Research map / Label-efficient and robust learning / Recognition and long-tail
Preserving Fairness Generalization in Deepfake Detection
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 a fairness-generalization method for deepfake detection combining disentangled demographic/forgery features, fair learning with a bi-level fairness loss, and sharpness-aware loss flattening, trained on FF++ and tested cross-domain.
Problem
Deepfake detectors show demographic performance disparities that do not persist fairly under cross-domain testing; existing fair loss methods (DAW-FDD) fail to generalize fairness to unseen forgeries.
Approach
Jointly disentangles domain-agnostic forgery and demographic features, fuses them via AdaIN for a fair classifier with a bi-level (inter- and intra-subgroup) fairness loss, and flattens the loss landscape with sharpness-aware minimization.
Key results
- On cross-domain Celeb-DF, FFPR reduced by 11.69% relative to DAW-FDD (per the paper's text), with best AUC on Xception across FF++, DFDC, Celeb-DF and DFD.
- Ablation shows loss flattening (Lf) reduces FFPR by 7.94% on Celeb-DF and 4.18% on DFDC versus no flattening.
- Max FPR gap across intersectional subgroups on Celeb-DF drops from 20.6 (DAW-FDD) to 9.3 (ours).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed framework instead of DAW-FDD | Xception deepfake detector · Celeb-DF | FFPR: 11.69% | Helped |
| Use the proposed framework instead of DAW-FDD | Xception deepfake detector · DFD | FFPR: 7.94% | Helped |
| Use the proposed framework instead of DAW-FDD | Xception deepfake detector · DFDC | FDP: 8.90% | Helped |
| Add reconstruction loss (ablation) | disentanglement learning module · Celeb-DF | FFPR: 5.47% | Helped |
| Add reconstruction loss (ablation) | disentanglement learning module · Celeb-DF | AUC: +2.69% | Helped |
| Add contrastive loss (ablation) | disentanglement learning module · Celeb-DF | FFPR: 6.32% | Helped |
| Add contrastive loss (ablation) | disentanglement learning module · Celeb-DF | AUC: +4.03% | Helped |
| Use demographic distribution-aware margin loss instead of cross-entropy (ablation) | disentanglement learning module · FF++ | FFPR: 5.95% | Helped |
| Use demographic distribution-aware margin loss instead of cross-entropy (ablation) | disentanglement learning module · FF++ | AUC: +0.23% | Helped |
| Add loss flattening (ablation) | proposed method · Celeb-DF | FFPR: 7.94% | Helped |
| Add loss flattening (ablation) | proposed method · DFDC | FFPR: 4.18% | Helped |
| Add feature fusion (ablation) | proposed method · Celeb-DF | FFPR: 4.10% | Helped |
Surprising findings
- The existing fair-loss method DAW-FDD underperforms the unfair baseline Ori on some cross-domain tests (e.g., 4.72% FDP decrease on DFDC), indicating fairness does not transfer across domains.
Limitations
- Depends on training datasets containing forgeries from multiple manipulation techniques; few datasets satisfy this.
- Evaluated only on video face-manipulation datasets, not diffusion or GAN-generated images.
Experimental setup
Datasets: FaceForensics++ DeepfakeDetection (DFD) DFDC Celeb-DF
Benchmarks: FaceForensics++ DFD DFDC Celeb-DF
Metrics: AUC FFPR FMEO FDP FOAE
Baselines: DAW-FDD UCF Ori (Xception with cross-entropy)
Builds on: DAW-FDD UCF sharpness-aware minimization AdaIN Xception
Closest papers by meaning
- Fair Deepfake Detectors Can Generalize (NeurIPS 2025 · similarity 0.93)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.87)
- Thinking Racial Bias in Fair Forgery Detection: Models, Datasets and Evaluations (AAAI 2025 · similarity 0.86)
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.86)
- DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion (NeurIPS 2024 · similarity 0.85)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.85)
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.