In this work, we propose a novel framework for linear classification, differentially private semi-supervised classification. The previous method in the classification problem, differentially private empirical risk minimization (ERM) only generates a classifier from labeled data. Inspired by semi-supervised learning, we propose two differentially private semi-supervised methods, which train a classifier by using both labeled and unlabeled data. We analyze the global sensitivity of the objective function and introduce differentially private ERM for semi-supervised prediction using output perturbation and objective perturbation. We experimentally evaluate the performance of the proposed methods and demonstrate that the proposed methods give more accurate prediction than regular differentially private ERM by increasing the number of unlabeled data used for training.