the determination result of the fake data DT2 by the data discrimination unit 115. Also, the data discrimination unit 115 performs repeated learning so as to be able to determine that the fake data DT2 generated by the data generation unit 114 is fake data by using information indicating whether or not the determination result of the fake data DT2 was valid.

[0020] This enables the worker to obtain, as new teacher data DT3, for example, fake data DT2 generated by the data generation unit 114 after sufficient learning has been performed, and which the data identification unit 115 after sufficient learning has determined to be genuine data, as shown in Figure 2.

[0021] However, the data generating unit 114 generates the fake data DT2 by using, for example, randomly generated noise. Therefore, the data generating unit 114 may take a long time to generate the fake data DT2 that can be used as the teacher data DT3 (the fake data DT2 that the data identifying unit 115 determines to be real data).

[0022] Therefore, the information processing device 1 in this embodiment generates multiple fake data DT2 (hereinafter also referred to as first data) from randomly generated noise, and calculates an evaluation value regarding the authenticity of each of the multiple generated fake data DT2.

[0023] Then, the information processing device 1 identifies common features in the plurality of fake data DT2 (hereinafter also referred to as second data) having an evaluation value lower than the reference data among the plurality of fake data DT2 generated. After that, the information processing device 1 regenerates the plurality of fake data DT2 that does not have the identified features.

[0024] That is, the information processing device 1 in this embodiment identifies a feature (a feature included in the fake data DT2) that is likely to have been at least one of the factors that caused the data identification unit 115 to determine that the fake data DT2 was fake data. Then, the information processing device 1 regenerates the fake data DT2 so that the identified feature is not included.

[0025] This enables the information processing device 1 in this embodiment to reduce the number of pieces of fake data DT2 that need to be generated and judged as genuine by the data generating unit 114 and the data identifying unit 115 before generating the teacher data DT3. This enables the information processing device 1 to reduce the time required to generate the teacher data DT3.

[0026] [Hardware configuration of information processing device] Next, a description will be given of the hardware configuration of the information processing device 1. FIG 3 is a diagram illustrating the hardware configuration of the information processing device 1.

[0027] 3, the information processing device 1 includes a CPU 101 which is a processor, a memory 102, a communication device 103, and a storage medium 104. Each unit is connected to one another via a bus 105.

[0028] The storage medium 104 has, for example, a program storage area (not shown) that stores a program 110 for performing the data generation process. The storage medium 104 also has, for example, an information storage area 130 that stores information used when performing the data generation process. The storage medium 104 may be, for example, a hard disk drive (HDD) or a solid state drive (SSD).

[0029] The CPU 101 executes a program 110 loaded from the storage medium 104 into the memory 102 to perform data generation processing.

[0030] Further, the communication device 103 communicates with the operation terminal 2 via, for example, the network NW.

[0031] [Functions of information processing device] Next, a description will be given of the functions of the information processing device 1. FIG.

[0032] As shown in FIG. 4, the information processing device 1 realizes various functions including, for example, a data receiving unit 111, a data management unit 112, a noise generating unit 113, a data generating unit 114, a data identification unit 115, a feature identifying unit 116, and a data output unit 117 by organic cooperation between hardware such as a CPU 101 and a memory 102 and a program 110.

[0033] Furthermore, the information processing device 1 stores, for example, real data DT1, fake data DT2, teacher data DT3, and feature information 131 in the information storage area 130, as shown in FIG.

[0034] The data receiving unit 111 receives, for example, the real data DT1 transmitted from the operation terminal 2. Then, the data management unit 112 stores, for example, the real data DT1 received by the data receiving unit 111 in the information storage area .

[0035] The noise generating section 113 generates noise, for example, randomly.

[0036] The data generating unit 114 generates fake data DT2 from, for example, the noise generated by the noise generating unit 113. The data generating unit 114 also generates fake data DT2 from, for example, the noise generated by the noise generating unit 113 and the real data DT1 stored in the information storage area 130. The data managing unit 112 then stores, for example, the fake data DT2 generated by the data generating unit 114 in the information storage area 130.

[0037] The data discriminating unit 115 judges the authenticity of each piece of fake data DT2 generated by the data generating unit 114. Then, the data discriminating unit 115 calculates an evaluation value for each piece of fake data DT2 generated by the data generating unit 114.

