Table 3 Comparison of AUC (%) for cross-dataset evaluation on CelebDF [8] and DFDC-P [29], including results of other methods cited from [11, 24, 37, 38, 60, 61]
From: Multi-attention-based approach for deepfake face and expression swap detection and localization
Method | FF++ | CelebDF | DFDC-P |
|---|---|---|---|
Two-stream [30] | 70.10 | 53.80 | 61.4 |
MesoNET [12] | 84.70 | 54.80 | 75.3 |
HeadPose [19] | 47.3 | 54.6 | 55.9 |
VA-MLP [18] | 66.4 | 55.0 | 61.9 |
FWA [57] | 80.10 | 56.90 | 72.7 |
Xception-raw [28] | 99.70 | 48.20 | 49.9 |
Xception-C23 [28] | 99.70 | 65.30 | 72.2 |
Xception-C40 [28] | 95.50 | 65.50 | 69.7 |
Capsule [10] | 96.60 | 57.50 | 53.3 |
Multi-task [41] | 76.30 | 54.30 | 53.6 |
Two-branch [11] | 93.18 | 73.41 | 64.0 |
\(F^3\)Net [16] | 98.10 | 65.17 | 70.1 |
EfficientNet [33] | 99.70 | 64.29 | 70.12 |
Sun et al. [38] | 99.3 | 64 | 69 |
GocNet [17] | 97.55 | 67.43 | – |
ADD [43] | 91.71 | 66.48 | – |
CViT [39] | 91.08 | 63.60 | 67.3 |
MaDD [24] | 99.80 | 67.44 | 67.1 |
FakePoI [37] | 94.7 | 61.2 | 72.5 |
Proposed method | 97.78 | 68.25 | 79.10 |