Location-Smooth Adversaries for Fairness Without Demographics

Many deployed models are forbidden from seeing who a person is (their income, ethnicity, or household circumstances) but are routinely told where a data point came from: the country of a photo, the coordinates of a satellite tile, the survey cluster behind a poverty estimate. Methods for "fairness without demographics" search for the people a model serves badly by letting an adversary upweight training examples where the model's loss is high. These methods ignore location, or at most treat it as one more input column. Geographic-bias research shows the underserved groups are strongly organized by location. Image models do worse on photos from Africa than on photos from Europe, and much worse in low-income homes. That research, however, answers with costly hand-collected data rather than training algorithms. We propose to join the two by restricting the reweighting adversary so that the location-dependent part of its weights must vary smoothly over space, at several spatial scales at once.

The bet is that nearby places tend to share hidden disadvantages. Statisticians call this spatial autocorrelation: near things are more alike than far things. If that holds, a smooth map of weights over location is a cheap and well-founded way to capture the unobserved sensitive attribute. It is also far less able than a free-form adversary to latch onto isolated noisy points. Summing over several spatial scales, instead of committing to one set of borders, sidesteps the arbitrary choice between countries, regions, or grid cells. In short, the adversary is allowed to say "this area, at this scale, is being failed". It is not allowed to say "these 40 scattered images are hard".

Core Claim

Adding a multi-scale, spatially smooth component to the adversary improves the worst-off hidden group more than three competitors on geographically skewed vision and remote-sensing tasks. The competitors are demographic-free reweighting that ignores location, reweighting that gets location as an unconstrained input, and Group DRO over a single fixed partition of space. Group DRO (group distributionally robust optimization) minimizes the loss of the worst predefined group. The hidden groups we score are income quartile on Dollar Street, region on GeoDE, and region or urban/rural stratum on the satellite tasks, and none of them is given to the learner. The claim is falsifiable in a specific way. If location-as-a-free-feature does as well as the smoothness-constrained version, the constraint is doing nothing and the idea fails.

Relation to Prior Work

Three demographic-free training methods supply the machinery we build on. Adversarially Reweighted Learning (ARL) plays a learner against an adversary that assigns weights as a function of features and labels. It showed that restricting the adversary to learnable functions ("computational identifiability") stops it from chasing outliers. It reported a 6.5-point worst-group AUC gain on UCI Adult but no gain on COMPAS, where the labels are noisy. Graph of Gradients (GoG) gets the same outlier resistance by smoothing the adversary over a nearest-neighbor graph of per-example gradients. Blind Pareto Fairness (BPF) protects any unknown subgroup above a minimum size and avoids needlessly hurting everyone else. Our proposal keeps ARL's core insight that restricting the adversary is what makes it useful. It then changes the restriction: smoothness over physical space, which these papers never use, in place of or alongside smoothness over features or gradients. A fourth method, RAD, finds minority examples with a deliberately biased linear probe. Its evidence that annotation-reliant methods collapse under 5–20% label noise is the main reason we do not simply run Group DRO on country codes.

Two spatial-fairness papers motivate and constrain the design. Evaluating Fairness in the Presence of Spatial Autocorrelation uses Moran's I, a standard measure of spatial clustering, and eigenvector spatial filtering to show that clustering can manufacture apparent disparities. It needs full demographic labels and only audits; it does not train. Spatial Fairness: The Case for its Importance argues that location proxies for protected traits. It names the modifiable areal unit problem (conclusions change with how borders are drawn) and recommends robustness across scales, but offers no algorithm. We turn that recommendation into an objective. The first paper's warning also becomes our main risk, discussed below: a smooth adversary may chase spatially clustered noise rather than harm. CELL is the closest paper to use region as a sensitive attribute, on the Pokec-n social graph. It requires labeled source sensitive attributes, a graph with homophilous edges (links mostly between similar nodes), and a binary region split. We need no sensitive labels anywhere, no graph, and we treat space as continuous.

The benchmarks come from geographic-bias work. GeoDE reports models scoring 7% worse on Africa than Europe and fixes this by commissioning images at about $0.87 each. Dollar Street reports 18.0% versus 57.4% top-5 accuracy between the lowest and highest income quartiles for pretrained models. WILDS supplies FMoW and PovertyMap, and reports that invariance methods rarely beat plain training.

Method

Let $s_i$ be the known location of example $i$: country centroid, cluster coordinates, or tile coordinates. Let $\ell_i(\theta)$ be the learner's loss. The adversary's weight is

$$\lambda_i = 1 + g_\phi(x_i, y_i) + \sum_{b=1}^{B} \sum_{j} \alpha_{bj}\, K_{h_b}(s_i, c_j),$$

where $g_\phi$ is an ARL-style feature adversary, $K_{h_b}$ is a Gaussian kernel of bandwidth $h_b$, $c_j$ are anchor locations, and $\alpha_{bj} \ge 0$ are the adversary's spatial coefficients. The bandwidths $h_b$ span city to continent scale, and the anchors are placed on a coarse global grid. Weights are normalized to mean one. The learner minimizes $\frac{1}{n}\sum_i \lambda_i \ell_i(\theta)$ and the adversary maximizes it, subject to a total budget on $\alpha$. That budget limits how much weight can sit in any one place, playing the role of BPF's minimum group size. Because each kernel covers a region, the adversary can only raise weight over whole neighborhoods. The mixture over bandwidths lets it choose the scale instead of the analyst. Distances are great-circle distances.