[0038] The feature identifying unit 116 identifies a common feature among a plurality of pieces of fake data DT2 having an evaluation value calculated by the data identifying unit 115 lower than the reference data, among the fake data DT2 generated by the data generating unit 114. The reference data is, for example, the genuine data DT1 used by the data generating unit 114 to generate the fake data DT2. Then, the data managing unit 112 stores, for example, feature information 131 indicating the feature generated by the feature identifying unit 116 in the information storage area 130.

[0039] For example, the data output unit 117 outputs, as teacher data DT3, fake data DT2 generated by the data generation unit 114, whose evaluation value calculated by the data discrimination unit 115 is equal to or greater than a threshold value. That is, for example, the data output unit 117 outputs, as teacher data DT3, fake data DT2 generated by the data generation unit 114, which the data discrimination unit 115 has determined to be genuine data DT1.

[0040] [Outline of the first embodiment] Next, an outline of the first embodiment will be described. Fig. 5 is a flow chart for explaining an outline of the data generation process in the first embodiment. Fig. 6 is a diagram for explaining an outline of the data generation process in the first embodiment.

[0041] 5, the information processing device 1 waits until it is time to generate data (NO in S11), for example. The data generation timing may be the timing when the operator inputs an instruction to generate the teacher data DT3.

[0042] Then, when it is time to generate data (YES in S11), the information processing device 1 generates a plurality of pieces of fake data DT2 from randomly generated noise (S12).

[0043] Next, the information processing device 1 calculates an evaluation value regarding the authenticity of each of the multiple pieces of fake data DT2 generated in the process of S12 (S13).

[0044] Furthermore, the information processing device 1 identifies common features among the plurality of pieces of fake data DT2 generated in the process of S12, the plurality of pieces of fake data DT2 having evaluation values ​​calculated in the process of S13 lower than the reference data (S14).

[0045] After that, the information processing device 1 regenerates a plurality of fake data DT2 that do not have the characteristics identified in the process of S14 (S15).

[0046] 6, the feature identifying unit 116 identifies a feature that is highly likely to have been at least one of the factors that caused the data identifying unit 115 to determine that the fake data DT2 was fake data, from the evaluation value acquired from the data identifying unit 115. Then, the data generating unit 114 regenerates the fake data DT2 that does not include the feature acquired from the feature identifying unit 116.

[0047] This allows the information processing device 1 in this embodiment to encourage the recreation of fake data DT2 with a high evaluation value. Therefore, the information processing device 1 can reduce the number of fake data DT2 that the data generating unit 114 and the data identifying unit 115 need to generate and judge before generating the teacher data DT3. Therefore, the information processing device 1 can shorten the time required to generate the teacher data DT3.

[0048] [Details of the first embodiment] Next, the details of the first embodiment will be described. Figures 7 to 11 are flow charts for explaining the details of the data generation process in the first embodiment. Figures 12 to 17 are diagrams for explaining the details of the data generation process in the first embodiment.

[0049] [Data learning processing] First, the process of learning data discrimination section 115 (hereinafter also referred to as data learning process) in the data generation process will be described. Fig. 7 is a flow chart for explaining the data learning process.

[0050] As shown in FIG. 7, the data receiving unit 111 waits until it receives, for example, a plurality of pieces of real data DT1 transmitted from the operation terminal 2 (NO in S21).

[0051] Then, when a plurality of pieces of genuine data DT1 are received (YES in S21), the data management unit 112 stores the plurality of pieces of genuine data DT1 received in the process of S21 in the information storage area 130 (S22).

[0052] Furthermore, the data identification unit 115 learns a plurality of pieces of genuine data DT1 stored in the information storage area 130 (S23).

[0053] This enables the information processing device 1 to perform authenticity determination on the fake data DT2 generated by the data generating unit 114, as will be described later.

[0054] [Main process of data generation] Next, the main processing of the data generation processing will be described with reference to Figures 8 to 11, which are flow charts illustrating the main processing of the data generation processing.

[0055] As shown in FIG. 8, the noise generating unit 113 waits until the data generation timing, for example (NO in S31).

[0056] Then, when it is time to generate data (YES in S31), the feature specification unit 116 specifies, for example, any of the genuine data DT1 stored in the information storage area 130 as reference data (S32).

[0057] Next, the noise generating section 113 randomly generates noise (S33).

