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

Table 1

Quantitative comparison results for different quality settings in FF++ dataset.

Methods	Level	HQ (c23)		LQ (c40)
	Metric	Acc	AUC	Acc	AUC
Steg. Features [44]		70.97	—	55.98	—
LD-CNN [45]		78.45	—	58.69	—
Constrained Conv [46]		82.97	—	66.84	—
MesoNet [4]		83.1	—	70.47	—
DSP-FWA [47]		—	57.49	—	62.34
Face X-ray [48]		—	87.4	—	61.6
Xception [20]		94.93	97.32	83.52	86.02
EfficientNetB4 [39]		95.84	98.31	85.14	87.12
Vit [15]		84.32	87.73	76.53	79.81
Swin-B [32]		90.64	92.32	81.68	83.74
SPSL [27]		91.5	95.32	81.57	82.82
GFFD [9]		96.18	98.56	86.16	87.94
MADD [7]		97.12	99.05	85.78	87.31
Our		97.37	99.34	88.21	89.84

The best results are marked in bold fonts.