SUV: Scalable Large Language Model Copyright Compliance with Regularized Selective Unlearning

Since DPO may hinder the LLM's performance in other unrelated tasks, we integrate gradient projection and Fisher information regularization to mitigate the degradation.