FD²Foremer: Thinking Face Forgery Detection in Midfrequency Geometry Details

Table 1

Test results (%) of the FD²Former and its variants on FF++(LQ), DFDC and Celeb-DF. The “image” indicates the model with cropped face images as input only. The “detail” indicates the model with facial geometry details as input only. The “swin” indicates the FD²Former of the swin transformer backbone. The “meta” is the FD²Former of the MetaFormer backbone. The “manifold” and the “hard” mean the manifold distillation and the hard-label distillation respectively. The metric on FF++(LQ), DFDC and Celeb-DF dataset is ACC.

Structure	FF++(LQ)	DFDC	Celeb-DF
Xception	80.32	85.60	61.25
Image(swin)	81.14	86.32	78.13
Detail(swin)	78.06	80.68	76.94
Img + detail(swin)	83.23	87.97	83.51
Img + detail + manifold(meta)	82.73	86.72	81.36
Img + detail + hard(meta)	81.67	86.03	79.91

The metric on FF++(LQ), DFDC and Celeb-DF dataset is ACC. Best results are shown in bold.