A Systematic Literature Review of Machine Unlearning Techniques in Neural Networks

Abstract

This review examines the field of machine unlearning in neural networks, an area driven by data privacy regulations such as the General Data Protection Regulation and the California Consumer Privacy Act. By analyzing 37 primary studies of machine unlearning applied to neural networks in both regression and classification tasks, this review thoroughly evaluates the foundational principles, key performance metrics, and methodologies used to assess these techniques. Special attention is given to recent advancements up to December 2023, including emerging approaches and frameworks. By categorizing and detailing these unlearning techniques, this work offers deeper insights into their evolution, effectiveness, efficiency, and broader applicability, thus providing a solid foundation for future research, development, and practical implementations in the realm of data privacy, model management, and compliance with evolving legal standards. Additionally, this review addresses the challenges of selectively removing data contributions at both the client and instance levels, highlighting the balance between computational costs and privacy guarantees.