Machine Unlearning: Concepts and Implementations

As "right to be forgotten" style regulations, data governance requirements, and security concerns expand worldwide, researchers and practitioners need methods that go beyond ad hoc retraining and provide effective deletion from models. Machine unlearning has emerged as a core capability for trustworthy artificial intelligence (AI), enabling trained models to remove the influence of specific data after deployment. This book offers a systematic, end to end guide to machine unlearning, from foundational problem formulations to practical design patterns for real world systems. It introduces the unlearning paradigm and key evaluation criteria, then presents a structured treatment of exact unlearning and approximate unlearning, highlighting when each is appropriate and what trade-offs arise in utility, efficiency, and reliability. A dedicated section on unlearning auditing and verification explains how to test and validate deletion claims, including protocol level schemes, model centric auditing approaches, and benchmark driven stress testing at scale. The book then extends unlearning to domain specific settings, covering graph unlearning, federated unlearning, and emerging techniques for large language models and diffusion models. Finally, it examines privacy and security risks such as leakage, backdoors, and poisoning, and surveys defenses and future directions for building dependable unlearning services. Written for graduate students, researchers, and engineers, the book provides a coherent taxonomy, practical insights, and a roadmap for developing, evaluating, and deploying unlearning in modern AI pipelines.