Information processor, method for processing information, and program

Abstract

To provide a technique of discretization to improve the quality of output of a data synthesis model which protects privacy.SOLUTION: An information processor for determining the number of sections which divide the range of a continuous variable to discretize data of the continuous variable includes an operation unit for determining the number of the sections so that the average value of the number of pieces of data in one section and the magnitude of noise applied to the number of pieces of data in one section will be in a predetermined ratio on the basis of the number of pieces of data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to privacy protection technology. [Background technology]

[0002] Various models have been proposed as privacy-preserving data synthesis models. Most of them can be classified into two types: Cont (continuous) and Disc (discrete).

[0003] (Cont) The model only accepts continuous variable data (numerical data) as input and output.

[0004] Disc (Disc) models have only data consisting of discrete variables (categorical data) as input and output. Examples of conventional techniques that belong to Disc include the techniques disclosed in Non-Patent Documents 1 to 3.

[0005] When you want to input a discrete variable to a data synthesis model belonging to Cont, you encode the data of the discrete variable to make it a continuous variable before inputting it. For example, one-hot encoding is often used to convert all three labels, A, B, and C, into (1,0,0), (0,1,0), and (0,0,1).

[0006] When you want to input continuous variables to Disc's data synthesis model, you must first discretize the data for that continuous variable. Discretization is a process in which the range of the variable is divided to create a sequence of real intervals (bins), and the data for each continuous variable is replaced with a discrete label depending on which interval it belongs to. The results of discretization vary greatly depending on how the sequence of real intervals (bins) is created.

[0007] Research has shown that the synthetic data output by some of the techniques in the Disc category is better than that output by any of the techniques in the Cont category. Therefore, if you want to perform privacy-preserving data synthesis, you can input discretized data (if necessary) from continuous variables into a data synthesis model that belongs to the Disc category. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Ryan Mckenna, Daniel Sheldon, and Gerome Miklau, "Graphical-Model Based Estimation and Inference for Differential Privacy," In: Proceedings of the 36th International Conference on Machine Learning. on 08 / 04 / 2022). [Non-Patent Document 2] Ryan McKenna et al, "AIM: An Adaptive and Iterative Mechanism for Differentially Private Synthetic Data," Jan. 29, 2022. doi: 10.48550 / arXiv.2201.12677. arXiv: 2201.12677 [cs]. url: http: / / arxiv.org / abs / 2201.12677 (visited on 07 / 25 / 2022). [Non-Patent Document 3] Jun Zhang et al, "PrivBayes: Private Data Release via Bayesian Networks," In: ACM Transactions on Database Systems 42.4 (Oct. 27, 2017), 25:1-25:41. issn: 0362-5915. doi: 10.1145 / 3134428. url:https: / / doi.org / 10.1145 / 3134428 (visited on 05 / 18 / 2021). Summary of the Invention [Problem to be solved by the invention]

[0009] The quality of the data output from a data synthesis model belonging to a Disc changes significantly depending on the discretization method combined with the model. However, in the techniques disclosed in Non-Patent Documents 1 to 3, discretization processing is selected without any basis. In addition, although the Sturges' rule is widely known as a method for determining a discretization method, these existing discretization methods do not take into account noise that is added later for privacy protection. Therefore, when the noise becomes large, the existing discretization methods are not necessarily optimal.

[0010] The present invention has been made in consideration of the above points, and aims to provide a discretization technique for improving the quality of the output of a data synthesis model that preserves privacy. [Means for solving the problem]

[0011] According to the disclosed technology, there is provided an information processing device that determines the number of intervals into which a range of a continuous variable is equally divided in order to discretize data of the continuous variable, the information processing device comprising: A calculation unit that determines the number of intervals based on the number of data in the data such that an average value of the number of data corresponding to one interval and a magnitude of noise added to the number of data corresponding to one interval have a predetermined ratio. An information processing device comprising: Effect of the Invention

[0012] The disclosed techniques provide discretization techniques for improving the quality of the output of privacy-preserving data synthesis models. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a configuration diagram of an information processing device 100. [Diagram 2] 2 is a flowchart illustrating an operation of the information processing device 100. [Diagram 3] FIG. 13 is a diagram illustrating an example of division into sections. [Figure 4] 1 is a diagram illustrating an example of a hardware configuration of an information processing device 100. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment of the present invention (the present embodiment) will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0015] (Outline of the embodiment) In this embodiment, an information processing device 100, which will be described later, performs discretization and data synthesis of input data that is a continuous variable.

