Assessment framework for deepfake detection in real-world situations

EURASIP Journal on Image and Video Processing

Table 3 AUC (%) scores of four selected deepfake detection methods on the distorted variants of the FFpp test set that are subject to different video processing operations. The notations C23 and C40 here refer to the two different compression rates using AVC/H.264 codec. The notation Resolution refers to reducing video resolution by a specific scale

Methods	TrainSet	Compression		Brightness		Grayscale	Contrast	Flipping		Resolution		Gaussian Noise	Vintage Filter	Overall Average
		C23	C40	Increase	Decrease			Horizontal	Vertical	x2	x4
CapsuleNet	FFpp-Raw	77.97	54.14	73.31	70.62	68.38	69.31	73.13	63.20	65.43	56.99	54.14	72.94	66.63
XceptionNet		69.49	55.70	65.92	66.40	65.51	65.32	65.26	57.36	57.23	55.90	50.50	66.90	61.79
SBIs		90.43	76.27	86.38	86.47	86.27	85.94	85.98	79.28	76.35	63.62	71.52	86.54	81.25
UIA-VIT		93.82	71.56	91.10	88.55	88.91	89.18	89.02	77.74	79.78	72.72	71.50	87.11	83.42
CapsuleNet	FFpp-C23	95.61	66.03	93.27	92.31	87.43	91.55	91.98	71.49	80.28	67.56	53.50	88.86	81.66
XceptionNet		98.34	70.71	97.07	96.65	93.17	96.34	96.20	66.82	83.42	72.03	51.04	94.99	84.73
SBIs		91.71	75.43	87.63	86.51	87.31	87.40	86.84	81.22	75.40	64.31	57.06	86.28	80.59
UIA-VIT		96.82	75.94	95.08	94.58	92.79	95.49	95.47	81.00	88.59	78.80	75.42	93.12	88.59
CapsuleNet	FFpp-C40	82.64	78.33	80.22	80.77	79.30	52.78	78.64	61.53	76.88	71.91	78.41	75.82	74.77
XceptionNet		83.25	80.69	80.85	82.83	80.65	51.74	81.39	55.70	80.62	74.99	71.30	78.43	75.20
SBIs		83.00	70.66	76.49	78.15	77.49	62.82	77.34	69.14	67.04	56.42	76.67	76.50	72.64
UIA-VIT		86.98	83.86	85.13	85.73	79.77	86.22	86.12	68.66	84.88	77.74	78.93	82.98	82.25