Deepfake Detection on Celeb-DF
0.981ROC-AUCAletheia
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Aletheia2026.04 | 0.981 | — | — | |
| DF40 baseline2026.04 | 0.975 | — | — | |
| LAA-Net2026.04 | 0.972 | — | — | |
| FACTOREvaluation Protocol=Ref. Set, Reference Set Source=50% subset of training videos2023.11 | 0.97 | — | — | |
| Xception baseline2026.04 | 0.965 | — | — | |
| LAA-Net (w/ SBI)Venue=CVPR24, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.954 | — | — | |
| M2F2-Det (w/ SBI)Training set (Real)=true, Training set (Fake)=false2025.03 | 0.951 | — | — | |
| FreqBlenderVenue=NeurIPS24, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.9459 | — | — | |
| CADDMVenue=CVPR23, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.9388 | — | — | |
| SBIVenue=CVPR22, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.9318 | — | — | |
| AUNetVenue=CVPR23, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.9277 | — | — | |
| SeeableVenue=ICCV23, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.873 | — | — | |
| UCFVenue=ICCV23, Training set (Real)=true, Training set (Fake)=true2025.03 | 0.824 | — | — | |
| GenDFCategory=Fine-tuned, Trainable Parameters=0.28M2025.12 | 0.82 | 0.258 | — | |
| UIA-ViT*Category=Custom-made, Trainable Parameters=81.83M2025.12 | 0.8047 | — | — | |
| ViT+LoRACategory=Fine-tuned, Trainable Parameters=0.27M2025.12 | 0.7967 | — | — | |
| EvolveReasonProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7841 | 0.284 | 76.5 | |
| LocalRLVenue=AAAI21, Training set (Real)=true, Training set (Fake)=false2025.03 | 0.7826 | — | — | |
| [Sun et al., 2025]Protocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7687 | 0.3138 | 73.91 | |
| SPSLTraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.765 | — | — | |
| FFAAProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7531 | 0.3095 | 74.08 | |
| UCFTraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.753 | — | — | |
| UCFCategory=Custom-made, Trainable Parameters=46.84M2025.12 | 0.7527 | — | — | |
| EfficientNet-B4Category=Pre-trained, Trainable Parameters=19M2025.12 | 0.7487 | — | — | |
| Forensics AdapterProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7455 | 0.3306 | 73.37 | |
| FFDTraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.744 | — | — | |
| SIDAProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7439 | 0.3058 | 73 | |
| FakeReasoningProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7394 | 0.3107 | 72.68 | |
| EffNetB4Training Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.739 | — | — | |
| XceptionTraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.737 | — | — | |
| F3NETTraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.735 | — | — | |
| RECCETraining Dataset=FF++(C23), Evaluation Protocol=Cross-Dataset2023.11 | 0.732 | — | — | |
| CorrDetailProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.728 | 0.3387 | 72.32 | |
| Deepfake+AdapterCategory=Fine-tuned, Trainable Parameters=16.92M2025.12 | 0.7174 | 0.34 | — | |
| RECCE (DD-VQA)Protocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7021 | 0.3563 | 69.46 | |
| UIA-VITProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.7015 | 0.3582 | 69.8 | |
| PFG-DDProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.6968 | 0.3594 | 68.51 | |
| HiFi-NetProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.688 | 0.3613 | 67.2 | |
| RECCE*Category=Custom-made, Trainable Parameters=24.69M2025.12 | 0.6871 | 0.357 | — | |
| RECCEProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.6871 | 0.3573 | 67.96 | |
| DE-AdapterCategory=Fine-tuned, Trainable Parameters=0.38M2025.12 | 0.6712 | — | — | |
| MultiAttCategory=Custom-made, Trainable Parameters=417.52M2025.12 | 0.6702 | 0.379 | — | |
| ViT*Category=Pre-trained, Trainable Parameters=1.5K2025.12 | 0.664 | 0.388 | — | |
| RFMProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.6564 | 0.388 | 64.42 | |
| RFMCategory=Custom-made2025.12 | 0.6563 | 0.385 | — | |
| XceptionCategory=Pre-trained, Trainable Parameters=20.81M2025.12 | 0.618 | 0.417 | — | |
| XceptionNetProtocol=Cross-Testing, Trained on=FF++ (LQ)2026.03 | 0.618 | 0.4173 | 60.11 | |
| F3-NetCategory=Custom-made, Trainable Parameters=20.75M2025.12 | 0.6151 | 0.42 | — | |
| DeepAgentModality=Audio + Video + Text, Architecture=CNN + DNN + RF, Fusion=Feature + Decision, AV Consistency=Explicit AV lexical & semantic, Cross-dataset=Yes2025.12 | — | — | 81.56 | |
| Ismail et al.Modality=Video, Architecture=YOLO + CNN + XGB, Fusion=None, AV Consistency=None, Cross-dataset=No2025.12 | — | — | 90.73 | |
| Saikia et al.Modality=Video, Architecture=CNN + LSTM, Fusion=None, AV Consistency=None, Cross-dataset=No2025.12 | — | — | 79.49 | |
| Shi et al.Modality=Video, Architecture=ViT (RFFR), Fusion=Feature-level, AV Consistency=None, Cross-dataset=No2025.12 | — | — | 88.98 |