Benchmarking Algorithmic Fairness Methods Under Violations of Sensitive Attribute Preconditions

The problem

Existing algorithmic fairness methods are almost universally evaluated under rigid preconditions: fully observed or missing-at-random discrete group labels, bounded minority group representations ($\rho$-fraction), and absence of feature-dependent annotation noise. In real-world deployments, sensitive attributes are often unannotated, continuous, severely underrepresented, or corrupted by feature-dependent noise. Because fairness methods have not been systematically evaluated across these non-idealized regimes, practitioners cannot determine which, if any, algorithmic debiasing approaches retain their fairness guarantees or utility tradeoffs when theoretical preconditions fail.

Why it matters

Provides the first standardized, empirical map of failure modes for fair learning methods under realistic sensitive attribute limitations, enabling practitioners to select robust methods and establishing rigorous baselines for future algorithmic development.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

1. 1

   Systematic empirical benchmark: Construct a standardized evaluation suite perturbing sensitive attribute regimes (continuous distributions, non-random missingness, low minority frequency below standard $\rho$ thresholds, and feature-dependent noise) across tabular, graph, and image datasets, measuring Pareto frontier shifts in demographic parity, equalized odds, and task accuracy across standard in-processing and post-processing fairness methods.

   Check for prior work
2. 2

   Robustness stress-testing of differentially private fair algorithms: Evaluate privacy-preserving fair learners under varying degrees of extreme demographic imbalance and unannotated proxy attributes, measuring empirical privacy-fairness-utility degradation curves.

   Check for prior work

Why it might fail

If concurrent large-scale empirical benchmarks release exhaustive stress-tests covering these exact missingness, continuous-attribute, and sparsity configurations before the evaluation pipeline is established.

Evidence

Each paper's own statement of the limitation, verbatim.

• Optimal Fair Aggregation of Crowdsourced Noisy Labels using Demographic Parity ConstraintsICML 2026

  Does not model feature-dependent annotator confusion matrices or interactions between non-sensitive input features and sensitive attributes.

• FairGE: Fairness-Aware Graph Encoding in Incomplete Social NetworksTheWebConference 2026

  Evaluated exclusively on node classification under synthetic uniform missing-at-random sensitive attribute distributions.

• Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMsICLR 2026

  Requires explicit, discrete demographic or group annotations to evaluate subgroup metrics and cannot handle continuous or unannotated sensitive attributes.

• Stochastic Differentially Private and Fair LearningICLR 2023

  Requires every sensitive attribute to appear in at least a ρ-fraction of the dataset; the authors note performance of fair DP learning without sample diversity is unknown.

Nearest existing work

• Fairness without Demographics through Adversarially Reweighted LearningNeurIPS 2020
• Differentially Private Fair LearningICML 2019
• On the Inductive Biases of Demographic Parity-based Fair Learning AlgorithmsUAI 2024
• Estimating and Controlling for Equalized Odds via Sensitive Attribute PredictorsNeurIPS 2023
• FFB: A Fair Fairness Benchmark for In-Processing Group Fairness MethodsICLR 2024
• Blind Pareto Fairness and Subgroup RobustnessICML 2021
• Stochastic Differentially Private and Fair LearningICLR 2023
• On the Fairness ROAD: Robust Optimization for Adversarial DebiasingICLR 2024
• Blind Justice: Fairness with Encrypted Sensitive AttributesICML 2018
• Fair Learning with Private Demographic DataICML 2020
• FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRANeurIPS 2025
• Self-Supervised Fair Representation Learning without DemographicsNeurIPS 2022
• Fair Classification with Noisy Protected Attributes: A Framework with Provable GuaranteesICML 2021
• Learning Fair RepresentationsICML 2013
• Fairness Risks for Group-Conditionally Missing DemographicsAISTATS 2025

Related open problems