Many classifiers now learn to say "I don't know." During training they are shown synthetic outliers, which are artificial examples placed just outside the region where real training data lives, so that at deployment they reject inputs unlike anything seen before. Current methods decide where "outside" begins by fitting simple shapes, such as Gaussians or principal subspaces, to each class's internal features. This proposal argues that the shape for a class is set mostly by the demographic majority within that class. Real members of a smaller demographic group then sit where that shape says the tail should be. When outliers are generated there, the model is in effect taught to treat legitimate minority inputs as unknown. The expected result is a detector that falsely rejects women, older applicants or darker-skinned faces more often than others, even though nobody asked it to make a fairness decision.
We propose to test this failure directly and to fix it at the level of the synthetic data rather than inside any particular detector. The fix has three parts. First, outlier geometry is estimated separately for each combination of class and demographic group. Second, any candidate outlier that lands inside another group's real support is discarded. Third, a small auditing loop, borrowed from the reductions approach to fair classification, adjusts how close to the data each group's outliers are placed until false-rejection rates are equal across groups. Because the intervention only changes which synthetic negatives the detector sees, it can be attached to existing outlier-exposure detectors without changing their losses. Protected attributes are needed only during training, never at test time.
Core Claim
The claim has two parts. The first is empirical and is the one most worth knowing. When the in-distribution data contains an under-represented demographic group, class-conditional synthetic outlier methods like VOS and Geometrically Constrained Outlier Synthesis (GCOS) should raise that group's in-distribution false-rejection rate more than the majority's. They should also do so more than a detector trained with no synthetic outliers at all, because a measurable fraction of their synthetic outliers land inside the minority group's real support.
The second part is methodological. Generating outliers per (class, group) cell, filtering them by support, and adjusting them with an auditor should remove most of this disparity while keeping out-of-distribution (OOD) detection quality close to that of the unconstrained method. Simpler fixes should not achieve the same result: reweighting real samples, per-group thresholds, or adding a parity penalty to the detector's loss.
The motivation comes from FAIROD. That work showed that density-based anomaly scoring confuses "statistical minority" with "anomaly" and over-flags protected groups. FAIROD studies autoencoder reconstruction scores on existing data and generates nothing. We ask whether the same confusion is actively manufactured by outlier synthesis, and whether it can be undone at the point of manufacture.
How this differs from the closest work
Four lines of work come near this question, and none addresses it. VMF-GOS is the closest in spirit. It shows that when tail classes are scarce, their boundaries blur and their real samples get rejected as OOD, and it repairs this with sampling on a hypersphere. It treats only label imbalance, however. Demographic groups inside a class are invisible to it, and it reports no fairness metric.
FAIROD supplies the fairness side. It contributes a statistical-parity loss on outlier scores and a within-group ranking ("group fidelity") loss that avoids the laziness trap, in which parity is met by flagging arbitrary minority members. Its constraint, though, lives inside one autoencoder's loss, which is exactly the model-specific coupling this proposal tries to avoid.
The antidote-data papers show that small, carefully generated synthetic sets can make unmodified learners fair. Fair Clustering Using Antidote Data does this for k-means and spectral clustering. Learning Antidote Data to Individual Unfairness does it for individual fairness with on-manifold tabular samples. Both place synthetic points inside the data. Outlier synthesis, by contrast, deliberately places them off-manifold within a narrow band. Bridging the two sides of that tension is the technical core here.
Finally, A Reductions Approach to Fair Classification supplies the auditing mechanism. In that work, an exponentiated-gradient player reweights groups and an unmodified learner responds. We repurpose it so that the auditor controls synthesis parameters rather than sample weights.
Research Program
Measurement. The first stage measures the failure before fixing it. We train standard VOS, GCOS and VMF-GOS detectors on data where one demographic group within each class is subsampled to a controlled fraction (50%, 20%, 10%, 5%). For each detector we record two quantities. The first is the group-wise in-distribution false-rejection rate at a threshold fixed so that 95% of all in-distribution validation points are accepted. The second is a mechanistic quantity: the fraction of generated outliers whose k-nearest-neighbor distance to real minority-group features is smaller than the typical within-group distance. If disparity grows as the minority fraction falls, and tracks this "encroachment fraction," the mechanism is confirmed. If it does not, the proposal is wrong, and we would report that as a negative result.
Per-cell geometry. The second stage builds the method. In the detector's penultimate feature space we fit a separate mean and covariance for every (class $y$, group $a$) cell. Small cells will have noisy covariance estimates, so each cell's covariance is shrunk toward its pooled class covariance with a data-driven weight.