The research program has three parts. The first is to build the objective above as a drop-in layer on standard fine-tuning. The second is to analyze what the adversary learns: map $\alpha$ over the globe and relate it to held-out income and region labels, which are used for scoring only. The third is to add a noise guard in the spirit of GoG's robustness to label noise. The guard compares each kernel's weighted loss against a held-out estimate of loss in the same area. It damps the spatial weight where high loss is not reproducible, which is the signature of a clustered sensor or label artifact and not of a group the model systematically fails.

Baselines

The naive reference point is plain empirical risk minimization with location withheld: standard fine-tuning of the same backbone on the same skewed training split. Every method gets the same backbone, the same training data, the same number of epochs, and the same hyperparameter search budget. That budget is a fixed number of trials, tuned on a validation split that is scored on average accuracy only, because tuning on hidden group labels would leak the sensitive attribute.

The strongest demographic-free competitors are ARL, GoG and RAD-UW, each run without location. RAD's reported figure is about 83.5% worst-group accuracy on CelebA, robust to annotation noise. GoG's is 20.46% worst-group accuracy versus 19.32% for its baseline on MIMIC-III. Neither is on our benchmarks, and no reported result was found on Lacuna for any of these three methods on GeoDE, Dollar Street, FMoW or PovertyMap, so we run them ourselves. They are strongest on the "no demographics" axis: if hidden groups are already findable from image features or gradients, they will find them without any use of location.

The strongest location-using competitor is Group DRO over a fixed geographic partition, run separately at region, country, and roughly 500 km grid-cell granularity. It is the method WILDS found most effective, raising CivilComments worst-group accuracy from 56.0% to 70.0% with true group labels. We report its best granularity chosen with oracle knowledge, which is a deliberately generous comparison.

Two ablations isolate the key ingredient. The first gives the ARL adversary raw coordinates as extra inputs with no smoothness constraint. The second uses our spatial term at a single bandwidth only.

Beating the field means a statistically reliable gain in worst hidden-group accuracy over the best of ARL, GoG and RAD across five seeds, with average accuracy at most one point below plain training. Two further conditions must hold. The method must match or beat single-scale Group DRO at its oracle granularity on the hidden attribute that does not align with that partition, namely income on Dollar Street. And it must beat the coordinates-as-features ablation. Reference numbers from the literature are the 7% Africa–Europe gap on GeoDE, the 18.0% versus 57.4% income-quartile top-5 gap on Dollar Street for pretrained models, and PovertyMap's out-of-distribution Pearson correlation of 0.45 for ERM, against 0.60 in-distribution. No reported result was found on Lacuna for FMoW-WILDS worst-region accuracy.

Why not just…?

Why not just run Group DRO on country or region labels? This is the real threat, because location labels are available in our setting. Two things argue against it. First, RAD shows annotation-reliant reweighting falls behind annotation-free methods at 5% annotation noise and drops from 86% to below 79% at 20% noise on CelebA, and geographic metadata is noisy. Second, the attribute we care about often cuts across any single partition: Dollar Street homes within one country span a wide income range. The oracle-granularity Group DRO baseline and the injected location-noise experiment below settle this directly. If Group DRO wins on Dollar Street income, the claim is wrong.

Why not just feed coordinates to ARL? ARL's own result is that the adversary's usefulness comes from restricting it. An unconstrained adversary given coordinates can memorize individual hard locations, which is standard DRO's outlier failure in a new place. The unconstrained-coordinates ablation tests exactly this. If it matches our method, we drop the smoothness claim.

Why not just collect balanced data, as GeoDE did? It works: GeoDE reports a 21% gain on its own test set and 9% on Dollar Street from augmenting ImageNet with its images. But it cost about $54,000 for 61,940 images and still filters out households without smartphones. Our method is complementary. We will also train on GeoDE-augmented data and ask whether the spatial adversary adds gains on top, particularly for the lowest Dollar Street income quartile that curated collections under-cover.

Proposed Experiments

The main vision experiment builds a controlled skew. We pool the GeoDE and Dollar Street training images, keep only 5% of the samples from Africa, Southeast Asia and the lowest two income quartiles, then fine-tune an ImageNet-pretrained ResNet and a CLIP image encoder. Evaluation is on balanced held-out splits. The metrics are worst-region accuracy on GeoDE and worst-income-quartile top-5 accuracy on Dollar Street, with the published 18.0% and 57.4% quartile figures as the pretrained reference and the roughly 75% supervised fine-tuning figure as a ceiling. The satellite experiments use FMoW-WILDS (worst-region accuracy) and PovertyMap-WILDS (worst urban/rural Pearson correlation on unseen countries, against the 0.45 ERM reference). Both use the official splits and location metadata only as the spatial input $s_i$; region labels are used only for scoring.

Three decisive checks determine whether the idea holds. The first is the income check on Dollar Street. Income is not recoverable from country alone, so gains there must come from the smooth spatial term combining with the feature adversary, not from simply relearning the region split. The second is a spatially clustered label-noise test. We corrupt 20% of labels inside a few randomly placed disks on FMoW. The spatial adversary without the noise guard should upweight those disks and lose accuracy, which is the confound Evaluating Fairness in the Presence of Spatial Autocorrelation warns about, and the noise guard should recover most of that loss. If it does not, the method is unsafe for sensor-noise-heavy domains such as the unvalidated satellite CO2 regions described in the XCO2 study. The third is the scale-robustness check. We score worst-group accuracy under region, country and grid partitions none of which were used in training. The multi-scale method should vary least across partitions, which is the concrete meaning of robustness to border choice.