Abstract
Standard Federated Learning (FL) often fails to account for demographic differences, resulting in models that perform poorly for minority groups. We introduce Provably Fair Federated Learning (PFFL), a framework that utilizes Bounded Group Loss (BGL) to provide rigorous fairness guarantees in decentralized environments. By optimizing a minimax objective, PFFL ensures that group-specific losses remain below defined thresholds. Our theoretical analysis proves convergence and generalization, while empirical results across diverse datasets demonstrate that PFFL achieves a superior accuracy-fairness trade-off compared to existing state-of-the-art methods.
Introduction
Federated Learning (FL) enables training models on decentralized data without compromising privacy by keeping data on-device. However, the "heterogeneity" of this data—where distributions vary significantly across devices—poses a significant challenge for fairness. Local fairness enforcement on individual devices does not guarantee a fair global model. Existing fair FL techniques often lack formal convergence proofs or fail to generalize to unseen data. PFFL addresses these gaps by providing a mathematically sound framework for fair decentralized optimization.
Methods
PFFL is designed to solve a "saddle point" (minimax) problem using a scalable optimization approach.
Bounded Group Loss (BGL)
Instead of the computationally difficult task of equalizing error rates across all groups, PFFL employs Bounded Group Loss (BGL). This approach ensures that the loss for every protected group remains below a pre-specified acceptable threshold. This formulation maintains the convexity of the optimization problem, allowing for stronger mathematical guarantees.
The PFFL Algorithm
The algorithm alternates between two primary phases:
- Model Update: Using standard FL techniques (e.g., FedAvg) to update global model weights.
- Fairness Adjustment: An "exponentiated gradient" step that adjusts fairness weights. This step prioritizes groups whose loss thresholds are currently being violated, forcing the model to improve its performance on those specific demographics.
Global vs. Local Constraints
A critical insight of this work is that enforcing a single global fairness constraint across the entire network is significantly more effective than requiring local fairness at each client. As shown in the figure above, global BGL constraints (blue) consistently achieve lower maximum group loss and lower error rates than local constraints (red) in heterogeneous environments.
Experimental Evaluation
The researchers evaluated PFFL on three benchmark datasets:
- ProPublica COMPAS: Recidivism prediction with gender as the protected attribute.
- ACS Employment: US Census data partitioned by state to simulate a federated environment, focusing on racial fairness.
- CelebA: A large-scale facial image dataset used to test PFFL with non-convex models like Convolutional Neural Networks (CNNs).
Key Results and Insights
- Optimal Trade-offs: On the ACS dataset, PFFL maintained higher overall accuracy while significantly reducing the "maximum group loss," ensuring the worst-performing group was better served than under vanilla FedAvg.
- Competitive Performance: Despite its general formulation, PFFL outperformed specialized algorithms like FedFB and FedFair, even on metrics those algorithms were specifically designed to optimize (e.g., Equal Opportunity).
- Theoretical Rigor: PFFL is the first fair FL framework to provide a "no-regret" analysis, proving that the algorithm converges to an optimal fair solution that generalizes to new data.
Limitations and Future Directions
While PFFL represents a significant advancement, it introduces a slight privacy overhead, as the server must know which groups are represented on which devices to calculate group losses. Additionally, while empirical results on CNNs are promising, the current formal convergence proofs are limited to convex loss functions. Future work will focus on extending these theoretical guarantees to the non-convex regimes typical of deep learning.
Summary updated