Face Forgery Detection with Long-Range Noise Features and Multilevel Frequency-Aware Clues

Table 2

Quantitative comparison results in terms of Acc and AUC on FF++ dataset with four manipulation methods.

Methods	Manipulations (LQ)	Deepfake		Face2Face		FaceSwap		NeuralTexture
	Metrics	Acc	AUC	Acc	AUC	Acc	AUC	Acc	AUC
Xception [20]		92.81	94.32	85.21	87.04	91.84	93.83	75.21	77.67
EfficientNet [39]		93.12	95.64	85.32	87.21	92.38	94.23	76.41	79.13
Vit [15]		79.86	82.78	67.91	69.34	76.65	79.18	65.43	68.78
Swin-B [32]		85.25	88.32	76.68	78.12	83.43	85.16	72.31	75.14
GFFD [9]		94.02	96.01	86.02	88.23	92.52	94.22	79.21	82.97
MADD [7]		94.47	96.43	86.87	89.42	93.66	95.45	81.21	83.61
Ours		95.58	97.01	88.23	90.81	94.74	96.16	83.46	86.02

The best results are marked in bold fonts.