Table 2 Quantitative results in terms of ACC (%) on the FF++ [28] dataset were obtained for four different manipulation methods, including Deepfake (DF), Face2Face (F2F), FaceSwap (FS), and NeuralTextures (NT)
From: Multi-attention-based approach for deepfake face and expression swap detection and localization
Method | Input | Mask | Face swap | Expression swap | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
DF (HQ) | DF (LQ) | FS (HQ) | FS (LQ) | NT (HQ) | NT (LQ) | F2F(HQ) | F2F (LQ) | |||
Steg.Features+SVM [54] | RGB | N | 77.12 | 65.58 | 79.51 | 68.93 | 76.94 | 60.69 | 74.68 | 60.58 |
Cozzolino et al. [51] | RGB | N | 81.78 | 68.26 | 85.69 | 73.79 | 80.60 | 62.42 | 85.32 | 62.08 |
Bayar and Stamm [59] | RGB | N | 90.18 | 80.95 | 93.14 | 82.52 | 86.04 | 72.38 | 94.93 | 76.83 |
MesoNet [12] | RGB | N | 95.26 | 89.52 | 81.24 | 61.17 | 85.95 | 75.74 | 95.84 | 83.56 |
XceptionNet [28] | RGB | N | 98.85 | 94.88 | 98.23 | 92.17 | 94.50 | 82.11 | 98.23 | 91.56 |
Multi-Task [41] | RGB | Y | 93.92 | 85.77 | – | – | 88.05 | 80.67 | 92.77 | 82.31 |
Sun et al. [38] | RGB | N | – | 69.1 | – | 68.1 | – | 60.8 | – | 65.7 |
SSTNet [14] | RGB | N | – | 95.33 | – | 94.09 | – | – | – | 90.48 |
SPSL [53] | FREQ | N | – | 93.48 | – | 92.26 | – | 76.78 | - | 86.02 |
ADD [43] | RGB | Y | 97.45 | – | 97.20 | – | 90.84 | – | 98.33 | – |
Proposed method | RGB+FREQ | Y | 99.97 | 96.47 | 97.88 | 93.88 | 96.06 | 90.55 | 95.97 | 90.92 |