Using Identity and Spatial Texture Information for Deepfake Detection

With the rapid advancement of synthetic media, deepfake detection has become a critical challenge in digital forensics. Existing methods struggle to generalize to unseen forgery techniques, limiting their practical effectiveness. In this paper, we propose a deepfake detection method that integrates identity and spatial texture inconsistencies to enhance detection robustness. Our approach employs an identity encoder to extract facial identity features and a spatial texture encoder to capture forgery artifacts. To improve generalization, we design a vector generation module that synthesizes forged feature representations, enhancing adaptability to unknown forgeries through contrastive learning. Additionally, we utilize the Mamba encoder for temporal modeling, effectively capturing identity and texture inconsistencies across frames. Extensive experiments on benchmark datasets demonstrate that our method outperforms existing approaches, achieving superior accuracy and robustness. Our model effectively detects manipulated videos across different forgery types, making it a promising solution for real-world deepfake detection.