Towards Privacy-Preserving and Secure Machine Unlearning: Taxonomy, Challenges and Research Directions

Liou Tang, James Joshi

Abstract: Machine Unlearning (MU) is a growing sub-field of Machine Learning (ML) that aims to update ML models efficiently following users’ requests to remove training data without retraining of the original ML model. While MU provides novel ways to preserve user privacy and ensure model integrity through the removal of compromised data, or the data that has been requested to be removed for privacy reasons, various privacy and security vulnerabilities introduced or attacks made possible by the MU process itself remain largely unexplored. In this paper, we examine challenges of Privacy-Preserving and Secure Machine Unlearning through a systematic and formal taxonomy of existing and possible threat models towards Machine Unlearning, and outline future directions for Machine Unlearning research in privacy and security.