[0058] Furthermore, the data generation unit 114 generates a plurality of pieces of fake data DT2, for example, from the noise generated in the process of S33 and the reference data identified in the process of S32 (S34).

[0059] Thereafter, the data identification unit 115 calculates an evaluation value indicating the probability that each of the multiple pieces of fake data DT2 generated in the process of S34 is not fake data (probability that it is real data) (S35).

[0060] Then, the feature identification unit 116 identifies, from among the multiple fake data DT2 generated in the process of S34 or the process of S52 (described later), multiple fake data DT2 whose evaluation value calculated in the process of S35 is lower than that of the reference data (S36).

[0061] Next, as shown in FIG. 9, for each of the multiple fake data DT2 identified in the process of S36, the feature identification unit 116 identifies a first pixel that has a greater brightness than the pixel at the same position in the reference data identified in the process of S32 (S41).

[0062] Furthermore, for each of the plurality of pieces of fake data DT2 identified in the process of S36, the feature identifying unit 116 identifies a second pixel having a lower luminance than the pixel at the same position in the reference data identified in the process of S32 (S42).

[0063] Furthermore, the feature identifying unit 116 identifies a first position identified as the location of the first pixel in a predetermined ratio of data included in the plurality of pieces of fake data DT2 identified in the process of S36 (S43).

[0064] Furthermore, the feature identifying unit 116 identifies a second position identified as the location of the second pixel in a predetermined ratio of data included in the plurality of fake data DT2 identified in the process of S36 (S44).

[0065] Thereafter, the feature identification unit 116 identifies information including information indicating that the first pixel is located at the first position identified in the process of S43 and information indicating that the second pixel is located at the second position identified in the process of S44 as feature information 131 (S45). A specific example of the processes from S41 to S45 will be described below.

[0066] [Specific example of processing from S41 to S45 (1)] 12 to 14 are diagrams for explaining a specific example of the processes of S41 to S45. In the example shown in FIG. 12 etc., "UP" is set to a position where a pixel with a higher brightness than the pixel arranged at the same position in the reference data is arranged. In the example shown in FIG. 12, "DN" is set to a position where a pixel with a lower brightness than the pixel arranged at the same position in the reference data is arranged.

[0067] Specifically, in the fake data DT2a shown in FIG. 12(A), "UP" is set in the positions corresponding to the first row, third column, the second row, second column, and the fourth row, first column, and "DN" is set in the positions corresponding to the first row, first column, the third row, third column, and the fourth row, third column.

[0068] Therefore, the feature identifying unit 116 identifies the pixels at positions corresponding to the first row, third column, the second row, second column, and the fourth row, first column in the fake data DT2a shown in Fig. 12(A) as first pixels (S41). Also, the feature identifying unit 116 identifies the pixels at positions corresponding to the first row, first column, the third row, third column, and the fourth row, third column in the fake data DT2a shown in Fig. 12(A) as second pixels (S42).

[0069] Similarly, the feature identifying unit 116 identifies pixels at positions corresponding to the first row, third column, the first row, fourth column, the second row, second column, and the third row, third column in the fake data DT2b shown in Fig. 12(B) as first pixels (S41). Also, the feature identifying unit 116 identifies pixels at positions corresponding to the fourth row, first column, and the fourth row, third column in the fake data DT2b shown in Fig. 12(B) as second pixels (S42).

[0070] Furthermore, the feature identifying unit 116 identifies pixels at positions corresponding to the third column of the first row, the second column of the second row, the fourth column of the third row, and the first column of the fourth row in the fake data DT2c shown in Fig. 12(C) as first pixels (S41). Also, the feature identifying unit 116 identifies pixels at positions corresponding to the first column of the second row, the first column of the third row, and the third column of the fourth row in the fake data DT2c shown in Fig. 12(C) as second pixels (S42).

[0071] Here, the positions corresponding to the first row, third column and the second row, second column are positions identified as the first pixel in all of the fake data DT2 shown in Figures 12(A), 12(B) and 12(C). Also, the position corresponding to the fourth row, third column is a position identified as the second pixel in all of the fake data DT2 shown in Figures 12(A), 12(B) and 12(C).

[0072] Therefore, in this case, the feature identifying unit 116 identifies the positions corresponding to the first row, third column and the second row, second column as the first position (S43), as shown in the underlined portions of Figures 13(A), (B), and (C). Also, in this case, the feature identifying unit 116 identifies the position corresponding to the fourth row, third column as the second position (S44), as shown in the underlined portions of Figures 13(A), (B), and (C).

