New Framework Boosts Machine Unlearning Efficiency and Accuracy

Current machine unlearning methods often rely on global, coarse-grained intervention strategies, lacking precise guidance signals to direct the unlearning process. This uniform approach leads to two significant problems: some data samples are "over-unlearned," negatively impacting model utility, while others are "under-unlearned," leaving residual information vulnerable to privacy attacks. The absence of differentiable guidance across diverse unlearning tasks further compounds these issues. To overcome these limitations, researchers introduce GSUO (Guidance-Signal-aware Unlearning Optimization), a novel framework designed to provide task-specific, fine-grained guidance signals. GSUO effectively steers the unlearning process, making it applicable to both random-subset and class-wise forgetting tasks. This targeted approach ensures that unlearning is neither excessive nor insufficient, striking a better balance between privacy and model utility. Extensive experiments demonstrated GSUO's superior performance compared to fourteen baseline methods. It achieved higher effectiveness in removing specific data, better generalization capabilities, and significant speedups. These results validate GSUO's effectiveness in delivering reliable machine unlearning, addressing critical concerns related to data privacy and model integrity.