Research map / Label-efficient and robust learning / Recognition and long-tail
Distributionally Generative Augmentation for Fair Facial Attribute Classification
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 DiGA, a two-stage generative framework that detects spurious attributes by combining biased latent-space editing directions and then trains a fair facial attribute classifier via uniformly-sampled generative augmentation of those attributes, without spurious-attribute labels.
Problem
Training fair facial attribute classifiers on biased data when spurious attribute labels (e.g., Male for Smiling) are unavailable, leading to accuracy inconsistencies across subpopulations.
Approach
Cancel the target-attribute component of two differently regularized latent classifier directions to isolate a spurious-attribute editing direction, then augment training images by editing spurious attributes with uniformly sampled magnitudes and enforce representation invariance via contrastive/BYOL-style learning.
Key results
- On CelebA (T=Smiling, S=Male), DiGA reaches 88.4% accuracy, 81.9% worst-group accuracy and 7.4% EO versus ERM's 88.2%, 70.1%, 25.3%.
- On UTKFace (T=Gender, S=Ethnicity), DiGA attains 92.7% accuracy and 6.8 EO... reported in Table 2 as the best EO among compared methods on the gender task.
- On Dogs and Cats, DiGA achieves 88.4% accuracy, 81.1% worst-group accuracy and 7.8 EO versus ERM's 87.5/67.8/26.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DiGA instead of ERM | FAC model · CelebA (T=Smiling, S=Male) | worst-group accuracy (%): 70.1 -> 81.9 | Helped |
| Apply DiGA instead of ERM | FAC model · CelebA (T=Smiling, S=Male) | EO (%): 25.3 -> 7.4 | Helped |
| Apply DiGA instead of ERM | FAC model · CelebA (T=Smiling, S=Male) | accuracy (%): 88.2 -> 88.4 | Helped |
| Apply DiGA instead of ERM | FAC model · UTK-Face (T=Gender, S=Ethnicity) | EO (%): 12.2 -> 6.8 | Helped |
| Apply DiGA instead of ERM | FAC model · Dogs and Cats (Species, Color) | worst-group accuracy (%): 67.8 -> 81.1 | Helped |
| Add fairness-aware generative augmentation to BYOL's traditional augmentation (ablation) | DiGA representation learning · CelebA (T=Smiling, S=Male) | accuracy and fairness | Helped |
| Use distributionally sampled generative augmentation instead of single-point augmentation (ablation) | DiGA representation learning · CelebA (T=Smiling, S=Male) | accuracy and fairness | Helped |
| Increase the augmentation-degree range (ablation) | DiGA representation learning · CelebA (T=Smiling, S=Male) | EO (%) | Helped |
| Use a smaller sampled subset to train the generative model (ablation) | DiGA · CelebA (T=Smiling, S=Male) | classification performance | Hurt |
| Set λ1=2e-4 and λ2=5e+3 (ablation) | DiGA bias detection · CelebA (T=Smiling, S=Male) | worst-group accuracy (%): 81.9 -> 82.7 | Helped |
| Set λ1=2e-5 and λ2=5e+4 (ablation) | DiGA bias detection · CelebA (T=Smiling, S=Male) | worst-group accuracy (%): 81.9 -> 85.1 | Helped |
| Set λ1=2e-5 and λ2=5e+4 (ablation) | DiGA bias detection · CelebA (T=Smiling, S=Male) | EO (%): 7.4 -> 4.8 | Helped |
Surprising findings
- Regularization-based fairness methods tend to degenerate toward ERM due to overfitting, per the authors' citations.
- Using only 10% of the training data to train the generative model causes little performance degradation.
Limitations
- Requires a reference model (e.g., CLIP) or human judgment to select the combination coefficients.
- Spurious attributes are limited to those detectable by the generative model's latent directions and the chosen target attribute; extension to general vision data is future work.
- Relies on a pretrained GAN whose quality bounds augmentation fidelity.
Experimental setup
Datasets: CelebA UTKFace Dogs and Cats
Benchmarks: CelebA UTKFace Dogs and Cats
Metrics: accuracy worst-group accuracy equalized odds
Baselines: ERM CVaR DRO EIIL LfF JTT MAPLE IRM Resampling BYOL FSCL
Builds on: BYOL InterFaceGAN JTT CLIP-based reference model
Closest papers by meaning
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.86)
- Do Counterfactually Fair Image Classifiers Satisfy Group Fairness? -- A Theoretical and Empirical Study (NeurIPS 2024 · similarity 0.84)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.84)
- AI-Face: A Million-Scale Demographically Annotated AI-Generated Face Dataset and Fairness Benchmark (CVPR 2025 · similarity 0.84)
- Attribute Based Interpretable Evaluation Metrics for Generative Models (ICML 2024 · similarity 0.84)
- AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.83)
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.