[0073] Thereafter, feature identification unit 116 generates feature information 131 including information indicating that the pixel at the position corresponding to the first row, third column is the first pixel, information indicating that the pixel at the position corresponding to the second row, second column is the first pixel, and information indicating that the pixel at the position corresponding to the fourth row, third column is the second pixel (S45). Then, data management unit 112 stores feature information 131 generated by feature identification unit 116 in information storage area 130, for example. A specific example of feature information 131 will be described below.

[0074] [Example of characteristic information (1)] 14 and 18 are examples of the feature information 131. FIG.

[0075] The feature information 131 shown in FIG. 14 and other figures has, as items, “position” to which the position identified in the processes of S43 and S44 is set, and “pixel” to which the pixel identified in the processes of S41 and S42 is set.

[0076] Specifically, in the feature information 131 shown in Fig. 14, the information in the first row has "first row, third column" set as the "position" and "first pixel" set as the "pixel." Additionally, the information in the second row has "second row, second column" set as the "position" and "first pixel" set as the "pixel." Furthermore, the information in the third row has "fourth row, third column" set as the "position" and "second pixel" set as the "pixel."

[0077] [Specific example of processing from S41 to S45 (2)] 15 to 17 are diagrams illustrating another specific example of the process from S41 to S45.

[0078] Specifically, in the fake data DT2d shown in FIG. 15(A), "UP" is set in the positions corresponding to the fourth column of the second row and the first column of the fourth row, and "DN" is set in the positions corresponding to the third column of the first row and the second column of the third row.

[0079] Therefore, the feature identifying unit 116 identifies the pixels at positions corresponding to the second row, fourth column and the fourth row, first column in the fake data DT2d shown in Fig. 15(A) as the first pixel (S41). Also, the feature identifying unit 116 identifies the pixels at positions corresponding to the first row, third column and the third row, second column in the fake data DT2d shown in Fig. 15(A) as the second pixel (S42).

[0080] Similarly, the feature identifying unit 116 identifies the pixels at positions corresponding to the first row, second column, the first row, fourth column, and the second row, fourth column in the fake data DT2e shown in Fig. 15(B) as the first pixels (S41). Also, the feature identifying unit 116 identifies the pixels at positions corresponding to the second row, second column, and the fourth row, third column in the fake data DT2e shown in Fig. 15(B) as the second pixels (S42).

[0081] Furthermore, the feature identifying unit 116 identifies pixels at positions corresponding to the first row, third column, the second row, first column, the second row, fourth column, and the third row, third column in the fake data DT2f shown in Fig. 15(C) as the first pixel (S41). Also, the feature identifying unit 116 identifies a pixel at a position corresponding to the third row, fourth column in the fake data DT2f shown in Fig. 15(C) as the second pixel (S42).

[0082] Here, the position corresponding to the second row and fourth column is the position identified as the first pixel in all of the fake data DT2 shown in FIG. 15(A), FIG. 15(B) and FIG. 15(C).

[0083] Therefore, in this case, the feature identifying unit 116 identifies the position corresponding to the second row, fourth column as the first position (S43), as shown in the underlined portions of Figures 16(A), (B), and (C). On the other hand, the feature identifying unit 116 does not identify the second position.

[0084] Thereafter, the feature identification unit 116 generates feature information 131 including information indicating that the pixel at the position corresponding to the second row and fourth column is the first pixel (S45).

[0085] Specifically, as shown in FIG. 17, the feature identification unit 116 generates feature information 131 consisting of information in which the “position” is set to “2nd row, 4th column” and the “pixel” is set to “1st pixel.”

[0086] 10, the noise generating unit 113 randomly generates noise (S51). Specifically, the noise generating unit 113 generates noise so that the rate at which fake data DT2 having the characteristics identified in the process of S45 is generated is equal to or lower than a threshold value.

[0087] Then, the data generation unit 114 generates a plurality of pieces of fake data DT2, for example, from the noise generated in the process of S51 and the reference data identified in the process of S32 (S52).

[0088] Furthermore, the data identification unit 115 calculates an evaluation value indicating the probability that each of the multiple pieces of fake data DT2 generated in the process of S52 is genuine data DT1 (S53).

