JP6978385B2 - Anonymization device, anonymization method and anonymization program - Google Patents

Description

The present invention relates to an anonymization device, anonymization method and anonymization program for anonymizing a data set.

Conventionally, when utilizing a data set containing user information, data anonymization is performed in order to protect personal privacy.
As anonymization methods, for example, the k-anonymization method is proposed in Non-Patent Documents 1 to 3, and the method of adding noise to data is proposed in Non-Patent Documents 4 to 6.

P. Samari and L. Sweeney, "Generalization data to provid anonymity when disclosing information," in Proc. of PODS 1998, 1998, p. 188. P. Samariti, “Protecting repondents' identities in microdata releases,” IEEE Trans. on Knowledge and Data Engineering, vol. 13, no. 6, pp. 1010-1027, 2001. L. Sweeney, "Achieving k-anonymity privacy promotion using generalization and support," in J. et al. Uncertainty, Fuzzines, and Knowledge-Base Systems, vol. 10 (5), 2002, pp. 571-588. K. Mivule, "Utilizing noise addition for data privacy, an overflow," arXiv preprint arXiv: 1309.3958, 2013. J. J. Kim, "A method for limiting discrosure in microdata based on random noise and transformation," in Proceedings of the second method. American Statistical Association, 1986, pp. 303-308. T. Yu and S. Jajodia, "Secure data management in decentralized systems," Springer Science & Business Media, 2007, vol. 33.

The purpose of data anonymization is to protect the privacy of individuals and to maintain the usefulness of the data. However, in the conventional anonymization method, if the anonymity of the data is improved, the usefulness may be greatly reduced.

An object of the present invention is to provide an anonymization device, anonymization method, and anonymization program capable of maintaining the usefulness of a data set and anonymizing it.

The anonymization device according to the present invention has an input unit that accepts an input of a tensor in which a data set of user information is described, and a tensor decomposition unit that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm. And, among the plurality of factors, the anonymization calculation unit that performs an anonymization calculation for a specific factor showing the characteristics of each user, and the anonymized factor replace the specific factor with the tensor. It includes an output unit that calculates and outputs an approximate anonymized tensor.

The anonymization calculation unit may perform a k-anonymization calculation for the specific factor.

The anonymization calculation unit may perform a calculation for adding noise to the specific factor.

The anonymization device may include a setting unit that accepts and sets the rank in the tensor decomposition and the strength of the anonymization as parameters.

The anonymization method according to the present invention includes an input step that accepts an input of a tensor in which a data set of user information is described, and a tensor decomposition step that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm. And, among the plurality of factors, the anonymization calculation step for performing an anonymization operation for a specific factor showing the characteristics of each user, and the anonymized factor replaces the specific factor with the tensor. The computer executes an output step that calculates and outputs an approximate anonymized tensor.

The anonymization program according to the present invention has an input step that accepts an input of a tensor in which a data set of user information is described, and a tensor decomposition step that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm. And, among the plurality of factors, the anonymization calculation step for performing an anonymization operation for a specific factor showing the characteristics of each user, and the anonymized factor replaces the specific factor with the tensor. The purpose is to have a computer execute an output step that calculates and outputs an approximate anonymized tensor.

According to the present invention, the usefulness of the data set can be maintained and anonymized.

It is a block diagram which shows the functional structure of the anonymization apparatus which concerns on embodiment. It is a figure which illustrates the algorithm of the anonymization method which concerns on embodiment. It is a figure which shows the comparative experiment result with the prior art of the anonymization method which concerns on embodiment.

Hereinafter, an example of the embodiment of the present invention will be described.
FIG. 1 is a block diagram showing a functional configuration of the anonymization device 1 according to the present embodiment.
The anonymization device 1 is an information processing device (computer) such as a server device or a personal computer, and includes a control unit 10, a storage unit 20, and various input / output devices.

The control unit 10 is a part that controls the entire anonymization device 1, and realizes the function in the present embodiment by appropriately reading and executing various programs stored in the storage unit 20. The control unit 10 may be a CPU.

The storage unit 20 is a storage area for various programs and various data for making the hardware group function as the anonymization device 1, and may be a ROM, RAM, flash memory, hard disk (HDD), or the like. Specifically, the storage unit 20 stores the data set to be processed, various parameters, and the like, in addition to the anonymization program for causing the control unit 10 to execute the function of the present embodiment.

