A Novel Model-Knowledge Transfer Approach for Effective Machine Unlearning and Performance Preservation

Machine unlearning has emerged as a vital approach to addressing data privacy and security concerns in the information age. This study presents a novel model-based machine unlearning algorithm, designed to operate under the assumption of complete model knowledge without requiring access to the data designated for removal. By transferring specific model layers and employing targeted data knowledge migration techniques, the algorithm effectively mitigates the influence of the data identified for unlearning. Experimental validation across two commonly used strategies demonstrates the algorithm's effectiveness, supplemented by an exploration of model usability through data augmentation. Additionally, the study investigates how the correlation between new data and the original dataset affects model accuracy post-unlearning.