[0089] Thereafter, the data identification unit 115 judges whether or not the plurality of fake data DT2 generated in the process of S51 includes data whose evaluation value calculated in the process of S52 is equal to or greater than a threshold value (S54).

[0090] As a result, when it is determined that there is no data whose evaluation value calculated in the process of S52 is equal to or greater than the threshold value (NO in S55), the information processing device 1 performs the processes from S36 onwards again.

[0091] That is, in this case, the feature identification unit 116 determines that the plurality of fake data DT2 generated in the process of S51 does not include fake data DT2 that can be used as training data DT3, and performs the processes from S36 onwards again.

[0092] On the other hand, if it is determined in the processing of S52 that data exists whose evaluation value is equal to or greater than the threshold value (YES in S55), the data output unit 117 outputs the data determined to exist in the processing of S54 as teacher data DT3 (S61), as shown in FIG. 11.

[0093] Then, the data output unit 117 determines whether or not the number of pieces of teacher data DT3 output in the process of S61 has reached a threshold value (S62).

[0094] As a result, when it is determined that the number of pieces of teacher data DT3 output in the process of S61 has not reached the threshold value (NO in S63), the information processing device 1 performs the processes from S36 onwards again.

[0095] On the other hand, when it is determined that the number of pieces of teacher data DT3 output in the process of S61 has reached the threshold value (YES in S63), the information processing device 1 ends the data generation process.

[0096] In addition, when the number of pieces of teacher data DT3 output in the process of S61 reaches a predetermined number, the information processing device 1 may perform the processes from S32 onward again. That is, the information processing device 1 may generate the teacher data DT3 while appropriately changing the reference data.

[0097] In this manner, the information processing device 1 in this embodiment generates multiple pieces of fake data DT2 (first data) from randomly generated noise, and calculates an evaluation value regarding the authenticity of each of the multiple pieces of generated first data DT2.

[0098] Then, the information processing device 1 identifies common features in the plurality of fake data DT2 (second data) having an evaluation value lower than the reference data among the plurality of fake data DT2 generated. After that, the information processing device 1 regenerates the plurality of fake data DT2 that does not have the identified features.

[0099] That is, the information processing device 1 in this embodiment identifies at least one feature that was the factor that caused the data identification unit 115 to determine that the fake data DT2 was fake data. Then, the information processing device 1 regenerates the fake data DT2 so that the identified feature is not included.

[0100] This enables the information processing device 1 in this embodiment to reduce the number of fake data DT2 that the data generating unit 114 and the data identifying unit 115 need to generate and judge before generating the teacher data DT3. Therefore, the information processing device 1 can reduce the time required to generate the teacher data DT3.

[0101] 17 indicates that the pixel at the position corresponding to the second row, fourth column is the first pixel. Therefore, for example, if each pixel constituting the reference data has 256 gradations and the gradation of the pixel corresponding to the second row, fourth column in the reference data is 128, the data generator 114 can reduce the generation frequency (number of generations) of fake data DT2 in which the gradation of the pixel corresponding to the second row, fourth column is less than 128.

[0102] In the above example, the comparison between the fake data DT2 and the reference data is described for each pixel, but the information processing device 1 may also compare the fake data DT2 with the reference data for each set of pixels (hereinafter also referred to as a block).

[0103] Specifically, in this case, for each of the fake data DT2 identified in the processing of S36, the feature identification unit 116 may identify, for each block consisting of multiple pixels arranged at adjacent positions in each fake data DT2, a block (hereinafter also referred to as a first block) having a greater luminance (e.g., average luminance) than the block at the same position in the reference data, and a block (hereinafter also referred to as a second block) having a lesser luminance (e.g., average luminance) than the block at the same position in the reference data.

[0104] Furthermore, in this case, the feature identification unit 116 may identify a position (hereinafter also referred to as a third position) identified as the position where the first block is located in a predetermined percentage of the plurality of fake data DT2 identified in the process of S36, and a position (hereinafter also referred to as a fourth position) identified as the position where the second block is located in the predetermined percentage of the plurality of second data.The feature identification unit 116 may identify, as feature information 131, information indicating that the first block is located in the third position and information indicating that the second block is located in the fourth position.

[0105] This enables the information processing device 1 to reduce the processing load on the feature identification unit 116, for example.

[0106] [Details of the second embodiment] Next, a second embodiment will be described with reference to Fig. 18, which is a flow chart illustrating a data generation process in the second embodiment.

