MOEA-SISA: Multiobjective Optimization to Improve Model Performance during Forgetting Data

With the proliferation of shared personal data online, users encounter difficulties in revoking data access permissions and requesting data deletions, thus increasing the risk of privacy breaches. Machine unlearning offers a “Sharded, Isolated, Sliced, and Aggregated” (SISA) solution, but this method produces models with high complexity and limited generalization ability. To address these issues, this study develops a feature decomposition-based differential grouping method, which improves optimization efficiency through dynamic grouping of decision variables. Three optimization objectives are proposed namely, the model accuracy, model complexity, and generalization ability. By optimizing these three objectives, the trained model becomes closer to the retrained results while preventing excessive forgetting. The experimental models include ViT, VGG-16, and ResNet-50, which are used for sub-class forgetting tasks. Compared with various state-of-the-art methods, MOEA-SISA offers advantages across these models, especially in terms of model complexity and generalization ability. Tests against membership inference attacks verify that the MOEA SISA effectively retains accuracy and enhances the generalization ability in sub-class forgetting scenarios.