Indícios fisiológicos e interpretáveis para a detecção de deepfake em vídeos: um estudo empírico com mapas PPG e características projetadas manualmente

Abstract

Deepfake videos have become increasingly realistic, reducing the reliability of detectors based solely on visible artifacts. This paper studies deepfake detection from two complementary viewpoints: (i) physiological consistency, captured through remote photoplethysmography (PPG) maps in the spirit of the popular FakeCatcher detector, and (ii) interpretable handcrafted descriptors related to motion, texture, chromatic variation, and compression artifacts. We perform an exhaustive ablation study over all handcrafted feature subsets (1023 combinations) and use SHAP to quantify feature relevance across datasets. Experiments on the CelebDF (v1 and v2) and FaceForensics++ datasets show that compact subsets of handcrafted descriptors consistently outperform using the full feature set, and that the most informative cues shift markedly with dataset characteristics. We also observe that PPG-based CNNs achieve strong intra-dataset performance, while cross-domain robustness remains challenging under heterogeneous conditions. Finally, we analyze the (often limited) complementarity between physiological and handcrafted cues under late-fusion strategies, and show that selective bias-correction can provide minor yet consistent gains in a small subset of near-boundary cases. Overall, our results clarify which explainable cues remain effective in low-artifact scenarios and provide practical guidance on when multimodal combination is beneficial.