[0107] In the second embodiment, when real data DT1 used for learning of the data identification unit 115 is added, all of the generated teacher data DT3 is regenerated. Note that only the points different from the first embodiment will be described below.

[0108] [Data re-learning process (1)] First, the process of learning the added real data DT1 (hereinafter, also referred to as re-learning process) in the data generation process will be described. Fig. 18 is a flow chart for explaining the data re-learning process.

[0109] 18, the data management unit 112 waits until one or more pieces of real data DT1 are received from the operation terminal 2 (NO in S71). That is, the data management unit 112 waits until new real data DT1 is added.

[0110] Then, when one or more pieces of real data DT1 are received (YES in S71), the data management unit 112 stores the one or more pieces of real data DT1 received in the processing of S71 in the information storage area 130 (S72).

[0111] Next, the data management unit 112 deletes the teacher data DT3 stored in the information storage area 130 (S73).

[0112] After that, the data identification unit 115 further learns one or more pieces of genuine data DT1 received in the process of S71 (S74).Then, the information processing device 1 executes, for example, the data generation process described with reference to Figs.

[0113] That is, when the added genuine data DT1 is learned, the criterion for determining whether the counterfeit data DT2 is genuine by the data discrimination unit 115 may change. Therefore, the information processing device 1 in the second embodiment regenerates the teacher data DT3 when genuine data DT1 used for learning by the data discrimination unit 115 is added.

[0114] This enables the information processing device 1 in the second embodiment to generate more accurate teacher data DT3 when real data DT1 is added.

[0115] [Details of the third embodiment] Next, a third embodiment will be described with reference to Fig. 19, which is a flow chart illustrating a data generation process in the third embodiment.

[0116] In the third embodiment, when real data DT1 used for learning of the data identification unit 115 is added, at least a part of the generated teacher data DT3 is regenerated. Note that only the points different from the first embodiment will be described below.

[0117] [Data re-learning process (2)] First, the re-learning process of the data generation process will be described with reference to a flow chart shown in FIG.

[0118] As shown in FIG. 19, the data management unit 112 waits until it receives one or more pieces of real data DT1 from the operation terminal 2 (NO in S81).

[0119] Then, when one or more pieces of real data DT1 are received (YES in S81), the data management unit 112 stores the one or more pieces of real data DT1 received in the processing of S81 in the information storage area 130 (S82).

[0120] Next, the data identification unit 115 further learns one or more pieces of genuine data DT1 received in the process of S81 (S83).

[0121] Furthermore, the data identifying unit 115 calculates an evaluation value indicating the probability that each piece of teacher data DT3 stored in the information storage area 130 is genuine data DT1 (S84).

[0122] After that, the data management unit 112 deletes the teacher data DT3 whose evaluation value calculated in the process of S84 is equal to or less than the threshold value from among the teacher data DT3 stored in the information storage area 130 (S85). Then, the information processing device 1 executes, for example, the data generation process described with reference to Figs. 8 to 11.

[0123] That is, the information processing device 1 in the third embodiment regenerates only the data that has been determined to be fake data by the data identification unit 115 that has been trained using the added genuine data DT1, out of the generated teacher data DT3.

[0124] This enables the information processing device 1 in the third embodiment to efficiently generate the teacher data DT3 even when the real data DT1 is added.

[0125] The information processing device 1 controls the processing from S35 in FIG. 8 onwards so that the processing is performed after the data re-learning processing is completed.

[0126] The above embodiment can be summarized as follows:

[0127] (Appendix 1) A data generation program that causes a computer to execute a process of generating data using a generative adversarial network, generating a plurality of first data from randomly generated noise; Calculating an evaluation value regarding authenticity of each of the generated first data; Identifying a common feature among a plurality of second data having the evaluation value lower than the reference data among the plurality of first data generated; regenerating the plurality of first data that do not have the identified characteristics; A data generation program that causes a computer to execute a process.

