KR20230056422A - Method and apparatus for generating synthetic data - Google Patents

Abstract

Disclosed are a method for generating synthetic data and an apparatus thereof. To this end, the apparatus for generating synthetic data according to one embodiment of the present invention may comprise: a data separating unit which separates original data into continuous data and categorical data; a data conversion unit which generates conversion data in which categorical data is converted into a continuous type; and a data analysis unit which deduces the latent vector of an input vector composed of the continuous data and the conversion data based on a normalizing flow to which differential privacy is applied, and the transformation function of the input vector and the latent vector.

Description

Method and apparatus for generating reproducible data {METHOD AND APPARATUS FOR GENERATING SYNTHETIC DATA}

The disclosed embodiments relate to reproduction data generation technology.

In general, a technique for generating reproducible data for tabular data has been developed based on statistical modeling or modeling using deep learning. In particular, the reproduction data generated by the deep learning model is widely used due to its high quality, but there is a problem in that sensitive information of the original data is revealed due to various attack techniques. As a way to solve this problem, differential privacy-based reproduction data generation technologies are being studied.

However, unlike image data, tabular data has a mixture of continuous and categorical variables, so when applying deep learning models targeting existing image data, it is not possible to simultaneously process the characteristics of different types of variables. quality may deteriorate. Accordingly, it is necessary to improve the model in order to apply it to tabular data in which continuous and categorical variables are mixed.

US Registered Patent No. US 9,672,364 (registered on June 6, 2017)

The disclosed embodiments are intended to provide a method and apparatus for generating reproduction data.

An apparatus for generating reproduced data according to an embodiment includes a data separator separating original data into continuous data and categorical data; a data conversion unit generating converted data obtained by converting categorical data into continuous data; And a latent vector of an input vector composed of continuous data and transformation data based on a normalizing flow with differential privacy applied, and data for deriving a transformation function of the input vector and the latent vector. An analysis unit may be included.

The reproduced data generating apparatus may further include a reproduced data generating unit generating reproduced data of the original data from the latent vector based on the inverse function of the transformation function.

The data conversion unit may estimate a probability distribution of the conditional random noise of the categorical data, and may generate converted data by adding the random noise obtained through the estimated probability distribution to the categorical data.

The data analysis unit may include a linear normalization flow unit including a first artificial neural network trained to analyze correlation characteristics between variables of an input vector based on a linear transformation.

Matrix decomposition may be applied to the weight matrix of the linear transformation.

The data analysis unit may include a spline normalization flow unit including a second artificial neural network trained to analyze multimodal characteristics of an input vector based on spline transformation.

When repeatedly performing spline transformation, the spline normalization flow unit performs learning only at the first transformation, and can share learned parameters in subsequent transformations.

The data analysis unit may include a variable exchange unit that changes the order of variables of the input vector.

The data analysis unit may include one or more artificial neural networks trained using gradient descent based on differential information protection to which gradient clipping and noise insertion are applied.

The data analysis unit may convert a value of an element having a size equal to or less than a predetermined reference value among elements of a gradient vector into 0 during learning using the differential information security based gradient descent method.

A method for generating reproduction data according to an embodiment includes a data separation step of separating original data into continuous data and categorical data; a data conversion step of generating converted data obtained by converting categorical data into continuous data; And a latent vector of an input vector composed of continuous data and transformation data based on a normalizing flow with differential privacy applied, and data for deriving a transformation function of the input vector and the latent vector. An analysis step may be included.

The reproduction data generation method may further include a reproduction data generation step of generating reproduction data of the original data from the latent vector based on the inverse function of the transformation function.

In the data conversion step, a probability distribution of conditional random noise of the categorical data is estimated, and the converted data may be generated by adding the random noise obtained through the estimated probability distribution to the categorical data.

The data analysis step may use a first artificial neural network trained to analyze correlation characteristics between variables of an input vector based on a linear transformation.

Matrix decomposition may be applied to the weight matrix of the linear transformation.

The data analysis step may use a second artificial neural network trained to analyze multimodal characteristics of an input vector based on a spline transformation.

In the data analysis step, learning is performed only at the first transformation when spline transformation is repeatedly performed, and learned parameters can be shared with subsequent transformations.

The data analysis step may perform variable exchange in which the order of variables of the input vector is changed.

The data analysis step may use one or more artificial neural networks learned using gradient descent based on differential information protection to which gradient clipping and noise insertion are applied.

