New Method Reduces Collateral Damage in Machine Unlearning.

Machine unlearning aims to remove the influence of specific training data from a model without requiring a full retraining, often using localization-based methods to target influential parameters. However, existing approaches typically focus only on the "forget-set" data, potentially causing unintended "collateral damage" to semantically similar data that should be retained. This paper addresses this by proposing a novel retain-aware localization method. The new method considers the importance of model parameters to *both* the forgotten and the retained datasets. To evaluate its effectiveness, the researchers also introduce a "retain-similar evaluation set," built using cosine similarity in the model's embedding space, to directly measure this collateral damage. Across multiple experiments on the CIFAR-10 dataset with a ResNet18 model, the proposed method consistently reduced collateral damage while improving standard unlearning metrics, demonstrating its effectiveness for more similarity-aware machine unlearning.