The proliferation of deepfake videos underscores the need to develop reliable detection methods to mitigate their negative consequences. Several deepfake video detection methods achieve satisfactory performance. However, the methods often focus on specific facial features and overlook the noise introduced during feature extraction. This chapter describes a deepfake video detection method that focuses on five key facial features and their changes in video frames and uses optical flows to fine-tune facial landmark points to reduce landmark errors. An efficient model is employed to compute optical flows in frame sequences and refine the actual and predicted landmark points. This facilitates the accurate identification of facial landmark points, including anomalous landmark points. Two recurrent neural networks are employed to capture temporal relationships between landmark points and their movement patterns. The proposed deepfake video detection method achieves high performance on raw and compressed videos.