Split learning is a new machine learning model, considering the distributed users' privacy. While preserving the privacy of user data, split learning can leverage an amount of training data from multiple users. This paper presents how split learning can efficiently trade privacy, prediction accuracy, and training overhead. We devise three practical implementation models of split learning with different levels of privacy. Our experiment shows that privacy, accuracy, and training overhead are differently presented according to the implementation model. The result supports that privacy-preserving layer in split learning enhances privacy with marginal processing overhead, and we can achieve reasonably high accuracy, compared with the local model with limited dataset.