Research Article
Face Forgery Detection with Long-Range Noise Features and Multilevel Frequency-Aware Clues
Table 3
Quantitative cross-dataset comparison results for AUC metric on Celeb-DF and DFDC, with training on FF++(c23).
| | Methods | Training Set | Celeb-DF | DFDC |
| | Xception [20] | FF++(c23) | 65.23 | 68.21 | | EfficientNetB4 [39] | FF++(c23) | 66.31 | 69.45 | | Vit [15] | FF++(c23) | 69.14 | 70.31 | | Swin-B [32] | FF++(c23) | 68.13 | 70.29 | | M2TR [34] | FF++(c23) | 65.70 | — | | SPSL [27] | FF++(c23) | 76.88 | 66.16 | | MD-CSDNet [49] | FF++(c23) | 68.77 | — | | F3-Net [10] | FF++(c23) | 65.17 | — | | MTD-Net [26] | FF++(c23) | 70.12 | — | | MADD [7] | FF++(c23) | 68.21 | 71.02 | | GFDD [9] | FF++(c23) | 70.13 | 72.17 | | Ours | FF++(c23) | 73.16 | 74.89 |
|
|
The best results are denoted in bold, and the second-best results are underlined.
|
