Signal-Guided Optimization for Machine Unlearning

Machine unlearning aims to remove the influence of specific training samples from a model without the prohibitive cost of retraining from scratch. Existing methods often use coarse-grained, uniform strategies that treat all samples equally, leading to a trade-off: some samples are over-unlearned, damaging model utility, while others are under-unlearned, leaving the model vulnerable to privacy attacks. This paper asks: can we design fine-grained, task-specific guidance signals to steer the unlearning process more precisely?

The authors propose the Guidance-Signal-aware Unlearning Optimization (GSUO) framework, which operates in two steps: Signal-Guided Optimization (SGO) and Compensatory Optimization (CO).

For random-subset forgetting, GSUO categorizes samples into 'Normal' (high-confidence) and 'Boundary' (low-confidence) groups. It then applies target distribution signals: Normal samples are guided toward a uniform distribution to reduce confidence, while Boundary samples are aligned with the predictive distribution of the most similar retained class. For class-wise forgetting, GSUO uses feature redistribution: it applies an intra-class dispersion loss to break up the clustering of the target class and an alignment loss to shift the forget class features toward the global center of the retained data. Finally, CO fine-tunes the model on the remaining data to restore performance.

GSUO consistently outperforms 14 state-of-the-art baselines across multiple datasets (CIFAR-10, CIFAR-100, Lacuna-10, Tiny-ImageNet) and architectures (ResNet, VGG, ViT). On random-subset tasks, it achieves the smallest accuracy gap between the forget set and the test set, indicating that the model behaves as if it never saw the forgotten data. It also demonstrates significant speedups (up to 31x) compared to retraining. In class-wise forgetting, GSUO achieves complete unlearning while maintaining superior accuracy on retained classes compared to existing methods.