• We propose a label-agnostic machine unlearning framework that does not require any class supervision during the unlearning phase, addressing practical and privacy-driven limitations of existing supervised MU methods. • We demonstrate that our method achieves effective unlearning by removing only the target-specific information while preserving the performance on retain data. • We validate the proposed method across various unlearning scenarios (e.g., class unlearning, sub-class unlearning, random sample unlearning, noisy label entry unlearning) and show that it outperforms existing baselines in terms of both privacy robustness and utility retention.
Cha et al. (Fri,) studied this question.