Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data

Asymmetric Langevin Unlearning (ALU) framework uses public data injection to suppress privacy costs in certified machine unlearning, addressing the noise-utility trade-off that degrades model performance during large-scale deletion requests.