In the data analysis step, when learning using the differential information security based gradient descent method, the value of an element having a magnitude less than or equal to a predetermined reference value among elements of a gradient vector may be converted to 0.

According to the disclosed embodiments, it is possible to generate reproduced data that reduces the risk of leakage of personal information of the original data while maintaining the characteristics of the original data.

1 is a block diagram of a device for generating reproduction data according to an embodiment;
Figure 2 is a block diagram of a reproduction data generating device according to an embodiment
3 is a configuration diagram of a data analysis unit according to an embodiment
4 is a flowchart of a method for generating reproduction data according to an embodiment
5 is a block diagram for illustrating and describing a computing environment including a computing device according to an exemplary embodiment;

Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The detailed descriptions that follow are provided to provide a comprehensive understanding of the methods, devices and/or systems described herein. However, this is only an example and the present invention is not limited thereto.

In describing the embodiments of the present invention, if it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to the intention or custom of a user or operator. Therefore, the definition should be made based on the contents throughout this specification. Terminology used in the detailed description is only for describing the embodiments of the present invention and should in no way be limiting. Unless expressly used otherwise, singular forms of expression include plural forms. In this description, expressions such as "comprising" or "comprising" are intended to indicate any characteristic, number, step, operation, element, portion or combination thereof, one or more other than those described. It should not be construed to exclude the existence or possibility of any other feature, number, step, operation, element, part or combination thereof.

1 is a configuration diagram of a device for generating reproduction data according to an exemplary embodiment.

Referring to FIG. 1 , the apparatus 100 for generating reproduced data may include a data separation unit 110 , a data conversion unit 120 and a data analysis unit 130 .

According to an embodiment, the data separator 110 may separate the original data into continuous data and categorical data.

As shown in FIG. 2 , original data x may be separated into continuous data and categorical data through the data separator 110 . Thereafter, the separated data may be input to the data conversion unit 120 and converted into continuous data.

According to an embodiment, the data conversion unit 120 may generate converted data obtained by converting categorical data into continuous data.

As a method to overcome the limitations of existing statistical modeling, a technology for generating reproducible data based on differential information protection using deep learning has been developed. However, most of the reproduction data technologies based on differential information security have been studied limited to image data composed only of continuous data. On the other hand, in the case of table data, unlike image data, continuous data and categorical data are mixed. When applying a deep learning model to image data, there is a limit to processing different types of data features, and thus, representation data Performance may deteriorate. Accordingly, a model for table data in which continuous data and categorical data are mixed is required.

For example, when a categorical variable is converted by one-hot encoding, the network size increases and the number of parameters required for learning increases (gradient size increases). Accordingly, the quality of the reproduced data may deteriorate due to a large number of related gradient clipping parts. For example, assuming that one categorical variable is school, there can be three classes such as elementary school, middle school, and high school. Here, when one-hot encoding is performed, categorical variables are changed into 3-dimensional variables. In other words, as the number of categorical variables and the number of classes for each variable increases, the dimension of the categorical variable changes to a higher dimension when one-hot encoding is performed.

According to an example, the data conversion unit 120 may perform a normalizing flow (NF) based on differential privacy (DP) so that categorical data can be treated as a continuous type without dimension change.

According to an embodiment, the data conversion unit 120 estimates a probability distribution of random noise conditional on categorical data, and adds the random noise obtained through the estimated probability distribution to the categorical data to generate transformed data. can do.

For example, assuming that the original data x is composed of d _n continuous variables (x ⁽ⁿ⁾ ) and d _c categorical variables (x ^(c) ), the data conversion unit 120 converts the categorical data into x ^(c) can be transformed (g _dpCNF : R ^(dc) → R ^(dc) ) into smooth continuous data (x ^(c,*) ). Here, the data conversion unit 120 learns Prob(ε|x ( ^{c)) based on the information on x ( c} ), adds ε obtained from the estimated distribution to x ^(c), and x ^{(c,* )} can be obtained.

According to an embodiment, the data analysis unit 130 generates a latent vector of an input vector composed of continuous data and converted data based on a normalizing flow to which differential privacy is applied, and A conversion function between an input vector and a latent vector can be derived.

Referring to FIG. 2 , the data analyzer 130 may receive continuous data from the data separator 110 and may receive converted data from the data converter 120 . Thereafter, the data analyzer 130 may generate an input vector based on the continuous data and the transformed data. Also, the data analyzer 130 may generate a latent vector z based on the input vector, and may derive a conversion function for converting the input vector into the latent vector.

