Title:	Privacy Solutions for Machine Learning
DNr:	Berzelius-2026-255
Project Type:	LiU Berzelius
Principal Investigator:	Buse Atli <buse.atli@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-09-23 – 2027-04-01
Classification:	10201
Keywords:

Abstract

The main objective of this project is to design and evaluate privacy solutions for machine learning applications, particularly around data privacy. The research project is structured around two main directions: (1) the development of theoretical foundations to understand privacy-utility tradeoffs of current privacy solutions and (2) designing attacks that can evade the verification of privacy protection mechanisms. Achieving these goals requires large-scale empirical studies with deep neural networks, including transformer architectures and high-dimensional representation models. The research will involve repeated training, different adversary models, and extensive hyperparameter searches to ensure statistically sound conclusions. GPU resources are therefore essential for fast training, scalable distributed experiments, and reproducible evaluations. High-performance computing infrastructure will enable systematic analyses that would otherwise be prohibitively expensive. Expected outcomes include novel algorithms, empirical benchmarks, and theoretical foundations for building trustworthy ML systems. The project will advance scientific understanding of ML privacy. Results will be shared through peer-reviewed publications, open-source tools when appropriate, and active engagement with the research community.