MeGU: A Machine-Guided Framework for Targeted Feature Disentanglement Unlearning

The study proposes MeGU, a machine unlearning framework that uses a multimodal large model to assign semantic perturbation labels to target samples to determine the realignment direction, aiming to alleviate insufficient forgetting and excessive forgetting. The method encodes inter-class concept similarity as a lightweight transition matrix and introduces positive and negative feature noise pairs: negative noise suppresses target-specific feature patterns, and positive noise strengthens the remaining associated features. This selectively destroys target representations while preserving shared semantic structure, achieving controllable forgetting.