FD²Foremer: Thinking Face Forgery Detection in Midfrequency Geometry Details

Table 7

Generalization evaluation (%) on the unseen dataset, Celeb-DF, and DFDC. The highlighted are best results.

Model	Training	Testing AUC
		Celeb-DF	DFDC
EfficientNetB4Ensemble [59]	FF++	55.8	63.0
Xception [54]		59.4	67.9
Face X-ray [55]		74.2	70.0
FD^2Former		79.2	73.6

Generalization evaluation (%) on the unseen dataset, Celeb-DF, and DFDC. Best results are shown in bold.