Face swapping detection based on identity spatial constraints with weighted frequency division

The generalization of face swapping detectors is necessary when applied to the practical applications. Although the most existing methods may achieve accuracy detection performance on the known forgeries, they fail to make the prediction when faced with unseen face manipulation methods. To alleviate the problem, we propose a novel and practical framework called Detection based on Identity Spatial Constraints with Weighted Frequency Division (DISC-WFD) through introducing the reference image, consisting of the backbone network, the shared Identity Semantic Encoder (ISE) and the corresponding Identity Spatial Constraint (ISC) branches. The ISE is utilized to measure the identity similarity between the input image and the reference image and generates identity spatial constraints. The constraints are imposed on ISC to focus on the high frequency and low-frequency identity-related areas for the discriminative information. The proposed method can significantly improve the performance and the generalization against the unseen manipulation methods. Furthermore, the cross-dataset experiments validate the superiority and the effectiveness of the DISC-WFD.