Researchers have published multiple approaches to address privacy concerns in large language model (LLM) deployments, where users typically must transmit sensitive prompts to cloud servers.
According to arxiv.org, a paper titled “Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation” introduces Privacy-Preserving Fine-Tuning (PPFT). The method operates in two stages: training a client-side encoder with a server-side projection module that allows servers to process k-pooled prompt embeddings instead of raw text, then fine-tuning on private data using noise-injected embeddings. The researchers report that PPFT “achieves a striking balance between privacy and utility, maintaining competitive performance with minimal degradation compared to noise-free upper bounds.”
Separately, arxiv.org describes ConfusionPrompt, another privacy framework that works by “decomposing the original prompt into smaller sub-prompts” and “generating pseudo-prompts alongside the genuine sub-prompts” sent to the LLM. According to the source, this approach “integrates seamlessly with existing black-box LLMs” and demonstrates “significantly higher utility than local inference methods using open-source models and perturbation-based techniques, while also reducing memory consumption.”
Both papers appear in the Cryptography and Security and Artificial Intelligence subject categories on arxiv.org, addressing the challenge of protecting user data in cloud-based LLM services.