Representation Alignment For Deepfake Detection

DeepFake detection faces growing challenges with the rapid advancement of generative models, enabling the creation of increasingly sophisticated forgeries. Existing methods often rely on heuristic features from spatial or frequency domains without effectively integrating them into backbone architectures to capture general forgery representations. To address this, we propose the Representation Alignment (RA) technique to enhance backbone design by incorporating high-quality external semantic features. By aligning intermediate representations, RA enables the detector to capture robust spatial attributes and discriminative frequency features, improving its ability to generalize across diverse forgery types. The proposed RA-enhanced Xception detector demonstrates superior performance in both within-domain and cross-domain evaluations. These results validate the effectiveness of RA in enhancing generalization and robustness, offering a lightweight and efficient approach for advancing DeepFake detection.