Abstract:

This study evaluates interpretable textural and edge descriptors for deepfake detection under real-world conditions. Using the new DFRW dataset with 46 371 clips from diffusion and face-swap models, classical features showed strong robustness to compression and re-encoding. The best-performing descriptors: CLBP, LBP, BSIF, HOG, structure tensor coherence, MBH, and checkerboard index, achieved Δp≈0.25–0.30 with stable accuracy above 720p. The results provide quantitative evidence that physics-based, explainable features can reliably separate fake from real content, advancing transparent forensic detection.