Data Privacy Examination against Semi-Supervised Learning

Semi-supervised learning, which learns with only a small amount of labeled data while collecting voluminous unlabeled data to aid its training, has achieved promising performance lately, but it also raises a serious privacy concern: Whether a user's data has been collected for use without authorization. In this paper, we propose a novel membership inference method against semi-supervised learning, serving to protect user data privacy. Due to involving both the labeled and unlabeled data, the membership patterns of semi-supervised learning's training data cannot be well captured by the existing membership inference solutions. To this end, we propose two new metrics, i.e., inter-consistency and intra-entropy, tailored specifically to the semi-supervised learning paradigm, able to respectively measure the similarity and calculate the cross-entropy among prediction vectors from the perturbed versions. By exploiting the two metrics for membership inference, our method can dig out membership patterns imprinted on prediction outputs of semi-supervised learning models, thus facilitating effective membership inference. Extensive experiments have been conducted for comparing our method with five rectified baseline inference techniques across four datasets on six semi-supervised learning algorithms. Experimental results exhibit that our inference method achieves over 80% accuracy under each experimental setting, substantially outperforming all baseline techniques.