This paper presents a novel, secure and dynamic mechanism to train machine learning models that achieve membership privacy. Our protocol falls in the two-server-aided model and allows one server to perform most of computations and allows another server to provide auxiliary computation. In addition, users distribute their private data among two non-colluding but untrusted servers who train neural network on the adaptively select data using secure multi-party computation (MPC). In our protocol, only the selected members can reconstruct a predefined secret and jointly decrypt the computation result. This protocol is proven to be secure in the semi-honest model and passive adversary settings, and show that security is maintained even if the unselected users’ drop out at any time.