AI Fairness through Robustness

Tutorial Slides: Please check the Outline Section

Conference: AAAI 2023

Date and Time: February 7th, 2023 (08:30 AM – 12:30 PM EST)

IBM Research

University of Michigan, Ann Arbor

IBM Research

Tutorial Outline and Slides

Description of the Tutorial

As machine learning (ML) models replace and/or assist humans in high-stakes decision-making and decision support roles, concern regarding the consequences of algorithmic bias is growing. As a response, researchers have developed a variety of formal fairness definitions and methods for enforcing them. A prevalent theme in these approaches is casting fair learning as a constrained optimization problem to enforce some notion of parity across protected groups, known as group fairness. A less well-studied approach to algorithmic fairness is based on casting the problem as robust optimization. The ``adversary's'' goal is to identify similar inputs (e.g. job applications with the same qualifications, but from applicants belonging to different demographic groups on which the ML model produces different outputs. This perspective has been adopted in the individual fairness literature but is also suitable for group fairness.

The goal of this tutorial is to elucidate the unique and novel connections between algorithmic fairness and the rich literature on adversarial machine learning. Compared to other tutorials on AI fairness, this tutorial will emphasize the connection between recent advances in fair learning and adversarial robustness literature. The range of the presented techniques will cover a complete fairness pipeline, starting from auditing ML models for fairness violations, post-processing them to rapidly alleviate bias, and re-training or fine-tuning models to achieve algorithmic fairness. Each topic will be presented in the context of adversarial ML, specifically, (i) connections between fair similarity metrics for individual fairness and adversarial attack radius, (ii) auditing as an adversarial attack, (iii) fair learning as adversarial training, (iv) distributionally robust optimization for group fairness. We will conclude with (v) a summary of the most recent advances in adversarial ML and its potential applications in algorithmic fairness. We are hoping this tutorial will inspire active collaborations and future work at the intersection of the mature field of adversarial ML and the younger field of algorithmic fairness.