A Survey on Federated Unlearning: Lifecycle, Taxonomy, and Insights

Federated Unlearning (FU) has emerged as a crucial extension of Federated Learning (FL), driven by stringent data privacy regulations, such as the Right to Be Forgotten (RTBF). While algorithmic research has progressed significantly, existing literature often treats unlearning as an isolated post-processing task, neglecting critical dependencies across the model's entire lifecycle. This fragmented view ignores the systemic dependencies of unlearning, specifically its reliance on the initial learning architecture and the subsequent verification requirements, leading to suboptimal solutions. To address this limitation, this survey proposes a holistic lifecycle perspective, structuring FU into a comprehensive three-phase workflow: the learning phase, unlearning phase, and verification phase. Within this framework, we provide a systematic taxonomy of state-of-the-art techniques. For the learning phase, we analyze proactive preparation strategies. The unlearning phase begins with the unlearning targets and unlearning participants to define the request's scope, followed by the execution of core algorithms, which we deconstruct into elimination, preservation, and enhancement modules. For the verification phase, we categorize mechanisms for checking unlearning success into alignment-based and attack-based paradigms. We then systematically compare existing FU algorithms based on this novel workflow. Furthermore, we comprehensively review practical applications and identify key future directions.