- This event has passed.
TEE-Shielded LLM Partition – Efficiently Protect Model Security on User Devices
Title: TEE-Shielded LLM Partition – Efficiently Protect Model Security on User Devices
Speaker: Ziqi Zhang
Date and Time: Monday, May 19, 11:00am
Location: EH 2430
Abstract:
In this talk, I will introduce recent work on TEE-based LLM protection (TSLP) on the user devices: TEESlice (S&P24) and ArrowCloak (Security25 revision). Existing TEE products provides high security guarantee, but usually lack GPU support and introduce additional inference latency. TSLP aims to incorporate untrusted GPU into TEE to reduce computation overhead. TSLP partition a LLM into two parts, offloading the privacy-insensitive part to the GPU while shielding the critical privacy-sensitive part within the TEE. In TEESlice, we systematically review and benchmark existing TSDP solutions and reveal that offloading model weights in plain-text leaks model privacy. Based on the observation, TEESlice uses a partition-before-training strategy for accurate separation between privacy-related weights from public weights. In ArrowCloak, we review existing lightweight weight obfuscation solutions in TSLP and find that these solutions cannot effectively protect vector direction in the weight matrix. We use a proof-of-concept attack to exploit this vulnerability and propose ArrowCloak that uses vector-matrix multiplication to protect vector direction.
Biography:
Ziqi Zhang is currently a Postdoc Researcher at University of Illinois Urbane-Champaign, working with Lingming Zhang. In 2023, he obtained Ph.D degree in the School of Computer Science from Peking University in China. Before that, he obtained the bachelor degree in the School of EECS of Peking University in 2018. His research interests lie in AI security, software engineering, and software security. His interested topics include trusted AI, AI with trusted hardware, and AI with SE. He has published several papers at top-tier conferences in various domains, including computer security (S&P, USENIX Security, CCS), software engineering (ESEC/FSE, ICSE, ISSTA, ISSRE), AI (ICML, WWW), and ubiquitous computing (Ubicomp). More information can be found at https://ziqi-zhang.github.io/
Hosted By: Prof. Zhou Li