Robust Fairness

A foundational mathematical framework for robust fairness that optimizes model performance against an adversarially chosen distribution within a defined uncertainty set around the empirical training distribution.

• Minimizes worst-case expected loss over a KL-divergence ball or Wasserstein ball
• Ensures fairness metrics like equalized odds hold even when the test distribution shifts
• Protects against subpopulation shift, where minority group proportions change at deployment
• Formulated as a min-max optimization: the model minimizes loss while an adversary maximizes it within constraints

Integrates adversarial training principles with fairness constraints to produce models whose equitable behavior resists intentional input perturbations designed to trigger discriminatory outputs.

• Defends against fairness-aware adversarial attacks that craft inputs to exploit model bias
• Uses projected gradient descent (PGD) to generate perturbed examples during training
• Ensures that a small change in non-sensitive features cannot flip a decision differentially across groups
• Critical for high-stakes domains like credit lending and criminal justice where adversaries may probe model boundaries

Addresses the problem of hidden stratification, where a model appears fair on aggregate demographic groups but performs inequitably on finer-grained subpopulations that were underrepresented or unobserved during training.

• Identifies and mitigates worst-case subgroup performance gaps using slice-based evaluation
• Employs distributionally robust importance weighting to upweight high-loss subgroups
• Prevents fairness gerrymandering, where coarse group definitions mask severe within-group disparities
• Example: A medical diagnosis model with equal accuracy across genders may still fail on older women of a specific ethnicity

Combines causal inference with robust optimization to ensure fairness guarantees hold under interventions and structural changes in the data-generating process, not just observational distributions.

• Models fairness using a structural causal model (SCM) rather than purely statistical associations
• Guarantees counterfactual fairness remains stable under shifts in the causal mechanisms
• Protects against spurious correlations that may appear in training data but vanish under distribution shift
• Uses invariant risk minimization (IRM) to learn predictors that are stable across multiple environments

A specific robustness scenario where the input feature distribution P(X) changes between training and deployment, but the conditional outcome distribution P(Y|X) remains stable.

• Applies importance weighting using the density ratio between source and target distributions
• Re-weights fairness constraints to reflect the target population rather than the training population
• Critical for models deployed across geographic regions with different demographic compositions
• Uses kernel mean matching or discriminative density ratio estimation to compute weights without explicit distribution modeling

Provides formal, verifiable bounds on fairness metric degradation under bounded input perturbations, analogous to certified adversarial robustness in classification accuracy.

• Uses interval bound propagation or linear relaxation-based perturbation analysis (LiRPA) to compute certified fairness radii
• Outputs a provable guarantee: for any input within an ℓ∞-ball of radius ε, the fairness metric difference between groups will not exceed a computed threshold
• Enables regulatory compliance by providing auditable, mathematical fairness warranties
• Contrasts with empirical evaluation, which can only test a finite set of perturbations

A causal definition of fairness where a decision is considered fair if it would remain the same in a counterfactual world where an individual's sensitive attributes were different. This approach uses structural causal models to simulate interventions, directly aligning with robust fairness by ensuring predictions are invariant to perturbations of protected attributes.

An in-processing technique that trains a model to simultaneously predict a target variable while an adversarial network attempts to predict the protected attribute from the model's representations. This creates a minimax game that maximizes accuracy while minimizing the leakage of sensitive information, a practical method for learning fair representations that are robust to an adversary.

The core problem robust fairness addresses. Standard fairness metrics often measure average performance, but a model can fail catastrophically on a minority subgroup. Techniques like distributionally robust optimization (DRO) explicitly minimize the maximum loss over a set of pre-defined subpopulations, guaranteeing a baseline of equitable performance even under the most adverse data conditions.

A pre-processing approach that learns a latent data representation that encodes useful information for prediction while obfuscating or removing information about sensitive attributes. By transforming the input data into a fairness-aware embedding space, the downstream model is inherently debiased, providing a robust foundation that is agnostic to the specific classifier used.

A fairness criterion requiring a classifier to have equal true positive and false positive rates across different protected groups. This ensures errors are evenly distributed, preventing a model from being highly accurate for a majority group while disproportionately misclassifying a minority group. It is a stricter metric than demographic parity and is often a target for robust optimization.