3 is a configuration diagram of a data analysis unit according to an embodiment.

Referring to FIG. 3 , the data analysis unit 130 may include a linear normalization flow unit 131, a spline normalization flow unit 133, and a variable exchange unit 135.

According to one embodiment, the data analysis unit 130 includes a linear normalization flow unit 131 including a first artificial neural network trained to analyze correlation characteristics between variables of an input vector based on a linear transformation. can do.

According to an example, a linear transformation-based normalization flow may be applied to learn correlation characteristics between variables of original data. The form of a linear transformation is:

[Equation 1]

z = h _lin ( x ) = Wx + b

Here, W ∈R ^(dХd) and b ∈R ^d are the weight matrix and bias vector for the linear transformation. That is, x = [x ⁽¹⁾ ,x ⁽²⁾ ,...,x ^(d) ], z = [z ⁽¹⁾ ,z ⁽²⁾ ,...,z ^(d) ] When converting from x to z using Equation 1, z ⁽¹⁾ is converted using all of x ⁽¹⁾ , x ⁽²⁾ , ... x ^(d) information, so between the variables of x Learning is possible while maintaining correlation.

According to one embodiment, matrix decomposition may be applied to the weight matrix of the linear transformation. For example, low rank decomposition can be applied to the weight matrix W as shown in Equation 2 below.

[Equation 2]

z = h _lin ( x ) = ( diag (s)+ AB ) x + b

Here, diag(s)∈R ^(d×d) is a diagonal matrix, A ∈R ^(d×r) and B ∈R ^(r×d) , where r is a small natural number. As above, if the weight matrix is transformed using matrix decomposition, the model is reduced, and the gradient clipping part is reduced during differential information security-based learning, so learning performance can be improved. For reference, the number of parameters in Equation 1 is dХd+d, whereas the number of parameters in Equation 2 is 2d×r+2d.

According to an embodiment, the data analysis unit 130 may include a spline normalization flow unit 133 including a second artificial neural network trained to analyze multimodal characteristics of an input vector based on a spline transformation. can

For example, the sample quality of a normalization flow depends on the depth of the flow used. For example, more complex distributions can be learned by stacking multiple spline flows than using a single spline flow layer. Alternatively, a more complex distribution can be learned with K iterations than using the spline regularization flow only once. However, when multiple layers are stacked or repeatedly performed, gradient clipping is performed in relatively large parts during differential information security-based learning, which makes learning difficult. Accordingly, a conditional spline flow can be applied so that one learned spline flow layer can serve as multiple layers.

According to an embodiment, when the spline normalization flow unit 133 repeatedly performs spline transformation, learning is performed only at the first transformation, and the learned parameters may be shared with subsequent transformations. For example, one-hot encoded layer index information may be provided as a condition to a network that generates spline transformation parameters. In other words, the spline normalization flow unit 133 learns the spline normalization flow only once when repeatedly performing the spline transformation, shares the parameters of the spline transformation with all network blocks that are generated, and only blocks information (order) condition can be reflected. In this case, it is possible to achieve the effect of stacking several flow layers without increasing the number of parameters.

According to an embodiment, the data analysis unit 130 may include a variable exchange unit that changes the order of variables of an input vector. For example, since autoregressive spline normalization flows are processed sequentially, the order in which variables are listed can have an effect. Therefore, randomly changing the order of variables can help identify relationships with other variables. For example, if the original data x consists of three variables (i.e., x=[x ⁽¹⁾ , x ⁽²⁾ , x ⁽³⁾ ]), performing variable permutation in the first block The order of each can be changed to [x ⁽³⁾ , x ⁽²⁾ , x ⁽¹⁾ ].

According to an embodiment, the data analysis unit 130 may include one or more artificial neural networks trained using gradient descent based on differential information protection to which gradient clipping and noise insertion are applied.

For example, the gradient descent method is an optimization that gradually updates the weights in the opposite direction of the gradient (g(x):=▽L _w (x)) of the cost function to reach the minimum value of the cost function (L _w (x)). It is an algorithm. Mini-batch gradient descent refers to applying gradient descent to a portion of training data (ex.{x ₁ ,...,x _m }) called a mini-batch. That is, given the current weight w _t , then update w _(t+1) as follows.

Calculate gradient over mini-batch

update w _(t+1)

Here, η is a parameter controlling the distance to move.