Further, the control unit 10 includes an input unit 11, a tensor decomposition unit 12, an anonymization calculation unit 13, an output unit 14, and a setting unit 15. These functional units are realized by the control unit 10 executing the anonymization program stored in the storage unit 20.

The input unit 11 accepts the input of the tensor in which the data set including the personal information of the user is described.
In the present embodiment, the input tensor is described as a matrix which is a second-order tensor as an example, but the present invention is not limited to this.

The data set to be anonymized is, for example, a user's movement history, purchase history, access history, etc. in a certain period. Such a data set is described as 1/0 of each element in a matrix of n rows and m columns with or without actual results among m positions, products, sites, etc. for n users.

The tensor decomposition unit 12 decomposes the input tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm.
Known algorithms are available for tensor decomposition. For example, in the case of a second-order tensor, a matrix factorization algorithm such as SVD (Singular Value Decomposition) or NMF (Non-negative Matrix Factorization) can be used.

As a result, the matrix M ∈ R ^{n × m} is divided into two matrices, ^{U ∈ R n × r} and V ∈ ^{R r × m.} U is a matrix showing the characteristics of the row, that is, the characteristics of each user. On the other hand, V indicates the characteristics of the column, that is, the characteristics such as historical data, and does not include user-specific information.
Here, as the rank r, which is a parameter of matrix factorization, becomes smaller, the number of feature quantities is reduced, and the approximation accuracy of the product UV of the decomposed matrices U and V with the original matrix M becomes lower.

The anonymization calculation unit 13 performs anonymization calculation for a specific factor showing the characteristics of each user among the plurality of decomposed factors.
For example, when the matrix M to be anonymized is decomposed into U and V, the anonymization calculation unit 13 performs the anonymization operation only on the matrix U showing the characteristics of each user.
For the anonymization operation, for example, k-anonymization or an operation for adding noise may be adopted.

The output unit 14 replaces a specific factor with an anonymized factor, calculates and outputs an anonymized tensor that is close to the original tensor.
For example, assuming that the anonymization function for the matrix is A (), the output unit 14 outputs a matrix in which only the user's characteristics are anonymized by calculating the approximate matrix A (U) V of the original matrix M.

The setting unit 15 accepts and sets the rank in the tensor decomposition and the strength of anonymization as parameters.
The intensity of anonymization is the value of k in k-anonymization, or the magnitude of noise, for example, the value ^{of variance 2φ 2 in the Laplace distribution, which is an example of noise.}

FIG. 2 is a diagram illustrating an algorithm of the anonymization method according to the present embodiment.
This example incorporates an anonymization operation into the NMF algorithm.
First, the rank r and the number of iterations I of the alternate optimization are given to the original matrix M to be anonymized.

In step 1-2, the tensor decomposition unit 12 as the initial value of the matrix U and V after the decomposition of the matrix M, randomly generates a ^{_{t = 0, U t ∈R n}} × r and _{V t} ∈R ^{r × m} do.

In step 3-7, the tensor decomposition section 12, by alternating ^{_{optimization, U t + 1 = U t}} · (MV t T) / (U t V t V t T) and _{V t + 1 = V t ·} (U t + 1 T M ) _/ a _{^{(U t + 1 T U t}} + 1 V t) is calculated and repeats the process of counting up the t, and calculates the _{U I} and _{V I.}

In step 8-9, the anonymizing calculation unit 13 performs calculation of the anonymous with respect to _{U I, t = I, U} 't + 1 = A (ano) (U t) and _{V' t + 1 = V t} · ( U calculates the ^{_{'t + 1 T M) /}} (U' t + 1 T U 't + 1 V t).

This algorithm 'and _{t + 1,} V for approximating the original matrix M' matrix U which is anonymous since the _{t + 1} is obtained, the output unit 14, U a _{'t + 1} V' _{t + 1,} anonymizing matrix M Is output as a result of.

FIG. 3 is a diagram showing the results of comparative experiments of the anonymization method according to the present embodiment with the conventional ones.
Here, when noise addition is adopted as the operation of anonymization, the conventional method of anonymizing the entire original matrix M and the present implementation of adjusting the magnitude and rank of noise to anonymize only the user matrix U. It shows the method of morphology.

