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

Table 1

Robustness experiments with various video preprocessing and different hash bits on FaceForensics++.

Video processing	FF++ raw (bits)					FF++ C23 (bits)					FF++ C40 (bits)
	64	128	256	512	1024	64	128	256	512	1024	64	128	256	512	1024
None	0.852	0.932	0.948	0.998	0.991	0.847	0.944	0.944	0.998	0.990	0.846	0.941	0.946	0.997	0.991
Sharpening	0.850	0.930	0.949	0.999	0.991	0.845	0.943	0.945	0.999	0.989	0.847	0.942	0.945	0.996	0.991
Noise	0.844	0.937	0.944	0.999	0.990	0.844	0.933	0.940	0.999	0.991	0.853	0.944	0.942	0.999	0.991
Blur	0.846	0.934	0.947	0.998	0.991	0.844	0.944	0.939	0.999	0.991	0.848	0.942	0.941	0.999	0.991
Median filter	0.850	0.935	0.948	0.998	0.991	0.844	0.945	0.941	0.998	0.991	0.851	0.942	0.946	0.997	0.992
Video crop	0.633	0.801	0.862	0.983	0.963	0.636	0.862	0.814	0.986	0.962	0.629	0.816	0.859	0.988	0.964

Bold values represent the best results in the correlation domain.