LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning

TL;DR

LoGoFair tackles fairness in federated learning by proposing a model-agnostic post-processing framework that achieves both local and global fairness via a Bayes-optimal classifier under joint constraints. It derives an explicit characterization of the Bayes-optimal federated fair classifier $h^*$ through a calibration function $F(oldsymbol{b4}^*,oldsymbol{b5}^*,x,a,c)$ and solves a convex program to obtain $(oldsymbol{b4}^*,oldsymbol{b5}^*)$, followed by a bi-level federated post-processing procedure that updates global fairness guidance while preserving local fairness. Empirically, LoGoFair demonstrates superior accuracy-fairness trade-offs on Adult, ENEM, and CelebA compared with baselines, and it supports flexible trade-offs via fairness budgets $(oldsymbol{b4}^l,oldsymbol{b4}^g)$ with low communication overhead. The work advances fair FL by enabling simultaneous local and global fairness under heterogeneity in a model-agnostic, privacy-preserving framework, paving the way for practical deployment in high-stakes decision-making systems.

Abstract

Federated learning (FL) has garnered considerable interest for its capability to learn from decentralized data sources. Given the increasing application of FL in decision-making scenarios, addressing fairness issues across different sensitive groups (e.g., female, male) in FL is crucial. Current research often focuses on facilitating fairness at each client's data (local fairness) or within the entire dataset across all clients (global fairness). However, existing approaches that focus exclusively on either local or global fairness fail to address two key challenges: (\textbf{CH1}) Under statistical heterogeneity, global fairness does not imply local fairness, and vice versa. (\textbf{CH2}) Achieving fairness under model-agnostic setting. To tackle the aforementioned challenges, this paper proposes a novel post-processing framework for achieving both Local and Global Fairness in the FL context, namely LoGoFair. To address CH1, LoGoFair endeavors to seek the Bayes optimal classifier under local and global fairness constraints, which strikes the optimal accuracy-fairness balance in the probabilistic sense. To address CH2, LoGoFair employs a model-agnostic federated post-processing procedure that enables clients to collaboratively optimize global fairness while ensuring local fairness, thereby achieving the optimal fair classifier within FL. Experimental results on three real-world datasets further illustrate the effectiveness of the proposed LoGoFair framework.

Figures (4)

• Figure 1: (a) Demonstrates a toy example of local and global fairness in the context of an one-dimensional classification problem. These fair classifiers ensure equal gender proportions in classification at local and global level (e.g. Local classifiers allocate 2/3 of samples from both genders to the left side for client 1). (b) Presents comparisons between local (FCFL) and global (FedFB) fair FL algorithms across varying levels of heterogeneity. A smaller $\alpha$ signifies more heterogeneity across clients, and a smaller $\Delta$ denotes a fairer model at local or global level.
• Figure 2: Overview of LoGoFair framework. The Bayes optimal federated fair classifier, which strikes the optimal accuracy-fairness balance, is identified as the objective. The federated post-processing procedure reformulates it as a bi-level problem incorporating local fairness optimization and global fairness guidance.
• Figure 3: Scalability and effectiveness analysis.
• Figure 4: Scalability and effectiveness analysis (EO).

Theorems & Definitions (7)

• Definition 1
• Example 1
• Theorem 1
• Lemma 1
• Lemma 2
• Lemma 3
• Theorem 2