Outliers are generated per cell following GCOS. We perturb features along the cell's lowest-variance principal directions inside a shell between two percentiles of that cell's own Mahalanobis distances, starting from the 95th to 99th band GCOS uses. Each candidate then passes a support check. It is kept only if it lies outside the high-density region of every other cell of the same class. This is the step that prevents a majority-group outlier from being placed on top of minority data.
Auditing loop. The auditor runs between training rounds. Let $\mathrm{FRR}_a$ be group $a$'s false-rejection rate on held-out in-distribution data and $\mathrm{FRR}$ the overall rate. We seek
$$|\mathrm{FRR}_a - \mathrm{FRR}| \le \epsilon \quad \text{for every group } a,$$
with $\epsilon$ a user-chosen tolerance. A multiplier per group is updated as $\lambda_a \leftarrow \lambda_a \exp\big(\eta(\mathrm{FRR}_a - \mathrm{FRR})\big)$, where $\eta$ is a step size. Each group's shell is then shifted outward in proportion to $\log \lambda_a$, which places its outliers farther from its real data, and its outlier budget is shrunk by the same factor.
The reductions approach yields a randomized mixture of models, and that is awkward in lending or hiring. We therefore deploy a single deterministic detector: the iterate with the best OOD performance among those meeting the constraint on validation data. We also report how far it falls from the mixture.
Ranking fidelity. Equalizing rejection rates could still produce FAIROD's laziness trap, so we monitor within-group ranking of OOD scores. Concretely, we compute the normalized discounted cumulative gain (NDCG) of each group's score ranking against the unconstrained detector's ranking on genuinely shifted inputs. We add FAIROD's group-fidelity term only if this ranking quality drops.
Portability. The method is meant to be detector-agnostic, which is the answer to the modularity problem of model-specific fairness constraints. The same generated outlier sets are therefore fed to three detectors with different training losses: an energy-based VOS head, GCOS's contrastive objective, and VMF-GOS's hyperspherical losses. The latter is run from the long-tail OOD codebases mala-lab/AdaptOD and mala-lab/COCL where the setup matches. Fairness metrics are computed with AI Fairness 360.
Benchmarks and metrics
We use three settings with increasing realism.
Colored MNIST. This is part of the GCOS near-OOD suite. We use color as a controllable "group" with a skewed color frequency within each digit. This gives exact control of the minority fraction and of how separated groups are in feature space.
CelebA. We use the "Heavy Makeup" task with Gender as the sensitive attribute, as in Navigating Towards Fairness with Data Selection. Out-of-distribution sets are Textures, SVHN and Places365, the far-OOD sets used in the CIFAR evaluations of DREAM-OOD. LFW faces serve as a covariate-shifted in-distribution set, which must not be rejected and gives a second test of false rejections.
Adult Census. Race and sex are the groups. Every clean test record is in-distribution, so any flag is a false rejection. This mirrors FAIROD's flag-rate parity evaluation. Real OOD inputs come from records with injected feature corruptions.
Metrics in all settings are OOD AUROC and FPR95, per-group in-distribution false-rejection rates with their maximum gap and ratio, per-group classification accuracy, within-group NDCG, and the encroachment fraction.
Baselines
Do nothing. The naive reference point is a detector trained with cross-entropy only and scored by maximum softmax or energy. It shows how much disparity exists before any synthesis. It anchors the claim that synthesis adds disparity, rather than merely failing to remove disparity that was already present.
Strongest existing synthesis method. This is GCOS, which reports an average AUROC of 93.47% across its four near-OOD datasets, against 91.97% for NCIS and 85.76% for DREAM-OOD. It also reports 0.00% FPR95 on Stanford Dogs, against 5.00% for VOS and 80.00% for an unregularized model (Geometrically Constrained Outlier Synthesis). For long-tailed data the strongest is VMF-GOS, with 77.05% AUROC on CIFAR100-LT (VMF-GOS). We found no reported result on Lacuna for group-wise false-rejection rates of any synthesis-based OOD detector, so the fairness half of this comparison has no prior number and will be established here.
We run GCOS and VMF-GOS with their published shell percentiles and losses, on identical backbones, data splits and epochs. Each method, including ours, gets the same hyperparameter search budget on the same validation split. Beating them means shrinking the maximum group false-rejection gap by at least half at the 10% and 5% minority settings, while losing no more than one AUROC point on OOD detection.
Strongest model-specific fairness competitor. This is a FAIROD-style regularizer: a statistical-parity penalty on OOD scores plus a group-fidelity ranking term, added directly to GCOS's loss (FAIROD). FAIROD reports a 9x statistical-parity improvement over its base autoencoder and 5x over the nearest fairness-aware baseline on Adult. Those figures are relative, and no absolute flag rates are given in the file. We tune its penalty weights with the same budget. Matching its fairness while transferring to the other two detectors without re-tuning their losses counts as a win on modularity. Beating its OOD AUROC at equal fairness counts as a win on accuracy.
