Fairness Without Demographics

Updated 13 April 2026

Fairness without demographics is a set of ML methods that achieve equitable outcomes without explicit group labels, using robust optimization and graph theory.
Techniques like DRO, adversarial reweighting, and variance minimization bound worst-case subgroup errors in applications ranging from social networks to healthcare.
These approaches offer provable fairness guarantees and practical insights while addressing privacy and legal constraints, though performance may vary with data quality.

Fairness Without Demographics denotes a class of machine learning fairness methodologies that do not rely on access to sensitive group labels (such as race, gender, or disability) at training or inference time. These approaches are motivated by legal, ethical, privacy, and practical constraints that prevent the collection or use of demographics, as well as by the recognition that sensitive group information can be high-dimensional, intersectional, incomplete, or noisy. The core technical challenge addressed is to achieve meaningful parity-based or Rawlsian group-fairness goals—traditionally dependent on observed groupings—without ever using or recovering sensitive attributes. Over the last several years, a suite of conceptually and mathematically distinct methods have emerged, including robust optimization, proxy-based clustering, graph-theoretic approaches, ethical regularizations, and techniques designed for federated or medical settings.

1. Foundations, Rationale, and Definitions

Traditional fairness in machine learning often presupposes that models have access to explicit group memberships and can directly optimize for metrics such as demographic parity, equalized odds, or worst-group risk. However, in practice, privacy regulations (GDPR, HIPAA), societal norms, or operational realities may preclude the use of demographics. Moreover, intersectional and unforeseen groupings render traditional group-based fairness brittle and potentially incomplete.

Fairness without demographics aims to develop machine learning algorithms that achieve parity-based, Rawlsian, individual, or preference-based fairness objectives without ever observing, collecting, or inferring sensitive group labels. Definitions central to the field include:

Rawlsian Max–Min Fairness: Optimize for the worst-off (latent) subgroup's utility over all possible (unknown) partitions of the population (Hashimoto et al., 2018, Barrainkua et al., 12 Feb 2026).
Group-Free Generalized-Entropy Fairness: Extend economic inequality decomposition to social networks using a homophily-based kernel rather than explicit group partitions (Liu et al., 2023).
Safe Fairness Guarantees: Provide minimax bounds on worst-case subgroup accuracy over distributional ambiguity sets defined in a spectral (Fourier feature) space (Barrainkua et al., 12 Feb 2026).
Preference-Based (No-Harm) Fairness: Ensure every (latent) group would explicitly prefer its subgroup-specific model over any pooled or other group-specific model, all without access to known groupings (Cai et al., 28 Sep 2025).

These definitions avoid demographic dependence by substituting data geometry, loss distributions, graph theory, or optimization over uncertainty sets for explicit group membership.

2. Principal Algorithms and Mathematical Methodologies

Approaches to fairness without demographics are diverse. Key methodologies include:

2.1 Distributionally Robust Optimization (DRO)

DRO upweights high-loss or systematically mis-served regions of the data, approximating "worst-case" groups without knowing their identities. For instance, the DRO formulation in (Hashimoto et al., 2018) maximizes the minimum group accuracy by solving:

$\min_{\theta} \sup_{Q: D(Q \| P) \leq \rho} \mathbb{E}_{z\sim Q}\big[\ell(\theta;z)\big]$

where $D(\cdot\|\cdot)$ is a divergence (e.g., $\chi^2$ ), and $\rho$ tunes the conservativeness. This approach is shown to prevent disparity amplification in repeated learning settings, as the worst-off group cannot be forgotten over time.

Extensions in survival analysis wrap the Cox partial likelihood inside a DRO objective to minimize the worst-case error for any large-enough subpopulation without demographic labels (Hu et al., 2022).

2.2 Adversarial and Graph-Based Instance Reweighting

ARL (Adversarially Reweighted Learning) (Lahoti et al., 2020) and subsequent works (Luo et al., 2024) design adversarial weight functions or graph neural networks that identify and upweight individuals that are analogs of high-loss groups. Notably, (Luo et al., 2024) proposes building a graph where nodes are connected by similarity in parameter gradients, enabling group-fair reweighting without explicit or proxy group membership. These techniques rely on the hypothesis that model errors and gradients correlate with latent demographic groupings.

