Network structure learning method based on differential privacy

Abstract

The invention provides a network structure learning method based on differential privacy, and belongs to the field of machine learning. The method comprises the steps: taking to-be-learned network data as an input; normalizing the network data to form a normalized data matrix XX; determining privacy parameters epsilon and delta delta of the differential privacy algorithm; determining a standard deviation beta beta of Gaussian distribution in a Gaussian noise mechanism; generating a symmetric noise matrix NN; and generating a disturbed sample covariance matrix, and solving a network structure based on the disturbed sample covariance matrix. According to the method, the differential privacy is satisfied through the network structure learned in each step, that is, the method has a privacy protection capability, and the privacy of individuals in the data cannot be leaked. While privacy is protected, a network structure learning result also has relatively high availability, and a real potential network structure of data can be reflected; the differential privacy solving problem of a classic Graphical lasso model in an undirected graph model is solved, and it is guaranteed that the privacy of individuals in data is not leaked in the network structure learning process.

Description

technical field [0001] The invention belongs to the field of machine learning, and in particular relates to a network structure learning method based on differential privacy. Background technique [0002] The problem of network structure learning has received extensive attention, and a series of model methods represented by graph models have been formed, including undirected graph models, directed graph models, etc. Among them, the most representative method in undirected graph models is Graphicallasso. Under the assumption that the network data obeys the Gaussian distribution, the Markov property can prove that the inverse of the covariance matrix is ​​equivalent to the conditional independence of pairwise random variables, and the network structure can be learned by estimating the inverse of the covariance matrix. Based on network-based sparse priors, Graphical lasso introduces l 1 l 1 The type penalty item efficiently realizes the structural learning of high-dimensional...

AI Technical Summary

Problems solved by technology

Based on network-based sparse priors, Graphical lasso introduces l 1 l 1 The type penalty item efficiently realizes the structural learning of high-dimensional sparse undirected networks. On the basis of Graphical lasso, a large number of research results of penalty likelihood estimation of undirected graphs have emerged, including models, theories, solutions, etc., but the existing methods The issue of privacy protection is not considered

Images