Table 3.
Performance on anti-forensic datasets (FF++-AF and Celeb-DF-AF).
| Method | FF++-AF | Celeb-DF-AF |
|---|---|---|
| Xception (Rössler et al., 2019) | 2.4 | 7.8 |
| FreqNet (Durall et al., 2020) | 5.7 | 8.2 |
| SRM (Guo et al., 2021) | 10.1 | 14.5 |
| LipForensics (Haliassos et al., 2021) | 12.3 | 15.8 |
| DCNetwork (Zhou et al., 2023) | 25.7 | 23.4 |
| IID (Huang et al., 2023) | 26.3 | 17.2 |
| DFS (Ye et al., 2024) | 9.8 | 20.1 |
| Clipping (Khan and Dang-Nguyen, 2024) | 16.5 | 22.1 |
| Ours | 75.3 | 68.9 |
Metrics are accuracy (%). The bold values indicate the best performance during a comparison.