ScholarWorks@숙명여자대학교

초록

The exponential growth in data volume has driven significant advances in machine learning, but the inclusion of private or unauthorized data in training dataset has led to a surge in deletion requests and privacy disputes. While retraining models without the contested data offers a fundamental solution, it incurs prohibitive computational costs. This challenge has given rise to machine unlearning, a process that selectively removes knowledge of specified data from a trained model. However, most unlearning methods require access to the complete original training dataset and often suffer from degraded generalization performance after unlearning. This limitation makes them impractical in scenarios where the full training dataset is no longer accessible. To address these challenges, we propose a novel unlearning technique for image classification and generation tasks that maintains effectiveness even with a small part of the original dataset. Experimental results demonstrate the superiority of our method, supporting the right to be forgotten.

키워드