MODEL GENERATION SYSTEM, SECURE COMPUTING DEVICE, ANALYSIS DEVICE, MODEL GENERATION METHOD, AND PROGRAM

Abstract

A model generation system according to the present invention is provided with a statistic value secure computation means, a synthetic data generation unit, and a model generation unit. N is a predetermined integer of 3 or greater. The statistic value secure computation means receives input of predetermined N shares of data, obtains a prescribed statistic value by using secure computation, and outputs plain text of the statistic value. The synthetic data generation unit generates synthetic data that is pseudo-data exhibiting a statistic value identical to the statistic value. The model generation unit generates a prediction model by using the synthetic data.

Description

[Technical field]

[0001] The present invention relates to a model generation system, a secure computing device, an analysis device, a model generation method, and a program for performing analysis without disclosing confidential information to an analyst. [Background technology]

[0002] In an environment where data distribution and utilization has become active, it is expected that data held by multiple organizations will be linked, the barriers between each organization will be eliminated, and then predictive models will be constructed using AI (Artificial Intelligence), regression analysis, etc. In such cases, the use of secure computing techniques disclosed in Non-Patent Documents 1 and 2 has attracted attention as a means of preventing information leaks between organizations. In addition, as another approach to data utilization, a method known as anonymous synthetic data generation technology, which generates pseudo data by inputting statistical values ​​generated from data disclosed in Non-Patent Document 3, is also known. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Kibuki Mishina, Hiroki Hamada, and Dai Igarashi, "Secure Computation Deep Learning Reproducing Processing Logic on Plain Text", FIT2019 (18th Forum on Information Science and Technology), Vol. 2, pp.81-88 (2019). [Non-Patent Document 2] Dai Igarashi, Hiroki Hamada, Ryo Kikuchi, Koji Senda, "Design and implementation of ultra-fast secure computation sorting: The day when secure computation will rival scripting languages", CSS2017, pp.600-607 (2017). [Non-Patent Document 3] Rina Okada, Akio Masaki, Satoshi Hasegawa, Satoshi Tanaka, "Privacy-preserving pseudo-data generation method using statistics", CSS2017, pp.1366-1372(2017). Summary of the Invention [Problem to be solved by the invention]

[0004] However, there is a problem that the processing cost increases when the secure computation techniques shown in Non-Patent Documents 1 and 2 are applied to all processing. In addition, in the anonymous synthetic data generation technique, in order to generate synthetic data by linking data from multiple organizations, it is necessary to obtain data in plain text, generate a table by combining data held by each organization, and then obtain necessary statistics. The present invention has been made in consideration of such problems, and aims to provide a prediction model generation technique that reduces processing costs while maintaining the confidentiality of data. [Means for solving the problem]

[0005] The model generation system of the present invention includes a statistical value secure calculation means, a synthetic data generation unit, and a model generation unit. N is a predetermined integer equal to or greater than 3. The statistical value secure calculation means receives N shares of predetermined data as input, calculates a predetermined statistical value using secure calculation, and outputs a plaintext of the statistical value. The synthetic data generation unit generates synthetic data, which is pseudo data having the same statistical value as the statistical value. The model generation unit generates a prediction model using the synthetic data. Effect of the Invention

[0006] According to the model generation system of the present invention, the input to the statistical secure calculation means is a share of data, so that the data itself can be kept secret. Furthermore, the statistical secure calculation means outputs statistics in plaintext, and the generation of synthetic data and prediction models is performed in plaintext, so that processing costs can be suppressed. [Brief description of the drawings]

[0007] [Figure 1] FIG. 2 is a diagram showing an example of the functional configuration of the model generation system of the present invention. [Diagram 2] FIG. 11 is a diagram showing a processing flow for secret sharing of shares of data recorded by a data providing device. [Diagram 3] FIG. 13 is a diagram showing a process flow for generating a predictive model. [Figure 4]FIG. 13 is a diagram showing a processing flow for secret sharing of a share of a prediction model recorded by an analysis device and a share of data recorded by a data providing device. [Diagram 5] FIG. 13 is a diagram showing a process flow for updating a prediction model. [Figure 6] FIG. 2 is a diagram showing an example of the functional configuration of a computer. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Hereinafter, an embodiment of the present invention will be described in detail. Note that components having the same functions are given the same reference numbers and duplicated explanations will be omitted. EXAMPLES

[0009] <Prediction model generation> Fig. 1 shows an example of the functional configuration of a model generation system of the present invention. Fig. 2 is a diagram showing a processing flow for secretly sharing shares of data recorded by a data providing device. Fig. 3 is a diagram showing a processing flow for generating a prediction model. In the following description, N is a predetermined integer of 3 or more, n is an integer between 1 and N, K is a predetermined integer between 2 and N, M is an integer of 1 or more, and m is an integer between 1 and M.

