FeatureTransfer: Unsupervised Domain Adaptation for Cross-Domain Deepfake Detection

Table 3

The image-level results of all compared methods in terms of AUC (%) and EER (%) on each dataset.

Method	Test set
	DF-TIMIT		FF-FS		DFD		DFDC-P		Celeb-DF
	AUC	ERR	AUC	ERR	AUC	ERR	AUC	ERR	AUC	ERR
Xception [16]	98.80	5.95	99.56	2.74	83.06	25.92	82.10	27.23	72.54	34.71
FSS [18]	97.33	—	—	—	—	—	—	—	76.26	—
X-Ray [22]	—	—	98.00	-	95.40	8.37	80.92	27.54	80.58	26.70
Se_Res [37]	90.61	16.22	84.52	22.83	89.02	21.06	97.99	6.25	78.21	29.80
FT (ours)	96.56	8.05	88.62	19.52	91.00	16.21	98.77	5.75	86.21	22.42

Note. The “FSS,” “X-Ray,” “Se_Res,” and “FT” are the short forms of “FSSpotter,” “Face X-Ray,” “se_resnext101_32 × 4 d,” and “FeatureTransfer,” respectively.