Differentially Private Semi-Supervised Learning With Known Class Priors
Anh Thi Lan Pham, Jing Xi · 2018
Machine learning algorithms benefit greatly from a large amount of data in the big data regime, however, the labeling cost for all training samples is prohibitively high. Semi-supervised learning is a designed framework to overcome the labeling challenge. In particular, in semi-supervised learning, there are only a small number of labeled samples and unlabeled samples are used as side information to leverage the overall performance. Additionally, preserving the privacy of the data is another challenge in big data analysis. This paper addresses challenges by designing a privacy-preserving method for semi-supervised learning. Assuming that class priors are available, the proposed method can achieve the optimal classifier. Moreover, the privacy and utility of the proposed method are theoretically guaranteed. Experiments on real UCI datasets illustrate the effectiveness of our proposed method.