Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FinP: Fairness-in-Privacy in Federated Learning by Addressing Disparities in Privacy Risk

Tianyu Zhao, Mahmoud Srewa, Salma Elmalaki

TL;DR

This work tackles fairness in privacy for federated learning by addressing disparities in privacy risk across clients, focusing on source inference attacks. It introduces FinP, a two-pronged framework combining server-side adaptive aggregation with client-side regularization to reduce overfitting and thus privacy vulnerability. Through experiments on HAR and CIFAR-10, FinP improves fairness in privacy (e.g., substantial equal-opportunity gains on CIFAR-10) while maintaining near-baseline utility. The approach offers a practical alternative to differential privacy, achieving stronger fairness in privacy without prohibitive utility loss and with adaptable server-side and client-side components for real-world FL deployments.

Abstract

Ensuring fairness in machine learning extends to the critical dimension of privacy, particularly in human-centric federated learning (FL) settings where decentralized data necessitates an equitable distribution of privacy risk across clients. This paper introduces FinP, a novel framework specifically designed to address disparities in privacy risk by mitigating disproportionate vulnerability to source inference attacks (SIA). FinP employs a two-pronged strategy: (1) server-side adaptive aggregation, which dynamically adjusts client contributions to the global model to foster fairness, and (2) client-side regularization, which enhances the privacy robustness of individual clients. This comprehensive approach directly tackles both the symptoms and underlying causes of privacy unfairness in FL. Extensive evaluations on the Human Activity Recognition (HAR) and CIFAR-10 datasets demonstrate FinP's effectiveness, achieving improvement in fairness-in-privacy on HAR and CIFAR-10 with minimal impact on utility. FinP improved group fairness with respect to disparity in privacy risk using equal opportunity in CIFAR-10 by 57.14% compared to the state-of-the-art. Furthermore, FinP significantly mitigates SIA risks on CIFAR-10, underscoring its potential to establish fairness in privacy within FL systems without compromising utility.

FinP: Fairness-in-Privacy in Federated Learning by Addressing Disparities in Privacy Risk

TL;DR

This work tackles fairness in privacy for federated learning by addressing disparities in privacy risk across clients, focusing on source inference attacks. It introduces FinP, a two-pronged framework combining server-side adaptive aggregation with client-side regularization to reduce overfitting and thus privacy vulnerability. Through experiments on HAR and CIFAR-10, FinP improves fairness in privacy (e.g., substantial equal-opportunity gains on CIFAR-10) while maintaining near-baseline utility. The approach offers a practical alternative to differential privacy, achieving stronger fairness in privacy without prohibitive utility loss and with adaptable server-side and client-side components for real-world FL deployments.

Abstract

Ensuring fairness in machine learning extends to the critical dimension of privacy, particularly in human-centric federated learning (FL) settings where decentralized data necessitates an equitable distribution of privacy risk across clients. This paper introduces FinP, a novel framework specifically designed to address disparities in privacy risk by mitigating disproportionate vulnerability to source inference attacks (SIA). FinP employs a two-pronged strategy: (1) server-side adaptive aggregation, which dynamically adjusts client contributions to the global model to foster fairness, and (2) client-side regularization, which enhances the privacy robustness of individual clients. This comprehensive approach directly tackles both the symptoms and underlying causes of privacy unfairness in FL. Extensive evaluations on the Human Activity Recognition (HAR) and CIFAR-10 datasets demonstrate FinP's effectiveness, achieving improvement in fairness-in-privacy on HAR and CIFAR-10 with minimal impact on utility. FinP improved group fairness with respect to disparity in privacy risk using equal opportunity in CIFAR-10 by 57.14% compared to the state-of-the-art. Furthermore, FinP significantly mitigates SIA risks on CIFAR-10, underscoring its potential to establish fairness in privacy within FL systems without compromising utility.

Paper Structure

This paper contains 32 sections, 3 theorems, 18 equations, 24 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Problem Statement
  4. Fairness-in-Privacy Framework in Federated Learning
  5. Formalizing Causes of Fairness in Privacy on Client Side
  6. Formalizing Symptoms of Fairness in Privacy on Server Side
  7. Evaluation
  8. Federated Learning System Setup
  9. SIA Attack
  10. Metrics for Comparison
  11. Evaluating FinP on HAR dataset
  12. Evaluating FinP on CIFAR-10 dataset
  13. Ablation experiments on local epochs and overfitting
  14. Summary of FinP in HAR and CIFAR-10 results
  15. Discussion and Limitations
  16. ...and 17 more sections

Key Result

Proposition 1

A model with a smaller Lipschitz constant is less prone to overfitting.

Figures (24)

  • Figure 1: Fairness in Privacy FinP in FL by addressing the causes and the symptoms to achieve.
  • Figure 2: CIFAR dataset profile for each client after Dirichlet sampling with $\alpha=0.1$
  • Figure 3: Disparity of SIA accuracy among clients using HAR dataset.
  • Figure 4: Disparity of prediction loss among clients using HAR dataset.
  • Figure 5: Global model classification accuracy using HAR dataset.
  • ...and 19 more figures

Theorems & Definitions (14)

  • Proposition 1
  • proof
  • Proposition 2
  • proof
  • Proposition 3
  • proof
  • Claim 1: Adaptive Overfitting Reduction
  • proof
  • Claim 2: Variance Minimization via Aggregation
  • proof
  • ...and 4 more