Fine-Grained and Efficient Self-Unlearning with Layered Iteration

As an increasing number of enterprises utilize machine learning (ML) models trained with customer data for decision-making and service provision, various privacy laws and regulations, such as the General Data Protection Regulation (GDPR), have been established to protect individual privacy rights, including the ‘right to be forgotten.’ Machine unlearning techniques can fulfill this right by removing specific data instances from ML models upon request. However, state-of-the-art unlearning methods, which often rely on adversarially misclassified data, lead to high computational costs and diminish the model’s utility by neglecting the broader context of retained data. These methods are typically designed for removal requests in a single scenario, limiting their versatility across different scenarios and data sizes. In this thesis, we address these challenges and propose a novel unlearning method, ‘Self-Unlearning with Layered Iteration’ (SULI), which operates uniquely by requiring only the target model and the requested data. To effectively diminish the influence of the requested data, we propose the ‘Soft Label Redistribution’ (SLR) technique to relabel the requested data. SLR modifies the target model’s output distribution for the requested data to resemble that of a model trained without encountering it, thereby treating the data as unseen. Our method have demonstrated superior efficiency, scalability, and effectiveness through rigorous evaluations against various state-of-the-art baseline methods, confirming their robust applicability across multiple scenarios. Moreover, to the best of our knowledge, SULI is distinguished as the most computationally time-effective unlearning method.