Standard approaches to algorithmic fairness operate under a comfortable assumption: that the data used to train and evaluate models contains explicit labels for protected attributes, such as race, gender, or age. With these labels in hand, engineers can directly measure disparities and apply constraints to equalize error rates across groups.
In practice, this assumption frequently falls apart. Legal frameworks like the GDPR, corporate privacy policies, and the sheer logistics of data collection mean that sensitive demographic attributes are often missing, noisy, or entirely unavailable. This creates a fundamental tension: how can an organization audit or enforce fairness when it is legally or ethically prohibited from knowing who belongs to which group?
To resolve this paradox, a growing body of research focuses on mitigating bias without explicit demographic labels. By leveraging internal model signals, robust optimization techniques, and probabilistic proxies, these methods make it possible to protect vulnerable subgroups while preserving user privacy.
The Intuition: Loss as a Proxy for Disadvantage
If a machine learning model is trained to minimize average loss across a population, it naturally becomes a "crowd-pleaser." It optimizes for the majority demographic because doing so yields the largest drop in average error. Consequently, underrepresented or systematically disadvantaged subgroups are left with significantly higher error rates.
This systematic failure mode provides a clever workaround. Even if we cannot see a subgroup's demographic label, we can see its performance footprint: disadvantaged groups manifest as clusters of high loss.
Standard ERM:
[ Low Loss Samples (Majority) ] ---> Model ignores minority nuances
[ High Loss Samples (Minority) ]
Adversarial Reweighting:
Adversary locates high-loss clusters ---> Upweights them ---> Forces model to learn fair representations
Preethi Lahoti operationalized this intuition through a framework called Adversarially Reweighted Learning (ARL). Instead of relying on demographic labels, ARL introduces an adversarial network that works in tandem with the primary learner. The adversary's goal is to maximize the model's weighted loss by assigning higher importance weights to specific samples. Because the adversary wants to make the model look as bad as possible, it naturally hunts for pockets of data where the model systematically fails. The primary model is then forced to minimize this worst-case weighted loss, effectively improving performance for the underserved subgroups without ever knowing their identities.
A similar philosophy underpins the Blind Pareto Fairness (BPF) framework proposed by Martin Andres Bertran et al. By combining minimax optimization with Pareto efficiency, the authors protect the most vulnerable segments of a population by preventing any single group from bearing a disproportionate share of the error, ensuring that optimization gains are shared equitably.
Taking this a step further, Jing Li 0009 et al. introduced an intrinsic reweighting method that looks at individual sample gradients. By tracking how much each data point contributes to the model's updates, their approach dynamically balances the training process to protect the worst-case $\alpha$-sized subsets of the population, providing robustness against both bias and outliers.
Soft Proxies and the Danger of Hard Imputation
When explicit labels are missing, a common industry fallback is to use an auxiliary "proxy model" to guess the missing attributes—for example, predicting a user's ethnicity based on their name or geography. However, treating these guesses as ground truth introduces severe mathematical distortions.
As demonstrated by Nathan Kallus et al., using "hard" imputation (assigning a definitive, discrete group label based on a proxy's highest-probability guess) systematically exaggerates or misstates disparities. Because proxy models have their own error distributions, hard imputation compounds measurement errors and can lead organizations to optimize for a completely distorted version of fairness. To fix this, the authors proposed a "soft" weighted estimator that treats group membership as a probability distribution rather than a certainty, recovering accurate fairness assessments from imperfect data.
Building on this concept, Zhaowei Zhu et al. proved that you do not need a single, highly accurate proxy model to audit fairness. Instead, their framework uses multiple weak proxies—models that are individually inaccurate or noisy. By applying a novel calibration algorithm across these weak signals, the authors demonstrated that the true underlying fairness metrics can be mathematically recovered, allowing organizations to audit their systems accurately while fully preserving individual privacy.
When the available labels are known to be noisy or corrupted, robust optimization offers a mathematical safety net. Serena Wang developed optimization techniques using Distributionally Robust Optimization (DRO) and soft group assignments. This approach guarantees that a model remains fair to the true underlying demographic groups even when the training process is forced to rely on highly unreliable or noisy proxy labels.
Blinding Models via Feature and Representation Space
Another way to achieve fairness without labels is to eliminate the model's ability to perceive demographic proxies in the first place. Even when sensitive attributes are removed from a dataset, models can easily reconstruct them through highly correlated features—such as text style, vocabulary, or consumer behavior.
To combat this in natural language processing, Alexey Romanov et al. targeted the implicit demographic signals embedded within text. In applications like screening resumes or biographies, names serve as powerful proxies for gender and race. The authors developed a method that penalizes the correlation between a model's predictions and the word embeddings of names during training. By scrubbing this proxy signal from the representation space, the model is prevented from using names as a crutch for predictions, mitigating multiple biases simultaneously with minimal impact on accuracy.
In computer vision, a similar strategy can be applied to representation spaces without retraining the entire core architecture. For face verification systems, Tiago Salvador et al. introduced FairCal, a post-processing method that relies on unsupervised clustering. By grouping face embeddings based on their intrinsic geometric properties in the feature space and locally calibrating the decision thresholds within those clusters, they significantly reduced the variance in false positive rates across demographics without ever requiring explicit ethnicity labels.
This idea of using internal model signals to maintain fairness also extends to sequential settings. In continual learning, models frequently suffer from "catastrophic forgetting" and latch onto shortcuts or spurious correlations over time. Giacomo Capitani et al. developed a framework called Learning without Shortcuts (LwS). By utilizing self-supervised clustering and a loss-based memory buffer, LwS ensures that a model's evolving representation space remains balanced and resilient to bias over time, matching the performance of fully supervised fairness systems without requiring privileged metadata.
The Limits of Blind Fairness
While these methods provide powerful tools for privacy-constrained environments, they are not a silver bullet. The primary limitation of label-free fairness techniques is their reliance on feature correlation.
As noted by Preethi Lahoti, if a disadvantaged demographic group is completely latent—meaning its members' features look identical to the rest of the dataset, and its disadvantage leaves no distinct footprint in the feature or loss space—the optimization process cannot address its specific plight. Label-free methods work because systemic bias usually leaves a statistical trail; if that trail is entirely absent, explicit demographic data remains the only path to equity.