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

Table 4

Ablation studies on different branches and proposed components.

Methods	Variant					FF++(c40)		Celeb-DF
	EfficientNetB4	Swin-B	ANTEM	MFAM	SC + CE	Acc	AUC	AUC
1						85.14	87.12	64.12
2						81.68	83.74	66.41
3						86.21	87.34	67.11
4						86.23	87.32	68.78
5						87.27	88.46	67.43
6						87.45	88.61	69.72
7						88.21	89.84	71.02

We progressively add each component and compare the detection performance in the intra-dataset (FF++(c40)) and cross-dataset (Celeb-DF) scenarios. The best performances are marked with bold fonts.