Cuong Tran

Date of Award

8-4-2023

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Ferdinando Fioretto

Subject Categories

Computer Sciences | Physical Sciences and Mathematics

Abstract

The availability of large datasets and computational resources has driven significant progress in Artificial Intelligence (AI) and, especially,Machine Learning (ML). These advances have rendered AI systems instrumental for many decision making and policy operations involving individuals: they include assistance in legal decisions, lending, and hiring, as well determinations of resources and benefits, all of which have profound social and economic impacts. While data-driven systems have been successful in an increasing number of tasks, the use of rich datasets, combined with the adoption of black-box algorithms, has sparked concerns about how these systems operate. How much information these systems leak about the individuals whose data is used as input and how they handle biases and fairness issues are two of these critical concerns. While some people argue that privacy and fairness are in alignment, the majority instead believe these are two contrasting metrics. This thesis firstly studies the interaction between privacy and fairness in machine learning and decision problems. It focuses on the scenario when fairness and privacy are at odds and investigates different factors that can explain for such behaviors. It then proposes effective and efficient mitigation solutions to improve fairness under privacy constraints. In the second part, it analyzes the connection between fairness and other machine learning concepts such as model compression and adversarial robustness. Finally, it introduces a novel privacy concept and an initial implementation to protect such proposed users privacy at inference time.

Access

Open Access

Download

Share

COinS