FD²Foremer: Thinking Face Forgery Detection in Midfrequency Geometry Details

Table 2

Results (%) of distillation for different manifold computing location. All trained with the image and the displacement map as inputs on low-quality FF++.

Teacher	Student	Teacher blocks extracted	Student blocks extracted	ACC
DF^2Former(swin)	DF^2Former(meta)	{11, 12}	{23, 24}	82.31
DF^2Former(swin)	DF^2Former(meta)	{1, 2}	{1, 2}	82.15
DF^2Former(swin)	DF^2Former(meta)	{1, 2, 11, 12}	{1, 2, 23, 24}	82.73
DF^2Former(swin)	DF^2Former(meta)	{1, 2, 3, 4, 11, 12}	{1, 2, 5, 6, 23, 24}	81.39

All trained with the image and the displacement map as inputs on low-quality FF++. Best results are shown in bold.