Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation
arXiv:2604.06831v1 Announce Type: cross
Abstract: Current LLM-based services typically require users to submit raw text regardless of its sensitivity. While intuitive, such practice introduces substantial privacy risks, as unauthorized access may expo…