Federated Fairness without Access to Sensitive Groups

TL;DR

This paper tackles fairness in federated learning when sensitive groups are unknown or unlabeled. It introduces Relaxed Conditional Value-at-Risk (RCVaR), a fairness objective that blends worst-tail performance with average utility through a trade-off parameter $\epsilon$ under a group-size constraint $\rho$, and extends it to a federated setting (FedSRCVaR). The authors provide theoretical convergence and excess-risk guarantees for the smoothing-based optimization and demonstrate, across multiple real-world datasets, that the approach improves the worst-performing subgroup while maintaining competitive overall performance, offering a continuum of fairness-utility trade-offs. This work enables robust, group-fair models in privacy-preserving, regulation-driven contexts without requiring predefined sensitive groups, with practical impact on high-stakes decision-making systems.

Abstract

Current approaches to group fairness in federated learning assume the existence of predefined and labeled sensitive groups during training. However, due to factors ranging from emerging regulations to dynamics and location-dependency of protected groups, this assumption may be unsuitable in many real-world scenarios. In this work, we propose a new approach to guarantee group fairness that does not rely on any predefined definition of sensitive groups or additional labels. Our objective allows the federation to learn a Pareto efficient global model ensuring worst-case group fairness and it enables, via a single hyper-parameter, trade-offs between fairness and utility, subject only to a group size constraint. This implies that any sufficiently large subset of the population is guaranteed to receive at least a minimum level of utility performance from the model. The proposed objective encompasses existing approaches as special cases, such as empirical risk minimization and subgroup robustness objectives from centralized machine learning. We provide an algorithm to solve this problem in federation that enjoys convergence and excess risk guarantees. Our empirical results indicate that the proposed approach can effectively improve the worst-performing group that may be present without unnecessarily hurting the average performance, exhibits superior or comparable performance to relevant baselines, and achieves a large set of solutions with different fairness-utility trade-offs.

Figures (4)

• Figure 1: We consider a federated learning setting with $K \in \mathcal{K}$ clients. Our method maximizes the performance of the worst possible group of size $\rho$ that can be formulated from the union of the local individuals/samples $z$, that is, within a model class, no other model performs better on its worst $\rho$ fraction of the samples. Equivalently, no other model has a lower $(1-\rho)$-th superquantile of its loss distribution. We achieve this objective at the lowest possible cost to the non-critical samples. We make no assumptions on the distribution of the worst-performing samples amongst clients, and note that (a) worst-performing sensitive groups might not align with a single conventional demographic, and (b) 'groups' and 'clients' are not synonyms in our setting.
• Figure 2: Comparison of worst group risk, utility risk and group risk disparity between the best and worst groups on different datasets. $\overline{h}$ denotes the uniform classifier. FedRCVaR recovers solutions equivalent to centralized settings, while improving both utility and fairness compared to FL baselines in many settings. $\rho$ is a hyperparameter of FedSRCVaR, DRO and BPF. Differences in average performance as a function of $\rho$ for ERM, FedAVG and AFL are due to the variation of the training hyperparameters, since for each $\rho$ we report the hyperparameter combination producing the model with the best performance for each method.
• Figure 3: Performance trade-offs among worst group and utility for different pairs of $(\epsilon,\rho) \in \{0.01,0.1,\dots,0.9,1.0\}\times \{0.1,\dots,0.9\}$ values on real datasets. The different colours indicate different $\epsilon$ values. $\overline{h}$ denotes the uniform classifier. A lower score indicates better performance. We report the worst group and average/utility risks, as a function of $\rho$.
• Figure 4: Performance comparison between FedSRCVaR for local epochs $\tau \in \{1,5,10\}$ and FedAVG. $\overline{h}$ denotes the uniform classifier. We report the worst group and average/utility risks, and the group risk disparity between the worst-performing samples and the remaining population, as a function of $\rho$.

Theorems & Definitions (21)

• Definition 3.1: Weak Pareto optimality
• Definition 3.2: Proper Pareto optimality
• Definition 4.1: Smooth Approximation smooth_plus_func
• Lemma 5.5
• Lemma 5.6: Convergence of FedSRCVaR
• Lemma 5.7: Excess Risk Analysis
• Remark A.1
• Definition B.1: Convex Set
• Definition B.2: Convex Function
• Definition B.3: Lipschitzness