Gradient descent based on differential information security requires gradient clipping to clip the gradient so that it does not exceed a certain threshold value (C) and a part to insert noise into the gradient. That is, the following process is added to the existing gradient descent method.

Calculate gradient over mini-batch

Added step with Differential Privacy: Gradient Clipping

가 C보다 크면

가 C가 되게끔 조정한다. 여기서

는 l₂-norm으로 gradient vector g_t(x_i)의 크기를 나타낸다. in other words,

is greater than C

Adjust so that becomes C. here

is the l ₂ -norm and represents the size of the gradient vector g _t (x _i ).

Added step with differential privacy: noise insertion

For example, differential information protection must be applied in all learning stages where original data is utilized. In other words, the final model designed with deep learning-based model A and model B must be trained using the gradient descent method based on differential information protection when learning models A and B so that the final model can satisfy the differential information protection.

According to an embodiment, the data analysis unit 130 may convert a value of an element having a size equal to or less than a predetermined reference value among elements of a gradient vector to 0 during learning using the differential information security based gradient descent method. .

According to an example, differential information protection should be applied in all learning stages in which original data is utilized. In other words, the artificial neural networks included in the data conversion unit 120, the linear normalization flow unit 131, and the spline normalization flow unit 133 are learned using the differential information security-based gradient descent method. In order to increase the learning efficiency of each step by minimizing the part where the gradient is clipped, the values of the elements having values less than a predetermined value among the elements of the gradient vector may be changed to 0 in all learning steps to reduce the size of the gradient.

According to an example, in order to perform a normalization flow, a gradient expression for each sample may be derived, and a GPU calculation advantage of a gradient descent method implemented with Tensorflow or Pytorch may be utilized. For example, deriving a gradient to update a parameter in a low-rank matrix decomposition-based linear transformation (Equation 2) is as follows.

[Equation 3]

,m은 batch size)을 빠르게 계산할 때, 각 샘플 별 최종 그래디언트(예, ∂L/∂A)는 확인이 불가능하지만, 중간 그래디언트(예, ∂L/∂z)는 확인 가능하다는 사실을 이용할 수 있다. 다시 말해, GPU의 이점을 유지하며, 각 샘플 별 그래디언트의 크기를 중간 그래디언트들로 계산하여 속도 저하 문제를 개선할 수 있다.Here, L is the objective function of the reproduction data generating device. As an example, the sum of the gradients in batches in parallel on the GPU (e.g.,

,m is the batch size), we can use the fact that the final gradient for each sample (e.g., ∂L/∂A) is not verifiable, but the intermediate gradient (e.g., ∂L/∂z) is verifiable. there is. In other words, while maintaining the advantage of the GPU, the slowdown problem can be improved by calculating the size of the gradient for each sample as intermediate gradients.

According to an embodiment, the reproduced data generating apparatus 100 may further include a reproduced data generator 140 generating reproduced data of original data from a latent vector based on an inverse function of a transform function.

Referring to FIG. 2 , the data generator 140 may receive the latent vector z output by the data analyzer 130, and obtain the latent vector z by using the inverse function of the conversion function that converts the original data into the latent vector. You can create reproducible data x ^* .

4 is a flowchart of a method for generating reproduction data according to an exemplary embodiment.

According to an embodiment, the apparatus for generating reproduction data may separate original data into continuous data and categorical data (410). For example, the original data ( x ) may be divided into continuous data and categorical data, and the separated data may be converted into continuous data through data conversion.

According to an embodiment, the apparatus for generating reproduction data may generate converted data obtained by converting categorical data into continuous data (420). According to an example, the reproduction data generation device may perform a normalizing flow (NF) based on differential privacy (DP) so that categorical data can be treated as a continuous type without dimension change.

According to an embodiment, the reproduction data generation apparatus may estimate a probability distribution of random noise conditional on categorical data, and generate transformation data by adding the random noise obtained through the estimated probability distribution to the categorical data. there is.

According to an embodiment, the apparatus for generating reproduced data includes a latent vector of an input vector composed of continuous data and transformed data based on a normalizing flow with differential privacy applied, and an input vector and a transform function of the latent vector can be derived (430).

For example, the reproduced data generating device may receive separated continuous data and converted converted data. Thereafter, the apparatus for generating reproduction data may generate an input vector based on the continuous data and the transformed data. Also, the apparatus for generating reproduction data may generate a latent vector z based on the input vector, and may derive a conversion function for converting the input vector into the latent vector.

