Applying Deep Learning to Preserve Data Confidentiality Keynote Address

摘要

Preserving data confidentiality is a crucial problem when releasing microdata for public-use. A lot of approaches have been proposed so far for preserving data confidentiality, and many of them are based on traditional probability and statistics which have the capability to mask the original data. However, their performance needs to be significantly improved in practice. In this paper, we approached this problem by using deep learning-based generative model, which can generate simulated data that are closely related to raw data but different for each item. Since the mechanism of generative model is to transform a distribution (like Uniform) sampled from a noise to another distribution (like Gaussian) sampled from a real dataset, it is hard to guarantee such generation that can represent the raw data in practice due to existing statistical variants between them. Despite deep learning's strong generative ability, the same issue still exists. In this study, we innovatively explore statistical similarity between two datasets via deep learning-based generative model. And we also introduced two statistical evaluation metrics to assess the similarity. We conducted extensive experiments to validate our idea with two real-world datasets, the census dataset and the environmental dataset.