Back to search

Boston University

Towards private and efficient optimization methods for machine learning

Abstract

dc:description.abstract

In the last couple of decades, we have seen tremendous progress in Machine Learning (ML) both in the scale and effectiveness of ML models. ML models are now used in every aspect of our life, from healthcare and finance to personalized recommendations and autonomous systems. As these models become more sophisticated, they are required to handle more data than ever. This has introduced a huge need to design ML algorithms that have good utility guarantees while ensuring privacy for users. This is where differentially private (DP) machine learning plays a critical role. Differential privacy provides a mathematically rigorous framework to ensure that individual data points cannot be reverse-engineered from the models outputs, even by someone with access to large amounts of auxiliary information. This thesis introduces novel differentially private (DP) algorithms for various settings, including Non-convex Empirical Risk Minimization (ERM), Stochastic Convex Optimization (SCO), and Stochastic Linear Bandits. We begin by exploring the general Non-convex Optimization landscape, which is central to modern ML, and discuss the challenges involved in making non-private algorithms differentially private. Following this, we introduce the binary tree mechanism, demonstrating how it can be leveraged to develop new private algorithms that achieve state-of-the-art (SOTA) utility guarantees in both ERM and SCO settings. These results are made possible by exploiting the sensitivity-reducing properties of the tree mechanism. Finally, we present a new private variant of the classic LinUCB algorithm (Abbasi-yadkori et al., 2011; Chu et al., 2011), which achieves asymptotically optimal regret bounds under specific settings for stochastic linear bandits.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Electrical & Computer Engineering
Grantor
Boston University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Tran, Hoang
Advisor dc:contributor.advisor
  • Cutkosky, Ashok
ORCID dc:identifier.orcid

Subjects

dc:subject × 1

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/50579

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Tran, Hoang. Towards private and efficient optimization methods for machine learning. doctoral thesis, Boston University, 2025. https://hdl.handle.net/2144/50579