2.3 Group-Free Homophily Measures

(Liu et al., 2023) introduces a group-free fairness index for settings with an explicit social network. The approach constructs a symmetric, column stochastic kernel $K$ that encodes network homophily. The group-free between-group inequality index

$\Delta_b(y; K) = F_{\alpha}(A(K, y), K1)$

aggregates disparities in outcomes as measured over the network structure, and can be directly plugged into optimization problems for classification, influence maximization, and recommendation, with theoretical guarantees aligning with decomposability properties from economics.

2.4 Variance Minimization and Harmless Rawlsian Updates

(Wang et al., 2024) formally proves that minimizing the variance of sample-wise loss is both necessary and sufficient (in the limit) for group-agnostic Rawlsian fairness: if all per-example losses are essentially equal, then every possible subgroup's worst-case gap vanishes. The VFair algorithm minimizes empirical variance of loss subject to near-optimal mean risk, ensuring fairness across all latent partitions.

2.5 Preference-Based Fairness Without Demographics

(Cai et al., 28 Sep 2025) describes the "Demographic-Agnostic Fairness without Harm" (DAFH) framework. DAFH constructs a learnable group partition $\theta$ , then trains decoupled classifiers for each (latent) group, directly maximizing a fairness-without-harm proxy:

$H(\theta, \{h_k\}) = \frac{1}{K} \sum_{k=1}^K [R_k(h_0) - R_k(h_k)] + \frac{1}{K^2}\sum_{k=1}^K\sum_{j=1}^K [R_k(h_j) - R_k(h_k)]$

The empirical surrogate and soft assignments facilitate differentiable, end-to-end optimization, augmented with KL regularization to avoid degenerate partitions.

Fairness without demographics has been realized in multiple domains:

Social Networks: Group-free homophily kernels demonstrate large reductions in between-group disparity for classification, influence maximization, and recommendation tasks on real social networks (e.g., PolBlogs, Email-EU, Lastfm-Asia) (Liu et al., 2023).
Federated Learning: Sharpness-aware minimax aggregation ensures equitable participant error/uncertainty without sensitive attributes in human-centered federated setups (Roy et al., 2024).
Medical Imaging: Methods that constrain feature entanglement (e.g., disease vs. skin) without using sensitive labels substantially improve demographic fairness in dermatologic diagnosis (Chiu et al., 2024).
Public Datasets: SPECTRE (Barrainkua et al., 12 Feb 2026) achieves minimax fairness on U.S. Census data across 20 states and up to 18 intersectional groups, outperforming both blind and group-aware baselines.
Preference-Based Scenarios: DAFH reliably produces non-harmful, envy-free assignments on datasets (COMPAS, Adult, German, Bank) without demographic access, often exceeding the performance of group-aware methods (Cai et al., 28 Sep 2025).

4. Theoretical Guarantees and Limitations

Most fairness-without-demographics methods admit provable, although sometimes conservative, fairness guarantees:

Minimax error guarantees: Robust optimization formulations bound the worst-case subgroup error over all possible latent partitions of sufficient size (Hu et al., 2022, Barrainkua et al., 12 Feb 2026, Hashimoto et al., 2018).
Decomposition axioms: Homophily-based fairness kernels inherit additive decomposability, scale invariance, and the transfer principle necessary for between-group inequality measures (Liu et al., 2023).
Variance–gap bounds: Loss variance minimization directly yields upper bounds on the maximum group gap for any unknown partition (Wang et al., 2024).
Preference-based generalization: DAFH supplies sample complexity bounds for fairness-without-harm and demonstrates that its latent partition can outperform even ideal demographic partitions under certain conditions (Cai et al., 28 Sep 2025).

However, critical limitations persist:

When data lacks a homophilous structure (e.g., zero assortativity) or group-structure is unidentifiable via gradients or features, methods may have no effect or fail outright (Liu et al., 2023, Luo et al., 2024).
All-variance or Rawlsian methods do not guarantee parity-based fairness metrics (e.g., equalized odds) if there exist partitions uncorrelated with the loss structure.
Proxy-group or clustering methods risk aligning with spurious, non-protected attributes unless rigorously validated.
Label and data noise can undermine performance, but adversarial and graph-based techniques show improved robustness over earlier DRO methods (Luo et al., 2024).
Theoretical bounds on subgroup error are generally conservative; practical group performance may vary.

