Lightweight motion-centric deepfake detection from the labial region through transferable and privacy-aware forensic representations

The growing realism of deepfake generation poses substantial challenges to multimedia security, authenticity verification, and digital forensics. Despite strong detection performance, many existing approaches rely on complex architectures and exhibit limited robustness across heterogeneous datasets and unseen manipulations. This work investigates lightweight motion-centric representations extracted from the labial region for deepfake detection. Spatio-temporal, gradient-based, and geometry-aware descriptors are employed to capture complementary local facial dynamics while limiting reliance on full-face imagery. The proposed methodology is evaluated on FaceForensics++ and CelebDF under intra- and cross-dataset protocols using multiple machine learning and deep learning classifiers. Results demonstrate that labial motion provides discriminative forensic cues, with gradient-based descriptors showing greater cross-dataset transferability than geometry-based representations. Overall, compact motion descriptors offer a favourable trade-off among detection performance, computational efficiency, interpretability, and privacy preservation, supporting localized motion analysis as a robust strategy for deepfake detection in heterogeneous multimedia environments.