FD²Foremer: Thinking Face Forgery Detection in Midfrequency Geometry Details

Table 6

Generalization evaluation (%) with previous methods on HQ (c23) FF++. AUC is used as metric of performance on the unseen manipulation technique. The highlighted are best results.

Training data	Model	Testing data (AUC)
		DF	F2F	FS	NT
DF	Xception [54]	99.3	73.6	49.0	73.6
	Face X-ray [55]	98.7	63.3	60.0	69.8
	FD^2Former	98.94	69.17	59.58	77.39
F2F	Xception [54]	80.3	99.4	76.2	69.6
	Face X-ray [55]	63.0	98.4	93.8	94.5
	FD^2Former	81.78	98.12	77.31	89.52
FS	Xception [54]	66.4	88.8	99.4	71.3
	Face X-ray [55]	45.8	96.1	98.1	95.7
	FD^2Former	70.21	97.83	99.27	93.12
NT	Xception [54]	79.9	81.3	73.1	99.1
	Face X-ray [55]	70.5	91.7	91.0	92.5
	FD^2Former	80.43	90.82	92.53	94.62

AUC is used as metric of performance on the unseen manipulation technique. Best results are shown in bold.