Research map / Label-efficient and robust learning / Recognition and long-tail
Fair Deepfake Detectors Can Generalize
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 DAID, a plug-and-play deepfake detection training framework that uses demographic-aware reweighting/normalization and demographic-agnostic alignment to improve both fairness (Skew) and cross-domain AUC, motivated by a back-door-adjusted causal claim that fairness improves generalization.
Problem
Deepfake detectors generalize poorly to unseen manipulations and are unfair across demographic groups; prior work treats fairness and generalization as a trade-off.
Approach
Uses back-door adjustment over data distribution and model capacity confounders to argue fairness causally improves generalization, and builds DAID from inverse-propensity reweighting, subgroup-wise feature normalization, and a cosine alignment loss across demographic groups with a low-rank orthonormal projection.
Key results
- On Celeb-DF, DAID (CADDM backbone) reaches Skew 0.289 and AUC 84.39 versus the best baseline CADDM at 0.391 / 81.75.
- On DFDC, DAID reaches AUC 66.85 and Skew 1.460, outperforming all listed baselines in both metrics.
- The ACE estimate shows an average gain of 2.35 percentage points from fairness intervention (bootstrap 95% CI [0.0186, 0.0280]).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DAID instead of the strongest fairness baseline | Deepfake detection · DFDC | Skew: 1.932 -> 1.460 | Helped |
| apply DAID instead of the strongest AUC baseline | Deepfake detection · DFDC | AUC: 65.89 -> 66.85 | Helped |
| apply DAID instead of the strongest fairness baseline | Deepfake detection · DFD | Skew: 0.351 -> 0.263 | Helped |
| apply DAID instead of the strongest AUC baseline | Deepfake detection · DFD | AUC: 90.08 -> 91.15 | Helped |
| apply DAID instead of the strongest fairness baseline | Deepfake detection · Celeb-DF | Skew: 0.391 -> 0.289 | Helped |
| apply DAID instead of the strongest AUC baseline | Deepfake detection · Celeb-DF | AUC: 82.71 -> 84.39 | Helped |
| add fairness intervention using rebalancing | Deepfake detector under confounder stratification · Cross-domain test sets | AUC: 2.35 percentage points; Δ = 0.0235, 95% CI [0.0186, 0.0280] | Helped |
| add sample reweighting (ablation) | CADDM on DFDC · DFDC | AUC: 63.77 -> 64.94 | Helped |
| add subgroup feature normalization to reweighting (ablation) | CADDM on DFDC · DFDC | AUC: 64.94 -> 65.96 | Helped |
| add feature alignment loss L_attr without data rebalancing (ablation) | CADDM on DFDC · DFDC | AUC: 63.77 -> 65.40 | Helped |
| add orthogonality loss L_ortho to feature alignment without rebalancing (ablation) | CADDM on DFDC · DFDC | AUC: 65.40 -> 64.96 | Hurt |
| add L_attr and L_ortho to reweighting and normalization (ablation) | CADDM on DFDC · DFDC | AUC: 65.96 -> 66.85 | Helped |
Surprising findings
- Fairness-oriented methods such as FG improve Skew but reduce AUC, while DAID improves both fairness and generalization simultaneously.
Limitations
- Relies on demographic annotations, which are not native to the evaluation datasets and must be annotated.
- Causal claim rests on a binary fairness treatment and a limited ACE estimate from a few backbones.
Experimental setup
Datasets: FaceForensics++ DFDC DFD Celeb-DF
Benchmarks: DFDC DFD Celeb-DF
Metrics: AUC Skew
Baselines: Xception EfficientNet F3-Net Face X-ray SBI RECCE GRU CADDM UCF ProDet VLFFD DAW-FDD
Builds on: back-door adjustment Xception EfficientNet F3-Net CADDM
Closest papers by meaning
- Preserving Fairness Generalization in Deepfake Detection (CVPR 2024 · similarity 0.93)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.86)
- Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales (NeurIPS 2026 · similarity 0.86)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.86)
- Towards Harmless Rawlsian Fairness Regardless of Demographic Prior (NeurIPS 2024 · similarity 0.86)
- Do Counterfactually Fair Image Classifiers Satisfy Group Fairness? -- A Theoretical and Empirical Study (NeurIPS 2024 · 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.