[0010] The model generation system 10 communicates with data providing devices 3001, . . . , 300 located inside or outside via a network 900. M The data providing device 300 is connected to m The data recording unit 390 records the data. m and data provider 310 m The data providing device 300 m is a database owned by an organization (such as a company or government) that provides data, and the data is stored in the data recording unit 390 m The data is recorded in the data provider 310. m generates N shares and provides the shares to the model generation system 10 (S311 m). A "share" is a fragment of data, and the data can be restored by obtaining K or more shares, but the data cannot be restored (the secret will not be leaked) if fewer than K shares are obtained. Details regarding shares are provided in non-patent documents 1 and 2, etc.

[0011] The model generation system 10 includes secure computing devices 1001, . . . , 100 connected via a network 900. N and an analysis device 200. n Statistical value secure calculation unit 110 n , Share Record Section 190 n The statistical value secure calculation unit 110 n takes one of N shares of predetermined data as input and processes it into other secure computing devices 1001, ..., 100 N The share recorder 190 obtains one of the shares of a predetermined statistical value by secure computation in cooperation with the scalar. Details of the secure computation technology are described in Non-Patent Documents 1 and 2. n data providing devices 3001, ..., 300 M Share obtained from Statistical Value Secret Calculation Section 110 n As shown in FIG. 2, when the model generation system 10 generates a prediction model, the data providing device 300 that holds data necessary for generating the prediction model is used. m generates N shares and provides the shares to the model generation system 10 (S311 m ). secret computing device 100 n obtains one of the N shares of the data and stores it in the share recorder 190 n Record in (S101 n More specifically, data providing devices 3001, . . . , 300 M However, each share is sent to the secure computing device 1001, ..., 100 N However, if the data can be concealed, the data providing unit 310 m The data providing device 300 m It may exist outside of the

[0012] The analysis device 200 includes a statistical plaintext generation unit 210, a synthetic data generation unit 220, a model generation unit 230, and an analysis and recording unit 290. The statistical plaintext generation unit 210 acquires K or more shares of a predetermined statistical value calculated using secure computation, and generates plaintext of the predetermined statistical value (S210). The synthetic data generation unit 220 generates synthetic data, which is pseudo data having the same statistical value as the aforementioned statistical value (S220). The model generation unit 230 generates a prediction model using the synthetic data (S230). The analysis device 200 is represented by secure computation devices 1001, ..., 100 in FIG. 1. N Although shown as a separate device, any of the secure computing devices 100 n may also serve as the analysis device 200.

[0013] 1 and 3, a process flow for generating a prediction model will be described. The model generation system 10 includes a statistical value secure calculation means 110, a synthetic data generation unit 220, and a model generation unit 230. The synthetic data generation unit 220 and the model generation unit 230 are included in an analysis device 200. The statistical value secure calculation means 110 is provided for each of the secure calculation devices 1001, ..., 100 N The statistical value secure calculation units 1101, ..., 110 N and a statistical plaintext generation unit 210 included in the analysis device 200. The statistical secure calculation means 110 receives N shares of predetermined data as input, calculates a predetermined statistical value using secure calculation, and outputs the plaintext of the statistical value (S110). More specifically, the statistical secure calculation unit 110 n takes one of N shares of predetermined data as input and processes it into other secure computing devices 1001, ..., 100 N Statistical value secret calculation unit 1101, ..., 110 N By secret computation in cooperation with the N ) The "statistical value" may be, but is not limited to, an average, variance-covariance, etc. for each attribute. Any statistical value necessary for generating a prediction model may be selected appropriately. The statistical value plaintext generation unit 210 obtains K or more shares of a predetermined statistical value calculated using secure computation, and generates a plaintext of the predetermined statistical value (S210). Steps 1101, ..., 110 NThe statistical secure computation step (S110) includes S210 and S210.

[0014] The synthetic data generation unit 220 generates synthetic data, which is pseudo data having the same statistical value as the statistical value output by the statistical value plaintext generation unit 210 (S220). The model generation unit 230 generates a prediction model using the synthetic data (S230). The synthetic data generation (S220) may be performed using the technology shown in Non-Patent Document 3. The prediction model generation (S230) may be performed using existing technology. For example, the prediction model may be a trained model trained using the synthetic data as training data, or may be a model obtained by regression analysis. The prediction model may be appropriately selected depending on the event to be predicted.

[0015] According to the model generation system 10, the input to the statistical value secure computation means 110 is a share of data, so the data itself can be kept secret. Also, the statistical value secure computation means 110 outputs statistical values ​​in plain text, and the generation of synthetic data and prediction models is performed in plain text, so processing costs can be suppressed. Furthermore, since the number of operations that utilize secure computation techniques is reduced, it is possible to reduce restrictions on operations that can be performed using secure computation techniques.

