In this paper, we propose an effective and real-time text steganalysis method based on multi-stage transfer learning to enhance inference efficiency and detection performance simultaneously.