Abstract:

This study investigates interpretable photometric and geometric features for deepfake detection under realistic conditions. The dedicated DeepFake RealWorld (DFRW) dataset, comprising 46 371 clips generated by diffusion, reenactment, and face-swap models, was used to evaluate lighting and shape consistency. Key descriptors, including light direction mismatch (Δθ), luminance deviation (ΔL), shading ratio (r_shade), shadow coherence (χ_shadow), and head-torso alignment, achieved Δp≈0.20–0.23 and PR up to 4.25. The results confirm that physically grounded descriptors of illumination and geometry enable reliable, explainable deepfake detection in forensic contexts.