Research map / Learning theory and algorithms / Fairness
Adversarial Fairness Network
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 FAIAS, an adversarial selector-predictor network that samples sensitive-irrelevant input features and minimizes the marginal contribution of the sensitive feature to improve group fairness without hurting accuracy.
Problem
Machine learning models amplify social bias by making group-dependent predictions tied to sensitive features such as race or gender, while fairness constraints often degrade predictive performance.
Approach
Jointly learn a probabilistic feature selector (trained adversarially to maximize a Jensen-Shannon sensitivity loss) that removes sensitive-relevant input features in the original feature space, and a single predictor trained to minimize the marginal effect of adding the sensitive feature, using generalized cross entropy as the classification loss.
Key results
- On Adult and COMPAS, FAIAS achieves the best accuracy-fairness trade-off among compared methods, reaching the lowest equalized odds difference at similar accuracy.
- On CelebA attractiveness classification with VGG19 features, FAIAS reduces fairness violation by almost 50% relative to the baseline while improving accuracy.
- Combining the FAIAS selector with AdvDeb improves over naive LfF+AdvDeb combination, and on COMPAS improves both fairness and accuracy over AdvDeb alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FAIAS fair feature selection and predictor instead of the full-feature baseline | FAIAS · Adult (Sex) | equalized odds difference | Helped |
| Use FAIAS fair feature selection and predictor instead of the full-feature baseline | FAIAS · COMPAS (Race) | equalized odds difference | Helped |
| Use FAIAS fair feature selection and predictor instead of the full-feature baseline | FAIAS · Adult and COMPAS | accuracy | No clear effect |
| Use FAIAS fair feature selection and predictor instead of the full-feature baseline | FAIAS · CelebA attractiveness classification | equalized odds difference | Helped |
| Use FAIAS fair feature selection and predictor instead of the full-feature baseline | FAIAS · CelebA attractiveness classification | accuracy | Helped |
| Add the FAIAS selector to AdvDeb (ablation) | AdvDeb · COMPAS (Race) | accuracy | Helped |
| Add the FAIAS selector to AdvDeb (ablation) | AdvDeb · Adult and COMPAS | equalized odds difference | Helped |
| Combine LfF with AdvDeb (ablation) | AdvDeb · Adult and COMPAS | accuracy | Hurt |
| Replace cross-entropy classification loss with generalized cross-entropy loss (ablation) | FAIAS predictor · CelebA attractiveness classification | accuracy | Helped |
| Replace cross-entropy classification loss with generalized cross-entropy loss (ablation) | FAIAS predictor · CelebA attractiveness classification | equalized odds difference | Helped |
Surprising findings
- Generalized cross entropy, usually used to train biased models, was found to help train a fair model within the FAIAS framework.
- Naively stacking two state-of-the-art fairness methods (LfF + AdvDeb) significantly harms accuracy despite improved fairness.
Limitations
- Proposed only for supervised learning; unsupervised extension is future work.
- Removing sensitive-relevant features sometimes harms accuracy because they can also be target-relevant.
- Requires a single binary sensitive attribute in the presented formulation (multi-attribute extension is only described informally).
Experimental setup
Datasets: Adult COMPAS CelebA
Benchmarks: Adult COMPAS CelebA
Metrics: equalized odds difference accuracy
Baselines: AdvDeb CEOP LAFTR LfF DIR Baseline (unconstrained predictor)
Builds on: AdvDeb LfF GSTAR generalized cross entropy
Closest papers by meaning
- On the Fairness ROAD: Robust Optimization for Adversarial Debiasing (ICLR 2024 · similarity 0.90)
- On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness (ICML 2025 · similarity 0.89)
- Can Private Machine Learning Be Fair? (AAAI 2025 · similarity 0.88)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.87)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.87)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 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.