[0128] (Appendix 2) Appendix 1 further states: repeating the process of calculating the evaluation value, the process of identifying the characteristics, and the process of regenerating the plurality of first data until the evaluation value for one or more data among the plurality of first data satisfies a predetermined condition; When the evaluation value for the one or more pieces of data satisfies the predetermined condition in the repeated processing, the one or more pieces of data are output. A data generation program comprising:

[0129] (Appendix 3) In Appendix 1: In the process of calculating the evaluation value, a probability that each of the plurality of first data is not data generated in a process of generating the first data or a process of regenerating the first data is calculated as the evaluation value corresponding to each of the plurality of first data. A data generation program comprising:

[0130] (Appendix 4) In Appendix 2: In the output process, when the evaluation value for the one or more pieces of data becomes equal to or greater than a threshold value, the one or more pieces of data are output. A data generation program comprising:

[0131] (Appendix 5) In Appendix 1: the plurality of first data and the reference data are image data, In the process of generating the plurality of first data, the plurality of first data are generated by changing luminance of a plurality of pixels included in the reference data. A data generation program comprising:

[0132] (Appendix 6) In Appendix 5: In the process of identifying the features, For each of the plurality of second data, a first pixel having a greater luminance than a pixel at the same position in the reference data and a second pixel having a less luminance than a pixel at the same position in the reference data are identified; identifying a first position identified as the first pixel being located in a predetermined ratio of the second data among a plurality of positions at which pixels are located in each of the plurality of second data, and a second position identified as the second pixel being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first pixel is disposed at the first position and information indicating that the second pixel is disposed at the second position; A data generation program comprising:

[0133] (Appendix 7) In Appendix 6: In the process of regenerating the plurality of first data, the plurality of first data are regenerated such that a ratio of data in which the first pixel is arranged at the first position and the second pixel is arranged at the second position is equal to or less than a threshold value. A data generation program comprising:

[0134] (Appendix 8) In Appendix 5: In the process of identifying the features, for each of the plurality of second data, identifying a first block having a higher luminance than the block at the same position in the reference data and a second block having a lower luminance than the block at the same position in the reference data, for each block consisting of a plurality of pixels; identifying a third position identified as the first block being located in a predetermined ratio of the second data among a plurality of positions at which blocks are located in each of the plurality of second data, and a fourth position identified as the second block being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first block is arranged at the third position and information indicating that the second block is arranged at the fourth position; A data generation program comprising:

[0135] (Appendix 9) An information processing device that generates data using a generative adversarial network, a data generating unit that generates a plurality of first data from randomly generated noise; Calculating an evaluation value regarding authenticity of each of the generated first data; a feature specifying unit that specifies a common feature in a plurality of second data items having an evaluation value lower than that of a reference data item among the plurality of first data items generated, The data generation unit regenerates the plurality of first data that do not have the identified characteristics. 23. An information processing apparatus comprising:

[0136] (Appendix 10) In Appendix 9: the plurality of first data and the reference data are image data, The data generating unit generates the plurality of first data by changing luminance of a plurality of pixels included in the reference data. 23. An information processing apparatus comprising:

[0137] (Appendix 11) In Appendix 10: The feature identification unit is For each of the plurality of second data, a first pixel having a greater luminance than a pixel at the same position in the reference data and a second pixel having a less luminance than a pixel at the same position in the reference data are identified; identifying a first position identified as the first pixel being located in a predetermined ratio of the second data among a plurality of positions at which pixels are located in each of the plurality of second data, and a second position identified as the second pixel being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first pixel is disposed at the first position and information indicating that the second pixel is disposed at the second position; 23. An information processing apparatus comprising:

[0138] (Appendix 12) A data generation method in which a computer executes a process of generating data using a generative adversarial network, generating a plurality of first data from randomly generated noise; Calculating an evaluation value regarding authenticity of each of the generated first data; Identifying a common feature among a plurality of second data having the evaluation value lower than the reference data among the plurality of first data generated; regenerating the plurality of first data that do not have the identified characteristics; A data generating method, characterized in that processing is executed by a computer.

