Source-linked AI summary
Fairness-aware Agnostic Federated Learning
Wei Du, Depeng Xu, Xintao Wu, Hanghang Tong
TL;DR
Fairness in federated learning remains under-explored when testing data differ from training data or are unknown. AgnosticFair uses kernel reweighting in the loss and fairness constraint, and experiments show accuracy and fairness under distribution shift, including fairness on local client data.
Problem
Fairness-aware federated learning lacks a solution for unknown testing distributions and client distribution shift.
Method
AgnosticFair uses kernel reweighting functions in both the agnostic loss and fairness constraint within a minimax formulation.
Results
Experiments on two datasets demonstrate good accuracy and fairness on unknown testing data, including fairness for each local client.
Takeaways & Limitations
The trained global model can be applied under distribution shift and deployed at local sites while retaining the stated fairness guarantees.
Takeaways & Limitations
Optimizing the fairness-constrained objective is challenging for many non-convex models, including deep learning models.
Abstract
from arXiv · showhide
Federated learning is an emerging framework that builds centralized machine learning models with training data distributed across multiple devices. Most of the previous works about federated learning focus on the privacy protection and communication cost reduction. However, how to achieve fairness in federated learning is under-explored and challenging especially when testing data distribution is different from training distribution or even unknown. Introducing simple fairness constraints on the centralized model cannot achieve model fairness on unknown testing data. In this paper, we develop a fairness-aware agnostic federated learning framework (AgnosticFair) to deal with the challenge of unknown testing distribution. We use kernel reweighing functions to assign a reweighing value on each training sample in both loss function and fairness constraint. Therefore, the centralized model built from AgnosticFair can achieve high accuracy and fairness guarantee on unknown testing data. Moreover, the built model can be directly applied to local sites as it guarantees fairness on local data distributions. To our best knowledge, this is the first work to achieve fairness in federated learning. Experimental results on two real datasets demonstrate the effectiveness in terms of both utility and fairness under data shift scenarios.
1 Introduction
Federated learning avoids uploading raw data but fairness remains under-explored when client and testing distributions differ or the testing distribution is unknown. AgnosticFair addresses this setting with kernel-based reweighting in both loss and fairness constraints, achieving fairness and accuracy under distribution shift.
- Motivation: Federated learning trains local models and exchanges parameters instead of collecting raw client data centrally.This reduces the privacy concern associated with uploading sensitive raw data to a central server.
- Problem: Fairness in federated learning is challenging because client training distributions differ and the testing distribution may be unknown.A global model must remain useful and fair despite distribution shift across clients and between training and testing data.
- Approach: AgnosticFair formulates fairness-aware federated learning as a two-player minimax game against possible unknown testing distributions.The adversary maximizes classifier loss, while the learner uses sample-level reweighting in the agnostic loss and fairness constraint.
- Contribution: The work presents fairness-aware federated learning as a first study of fairness under client data distribution shift.Its framework generalizes related formulations that minimize over a single client or group.
- Approach: Kernel function parametrization makes the loss function and fairness constraints agnostic to data distribution shift.The approach exchanges parameters and coefficients rather than raw data during optimization.
- Results: Experiments on two public datasets show fair prediction under distribution shift while maintaining high accuracy.The resulting global model can also be deployed at local sites because it guarantees fairness on local data distributions.
2 Related Work
Prior federated-learning research addresses privacy leakage, limited bandwidth, communication cost, and non-IID client data, while fairness research uses data processing or optimization constraints. This work distinguishes itself by treating fairness-aware federated learning under client distribution shift with individual-sample reweighting.
- Federated Learning: Federated-learning research has addressed bandwidth limits, privacy leakage, non-IID data, communication rounds, parameter sparsification, and gradient quantization.These works primarily target system and statistical challenges in distributed training.
- Distribution Shift: Distribution-shift methods reweight training data to represent testing data or correct mismatched distributions.Related approaches use density-ratio estimation, sample reweighting, or reweighting values in loss optimization.
- Fairness Learning: Fairness-aware learning commonly removes sensitive information during preprocessing, post-processes classifiers, or incorporates fairness constraints during optimization.This work follows the constraint-based direction.
- Research Gap: Earlier fairness-aware federated formulations minimize over a single client or group, limiting their scale.The proposed framework instead uses an individual-sample reweighting function to address distribution shift among clients.
