Local Corner and Motion Key Point Trajectory Extraction for Facial Forgery Identification

Table 5

Cross-dataset experiments (AUC (%)).

Methods	Celeb-DF	DFDC Preview
Train on FF++ (raw)
Xception [18]	48.2	49.9
LRNet [11]	56.9	—
DSP-FWA [24]	64.6	—
Capsule [25]	57.5	53.3
FWA [7]	56.9
Ours (raw)	64.8	63.5
Train on FF++ (c23)
FWA [7]	53.9
LipForensics [31]	82.4	—
Ours (c23)	65.1	64.1
Train on FF++ (c40)
STIL [33]	75.58	—
ADDNet-3D [34]	60.85
Ours (c40)	64.7	63.8

The bold values are used to highlight the experimental records that represent the most optimal performance within each experimental group. Specifically, the bold values labeled as “Ours” indicate the results obtained from the experiments conducted using our proposed method.