[0139] (Appendix 13) In Appendix 12: the plurality of first data and the reference data are image data, In the process of generating the plurality of first data, the plurality of first data are generated by changing luminance of a plurality of pixels included in the reference data. A data generation method comprising:

[0140] (Appendix 14) In Appendix 13: In the process of identifying the features, For each of the plurality of second data, a first pixel having a greater luminance than a pixel at the same position in the reference data and a second pixel having a less luminance than a pixel at the same position in the reference data are identified; identifying a first position identified as the first pixel being located in a predetermined ratio of the second data among a plurality of positions at which pixels are located in each of the plurality of second data, and a second position identified as the second pixel being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first pixel is disposed at the first position and information indicating that the second pixel is disposed at the second position; A data generation method comprising: [Explanation of symbols]

[0141] 1: Information processing device 2: Operation terminal 10: Information processing system 113: Noise generating section 114: Data generation unit 115: Data identification unit DT1: Real data DT2: Fake data DT3: New training data GA: Generative adversarial network NW: Network

Claims

1. A data generation program that causes a computer to execute a process of generating training data using a generative adversarial network, generating a plurality of first data by adding randomly generated noise to each of the plurality of training data; Calculating an evaluation value regarding authenticity of each of the generated first data; Identifying a common feature among a plurality of second data having the evaluation value lower than the reference data among the plurality of first data generated; generating the training data by adding noise capable of generating the plurality of first data not having the identified characteristic to each of the plurality of training data; A data generation program that causes a computer to execute a process.

2. In claim 1, further comprising: The process of calculating the evaluation value, the process of identifying the characteristics, and the process of generating the teacher data are repeated until the evaluation value for one or more data among the plurality of first data satisfies a predetermined condition; When the evaluation value for the one or more data satisfies the predetermined condition in the repeated processing, the one or more data are output as the teacher data. A data generation program comprising:

3. In claim 1, In the process of calculating the evaluation value, a probability that each of the plurality of first data is not data generated in a process of generating the first data is calculated as the evaluation value corresponding to each of the plurality of first data. A data generation program comprising:

4. In claim 2, In the output process, when the evaluation value for the one or more pieces of data becomes equal to or greater than a threshold value, the one or more pieces of data are output. A data generation program comprising:

5. In claim 1, the plurality of first data and the reference data are image data; In the process of generating the plurality of first data, the plurality of first data are generated by changing luminance of a plurality of pixels included in the reference data. A data generation program comprising:

6. In claim 5, In the process of identifying the features, identifying, for each of the plurality of second data, a first pixel having a luminance greater than that of a pixel at the same position in the reference data, and a second pixel having a luminance less than that of a pixel at the same position in the reference data; identifying a first position identified as the first pixel being located in a predetermined ratio of the second data among a plurality of positions at which pixels are located in each of the plurality of second data, and a second position identified as the second pixel being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first pixel is disposed at the first position and information indicating that the second pixel is disposed at the second position; A data generation program comprising:

7. In claim 5, In the process of identifying the features, for each of the plurality of second data, a first block having a higher luminance than the block at the same position in the reference data and a second block having a lower luminance than the block at the same position in the reference data, for each block consisting of a plurality of pixels; identifying a third position identified as the first block being located in a predetermined ratio of the second data among a plurality of positions at which blocks are located in each of the plurality of second data, and a fourth position identified as the second block being located in the predetermined ratio of the second data among the plurality of second data; identifying, as the feature, information indicating that the first block is arranged at the third position and information indicating that the second block is arranged at the fourth position; A data generation program comprising:

8. An information processing device that generates training data using a generative adversarial network, a data generating unit that generates a plurality of first data by adding randomly generated noise to each of the plurality of teacher data; a data identification unit that calculates an evaluation value regarding authenticity of each of the generated first data; a feature specifying unit that specifies a common feature among a plurality of second data pieces having the evaluation value lower than the reference data, among the plurality of first data pieces generated; The data generation unit generates the teacher data by adding noise capable of generating the plurality of first data not having the identified characteristic to each of the plurality of teacher data.

23. An information processing apparatus comprising:

9. A data generation method in which a computer executes a process of generating training data using a generative adversarial network, generating a plurality of first data by adding randomly generated noise to each of the plurality of training data; Calculating an evaluation value regarding authenticity of each of the generated first data; Identifying a common feature among a plurality of second data having the evaluation value lower than the reference data among the plurality of first data generated; generating the training data by adding noise capable of generating the plurality of first data not having the identified characteristic to each of the plurality of training data; A data generating method, characterized in that processing is executed by a computer.

Citation Information

Patent Citations

  • Image processing device, image processing method, image processing program, and teacher data generating method

    JP2018163554A

  • Learning method, computer program, classifier, generator, and processing system

    JP2020205030A

  • Shooting condition suggestion system

    JP6687962B1

  • JPP7424595B

  • Privacy Preserving Artificial Intelligence System For Dental Data From Disparate Sources

    US20200364624A1