Research map / Detection and segmentation / CNN features
Can We Leave Deepfake Data Behind in Training Deepfake Detector?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Oriented Progressive Regularizer and feature bridging to train a deepfake detector on real, blendfake (SBI, CBI) and deepfake data arranged as progressive anchors in latent space, outperforming blendfake-only training.
Problem
Blendfake-only training is standard for generalizable deepfake detection, but naively combining deepfake and blendfake data (vanilla hybrid training) degrades performance; the question is how to effectively use deepfake data too.
Approach
Treat real, blendfake (SBI, CBI), and deepfake as ordered anchors of a progressive real-to-fake transition, encode accumulated forgery attributes with triplet binary labels, bridge adjacent anchors via feature mixup with mixed labels, and add a noise-based transition loss to organize the latent space.
Key results
- Cross-dataset average AUC of 0.8225 on CDFv1/CDFv2/DFDCP/DFDC versus 0.7814 for prior SoTA SBI, a 5.26% relative gain.
- Triplet binary classification achieves C-Avg AUC 0.8553 versus 0.8229 (multi-class) and 0.8308 (multi-label) in ablation.
- Vanilla hybrid training (VHT) underperforms blendfake-only training on cross-dataset metrics (C-Avg AUC 0.7811 vs 0.8070).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OPR's progressive organization to train with both blendfake and deepfake data instead of vanilla hybrid training | Deepfake detector · C-Avg. cross-dataset evaluation from FF++ | AUC: 0.7811 -> 0.8225 | Helped |
| Add blendfake-only training rather than vanilla hybrid training | Deepfake detector · C-Avg. cross-dataset evaluation from FF++ | AUC: 0.7811 -> 0.8070 | Helped |
| Add oriented loss L_o to progressively organize latent-space anchors (ablation) | OPR · C-Avg. cross-dataset evaluation | AUC: 0.8087 -> 0.8553 | Helped |
| Add feature bridging between adjacent anchors (ablation) | OPR · C-Avg. cross-dataset evaluation | AUC: 0.8337 -> 0.8553 | Helped |
| Add transition loss L_t (ablation) | OPR · C-Avg. cross-dataset evaluation | AUC: 0.8473 -> 0.8553 | Helped |
| Replace triplet binary classification with multi-class classification (ablation) | OPR · FF++ and cross-dataset benchmarks | AUC: FF++: 0.9591 -> 0.9677; C-Avg.: 0.8553 -> 0.8229 | Mixed |
| Replace triplet binary classification with multi-label classification (ablation) | OPR · CDFv1 | AUC: 0.9094 -> 0.8757 | Hurt |
| Use triplet binary rather than multi-label classification to distinguish forgery attributes (ablation) | OPR · CDFv1 | AUC: 0.8757 -> 0.9094 | Helped |
| Use OPR rather than VHT under unseen perturbations | Deepfake detector · FF++ test set under unseen perturbations (Avg.) | AUC: 0.8353 -> 0.8713 | Helped |
Surprising findings
- Vanilla hybrid training combining deepfake and blendfake data performs worse than blendfake-only training across cross-dataset metrics.
- Multi-class classification gives the best in-dataset FF++ AUC but a marginal cross-dataset gain.
Limitations
- Relies on the assumption that deepfakes are easier to distinguish than blendfakes, which may weaken as deepfake techniques advance.
- Evaluated only with FF++ as training data and a single backbone (EfficientNet-B4).
- t-SNE fails to show progressive transitions, so analysis relies on 2D toy models.
Experimental setup
Datasets: FaceForensics++ Celeb-DF-v1 Celeb-DF-v2 DFDC-Preview DFDC DF40 DiffusionFace
Benchmarks: FaceForensics++ Celeb-DF-v1 Celeb-DF-v2 DFDC-Preview DFDC DF40 subsets
Metrics: AUC EER video-level AUC mPD
Baselines: Xception EfficientB4 SPSL SRM SLADD SBI IID UCF X-ray VHT
Builds on: SBI CBI DeepFakeBench EfficientNet
Closest papers by meaning
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.88)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.88)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.88)
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.87)
- DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion (NeurIPS 2024 · similarity 0.87)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.