Ablations. Three ablations isolate the key ingredient. Per-cell geometry without the support filter tests whether separate estimation alone suffices. Per-cell geometry with the filter but without the auditor, using fixed 95th to 99th percentile shells, tests whether the adaptive loop is necessary. Class-level geometry with only the support filter tests whether filtering alone suffices. If the auditor-free variant matches the full method, we drop the auditor and report the simpler method.
Why not just…?
Why not just set per-group rejection thresholds after training? Post-processing in the style of Optimized Score Transformation for Consistent Fair Classification can equalize false-rejection rates exactly, so it is a real threat whenever protected attributes are available at deployment. It has two problems. It requires the attribute at test time, which is often prohibited in lending and hiring, and our method avoids that. It also cannot repair a feature space in which minority inputs and genuine outliers have been pushed together. Per-group thresholds equalize how many minority points are rejected, but not which ones, so the OOD AUROC computed within the minority group should stay low.
We run per-group thresholding on the unconstrained GCOS detector as an attribute-aware oracle and compare within-group OOD AUROC directly. If thresholding matches our within-group AUROC, our method's only advantage is test-time attribute-freedom, and we will say so.
Why not just reweight or oversample minority-group samples while keeping standard synthesis? This is the cheapest serious alternative, and it is the one most likely to work. Upweighting minority features when fitting the class Gaussian or principal subspace widens that shape around the minority group, which could pull outliers off its support. The reason to doubt that it suffices is geometric. When two groups form separate clusters within a class, a single reweighted Gaussian places much of its tail mass in the gap between them and along directions set by the pooled covariance. The encroachment fraction tells us directly whether that happens.
Optimising Fairness Through Parametrised Data Sampling supplies a principled resampling sweep that we use as this baseline. The reductions paper's finding that heuristic reweighting is dominated by optimized trade-offs (A Reductions Approach to Fair Classification) concerns in-distribution classification. It is suggestive here, not conclusive. If reweighted GCOS closes the gap on Colored MNIST and CelebA, we will reframe the contribution as the diagnostic plus that simple fix.
Why not just add a parity penalty to the detector's loss? This is the FAIROD-style baseline above, and it may well match us on any single detector. Its costs are the modularity crisis this research direction exists to escape. The penalty must be re-derived and re-tuned for each loss: energy, contrastive, or hyperspherical. It also invites the laziness trap unless a ranking term is added. The portability experiment, in which outlier sets are generated once and given to three detectors, together with the NDCG measurement, decides whether moving the constraint into the data buys anything beyond convenience.
Why not just generate outliers from random noise, far from all groups? Outliers far from every group cannot encroach on anyone, but they also do not teach the detector its boundary. GCOS reports that Gaussian VOS outliers leave a 5.00% FPR95 on Stanford Dogs where geometric synthesis reaches 0.00%, and DREAM-OOD makes the same point for distant samples. We include noise-only exposure to show that the fairness gain does not come from simply giving up near-boundary supervision.
Decisive Checks
The proposal rests on three results.
The first is whether the encroachment mechanism exists. On Colored MNIST, the group false-rejection gap of standard GCOS must increase as the minority fraction falls from 50% to 5%, must exceed the gap of the do-nothing detector, and must correlate with the encroachment fraction. If any of these fails, the problem is not created by synthesis, and the method is unnecessary.
The second is whether reweighted standard synthesis falls short. On CelebA, the full method must reduce the gender false-rejection gap meaningfully more than reweighted GCOS, at matched OOD AUROC and FPR95 on Textures, SVHN and Places365, and without rejecting more LFW faces.
The third is portability. On Adult, outlier sets generated once must cut the race and sex flag-rate gaps for all three detector families without per-detector fairness tuning. They must also keep within-group NDCG near the unconstrained detector's, so that fairness is not bought through the laziness trap.
Together these checks determine whether fair outlier exposure is a genuine new tool or whether simpler data-level fixes already suffice.
References
- VOS: Learning What You Don't Know by Virtual Outlier Synthesis
- Geometrically Constrained Outlier Synthesis
- FairOD: Fairness-aware Outlier Detection
- VMF-GOS: Geometry-guided virtual Outlier Synthesis for Long-Tailed OOD Detection
- Fair Clustering Using Antidote Data
- Learning Antidote Data to Individual Unfairness
- A Reductions Approach to Fair Classification
- AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias
- Navigating Towards Fairness with Data Selection
- Dream the Impossible: Outlier Imagination with Diffusion Models
- Optimized Score Transformation for Consistent Fair Classification
- Optimising Fairness Through Parametrised Data Sampling