[0016] For example, it is assumed that the data providing device 3001 holds an identifier for identifying an individual (e.g., a number for identifying an individual) and data on a purchase amount, and the data providing device 3002 holds an identifier for identifying an individual and data on a gender. N obtains the identifier and the share of the purchase amount from the data providing device 3001, and obtains the identifier and the share of the gender from the data providing device 3002 (S3111, S3112, S1011, ..., S101 N ) And the statistical value secure calculation units 1101, ..., 110 N Using secret calculation, calculate the share of the statistical value of the purchase amount limited to men and the share of the statistical value of the purchase amount limited to women (S1101, ..., S110 N). The statistical plaintext generation unit 210 acquires K or more shares of the statistical value for each gender calculated using secure computation, and generates plaintext of the statistical value for each gender (S210). The synthetic data generation unit 220 generates synthetic data, which is pseudo data that has the same statistical value as the generated statistical value (S220). The model generation unit 230 generates a prediction model using the synthetic data (S230). Since a prediction model for each gender can be generated by linking the data of the data providing device 3001 and the data of the data providing device 3002, the prediction accuracy can be improved compared to a prediction model using only the data of the data providing device 3001.

[0017] <Update forecast model> 4 is a diagram showing a processing flow for secretly sharing the share of the prediction model recorded by the analysis device and the share of the data recorded by the data providing device. FIG. 5 is a diagram showing a processing flow for updating the prediction model. When updating the prediction model, the secure computing device 100 n Model update secret calculation unit 150 n The analysis device 200 also includes a model providing unit 240 and a model plaintext generating unit 250.

[0018] The model providing unit 240 transmits information about existing prediction models (prediction models recorded by the analysis recording unit 290) to the secure computing devices 1001, . . . , 100 N (S240). The "information about the prediction model" may be the prediction model itself, or one of the N shares of the prediction model. In addition, the data providing device 300 that has additional data required for generating the prediction model may transmit the information about the prediction model to the data providing device 300. m generates N shares of the additional data and provides the shares to the model generation system 10 (S312 m ). secret computing device 100 n obtains information about the prediction model and one of the N shares of the additional data, and stores the information in the share recorder 190; n (S102 n ). Data providing devices 3001, ..., 300 M However, each additional data share is sent to the secure computing device 1001, ..., 100 NAdditional data can be anonymized by sending it separately to

[0019] Model Update Secret Calculation Unit 150 n takes as input information about the prediction model and one of N shares of predetermined additional data, and transmits the information to other secure computing devices 1001, ..., 100 N Model update secret calculation unit 1501, ..., 150 N By secret computation in cooperation with the n The model plaintext generation unit 250 obtains K or more shares of the prediction model updated using secure computation, and generates plaintext for the updated prediction model (S250). The model update secure computation means 150 obtains K or more shares of the prediction model updated using secure computation, and generates plaintext for the updated prediction model (S250). N The model update secret calculation units 1501, ..., 150 N and a model plaintext generation unit 250 included in the analysis device 200. The model update secure computation means 150 receives an existing prediction model and N shares of predetermined additional data as input, and updates the prediction model using secure computation (S150). Steps 1501, ..., 150 N The model updating secure computation step (S150) is composed of S250 and S250. More specifically, the model may be updated using the algorithm shown in Non-Patent Document 1. Alternatively, the prediction model may be updated using regression analysis with added data.

[0020] According to the model generation system 10, the input to the model update secure computation means 150 is a share of the additional data, so that the prediction model can be updated while keeping the additional data itself secret. Furthermore, in the process of generating the prediction model, the data itself can be kept secret and processing costs can be suppressed. However, the calculations required to update the prediction model are limited to those that can be performed using secure computation technology.

[0021] For example, the data providing device 3001 holds data on an identifier for identifying an individual (e.g., a number for identifying an individual) and a purchase amount, and the data providing device 3002 holds data on an identifier for identifying an individual and gender. It is assumed that a prediction model for men and a prediction model for women are generated in the process flow of generating a prediction model, and the data providing device 3001 holds additional data on an identifier for identifying an individual and a purchase amount acquired after the generation of the prediction model, and the data providing device 3002 holds additional data on an identifier for identifying an individual and gender. The secure computing devices 1001, ..., 100 N obtains a share of additional data of identifier and purchase amount from the data providing device 3001, and obtains a share of additional data of identifier and gender from the data providing device 3002 (S3121, S3122, S1021, ..., S102 N ) Then, the model update secret calculation units 1501, ..., 150 N Using secret calculation, the share of the prediction model of the purchase amount limited to men and the share of the prediction model of the purchase amount limited to women are calculated (S1501, ..., S150 N ). The model plaintext generation unit 250 acquires K or more shares of the prediction model for each gender that have been updated using secure computation, and generates plaintext for the updated prediction model for each gender (S250). By linking the additional data of the data providing device 3001 and the additional data of the data providing device 3002, the prediction model for each gender can be updated with the additional data, thereby improving the accuracy of prediction.

