Exploring the Unfairness of DP-SGD Across Settings

TL;DR

This paper investigates how Differentially Private Stochastic Gradient Descent (DP-SGD) affects fairness across different machine learning tasks, revealing a negative correlation between privacy and fairness in some cases and minimal impact in others.

Contribution

The study provides a comprehensive evaluation of DP-SGD's impact on fairness across multiple implementations and tasks, highlighting the nuanced relationship between privacy and fairness.

Findings

Negative logarithmic correlation between privacy and fairness in classification and deep learning.

DP-SGD had no significant impact on fairness for PCA.

DP-SGD did not seem to produce private representations in PCA.

Abstract

End users and regulators require private and fair artificial intelligence models, but previous work suggests these objectives may be at odds. We use the CivilComments to evaluate the impact of applying the {\em de facto} standard approach to privacy, DP-SGD, across several fairness metrics. We evaluate three implementations of DP-SGD: for dimensionality reduction (PCA), linear classification (logistic regression), and robust deep learning (Group-DRO). We establish a negative, logarithmic correlation between privacy and fairness in the case of linear classification and robust deep learning. DP-SGD had no significant impact on fairness for PCA, but upon inspection, also did not seem to lead to private representations.