Abstract: Fine-tuning large language models (LLMs) on down-stream tasks has become a standard approach to adapt their capabilities. However, the process raises privacy concerns when using sensitive ...