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 solution and is useful for integrated sensing and computing integrated chips and systems, but the Sharded, Isolated, Sliced, and Aggregated (SISA) method as one approach produces models with high complexity and limited generalization ability. This study proposes a novel framework, named MOEA-SISA, which integrates feature decomposition-based differential grouping (FDbDG) to improve optimization efficiency through dynamic grouping of decision variables. Three optimization objectives can hence be proposed: 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. Experimental models including ViT, VGG-16, and ResNet-50 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 demonstrate that MOEA-SISA effectively retains accuracy and enhances the generalization ability in sub-class forgetting scenarios. Additionally, the proposed approach offers significant advantages for improving the efficiency and performance of integrated sensing and computing chips and systems.