5. Comparison to Proxy-based and Partial-Demographics Methods

Proxy-based grouping schemes (feature clustering, causal proxy inference, or foundation-model embeddings) are not strictly fairness without demographics. These operate by inferring "pseudo-group" labels via unsupervised clustering or domain adaptation, upon which classical group fairness constraints are then imposed (Queiroz et al., 2024, Wang et al., 17 Nov 2025, Jiang et al., 2024). While proxy approaches can reduce representational imbalance when proxies are highly informative (e.g., FM backbones capture gender), they are error-prone when protected group information is not well aligned with chosen clusters.

Semi-supervised and limited-demographics regimes—where a small subset of labels is available—ease the transition between group-agnostic and group-aware fairness (Ozdayi et al., 2021, Jiang et al., 2024). Even 0.1% labeled samples suffice to reduce standard bias metrics (SPD, AOD, EOD) by 25–50% compared to pure Rawlsian or variance-minimization strategies.

The critical distinction remains: "fairness without demographics" methods fundamentally do not rely on demographic inference, estimation, or partial labeling, and instead optimize for robustness or equitable loss directly over data geometry, loss distributions, graph structure, or output space.

6. Open Problems and Future Directions

Research on fairness without demographics identifies multiple open directions:

Adapting to non-binary, intersectional, or continuous group structures: Many methods are developed under binary or categorical assumptions; generalization to granular, multi-dimensional groupings remains incomplete (Zhou, 2024, Wang et al., 17 Nov 2025).
Improving robustness to distributional drift, label noise, and adversarial data: Combining robust optimization with causality or outlier-resistance is a current focus (Barrainkua et al., 12 Feb 2026, Luo et al., 2024).
Integrating differential privacy: Designing homophily kernels or graph-theoretic structures that are private and robust to attribute inference is an open challenge (Liu et al., 2023).
Transfer learning and third-party fairness: Exploiting domains where partial group information is available or leveraging external datasets via multi-task or adversarial transfer require further exploration (Wang et al., 17 Nov 2025).
Normative and ethical extensions: Aligning fairness-without-demographics with ethical principles (beneficence, non-maleficence, justice) and practical constraints in high-stakes domains demands blended methodological and social science insights (Roy et al., 10 Mar 2026).
Developing realistic benchmarks and causal evaluation frameworks: Current benchmarks resort to synthetic or manually-masked demographics; naturally missing, non-random, or strategically withheld demographics remain largely unstudied (Wang et al., 17 Nov 2025), and the fundamental causal implications of demographic–free fairness are not fully understood (Silva, 2024, Wang et al., 17 Nov 2025).

7. Representative Methods and Empirical Results

Method	Principle	Key Setting	Notable Outcomes
DRO (Hashimoto et al., 2018 Hu et al., 2022)	Robust max-min	Classification, Survival	Controls worst-case group risk without demographics; proven on text, survival data
ARL (Lahoti et al., 2020), GoG (Luo et al., 2024)	Adversarial reweighting / graph	Tabular, sequential	Improves worst-group accuracy, robust to label noise
Homophily kernel (Liu et al., 2023)	Group-free network smoothing	Social networks	Achieves fair classification, influence, recommendation
VFair (Wang et al., 2024)	Variance minimization	Classification, regression	Sharp reduction in group gap for regression, flattened loss distributions
DAFH (Cai et al., 28 Sep 2025)	Decoupled models w/o harm	Tabular (no group info)	Matches/ outperforms group-aware decoupling in rationality/ envy-freeness
Flare (Roy et al., 10 Mar 2026)	Ethical regularization	Healthcare, sensor data	No-harm and increased per-subgroup F1; cluster-based evaluation
Proxy-FM (Queiroz et al., 2024)	FM embedding-based proxies	Medical imaging	Cuts gender imbalance in half with cluster sampling

Empirical studies generally report the highest worst-group performance among demographic-agnostic methods, with modest tradeoffs (<5% overall accuracy) in most settings.

Fairness without demographics constitutes a mathematically principled, empirically validated, and rapidly advancing domain within algorithmic fairness. Its solutions blend robust optimization, network theory, deep learning, adversarial training, and ethical principles to ensure equitable outcomes even when direct demographic labeling is unavailable, precluded, or undesirable. The field will continue to evolve as new theoretical, empirical, and ethical challenges emerge in high-stakes, privacy-sensitive domains.