ManiF-SMC: New Manifold Forgetting Method Boosts Machine Unlearning

A new paper on arXiv presents ManiF-SMC (Manifold Forgetting with Self Mode Connectivity), an approximate machine unlearning technique that addresses key limitations of existing approaches. Traditional unlearning methods often rely on label manipulation or task-gradient reversal, which can degrade model performance and fail to match retraining-based unlearning. ManiF-SMC instead recasts unlearning as pushing each erased sample away from its original learned manifold representation centroid toward its nearest semantic neighbors in the retained data. This purely representation-space operation reduces dependence on labels and task-specific gradients, aligning unlearning behavior more closely with retraining.

The method combines a margin-based triplet loss for unlearning and representation preservation, along with a self-mode-connectivity module that rapidly reconstructs the local manifold to generate adaptive margins for each unlearning case. This avoids the difficulty of manually setting margins. Extensive experiments on four representative datasets show ManiF-SMC achieves unlearning effectiveness comparable to state-of-the-art approximate methods while operating solely within the model's representation space. This could enable more reliable and efficient data deletion for AI systems, supporting regulatory compliance like the right to be forgotten.