Abstract: Insufficient generalization and complexity in face anti-spoofing detection models results in a poor performance targeting on new face attack algorithm. Therefore, a face recognition model in vivo (FT-CNN) is proposed based on optical flow estimate and micro-expression in face. The model consisted of TVNet-DTSCNN and Attention CNN-LSTM. TVNet-DTSCNN performed optical flow prediction and micro-expression extraction on the input time-series face frames, and attention CNN-LSTM extracted and magnified the motion detail cues in the face video, which made the model to learn the robust feature for both live and prosthetic faces. Experiments on CASIA, CASIA-SURF and MSU-MFSD datasets indicate that the performance of FT-CNN in accuracy (Acc), average error rate (HTER) and generalization is significantly improved compared with the previous models.