Data-limited Class Unlearning: from Methodology to Evaluation

The School of EECS is hosting the following HDR Progress Seminar:

Data-limited Class Unlearning: from Methodology to Evaluation

Speaker: Mr Chenhao Zhang
Host/Chair: Professor Hongzhi Yin

Abstract

Machine unlearning has emerged as a critical paradigm for removing the influence of specific training data from deployed models, enabling compliance with privacy regulations, correcting erroneous data, and adapting to evolving real-world demands. However, most existing approaches assume full access to the original training dataset—an assumption that rarely holds in practice, where data availability is often constrained by privacy requirements, storage costs, or organizational policies.

This thesis advances the study of class unlearning under data-limited settings from both methodological and evaluation perspectives. First, we propose a perception-revising framework that enables unlearning with only the forgetting data—the data whose influence must be removed—while mitigating the risk of the forgotten class re-emerging when the model continues to learn from retained-class data. Second, we develop an efficient data-free unlearning method that operates without access to either forgetting or retained samples, instead synthesizing and leveraging pseudo-samples to achieve effective and scalable unlearning. Finally, we introduce a novel evaluation framework that provides a post-hoc, data-efficient, and model-agnostic criterion for verifying unlearning success beyond conventional accuracy- or retraining-based metrics.

Bio

Mr. Chenhao Zhang received his master's degree in computer science from the School of Electrical Engineering and Computer Science (EECS), University of Queensland, Australia, in 2022. He is currently pursuing a Ph.D. degree in the same school under the supervision of Dr Miao Xu and Dr Weitong Chen. His research interests include machine unlearning, deep learning, and trustworthy machine learning.

About Data Science Seminar

This seminar series is hosted by EECS Data Science.

Venue

Room 631, Building 78
Zoom Link: https://uqz.zoom.us/j/7151984997