Physiological and morphometric biomarkers for synthetic media detection

The growing realism of synthetic media generated by diffusion and neural rendering models poses a critical challenge for digital forensics and medical data integrity. Synthetic videos, often referred to as deepfakes, may compromise biometric verification systems, patient identification, clinical documentation, and telemedical imaging. This study introduces a dual-framework approach that combines physiological and morphometric anomalies as interpretable and degradation-aware forensic indicators of synthetic content. The goal is to evaluate whether biologically grounded descriptors can provide interpretable anomaly evidence for differentiating authentic and synthetic audiovisual data within the evaluated dataset and degradation conditions. Experiments were conducted using the DeepFake RealWorld (DFRW) dataset comprising 46,371 audiovisual clips (229 h total duration). Physiological features included remote photoplethysmographic (rPPG) variability, oculomotor dynamics, and speech–motion synchrony, whereas morphometric and topological features captured curvature variance, bilateral symmetry, and persistent homology. Each feature was analyzed independently under a fixed-threshold evaluation protocol to compute interpretable effect size measures, namely, the probability difference (Δp) and prevalence ratio (PR), and to derive standard benchmarking metrics, including sensitivity, specificity, accuracy, F1 score, and receiver operating characteristic area under the curve (ROC AUC), with robustness evaluated under compression and rescaling. Both physiological and morphometric markers showed measurable descriptor-level discrimination within the evaluated corpus (mean Δp ≈ 0.21, PR up to 4.7), while the added classification metrics indicated a high-specificity but moderate-sensitivity operating profile, supporting their interpretation as conservative triage and validation signals rather than as stand-alone detectors. Physiological descriptors provided low-cost, high-specificity triage signals, whereas morphometric and topological metrics showed slightly stronger descriptor-level discrimination and greater stability under the tested distortions. Physiological and morphometric anomalies should therefore be interpreted as complementary and degradation-aware indicators of synthetic media within the evaluated experimental conditions, not as evidence of a deployment-ready detection system. The proposed dual-framework bridges digital security engineering and biomedical signal analysis, offering a methodological basis for further research on synthetic content detection, telemedical data integrity, and quality assurance. From a biosafety and biosecurity perspective, these findings suggest that interpretable detection of synthetic media may help inform future mitigation strategies for AI-enabled risks to telemedical systems and clinical data integrity.