[0022] [Programs, recording media] The various processes described above can be implemented by loading a program for executing each step of the above method into the recording unit 2020 of the computer 2000 shown in FIG. 6, and operating the control unit 2010, input unit 2030, output unit 2040, display unit 2050, etc.

[0023] The program describing the processing contents can be recorded on a computer-readable recording medium, which may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other type of recording medium.

[0024] The program may be distributed, for example, by selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of a server computer and transferring the program from the server computer to other computers via a network.

[0025] A computer that executes such a program, for example, first stores the program recorded on a portable recording medium or the program transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from the portable recording medium and execute a process according to the program, or may execute a process according to the received program each time a program is transferred from the server computer to the computer. The above-mentioned process may be executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by issuing an execution instruction and obtaining a result, without transferring a program from the server computer to the computer. Note that the program in this embodiment includes information used for processing by an electronic computer and equivalent to a program (data that is not a direct command to the computer but has a nature that specifies the processing of the computer, etc.).

[0026] In addition, in this embodiment, the present device is configured by executing a specific program on a computer, but at least a part of the processing contents may be realized by hardware. [Explanation of symbols]

[0027] 10 Model Generation System 100 n secret computing device 110 Secret calculation method for statistical values ​​110 n Statistical Value Secret Calculation Department 150 Model Update Secret Computation Method 150 n Model Update Secret Calculation Unit 190 n Share recorder 200 Analysis equipment 210 statistical plaintext generator 220 synthetic data generator 230 Model generation unit 240 Model provision unit 250 Model plaintext generation unit 290 Analysis and recording unit 300 m Data providing device 310 m Data Provision Department 390 m Data recording unit 900 Network

Claims

1. A model generation system having N secure computing devices and generating models for data held by a plurality of data providing devices, N is a predetermined integer of 3 or more, a statistical value secure computation means for inputting each of the N shares of the data held by each of the data providing devices into each of the secure computation devices, calculating a predetermined statistical value using secure computation, and outputting a plaintext of the statistical value; a synthetic data generation unit that generates synthetic data that is pseudo data having the same statistical value as the statistical value; a model generation unit that generates a predictive model using the synthetic data; A model generation system comprising:

2. 2. The model generation system of claim 1, a model updating secure computation means for inputting information about an existing prediction model and each of the N shares of additional data generated and held by each of the data providing devices into each of the secure computation devices, and updating the prediction model using secure computation; A model generation system that also has

3. One of the N secure computing devices included in the model generation system according to claim 2, a statistical value secure computation unit that receives one of the N shares of the additional data as an input and obtains one of the shares of the predetermined statistical value by secure computation in cooperation with other secure computation devices; a model updating secure computation unit that receives information about an existing prediction model and one of the N shares of the predetermined additional data as input, and obtains a share of a prediction model obtained by updating the prediction model through secure computation in cooperation with other secure computation devices; A secure computing device comprising:

4. An analysis device included in the model generation system according to claim 1 or 2, K is a predetermined integer between 2 and N, a statistical plaintext generation unit that acquires K or more shares of the predetermined statistical value calculated using secure computation and generates a plaintext of the predetermined statistical value; The synthetic data generation unit and the model generation unit An analytical device comprising:

5. The analysis device according to claim 4, a model provider for generating N shares of a predictive model; a model plaintext generation unit that acquires K or more shares of the prediction model updated using secure computation and generates a plaintext of the prediction model; An analytical device comprising:

6. A model generation method for generating a model for data held by a plurality of data providing devices, using N secure computing devices and an analysis device, comprising: N is a predetermined integer of 3 or more, a statistical secure computation step in which each of the N shares of the data held by each of the data providing devices is input to each of the secure computation devices, and the N secure computation devices calculate a predetermined statistical value using secure computation and output a plaintext of the statistical value; a synthetic data generating step in which the analysis device generates synthetic data that is pseudo data having the same statistical value as the statistical value; a model generation step in which the analysis device generates a predictive model using the synthetic data; A model generation method to perform.

7. 7. The model generation method according to claim 6, further comprising the steps of: a model updating secure computation step in which information about an existing prediction model and each of the N shares of additional data generated and held by each of the data providing devices are input to each of the secure computation devices, and the N secure computation devices update the prediction model using secure computation; A model generation method that also performs

8. A program for causing a computer to function as the secure computing device according to claim 3 or the analysis device according to claim 4 or 5.

Images