Protecting the Underrepresented: Data Debiasing via Asymmetric Adversarial Encoding for Fair Machine Learning

Abstract

Machine learning systems deployed in high-stakes domains often disadvantage underrepresented demographic groups because the training data itself encodes historical bias. Existing preprocessing methods typically transform all groups symmetrically or rebalance sample counts, without accounting for a basic asymmetry: the unprivileged group is often represented by fewer samples and may therefore be more vulnerable to additional distortion. We propose AdvEncoder, a training-time data debiasing framework based on asymmetric alignment: the latent reference space is learned from the unprivileged group, and the privileged group is adversarially aligned toward that reference. The goal is to reduce group-discriminative structure in the training data while minimizing additional distortion to the data-scarce group. AdvEncoder has two variants distinguished only by the discriminator's conditioning. AdvEncoder-SP uses an unconditional discriminator and targets marginal alignment related to statistical parity, while AdvEncoder-EO uses a label-conditional discriminator and targets class-conditional alignment related to equal opportunity and equalized odds. We provide formal implications under exact and approximate alignment, while emphasizing that these are alignment-level guarantees rather than optimization or deployment guarantees. AdvEncoder is used only at training time: it outputs a debiased training set in the original feature space, after which any downstream classifier can be trained normally and evaluated on ordinary, untransformed test or deployment inputs. Across four tabular benchmarks (UCI Adult, LSAC, ACSIncome, and German Credit) and five downstream classifier families, AdvEncoder provides a strong fairness-utility trade-off. It does not uniformly dominate every baseline on every individual metric; rather, it consistently achieves near-best fairness while retaining competitive accuracy and avoiding the severe utility loss or single-metric failure modes observed for several baselines. Ablations confirm that both adversarial alignment and the direction of asymmetric anchoring toward the unprivileged group are important to the method's effectiveness.