According to an embodiment, the reproduced data generating apparatus may generate reproduced data of the original data from the latent vector based on the inverse function of the transformation function.

According to an example, the reproduced data generating apparatus may generate reproduced data x ^* from the latent vector z by using an inverse function of a conversion function that converts original data into latent vectors.

According to an embodiment, the apparatus for generating reproduced data may use a first artificial neural network trained to analyze correlation characteristics between variables of an input vector based on a linear transformation to analyze data.

According to an example, a linear transformation-based normalization flow may be applied to learn correlation characteristics between variables of original data. The form of the linear transformation is as shown in Equation 1 above.

According to one embodiment, matrix decomposition may be applied to the weight matrix of the linear transformation. For example, as shown in Equation 2 above, low rank decomposition can be applied to the weight matrix W.

According to an embodiment, the reproduction data generating apparatus may use a second artificial neural network trained to analyze multimodal characteristics of an input vector based on a spline transformation to analyze data.

According to an example, the apparatus for generating reproduction data may apply a conditional spline flow so that one learned spline flow layer serves as multiple layers.

According to an embodiment, when performing repeated spline transformation in order to analyze data, the apparatus for generating reproduced data performs learning only at the first transformation, and may share learned parameters in subsequent transformations. For example, the reproduction data generation device learns the spline normalization flow only once when repeatedly performing the spline transformation, shares the parameters of the spline transformation with all network blocks that generate it, and reflects only the block information (order) as a condition. can

According to an embodiment, the apparatus for generating reproduced data may perform variable exchange in which the order of variables of an input vector is changed in order to analyze data. For example, if the original data x consists of three variables (i.e., x=[x ⁽¹⁾ , x ⁽²⁾ , x ⁽³⁾ ]), performing variable permutation in the first block The order of each can be changed to [x ⁽³⁾ , x ⁽²⁾ , x ⁽¹⁾ ].

According to an embodiment, the apparatus for generating reproduced data may use one or more artificial neural networks trained using gradient descent based on gradient descent to which gradient clipping and noise insertion are applied to analyze data.

According to an embodiment, the apparatus for generating reproduced data converts the value of an element having a size equal to or less than a predetermined reference value among elements of a gradient vector to 0 when learning using a gradient descent method based on differential information security to analyze data. can do. For example, in order to increase the learning efficiency of each step by minimizing the part where the gradient is clipped, the reproduction data generation device changes the values of the elements having values less than a predetermined value among the elements of the gradient vector to 0 in every learning step, size can be reduced.

5 is a block diagram illustrating a computing environment including a computing device according to an exemplary embodiment.

In the illustrated embodiment, each component may have different functions and capabilities other than those described below, and may include additional components other than those described below.

The illustrated computing environment 10 includes a computing device 12 . In one embodiment, computing device 12 may be one or more components included in reproduction data generating device 120 . Computing device 12 includes at least one processor 14 , a computer readable storage medium 16 and a communication bus 18 . Processor 14 may cause computing device 12 to operate according to the above-mentioned example embodiments. For example, processor 14 may execute one or more programs stored on computer readable storage medium 16 . The one or more programs may include one or more computer-executable instructions, which when executed by processor 14 are configured to cause computing device 12 to perform operations in accordance with an illustrative embodiment. It can be.

Computer-readable storage medium 16 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. Program 20 stored on computer readable storage medium 16 includes a set of instructions executable by processor 14 . In one embodiment, computer readable storage medium 16 includes memory (volatile memory such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash memory devices, other forms of storage media that can be accessed by computing device 12 and store desired information, or any suitable combination thereof.

Communications bus 18 interconnects various other components of computing device 12, including processor 14 and computer-readable storage medium 16.

Computing device 12 may also include one or more input/output interfaces 22 and one or more network communication interfaces 26 that provide interfaces for one or more input/output devices 24 . An input/output interface 22 and a network communication interface 26 are connected to the communication bus 18 . Input/output device 24 may be coupled to other components of computing device 12 via input/output interface 22 . Exemplary input/output devices 24 include a pointing device (such as a mouse or trackpad), a keyboard, a touch input device (such as a touchpad or touchscreen), a voice or sound input device, various types of sensor devices, and/or a photographing device. input devices, and/or output devices such as display devices, printers, speakers, and/or network cards. The exemplary input/output device 24 may be included inside the computing device 12 as a component constituting the computing device 12, or may be connected to the computing device 12 as a separate device distinct from the computing device 12. may be

Although the present invention has been described in detail through representative examples above, those skilled in the art can make various modifications to the above-described embodiments without departing from the scope of the present invention. will understand Therefore, the scope of the present invention should not be limited to the described embodiments and should not be defined, and should be defined by not only the claims to be described later, but also those equivalent to these claims.

