Vision Transformer-Based Video Hashing Retrieval for Tracing the Source of Fake Videos

Table 4

Comparison experiment of fine-grained accuracy (ACC) with recent works on FaceForensics++ high-quality (HQ) and low-quality (LQ) datasets.

Methods	FF++ (HQ)				FF++ (LQ)				Celeb-DF
	DF	F2F	FS	NT	DF	F2F	FS	NT
Xception	98.9	98.9	99.6	95.0	96.8	91.1	94.6	87.1	99.4
I3D	92.9	92.9	96.4	90.4	91.1	86.4	91.4	78.6	99.2
LSTM	99.6	99.3	98.2	93.9	96.4	88.2	94.3	88.2	95.7
TEI	97.9	97.1	97.5	94.3	95.0	91.1	94.6	90.4	99.1
ADDNet-3d	92.1	83.9	92.5	78.2	90.4	78.2	80.0	69.3	95.2
S-MIL	98.6	99.3	99.3	95.7	96.8	91.4	94.6	88.6	99.2
S-MIL-T	99.6	99.6	100.0	94.3	97.1	91.1	96.1	86.8	98.8
STIL	99.6	99.3	100.0	95.4	98.2	92.1	97.1	91.8	99.8
VTN	99.6	99.3	99.6	95.4	97.9	92.1	95.7	90.4	99.3
ISTVT	99.6	99.6	100.0	96.8	98.9	96.1	97.5	92.1	99.8
Ours	99.9	99.9	99.9	0.999	99.9	100.0	99.9	99.9	99.4