Deepfake Detection Using Robust Spatial and Temporal Features from Facial Landmarks

Most current deepfake detectors may suffer the decrease of detection accuracy under common video processing like compression. To deal with this issue, we proposed a new way of using biometric features for robust deepfake detection. Our biometric features are derived from a set of selected facial landmarks. We first presented a metric to select robust facial landmarks, and then constructed facial feature vectors with the selected landmarks. The spatial angles and the temporal rotation angles are introduced to facilitate the construction of the SVM feature vector. In essence, we use the spatial angles and temporal rotation angles to characterize the inherent consistency of facial landmarks at both frame level and video level. Experimental results have demonstrated that our detector has the best robustness compared with 6 current methods. It also has good scores in AUC (Area Under the Receiver Operating Characteristic Curve) and detection accuracy.