The object of anonymization used in this experiment is a matrix M of n × m = 200 × 1000. When the noise of φ = 1.5 is added to the matrix M, when the noise of φ = 1.5 is added to the matrix U of rank r = 40, the noise of φ = 2.5 is added to the matrix U of rank r = 80. When the noise of φ = 3.5 was added to the matrix U of rank r = 120, the usefulness (D) of the four patterns was calculated.
It should be noted that these four patterns of data sets have the same probability (about 0.63) that the same user can be matched, that is, re-identified before and after anonymization.

For the evaluation of usefulness, the F value, which is the prediction accuracy by machine learning using the data set as the teacher data, was used, and the ratio of the original matrix M before anonymization to the F value was defined as the evaluation value Utility (D).
Compared with the conventional method, the method of the present embodiment is particularly useful when noise of φ = 1.5 is added to the matrix U of rank r = 40.

By reducing the rank, the approximation accuracy from the original matrix becomes low, so the tensor decomposition itself has the effect of anonymization. By properly selecting these parameters in combination with anonymization operations such as noise addition, a highly useful anonymization dataset with the same degree of security was obtained.

According to the present embodiment, the anonymization device 1 extracts a specific factor (tensor such as a matrix) showing characteristics for each user by decomposing a data set described by a tensor such as a matrix into a tensor. The anonymization device 1 performs anonymization calculation only for this specific factor, so that only the user's characteristics are maintained while maintaining the correlation between other information (features) such as historical data which is user information. Can be anonymized.
Therefore, the anonymization device 1 can maintain the usefulness of the data set and anonymize it.

The anonymization device 1 can apply a known method as an anonymization operation, and for example, a method such as k-anonymization or noise addition can be easily combined and used.
Further, the anonymization device 1 accepts the rank in the tensor decomposition and the strength of anonymization as parameters, so that it can be calculated with appropriate settings according to the data set to be anonymized, and the anonymity of the data. And usefulness are highly compatible.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments. Moreover, the effects described in the above-described embodiments are merely a list of the most suitable effects resulting from the present invention, and the effects according to the present invention are not limited to those described in the embodiments.

The anonymization method by the anonymization device 1 is realized by software. When realized by software, the programs that make up this software are installed in the information processing device (computer). Further, these programs may be recorded on a removable medium such as a CD-ROM and distributed to the user, or may be distributed by being downloaded to the user's computer via a network. Further, these programs may be provided to the user's computer as a Web service via a network without being downloaded.

1 Anonymization device 10 Control unit 11 Input unit 12 Tensor decomposition unit 13 Anonymization calculation unit 14 Output unit 15 Setting unit 20 Storage unit

Claims (6)

An input unit that accepts input from a tensor that describes a data set of user information,
A tensor decomposition unit that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm, and a tensor decomposition unit.
Among the plurality of factors, the specific factor indicating the characteristics of each user is anonymized, and the specific factor is used to approximate the tensor by the product of the anonymized factor. Anonymization calculation unit that converts other factors except
An anonymization device including an output unit that calculates and outputs an anonymized tensor similar to the tensor by the product of the anonymized factor and the converted other factors. The anonymization device according to claim 1, wherein the anonymization calculation unit performs a k-anonymization calculation for the specific factor. The anonymization device according to claim 1, wherein the anonymization calculation unit performs an operation of adding noise to the specific factor. The anonymization device according to any one of claims 1 to 3, further comprising a setting unit that accepts and sets the rank in the tensor decomposition and the strength of the anonymization as parameters. An input step that accepts input from a tensor that describes a dataset of user information,
A tensor decomposition step that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm, and
Among the plurality of factors, the specific factor indicating the characteristics of each user is anonymized, and the specific factor is used to approximate the tensor by the product of the anonymized factor. Anonymization operation steps that transform other factors except
An anonymization method in which a computer executes an output step of calculating and outputting an anonymized tensor similar to the tensor by the product of the anonymized factor and the converted other factors. An input step that accepts input from a tensor that describes a dataset of user information,
A tensor decomposition step that decomposes the tensor into a product of a plurality of factors by a predetermined tensor decomposition algorithm, and
Among the plurality of factors, the specific factor indicating the characteristics of each user is anonymized, and the specific factor is used to approximate the tensor by the product of the anonymized factor. Anonymization operation steps that transform other factors except
The anonymized factor, and the product of the transformed the other factors, anonymizing program to execute an output step, to a computer which calculates and outputs the anonymous tensor approximate to the tensor.

Images