A Bio-Kinematic Index Framework for Deepfake Video Detection

Modern deepfakes increasingly suppress visible synthesis artifacts, while social-media processing further changes resolution, illumination, frame rate, and compression traces. This paper presents the Bio-Kinematic Index (BKCI), a lightweight and interpretable video-level framework that measures whether physiological color variation, regional facial micro-motion, mouth-face coupling, and their temporal relationships remain mutually consistent. The method converts regional evidence into physiological disorder, kinematic disorder, and perturbation-based temporal-instability descriptors and classifies the resulting vector using lightweight learners. On the held-out Celeb-DF test partition, BKCI-SVM obtained 77.46% accuracy, 78.45% F1score, and 82.83% AUC. A fusion model increased AUC to 84.68%, 3.65 percentage points above the artifact-RF baseline. Under low-resolution recompression, 15-fps reduction, and lowlight degradation, the selected BKCI variants retained AUC values of 80.45%, 81.70%, and 76.84%, respectively. These results indicate that bio-kinematic is most useful as complementary forensic evidence; low illumination remains the principal failure condition.