Machine Unlearning on Pre-trained Models by Residual Feature Alignment Using LoRA