Table 4.
Cross-dataset in depth analysis: the architectures are trained on each of the 8 kinds of ForgeryNet deepfake manipulations and then tested on the DFDC Preview test set. AUC is given accordingly.
| Model | Train Set (ForgeryNet) | AUC |
|---|---|---|
| EfficientNet-V2-M | Method 1 | 51.0 |
| Method 2 | 50.3 | |
| Method 3 | 47.0 | |
| Method 4 | 49.7 | |
| Method 5 | 50.3 | |
| Method 6 | 47.0 | |
| Method 7 | 52.5 | |
| Method 8 | 50.0 | |
| ViT-Base | Method 1 | 53.3 |
| Method 2 | 52.5 | |
| Method 3 | 43.0 | |
| Method 4 | 52.0 | |
| Method 5 | 52.3 | |
| Method 6 | 51.3 | |
| Method 7 | 50.5 | |
| Method 8 | 49.8 | |
| Swin-Small | Method 1 | 53.0 |
| Method 2 | 65.3 | |
| Method 3 | 58.0 | |
| Method 4 | 59.5 | |
| Method 5 | 58.0 | |
| Method 6 | 55.3 | |
| Method 7 | 59.3 | |
| Method 8 | 56.7 |