3 Fair Agnostic Federated Learning
This section formulates fairness-aware agnostic federated learning for unknown testing distributions and describes its kernel-reweighted optimization and federated implementation. AgnosticFair combines robust loss and fairness constraints so the learned classifier targets accuracy and fairness under distribution shift, including at local clients.
- Problem Formulation: Federated learning trains local models and exchanges parameters to generate a global model, but standard fairness constraints assume training and testing distributions coincide.When distributions differ, training fairness does not guarantee testing fairness.
- Agnostic Loss Function: Without testing data, reweighing values cannot be estimated directly, motivating an agnostic objective favorable to unknown testing distributions.The agnostic loss minimizes worst-case loss and is equivalent to empirical risk minimization under IID data.
- Kernel Function Parametrization: The reweighing function uses kernel basis functions with nonnegative bounded coefficients to model possible distribution shifts and control adversarial capacity.Multiple basis functions capture different uncertainties, while the linear form supports linear or convex optimization.
- Kernel Function Parametrization: Individual sample-level reweighing models unknown shifts more flexibly than client-level weights because samples within a client can differ and samples across clients can resemble one another.Client-level AgnosticFL assigns one value to every sample from a client, whereas AgnosticFair assigns weights per data point.
- Agnostic Fairness Constraint: AgnosticFair applies the same reweighing function to the fairness constraint, combining it with agnostic loss to preserve fairness on unknown testing data and local client distributions.Clients optimize local w while the server optimizes α through iterative coordination.
- Variants of AgnosticFair: AgnosticFair-a removes the fairness constraint to evaluate accuracy under distribution shift, whereas AgnosticFair-b is expected to fail fairness under unknown shift.The full AgnosticFair formulation is intended to maintain fairness and high accuracy under unknown distribution shift.
4 Experiments
Experiments on Adult and Dutch evaluate AgnosticFair under distribution shift and IID settings, including accuracy, fairness, and performance as the number of local clients changes. The results report improved shifted-distribution accuracy and fairness guarantees, while IID accuracy remains comparable to FL.
- Experimental setup: The experiments use Adult and Dutch datasets, with distribution-shift scenarios created by separating training and testing populations and two local clients.Adult has 45,222 samples and 40 features; Dutch has 60,420 samples and 35 features.
- Unknown data shift: 0.8264 and 0.8162 are AgnosticFair-a's testing accuracies on Adult and Dutch, versus 0.7998 and 0.7951 for FL under distribution shift.The corresponding improvements over FL are 0.0296 on Adult and 0.0211 on Dutch.
- Unknown data shift: AgnosticFair outperforms baselines on fairness under unknown testing distributions, whereas training-data fairness constraints alone do not guarantee testing-data fairness.FL has RD 0.1477, while a fair learning model usually requires RD below 0.05; FairFL achieves fairness on training data but fails on testing data.
- Multiple local clients: LocalFair's accuracy degrades significantly as the number of clients increases because it adds a separate local fairness constraint for each client.LocalFair can guarantee fairness for each local client, but the added constraints reduce global-model utility when more clients are included.
- IID setting: Under IID data, AgnosticFair-a has nearly the same testing accuracy as FL: 0.8111 versus 0.8130 on Adult and 0.8089 versus 0.8096 on Dutch.The authors relate this parity to the equivalence between robust reweighted loss minimization and empirical risk minimization under IID data.
- IID setting: Under IID data, AgnosticFair-a exceeds the accuracy of, while AgnosticFair preserves fairness guarantees with good accuracy.AgnosticFair-a reports 0.7785 on Adult and 0.7892 on Dutch for the comparison with.
5 Conclusions and Future Work
The proposed framework addresses unknown testing distributions with kernel-reweighted loss and fairness constraints, achieving accuracy and fairness across shifted testing and local client distributions.
- AgnosticFair uses kernel reweighing functions to parameterize both the loss function and fairness constraint.
- The framework improves prediction accuracy when testing data shifts from the training distribution.
- It guarantees fairness on unknown testing data and on each local client.
- Experiments on two datasets demonstrated these benefits for the trained centralized model.
- Future work will extend the framework to additional fairness notions, surrogate functions, and alternative kernel basis functions.