Towards Auditability for Fairness in Deep Learning

TL;DR

This paper introduces smooth prediction sensitivity, a new efficient measure inspired by interpretability techniques, to audit individual fairness in deep learning models and detect unfair predictions even when models pass group fairness metrics.

Contribution

It proposes a novel measure called smooth prediction sensitivity for auditing individual fairness in deep learning models, addressing limitations of group fairness metrics.

Findings

Preliminary results show smooth prediction sensitivity can distinguish fair from unfair predictions.

It may help identify blatantly unfair predictions in models that are group-fair.

The method is computationally efficient and inspired by interpretability techniques.

Abstract

Group fairness metrics can detect when a deep learning model behaves differently for advantaged and disadvantaged groups, but even models that score well on these metrics can make blatantly unfair predictions. We present smooth prediction sensitivity, an efficiently computed measure of individual fairness for deep learning models that is inspired by ideas from interpretability in deep learning. smooth prediction sensitivity allows individual predictions to be audited for fairness. We present preliminary experimental results suggesting that smooth prediction sensitivity can help distinguish between fair and unfair predictions, and that it may be helpful in detecting blatantly unfair predictions from "group-fair" models.