Multisemantic Path Neural Network for Deepfake Detection

Table 2

Y, Cr, Cb, and RGB channels used as input for training and test results for C23 and C40 datasets.

	Acc (%) on FF++(HQ)				Acc (%) on FF++(LQ)
	DeepFakes	Face2Face	FaceSwap	NeuralTextures	DeepFakes	Face2Face	FaceSwap	NeuralTextures
RGB	99.73	99.09	99.17	91.2	93.83	91.15	92.47	74.41
Cr channel	84.46	85.58	79.04	83.35	79.67	75.73	75.32	63.32
Cb channel	87.19	85.17	77.08	80.11	80.45	72.91	69.65	59.83
Y channel	99.57	99.21	99.34	91.08	94.35	91.01	92.55	74.92

Bold values indicate the best results for four different channels.