10: Computing environment
12: computing device
14: Processor
16: computer readable storage medium
18: communication bus
20: program
22: I/O interface
24: I/O device
26: network communication interface
100: reproduction data generating device
110: data separation unit
120: data conversion unit
130: data analysis unit
131: linear normalization flow part
133: spline normalization flow unit
135: variable exchange unit
140: reproduction data generation unit

Claims (20)

a data separation unit that separates the original data into continuous data and categorical data;
a data conversion unit generating converted data obtained by converting the categorical data into continuous data; and
Based on a normalizing flow to which differential privacy is applied, a latent vector of an input vector composed of the continuous data and the transformed data, and a conversion function of the input vector and the latent vector Including a data analysis unit to derive, reproduction data generating device.
The method of claim 1,
and a reproduction data generation unit configured to generate reproduction data of the original data from the latent vector based on an inverse function of the transformation function.
The method of claim 1,
The data conversion unit
Apparatus for generating conversion data by estimating a probability distribution of the conditional random noise of the categorical data and adding random noise obtained through the estimated probability distribution to the categorical data.
The method of claim 1,
The data analysis department
A linear normalization flow unit including a first artificial neural network learned to analyze correlation characteristics between variables of the input vector based on linear transformation.
The method of claim 4,
The weight matrix of the linear transformation is a matrix decomposition applied, reproduced data generating device.
The method of claim 1,
The data analysis department
and a spline normalization flow unit including a second artificial neural network learned to analyze multimodal characteristics of the input vector based on spline transformation.
The method of claim 6,
The spline normalization flow part
Reproducible data generation device that performs learning only at the first conversion when repeatedly performing spline conversion and shares learned parameters with subsequent conversions.
The method of claim 1,
The data analysis department
A device for generating reproduction data comprising a variable exchange unit that changes an order between variables of the input vector.
The method of claim 1,
The data analysis department
An apparatus for generating reproduction data, including one or more artificial neural networks trained using gradient descent based on differential information protection with gradient clipping and noise insertion.
The method of claim 9,
The data analysis department
When learning using the differential information security-based gradient descent method, the reproduction data generation device converts the value of an element having a size of a predetermined reference value or less among elements of a gradient vector to 0.
a data separation step of separating original data into continuous data and categorical data;
a data conversion step of generating converted data obtained by converting the categorical data into continuous data; and
Based on a normalizing flow to which differential privacy is applied, a latent vector of an input vector composed of the continuous data and the transformed data, and a conversion function of the input vector and the latent vector A method for generating reproducible data, comprising a data analysis step to derive.
The method of claim 11,
and a reproduction data generation step of generating reproduction data of the original data from the latent vector based on an inverse function of the transformation function.
The method of claim 11,
The data conversion step
The reproduction data generation method of estimating a probability distribution of the conditional random noise of the categorical data, and generating transformed data by adding random noise obtained through the estimated probability distribution to the categorical data.
The method of claim 11,
The data analysis step is
A method for generating reproduction data using a first artificial neural network learned to analyze correlation characteristics between variables of the input vector based on a linear transformation.
The method of claim 14,
Method for generating reproduction data, wherein matrix decomposition is applied to the weight matrix of the linear transformation.
The method of claim 11,
The data analysis step is
A method for generating reproduction data using a second artificial neural network learned to analyze multimodal properties of the input vector based on a spline transformation.
The method of claim 16
The data analysis step is
A method of generating reproducible data in which learning is performed only at the first conversion when repeatedly performing spline conversion, and the learned parameters are shared with subsequent conversions.
The method of claim 11,
The data analysis step is
A method of generating reproduction data, performing variable exchange in which the order of variables of the input vector is changed.
The method of claim 11,
The data analysis step is
A method for generating reproduction data using one or more artificial neural networks learned using gradient descent based on differential information protection applied with gradient clipping and noise insertion.
The method of claim 19
The data analysis step is
When learning using the differential information security-based gradient descent method, converting a value of an element having a size of a predetermined reference value or less among elements of a gradient vector into 0, a reproduction data generation method.

Images