Principal component analysis-oriented differential privacy protection method

Abstract

The invention discloses a principal component analysis-oriented differential privacy protection method, which comprises the following steps of: centralizing a data matrix, namely subtracting a mean value of the dimension from data of each dimension; calculating a covariance matrix (shown in the specification) of the data matrix, and calculating a characteristic value and a characteristic vector Vof the covariance matrix A; calculating a characteristic value and a characteristic vector V of the covariance matrix A; calculating the number k of reserved main components; mapping the original datato a principal component space to obtain a projection matrix Z; distributing a privacy budget j to each column of elements of the projection matrix Z, and calculating the added random noise; adding noise to the projection matrix Z to obtain a noise-added projection matrix Z '; and calculating an error between the original data and the low-rank approximate data. According to the method, the dimension of the data set can be effectively reduced, data simplification can be achieved, noise can be prevented from being added to unimportant data, privacy budget waste can be reduced, and therefore thedata availability can be improved, the released data can reflect real data as much as possible, and meanwhile the privacy of the data can be protected.

Description

technical field [0001] The invention relates to a differential privacy protection method oriented to principal component analysis, and belongs to the technical field of information security. Background technique [0002] With the continuous development of big data technology, the data stored in various information systems is becoming more and more abundant, which increases the complexity of data analysis and processing. As one of the important methods of data analysis, principal component analysis can convert multiple variables into several main variables, which can represent most of the information of the original data and reveal the essence of the data. Principal component analysis simplifies the data, makes the data easier to use and reduces the computational overhead of the algorithm. Data sets usually contain a lot of private information. If machine learning or data mining algorithms are used to analyze the data directly, it will cause privacy leakage. The differentia...

AI Technical Summary

Problems solved by technology

The differential privacy protection method is currently a popular privacy protection technology. It is realized through the noise mechanism, that is, adding random noise to the output result to protect the data security. The greater the noise added, the more secure the data, however, the lower the availability of the data. vice versa

[0003] For multi-attribute data, the traditional Laplacian mechanism allocates the same size privacy budget to all attributes. This scheme is simple and easy to operate, but it will cause too much noise added, and the data availability will decrease sharply. Allocation of privacy budget, wasting part of the privacy budget, so the effect is not ideal

Images