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

Summary

PI-SAFE is a framework for privacy-preserving LLM inference that splits a pre-trained model between client and server, so clients send obfuscated intermediate representations instead of raw text and the server-side model needs no changes. It adds Dual-constraint Adversarial Fine-Tuning (AdvFT) and a random prefix-based obfuscation (PI-SAFE+) to resist reconstruction attacks. The authors report that PI-SAFE+ cuts the success rate of optimization-based reconstruction attacks to near zero and drives attribute inference to random-guessing levels.

Related items

• MediumCVE-2026-125386: langchain4j-aideepin insecure direct object reference in message endpointsSimilar attack · NVD/CVE Database
• InfoSNAPRAG: Cost-Efficient Knowledge Extraction Attack on RAG Systems via Adversarial PerturbationsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• LowMGE: A Membership Guided Data Extraction Attack against Diffusion ModelsSimilar attack · Arxiv (cs.RO + cs.CV security)
• InfoSecuring Locally Deployed LLM-RAG Assistants: A Cross-Model and Cross-Lingual Evaluation of Indirect Prompt Injection DefensesSimilar attack · OpenAlex (peer-reviewed AI security)
• InfoGem5-Shark: Smaug-based Heterogeneous Deep Learning Accelerator Security and Robustness Evaluation and Characterization FrameworkSimilar attack · OpenAlex (peer-reviewed AI security)