Deepfake Detection on CelebDF (CDF) v2 (test)
95.4AUCLAA-Net (Ours w/ SBI)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| LAA-Net (Ours w/ SBI)Venue=CVPR’24, Training set (Real)=✓, Training set (Fake)=×2026.04 | 95.4 | — | 97.64 | |
| LAA-Former (Ours w/ SBI)Venue=not specified, Training set (Real)=✓, Training set (Fake)=×2026.04 | 94.45 | — | 97.15 | |
| Ours†Training Dataset=FF++, Feature Extracting Backbone=Swin Transformer2026.01 | 94.01 | — | — | |
| FA-ViTVenue=TCSVT’24, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 93.83 | — | — | |
| SBIsTraining Dataset=FF++, Feature Extracting Backbone=EfficientNet-B42026.01 | 93.18 | — | — | |
| UDDVenue=AAAI’25, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 93.1 | — | — | |
| AUNetVenue=CVPR’23, Training set (Real)=✓, Training set (Fake)=×2026.04 | 92.77 | — | — | |
| Ours*Training Dataset=FF++, Feature Extracting Backbone=EfficientNet-B42026.01 | 92.51 | — | — | |
| FakeSTormerVenue=ICCV’25, Training set (Real)=✓, Training set (Fake)=×2026.04 | 92.4 | — | — | |
| TALL-SwinVenue=ICCV’23, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 90.79 | — | — | |
| LAA-Former (Ours w/ BI)Venue=not specified, Training set (Real)=✓, Training set (Fake)=×2026.04 | 90.34 | — | 94.9 | |
| PCL+I2GVenue=ICCV’21, Training set (Real)=✓, Training set (Fake)=×2026.04 | 90.03 | — | — | |
| LSDAVenue=CVPR’24, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 89.8 | — | — | |
| FreqDebiasVenue=CVPR’25, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 89.6 | — | — | |
| AltFreezingVenue=CVPR’23, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 89.5 | — | — | |
| SBIsTraining Dataset=FF++, Feature Extracting Backbone=Swin Transformer2026.01 | 89.12 | — | — | |
| AltFreezeTraining Dataset=FF++, Training Compression=c23, Weight Source=reproduced using officially provided weights, Model Type=Video-based2024.03 | 89 | 95.2 | — | |
| OursTraining Dataset=FF++, Training Compression=c23, Model Type=Video-based2024.03 | 89 | 95.8 | — | |
| DFDTVenue=Appl.Sci.’22, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 88.3 | — | — | |
| FTCNTraining Dataset=FF++, Training Compression=c23, Weight Source=reproduced using officially provided weights, Model Type=Video-based2024.03 | 86.9 | 94.7 | — | |
| LAA-Net (Ours w/ BI)Venue=CVPR’24, Training set (Real)=✓, Training set (Fake)=×2026.04 | 86.28 | — | 91.93 | |
| SBI-R50Training Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Backbone=ResNet-50, Model Type=Image-based2024.03 | 85.7 | 87.3 | — | |
| RealForensicsTraining Dataset=FF++, Training Compression=c23, Weight Source=reproduced using officially provided weights, Model Type=Video-based2024.03 | 85.6 | 91.8 | — | |
| SBI*Venue=CVPR’22, Training set (Real)=✓, Training set (Fake)=×2026.04 | 85.55 | — | 77.81 | |
| ISTVTVenue=TIFS’23, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 84.1 | — | — | |
| LipForensicsTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Video-based2024.03 | 82.4 | — | — | |
| DCLTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 82.3 | — | — | |
| CADDM*Venue=CVPR’23, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 80.7 | — | 87.72 | |
| Face-X-RayTraining Dataset=FF++2026.01 | 80.58 | — | — | |
| SLADDTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 79.7 | — | — | |
| Face X-rayTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 79.5 | 88.6 | — | |
| FaceXRay (w/ BI)Venue=CVPR’20, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 79.5 | — | — | |
| LRLTraining Dataset=FF++2026.01 | 78.26 | — | — | |
| SPSLTraining Dataset=FF++2026.01 | 76.88 | — | — | |
| Two-branchTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 76.7 | — | — | |
| NoiseDFTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 75.9 | — | — | |
| SFDGVenue=CVPR’23, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 75.83 | — | — | |
| Multi-taskTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 75.7 | — | — | |
| CNN-augTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 75.6 | — | — | |
| STILTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Video-based2024.03 | 75.6 | — | — | |
| XceptionTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 73.7 | — | — | |
| RECCE*Venue=CVPR’22, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 70.93 | — | 70.35 | |
| CNN-GRUTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Video-based2024.03 | 69.8 | — | — | |
| PatchForensicsTraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 69.6 | — | — | |
| FWATraining Dataset=FF++, Training Compression=c23, Weight Source=obtained from AltFreeze [55], Model Type=Image-based2024.03 | 69.5 | — | — | |
| RECCETraining Dataset=FF++2026.01 | 68.71 | — | — | |
| Multi-attentional*Venue=CVPR’21, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 68.26 | — | 75.25 | |
| Multi-AttTraining Dataset=FF++2026.01 | 67.44 | — | — | |
| EfficientNet-B4Training Dataset=FF++2026.01 | 64.29 | — | — | |
| StyleGRUTraining Dataset=FF++, Training Compression=c23, Stage=Stage 1 training, Model Type=Video-based2024.03 | 62.5 | 79.8 | — | |
| XceptionTraining Dataset=FF++2026.01 | 61.8 | — | — | |
| Xception*Venue=ICCV’19, Training set (Real)=✓, Training set (Fake)=✓2026.04 | 61.18 | — | 66.93 |