Advancing Machine Unlearning in Deep Models: Effectiveness, Efficiency, and Reliability

The School of EECS is hosting the following PhD Progress Review 3 Seminar:
 

Advancing Machine Unlearning in Deep Models: Effectiveness, Efficiency, and Reliability

Speaker: Shaofei Shen

Host: Prof Hongzhi Yin

Abstract:

In the current context of heightened data privacy awareness, machine unlearning has gained increasing attention. This process aims to remove information derived from designated forget data in a well-trained model while preserving the knowledge acquired from the remaining retain data. To handle unlearning requests, an unlearning algorithm should achieve the following three objectives: 1) removing the influence of forget data without affecting the normal functionality supported by retain data; 2) improving data and time efficiency when processing unlearning requests; and 3) reliably removing the designated information, particularly from hard-to-unlearn data. Nonetheless, previous methods often encounter obstacles that prevent these objectives from being achieved, including performance degradation on retain data, reliance on label information and repeated data access, high computational costs, and unreliable forgetting of highly memorized examples. 

To address these issues, this PhD research explores three key aspects: 1) investigating the causes of performance degradation and preserving performance on retain and test data; 2) reducing reliance on label information and improving the efficiency of processing sequential unlearning requests; and 3) introducing adaptive, instance-specific target estimation to improve unlearning reliability for highly memorized data. The overall aim of this research is to advance machine unlearning in terms of effectiveness, efficiency, and reliability.

Biography:

Mr Shaofei Shen received a B.Sc. degree in mathematics and applied mathematics from the University of Science and Technology of China and a M.Sc. degree in data science from the University of Queensland. He is currently pursuing a Ph.D degree at the School of EECS under the supervision of Dr Miao Xu, Dr Alina Bialkowski, and Dr Weitong Tony Chen. His research interests encompass machine learning and unlearning.