Paper: Unlearning Aware Minimization

Authors: Hoki Kim (Chung-Ang University), Gunwoo Kim (Naver Cloud), Seongwon Chae (Seoul National University, PhD candidate), Sangwon Yoon (Supreme Prosecutors’ Office, public interest attorney)

Venue: NeurIPS 2025

URL: https://neurips.cc/virtual/2025/loc/san-diego/poster/116406/

Original JoongAng Ilbo article

Chung-Ang University (President Sangkyu Park) has announced research findings showing that AI can also forget.

Chung-Ang University announced that a paper by a research team led by Prof. Hoki Kim from the Department of Industrial Security as first author has been accepted at ‘NeurIPS (Neural Information Processing Systems) 2025,’ one of the world’s most prestigious AI conferences.

This research was conducted as a joint study with Naver Cloud, the Supreme Prosecutors’ Office, and others. The research team proposed a new optimization framework that enables AI models to forget (unlearn) specific training data. Previous techniques had limitations in that they either failed to completely remove ‘data to be forgotten’ or degraded the performance of ‘data to be retained.’

To overcome this, the research team introduced a new min-max optimization method called ‘Unlearning-Aware Minimization (UAM).’ The method identifies weights that cause high loss on the data targeted for forgetting and uses them to adjust learning in the direction that minimizes loss on the retained data. Experimental results demonstrated superior performance compared to existing methods on image datasets such as CIFAR-10, CIFAR-100, and TinyImageNet, as well as large language model benchmarks like WMDP-Bio and WMDP-Cyber.

Prof. Hoki Kim stated, “This research is a technology that can realize the user’s ‘right to be forgotten,’ and we expect it to develop into a core technology in the field of personal information protection in the future.”

The research team plans to continue research with the goal of implementing trustworthy AI through the adoption of unlearning in practical application environments and the development of explainable AI.

The paper “Unlearning-Aware Minimization” is scheduled to be presented at NeurIPS 2025, held in San Diego, USA, in December.