Research map / Architectures and efficiency / RNNs and normalization
LoRO: Real-Time on-Device Secure Inference for LLMs via TEE-Based Low Rank Obfuscation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LoRO, a TEE-based low-rank obfuscation scheme for real-time on-device secure LLM inference, motivated by a statistical weakness in existing protection methods that enables model stealing.
Problem
LLMs deployed on untrusted edge devices risk model theft; existing TEE-based secure inference methods have a statistical vulnerability that undermines their security guarantees.
Approach
Identifies a statistical vulnerability in existing TEE-based protection methods, shows it can be exploited by a model stealing attack, and proposes a low-rank obfuscation scheme executed within a TEE to protect model weights while keeping inference real-time on device.
Surprising findings
- Existing TEE-based protection methods have a statistical vulnerability that compromises their security guarantees, as shown by a model stealing attack.
Closest papers by meaning
- BiShield-TEE: On the (In-)Security of Unilateral Weight Obfuscation in On-Device TEE-Shielded LLM Partition (NeurIPS 2026 · similarity 0.90)
- GroupCover: A Secure, Efficient and Scalable Inference Framework for On-device Model Protection based on TEEs (ICML 2024 · similarity 0.87)
- CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment (NeurIPS 2025 · similarity 0.87)
- TensorCommitments: A Lightweight Verifiable Inference for Large Language Models (NeurIPS 2026 · similarity 0.87)
- Fingerprinting Inference Systems of Large Language Models (NeurIPS 2026 · similarity 0.86)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.