Google have their "Machine Unlearning" challenge to address this specific issue - removing the influence of given training data without retraining from scratch. Seems like a hard problem.
https://blog.research.google/2023/06/announcing-first-machin...