Multisemantic Path Neural Network for Deepfake Detection

Table 5

Quantitative detection results of ACC (%) using FF++ dataset on high quality (C23 light compression) and low quality (C40 heavy compression) videos and AUC on TIMIT datasets. Bold font indicates the best result.

	Acc on FF++(HQ)	Acc on FF++(LQ)	AUC on TIMIT(HQ)	AUC on TIMIT(LQ)
Bayar and stamm [47]	88.68	61.6	86.50	88.33
InMesonet [4]	57.81	69.75	81.15	82.63
Rahmouni et al. [48]	—	58.10	—	—
Mesonet [4]	54.91	50.28	63.68	77.44
Zhou et al. [49]	—	—	73.5	83.5
Chollet [6]	91.87	72.93	93.64	88.24
Nirkin et al. [50]	—	75.00	—	—
Ours	94.21	76.31	99.12	99.52