Multisemantic Path Neural Network for Deepfake Detection

Table 6

Face cross test results, using frame-level AUC (%) to compare our method with others on both benchmarks.

Method	FF-DF	Celeb-DFv2
Two-stream [49]	70.1	53.8
Meso4 [4]	84.7	54.8
MesoInception4 [4]	83.0	53.6
HeadPose [51]	47.3	54.6
FWA [52]	80.1	56.9
DSP-FWA [52]	93.0	64.0
VA-MLP [53]	66.4	55.0
VA-LogReg [53]	78.0	55.1
Xception [6]	93.65	64.52
Multitask [54]	76.3	54.3
Two branch [55]	93.18	73.41
Capsule [56]	96.6	57.5
Ours	96.7	66.7

Bold values indicate higher results in two experiments.