Deepfake Detection on Universal Deepfake Detection Evaluation Suite
100ProGAN AUROCOjha et al.
Evaluation results
| Method | Links | |||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Ojha et al.Backbone=CLIP-RN502025.12 | 100 | 98.98 | 97.045 | 89.62 | 99.74 | 98.3 | 74.525 | 62.16 | 68.26 | 78.11 | 85.5 | 78.9 | 96.665 | 82.955 | 96.765 | 86.905 | 87.595 | 86.355 | 91.965 | |
| Tan et al.Backbone=S-ResNet50 (Shallow ResNet architecture)2025.12 | 100 | 86.3 | 82.1 | 98 | 70.3 | 100 | 94.6 | 97.9 | 91.8 | 84.5 | 85.1 | 92.9 | 93.6 | 91.9 | 94 | 89.3 | 92.8 | 95.1 | 71.1 | |
| Wang et al.Backbone=ResNet50, Data Processing=50% probability of Gaussian blur and JPEG compression2025.12 | 100 | 93.1 | 85.7 | 98.9 | 94.8 | 96.4 | 85 | 97.2 | 75.1 | 98 | 98.9 | 77.5 | 75 | 72.1 | 75.6 | 83.8 | 87.1 | 85.3 | 59.9 | |
| Wang et al.Backbone=VGG112025.12 | 100 | 89.5 | 86.1 | 99.5 | 87 | 96.1 | 82 | 97.7 | 68.1 | 90.2 | 91.2 | 76.8 | 64.3 | 66.8 | 64.5 | 78.1 | 83.8 | 81 | 62.9 | |
| Wang et al.Backbone=ResNet502025.12 | 100 | 91.4 | 84.4 | 97.4 | 94.2 | 96.4 | 77 | 93.7 | 76.7 | 96 | 96.8 | 75.7 | 78.9 | 74.9 | 80.6 | 81 | 84.1 | 80.3 | 57.5 | |
| Wang et al.Backbone=MobileNetv22025.12 | 100 | 92.4 | 87.3 | 99.7 | 89.6 | 98.7 | 96 | 92.4 | 75.2 | 92.2 | 97.9 | 81 | 88.4 | 85.9 | 88 | 90.1 | 92.9 | 91 | 65.5 | |
| Frequency-Domain Masking (Wang et al. + Ours)Backbone=ResNet50, Data Processing=50% probability of Gaussian blur and JPEG compression2025.12 | 100 | 92.8 | 91.3 | 98.1 | 97.9 | 86.5 | 72.7 | 89.2 | 75 | 97.1 | 98.1 | 72.8 | 80.2 | 74.4 | 80.7 | 80.2 | 84.1 | 80.4 | 80 | |
| Frequency-Domain Masking (Wang et al. + Ours)Backbone=VGG112025.12 | 100 | 92 | 93.3 | 98.1 | 92.5 | 98.9 | 91.5 | 95.4 | 79 | 96.6 | 96.1 | 81.1 | 81 | 80.2 | 81.1 | 85.5 | 88 | 87.4 | 79.8 | |
| Frequency-Domain Masking (Wang et al. + Ours)Backbone=ResNet502025.12 | 100 | 93.2 | 92.9 | 98.9 | 97.2 | 94.3 | 80.5 | 95.5 | 75.6 | 98.7 | 98.4 | 74 | 83 | 79.4 | 83.1 | 84.8 | 88.6 | 85.6 | 82 | |
| Frequency-Domain Masking (Wang et al. + Ours)Backbone=MobileNetv22025.12 | 100 | 92.3 | 87.3 | 95.9 | 86.2 | 100 | 93.1 | 91.5 | 86.2 | 92.2 | 92.9 | 84.5 | 92.8 | 92.1 | 92.2 | 90.8 | 93.9 | 92.3 | 80.5 | |
| Gragnaniello et al.Backbone=ResNet50, Downsampling Protocol=First conv layer not downsampled2025.12 | 100 | 87.2 | 95.5 | 99.9 | 94 | 100 | 89.2 | 94.7 | 74.3 | 98.9 | 98.5 | 79.8 | 87.3 | 86.7 | 88.2 | 87.9 | 92.2 | 89.3 | 91.8 | |
| Frequency-Domain Masking (Gragnaniello et al. + Ours)Backbone=ResNet50, Downsampling Protocol=First conv layer not downsampled2025.12 | 100 | 93.7 | 98 | 99.9 | 98.7 | 100 | 93 | 95.7 | 81.8 | 97.3 | 96.2 | 79.5 | 94.2 | 93.8 | 94.8 | 90 | 93.1 | 90.9 | 96.4 | |
| Natraj et al.2025.12 | 98.72 | 72.05 | 52.18 | 95.565 | 52.105 | 61.345 | 58.12 | 66.02 | 58.135 | 69.355 | 76.505 | 65.35 | 80.955 | 79.79 | 81.695 | 79.785 | 78.975 | 76.345 | 74.26 | |
| Chai et al.Backbone=ResNet502025.12 | 96.62 | 69.71 | 66.705 | 87.61 | 56.545 | 86.175 | 57.75 | 65.205 | 52.055 | 61.515 | 61.35 | 67.595 | 83.465 | 80.555 | 83.73 | 70.8 | 72.415 | 71.94 | 73.26 | |
| Chai et al.Backbone=Xception2025.12 | 77.955 | 70.905 | 70.065 | 82.455 | 65.11 | 66.595 | 76.045 | 75.665 | 75.81 | 73.425 | 61.91 | 71.22 | 81.8 | 81.41 | 81.085 | 80.09 | 78.505 | 73.45 | 71.79 | |
| Zhang et al.Backbone=CycleGAN2025.12 | 52.65 | 99.95 | 62.79 | 52.505 | 58.19 | 99.85 | 47.64 | 48.73 | 52.56 | 52.105 | 50.54 | 54.31 | 64.06 | 63.825 | 63.385 | 60.14 | 57.99 | 56.16 | 58.885 |