FairGAN

FairGAN introduces a fairness discriminator that competes with the generator alongside the standard discriminator. This auxiliary network attempts to predict the sensitive attribute (e.g., race, gender) from the generated data. The generator is penalized when the fairness discriminator succeeds, forcing it to produce representations where sensitive attributes are statistically independent from the generated features.

• The generator learns to fool both the quality discriminator and the fairness discriminator simultaneously
• This min-max game results in a fair representation that cannot be used to reconstruct protected attributes
• The architecture enforces demographic parity by ensuring equal probability of favorable outcomes across groups

FairGAN enforces statistical parity by ensuring that the distribution of generated outcomes is independent of sensitive attributes. The generator is trained to produce data where the probability of a positive prediction is equal across all protected groups.

• The fairness loss term penalizes deviations from P(ŷ=1|S=0) = P(ŷ=1|S=1)
• This constraint is implemented as a Lagrangian multiplier in the generator's objective function
• Statistical parity is particularly suitable for applications where historical discrimination must be actively corrected, such as loan approval or hiring datasets

Beyond statistical parity, FairGAN supports equalized odds constraints, which require that the false positive rate and true positive rate are equal across protected groups. This ensures fairness in both positive and negative prediction outcomes.

• The generator optimizes for P(ŷ=1|Y=1, S=0) = P(ŷ=1|Y=1, S=1) (equal true positive rate)
• Simultaneously enforces P(ŷ=1|Y=0, S=0) = P(ŷ=1|Y=0, S=1) (equal false positive rate)
• Equalized odds is preferred when the ground truth label Y is considered reliable and unbiased, as in clinical diagnostic datasets where disease prevalence may legitimately differ across populations

FairGAN's objective function combines the standard adversarial loss with a tunable fairness regularization term, allowing practitioners to balance data utility against fairness requirements. The hyperparameter λ controls the trade-off between generation fidelity and bias mitigation.

• The composite loss: L_total = L_GAN + λ * L_fairness
• When λ = 0, FairGAN reduces to a standard GAN with no fairness guarantees
• As λ increases, the generator prioritizes demographic balance over exact distribution matching
• This tunability enables compliance with varying regulatory requirements across jurisdictions, from the EU AI Act to HIPAA

A key property of FairGAN is that models trained on its synthetic data inherit the fairness properties without requiring modification to the downstream training pipeline. The debiasing is embedded in the data representation itself.

• A classifier trained on FairGAN-generated data exhibits lower disparate impact compared to one trained on original biased data
• This property is validated using the Train-Synthetic-Test-Real (TSTR) evaluation paradigm
• Fairness transfer eliminates the need for in-processing or post-processing debiasing techniques in production models
• The approach is model-agnostic: any downstream algorithm benefits from the fair representation

FairGAN natively supports intersectional fairness by conditioning on multiple sensitive attributes simultaneously. The fairness discriminator can be structured to enforce parity across compound subgroups (e.g., race × gender × age).

• The architecture uses multi-head fairness discriminators, each targeting a different sensitive attribute or intersection
• Prevents fairness gerrymandering, where a model appears fair on individual attributes but discriminates against intersectional subgroups
• Critical for healthcare applications where compounding biases affect diagnosis rates for minority women or elderly patients from specific ethnic backgrounds
• Supports both categorical and continuous sensitive attributes through appropriate discriminator architectures