[0016] In this embodiment, the range [X min ,X max ](∋X) and the number n of input data are assumed to be public, and the information processing device 100 determines how many bins (real intervals (intervals of real numbers)) the range should be equally divided into.

[0017] In addition, X min It is sufficient that X is smaller than the minimum value of the input data, and it does not have to be equal to the minimum value of the input data. max It is also sufficient that is larger than the maximum value of the input data.

[0018] The information processing device 100 performs data synthesis that takes privacy protection into consideration. In the data synthesis, the information processing device 100 counts discrete variables and then adds noise to the counting results in order to protect privacy. In this embodiment, the counting results are the number of data for each section (i.e., a histogram).

[0019] In this embodiment, the information processing device 100 improves the quality of the output data by creating bins (division into intervals) that take this noise into consideration.

[0020] The expected absolute value of the noise added for privacy protection (hereinafter referred to as the “noise magnitude”) is determined by various factors such as the number of data n and the strength of privacy protection, but in this embodiment, only the number of bins d can be changed, and the noise magnitude is written as μ(d) (≧0).

[0021] The configuration and operation of the information processing device 100 will be described in detail below.

[0022] (Configuration and Operation Overview of Information Processing Device 100) 1 shows an example of a functional configuration of an information processing device 100 according to the present embodiment. As shown in FIG. 1, the information processing device 100 includes an input unit 110, a calculation unit 120, an output unit 130, and a storage unit 140.

[0023] An example of the overall operation of the information processing device 100 will be described with reference to the flowchart in Fig. 2. In S101, data to be subjected to data synthesis (data of continuous variables) is input from the input unit 110. The input data is stored in the storage unit 140.

[0024] In S102, the calculation unit 120 reads out input data from the storage unit 140 and performs discretization processing on the data. The details of the discretization processing will be described later. The data that has been subjected to the discretization processing (specifically, data with discrete labels) is stored in the storage unit 140.

[0025] In S103, the calculation unit 120 reads out the discretized data from the storage unit 140, performs counting on the data, and adds noise to the counting result. S103 corresponds to calculation using a data synthesis model.

[0026] In S104, the output section 130 outputs the counting result to which noise has been added.

[0027] The information processing device 100 may perform up to the discretization process, and other devices may perform the processes after the discretization process. Also, the information processing device 100 may perform the process of determining d based on the number of data n, and other devices may perform the other processes.

[0028] (Details of discretization process) The discretization process in S102 will now be described in detail.

[0029] As described above, in the discretization process, the calculation unit 120 divides the range of the input data (continuous variables) to create a sequence of real intervals (bins), and replaces the data with discrete labels depending on which interval each piece of data of the continuous variables belongs to.

[0030] For example, the range of a continuous variable X [X min ,X max An image of dividing [ into three equal parts is shown in Figure 3. In the example in Figure 3, the labels A, B, and C are associated with the three sections.

[0031] <Determining the number of bins> The method for determining the number of bins will now be described in detail.

[0032] If the number of bins d is reduced, the distribution of the original continuous variable X is lost through discretization. For example, if there is only one bin, information other than the original number of data points and the range of the data is lost. On the other hand, if the number of bins is increased, the number of data points that fall into one bin will decrease. For example, if the number of bins d is set to the number of data points n, the number of data points that fall into one bin will be around 1 or 2. In that case, the information about the number of data points will be drowned out by the noise added later.

[0033] To prevent such a situation, the calculation unit 120 selects the number of bins d as follows: When the number of data n (>0) and the noise level μ(d) (≧0) are known, the number of bins d (>0) should be selected so that the following value is closest to the constant α (>0). Empirically, it is known that a constant of around 0.1 is good.

[0034] μ(d) / (n / d) That is, the calculation unit 120 in this embodiment The average number of data points in one bin, n / d, The amount of noise added to the number of data points in one bin, μ(d), is The number of bins is chosen to give an appropriate ratio α.

