Faculty/School
Faculty of Science
Topic status
We're looking for students to study this topic.
Primary Supervisor
- Position
- Professor
- Division / Faculty
- Faculty of Science
Overview
As AI-based systems increasingly rely on large-scale, sensitive data, ensuring data privacy while maintaining high model performance has become a critical challenge. This project addresses the fundamental tension between privacy protection and machine learning utility by developing methods that enable accurate AI models to be trained and deployed on protected data.
The central research question is how to effectively integrate advanced privacy-preserving techniques, particularly differential privacy (DP), into modern machine learning pipelines. While DP provides strong, quantifiable privacy guarantees, it often introduces noise that can degrade model accuracy. This project aims to systematically bridge this gap by designing solutions that optimise the privacy–utility trade-off.
Research engagement
Literature review, data collection, algorithm design and a system demo of an AI-based model (prototype).
Research activities
Survey of Recent Advances
Conduct a literature review of recent developments in privacy-preserving machine learning, with a focus on differential privacy and related techniques. The survey will also examine existing open-source tools and established benchmarks to support reproducible research in privacy-preserving AI.
Quantifying the Privacy–Utility Trade-off
Develop a framework to quantify the impact of privacy-preserving mechanisms on model performance. This includes evaluating how different privacy budgets and mechanisms influence accuracy, robustness, and generalisation across diverse tasks and datasets.
Design of High-Accuracy DP-based Algorithms
Design and implement a differential privacy (DP)-based training strategy aimed at minimising utility degradation. Potential approaches may include adaptive noise injection, advanced gradient clipping techniques, privacy-aware optimisation methods, or data granularity-level integration to enhance learning efficiency under strict privacy constraints.
Benchmarking and Evaluation
Establish an evaluation framework using publicly available datasets (e.g., UCI Machine Learning Repository) to assess model performance, privacy guarantees, or computational scalability.
Research skills
Students will gain hands-on experience with advanced AI methodologies, including federated learning and AI-driven model design and evaluation. They will develop practical expertise in integrating DP with machine learning algorithms, as well as implementing privacy-preserving mechanisms and conducting performance and privacy evaluations.
Outcomes
This project aims to develop an AI-based prototype that balances data privacy and machine learning utility, demonstrating the practical application of modern AI and machine learning techniques in real-world scenarios.
Skills and experience
Applicants should have a strong academic background in computer science or a related discipline. Prior programming experience (e.g., Python) is required, and familiarity with machine learning or privacy-preserving techniques is highly desirable.
Start date
2 November, 2026End date
19 February, 2027Location
GP S Block
Keywords
Contact
Professor Yuefeng Li, School of Cumputer Science
07 31385212
y2.li@qut.edu.au