FedSVD mitigates noise amplification in differentially private federated fine-tuning with LoRA by periodically refactorizing LoRA matrices via SVD, improving stability and performance under DP-SGD.