Form data privacy protection method fusing differential privacy GAN model and PATE model

Abstract

The invention relates to a form data privacy protection method fusing a differential privacy GAN model and a PATE model. The method comprises the steps of 1, training a differential privacy generationmodel by using original table data; 2, training a teacher classifier under the differential privacy budget by using the original table data; Step 3, generating 'false' table data by using the generation model, predicting labels of the 'false' table data by using a teacher classifier, selecting data with consistent prediction labels and generated labels, defining an 'available' data set, and training a student classifier by using the 'available' data set; and step 4, releasing the generation model and the student classifier, synthesizing data by using the generation model, selecting the data by using the student model, and finishing a data analysis task. According to the method, privacy protection is carried out on the table data in the data release stage, a data analyst cannot restore original training data through a generation model and cannot speculate the original training data through a student model, protection on the original table data is achieved, and the requirement of the data analyst for the data is met.

Description

technical field [0001] The present invention relates to a method for protecting privacy of table data in the publishing stage, and in particular to a method for protecting privacy of table data that integrates differential privacy GAN and PATE models. Background technique [0002] In the era of big data, machine learning extracts useful information from data sets to train models to facilitate people's lives. At the same time, data security and user privacy issues also need to be paid attention to. Data privacy protection methods are used to reconcile the conflict of interest between the two . [0003] The generative adversarial network has the characteristics of rapidly and large-scale data generation, but individual samples in the training set have a greater impact on the generative model, and the generated data may expose the privacy of individual samples. The attacker generates data by using the generative model multiple times, combined with special calculation methods, ...

AI Technical Summary

Problems solved by technology

[0003] Generative adversarial networks have the characteristics of fast and large-scale data generation, but individual samples in the training set have a greater impact on the generative model, and the generated data may expose the privacy of individual samples

The attacker generates data by using the generative model multiple times, combined with special calculation methods, may restore individual samples in the training set

Images