Research map / Learning theory and algorithms / Fairness
On the Fairness ROAD: Robust Optimization for Adversarial Debiasing
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 ROAD, a DRO-reweighted adversarial debiasing method that upweights instances where a sensitive-attribute adversary is most successful, to enforce local fairness in unknown subgroups.
Problem
Global group fairness criteria (Demographic Parity, Equalized Odds) can leave persistent local disparities within subregions of the feature space, and predefined comparable subgroups are hard to specify; fairness also degrades under distribution shift.
Approach
Combine distributionally robust optimization with fair adversarial learning: an instance-level weight r(x,s), constrained by KL-ball regularization and conditional (per-sensitive-group) normalization, upweights samples where the adversary reconstructs S most easily. Two implementations: non-parametric BROAD (Boltzmann weights) and parametric ROAD (neural network weights for smoothness).
Key results
- On Adult, Compas, Law and German datasets, ROAD mostly dominates baselines on the accuracy vs worst-1-DI Pareto front under global DI < 0.05.
- On Folktables 2014/2015 drift tests, ROAD achieves significantly better accuracy-EO tradeoffs than other methods including DRO-based fairness approaches, while CUMA slightly leads in-distribution.
- Ablation on Compas shows subgroups with higher local DI receive higher average r weights, supporting the adversary-focus assumption.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ROAD instead of competing methods under the global DI constraint | Fair classification · Compas, Law, German Credit | Worst-1-DI and global accuracy | Helped |
| Use the parametric ROAD implementation instead of BROAD (ablation) | Fair classification · Compas, Law, German Credit | Worst-1-DI and global accuracy | Helped |
| Use ROAD instead of competing fairness methods | Fairness generalization on Adult · Folktables 2014 | Accuracy and Equalized Odds | Helped |
| Use ROAD instead of competing fairness methods | Fairness generalization on Adult · Folktables 2015 | Accuracy and Equalized Odds | Helped |
| Increase temperature τ from zero (ablation) | ROAD · Compas | Worst-1-DI | Mixed |
| Use conditional normalization for each sensitive group instead of single global normalization (ablation) | ROAD · Law | Local fairness | Helped |
| Increase the complexity of adversary g (ablation) | ROAD · Law | Fairness result quality | Helped |
| Increase the complexity of weight network r (ablation) | ROAD · Law | Fairness result quality | Hurt |
| Vary the subgroup definition used to evaluate local fairness (ablation) | ROAD · Law | Worst local DI | Mixed |
Surprising findings
- Globally fair adversarial baselines still slightly reduce local bias through their global mechanisms.
- A global (non-conditional) normalization of r concentrates fairness effort on the majority sensitive group, motivating the conditional normalization.
Limitations
- Requires access to the sensitive attribute at training time.
- Performance depends on the temperature tau and network complexity of the weighting network; infinite-capacity networks reduce to BROAD's pessimistic solutions.
- Local fairness is evaluated only on arbitrarily chosen subgroups defined from features.
Experimental setup
Datasets: Compas Law School German Credit Adult Folktables 2014 Folktables 2015
Benchmarks: Compas Law School German Credit Adult Folktables 2014 Folktables 2015
Metrics: accuracy worst-1-DI (local disparate impact) global disparate impact Equalized Odds
Baselines: adversarial debiasing (Zhang et al.) Adversarial fairness with EO adversary FairLR Robust FairCORELS CUMA BROAD
Builds on: adversarial debiasing (Zhang et al. 2018) distributionally robust optimization Adversarial fairness with KL-regularized DRO (Hu et al.) Conditional Demographic Disparity
Closest papers by meaning
- On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness (ICML 2025 · similarity 0.91)
- DAFA: Distance-Aware Fair Adversarial Training (ICLR 2024 · similarity 0.91)
- Adversarial Fairness Network (AAAI 2024 · similarity 0.90)
- Towards Fairness-Aware Adversarial Learning (CVPR 2024 · similarity 0.89)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.88)
- Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference (NeurIPS 2025 · similarity 0.88)
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.