PI-SAFE: Practical Privacy-Preserving LLM Inference With Adversarial Fine-Tuning for Optimized Utility

Cloud-based LLM services usually require users to send their text inputs unencrypted, which creates privacy risks. This paper presents PI-SAFE, a framework that protects privacy during LLM inference (the process of running a model to generate outputs) by splitting the model between client and server and sending obfuscated intermediate representations (partially hidden data states) instead of raw text, combined with adversarial fine-tuning (training the model with specially crafted examples to make it harder to attack) to prevent attackers from reconstructing the original input.