[0035] After the number of bins d (>0) is determined, the calculation unit 120 determines the range of the variable [X min ,X max ] is divided into d equal intervals, and each data point is labeled according to which interval it falls in, completing the discretization.

[0036] The above α is not limited to 0.1, and any value satisfying 0<α≦1 can be used as α.

[0037] μ(d) for each d is, for example, stored in advance in the storage unit 140. That is, the calculation unit 120 can determine d such that “μ(d) / (n / d)” is closest to the constant α (>0) by referring to the storage unit 140.

[0038] <Estimation of noise magnitude> Even if the noise magnitude μ(d) is not known in advance, it can be experimentally investigated. Below, a method for experimentally estimating μ(d) will be described. Although any device may perform the calculation for the estimation, here, as an example, it is assumed that the information processing device 100 has a data synthesis model and performs the calculation for the estimation.

[0039] In data synthesis that protects privacy, μ(d) may depend on factors such as the number of data n, but it usually does not change depending on the distribution of the data. Therefore, the type of function μ(d) is can be investigated in advance experiments using appropriate data. Specifically, the noise level μ(d) is investigated using the following procedure.

[0040] (1) Create data D0 that follows a uniform distribution. Specifically, follow steps 1-1 and 1-2 below.

[0041] <1-1> First, the information processing device 100 calculates d types of values ​​{v1,...,v d} is prepared.

[0042] <1-2> Next, the information processing device 100 assigns each value to n / d data points. As an example, if n=9, d=3, {v1, v2, v3}={A, B, C}, and the data points are 1, 2, 3, 4, 5, 6, 7, 8, and 9, then A is assigned to 1, 2, and 3, B is assigned to 4, 5, and 6, and C is assigned to 7, 8, and 9.

[0043] The set of data points that are assigned values ​​is D0.

[0044] (2) Repeat steps 2-1, 2-2, and 2-3 below to calculate M multiple times and calculate the average value.

[0045] <2-1> The information processing device 100 inputs the data D0 into a data synthesis model to obtain synthesized data D1.

[0046] <2-2> The information processing device 100 calculates a value v i The number of data points with i Count the number.

[0047] In the above example, adding noise may cause, for example, the number of data points assigned B to change from 3 to 4, and the number of data points assigned C to change from 3 to 2, etc. Also, the total numbers may change before and after adding noise.

[0048] <2-3> The information processing device 100 is d i=1 |c i Calculate |-n / d|. This calculation allows you to understand the deviation of the data (absolute value of noise) before and after inputting it into the data synthesis model.

[0049] The average value (expected value) of M obtained from the experiment consisting of steps (1) and (2) above is the estimated value of μ(d). μ(d) is not related to privacy information, so it may be made public.

[0050] In addition, in the differentially private data synthesis model proposed in Non-Patent Documents 1 and 2, it is known that μ(d) is approximately proportional to log(d / ε) when the privacy budget (strength of privacy protection) is ε.

[0051] For example, when the information processing device 100 uses the differentially private data synthesis model proposed in Non-Patent Documents 1 and 2 as the data synthesis model, C log(d / ε) can be used as μ(d) in the discretization process in S102, where C is a certain constant.

[0052] That is, the information processing device 100 can use, as d, a value that makes "C·log(d / ε) / (n / d)" closest to the constant α (>0).

[0053] (Hardware configuration example) The information processing device 100 described in this embodiment can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0054] That is, the device can be realized by executing a program corresponding to the processing performed by the device, using hardware resources such as a CPU and memory built into a computer. The program can be recorded on a computer-readable recording medium (such as a portable memory) and stored or distributed. The program can also be provided via a network such as the Internet or email.

[0055] Fig. 4 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 4 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., which are mutually connected by a bus BS. The computer may further include a GPU.

[0056] The program for realizing the processing on the computer is provided by a recording medium 1001 such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0057] The memory device 1003 reads out and stores the program from the auxiliary storage device 1002 when an instruction to start the program is received. The CPU 1004 realizes functions related to the information processing device 100 in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network or the like. The display device 1006 displays a GUI (Graphical User Interface) or the like according to a program. The input device 1007 is composed of a keyboard and mouse, buttons, a touch panel, or the like, and is used to input various operation instructions. The output device 1008 outputs the results of calculations.

