Federated Learning Beyond Training: Optimization, Personalization, Security, and Unlearning

Federated learning (FL) leaves training records with participating clients, but restrictions on the visibility and retention of model updates can limit later verification, robustness checks, and deletion. This review examines these dependencies across distributed optimization, personalization, parameter-efficient foundation-model adaptation, privacy, Byzantine robustness, and federated unlearning. For a specified actor and authorized observation, an exact downstream operation is obstructed when two admissible states produce identical observations but require different outputs. Aggregate-sum and merged low-rank examples establish this information limit, while a two-round example shows why replaying historical updates need not reproduce retraining without a deleted client. A conditional screening procedure examines information access, common assumptions, and declared resource budgets. A specified cross-silo case separates demonstrated obstructions, alternatives that change the available operation, and claims not established for complete protocols. This review distinguishes component-level proofs and experiments from evidence for joint implementations and identifies evaluation requirements under changing client distributions, correlated participation, adaptive attacks, and sequential deletion. Its conclusions are limited to a selected narrative literature and analytical examples; no general joint security or unlearning guarantee is claimed.