[0058] (Summary of the embodiment, effects, etc.) As described above, in the technology described in this embodiment, the information processing device 100 determines how many small intervals to divide the range of a continuous variable into, taking into consideration the magnitude of noise that will be added to a histogram created by discretizing the continuous variable. Specifically, the information processing device 100 determines the number of small intervals so that the average number of data corresponding to one interval and the expected value of the absolute value of the noise (noise magnitude) added to the number of data corresponding to one interval have a predetermined ratio. By determining the number of small intervals in this manner, the number of small intervals can be optimized, and it is possible to prevent noise from drowning out information.

[0059] When using a data synthesis model that handles only discrete variables, by determining the number of small intervals in the discretization of continuous variables using the method described above, it is possible to obtain high-quality synthetic data while protecting privacy.

[0060] The following supplementary notes are further disclosed regarding the above embodiment.

[0061] <Additional Notes> (Additional note 1) 1. An information processing apparatus for determining the number of intervals into which a range of a continuous variable is equally divided in order to discretize data of the continuous variable, comprising: Memory, at least one processor coupled to the memory; Including, The processor, The number of intervals is determined based on the number of data in the data such that the average value of the number of data corresponding to one interval and the magnitude of noise added to the number of data corresponding to one interval have a predetermined ratio. Information processing device. (Additional note 2) If the ratio is α, then α is a value that satisfies 0<α≦1. Item 1. An information processing device according to claim 1. (Additional note 3) The processor uses, as the noise magnitude, an estimated value based on an expected value of the absolute value of noise calculated from data input to a data synthesis model and data output from the data synthesis model. Item 1. An information processing device according to claim 1. (Additional note 4) When using a differentially private data synthesis model as a data synthesis model to which data obtained by discretizing continuous variable data is input, The processor uses C log(d / ε) as the noise magnitude, where ε is the privacy budget and C is a constant. Item 1. An information processing device according to claim 1. (Additional note 5) 1. An information processing method for determining the number of intervals into which a range of a continuous variable is equally divided in order to discretize data of the continuous variable, comprising: The number of intervals is determined based on the number of data in the data such that the average value of the number of data corresponding to one interval and the magnitude of noise added to the number of data corresponding to one interval have a predetermined ratio. Information processing methods. (Additional note 6) A non-transitory storage medium storing a program for causing a computer to function as a calculation unit in the information processing device according to any one of claims 1 to 4.

[0062] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and variations are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0063] 100 Information processing device 110 Input section 120 Arithmetic section 130 Output section 140 Storage section 1000 Drive Device 1001 Recording media 1002 Auxiliary storage 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input Device 1008 Output device

Claims

1. 1. An information processing apparatus for determining the number of intervals into which a range of a continuous variable is equally divided in order to discretize data of the continuous variable, comprising: A calculation unit that determines the number of intervals based on the number of data in the data such that an average value of the number of data corresponding to one interval and a magnitude of noise added to the number of data corresponding to one interval have a predetermined ratio. An information processing device comprising:

2. If the ratio is α, α is a value that satisfies 0<α≦1. The information processing device according to claim 1 .

3. The calculation unit uses, as the magnitude of the noise, an estimated value based on an expected value of the absolute value of the noise calculated from the data input to a data synthesis model and the data output from the data synthesis model. The information processing device according to claim 1 .

4. When using a differentially private data synthesis model as a data synthesis model to which data obtained by discretizing continuous variable data is input, When the privacy budget is ε and the constant is C, the calculation unit uses C·log(d / ε) as the noise magnitude. The information processing device according to claim 1 .

5. 1. An information processing method for determining the number of intervals into which a range of a continuous variable is equally divided in order to discretize data of the continuous variable, comprising: The number of intervals is determined based on the number of data in the data such that the average value of the number of data corresponding to one interval and the magnitude of noise added to the number of data corresponding to one interval have a predetermined ratio. Information processing methods.

6. A program for causing a computer to function as the arithmetic unit in the information processing device according to any one of claims 1 to 4.

Images