#15 No Privacy Left Outside Ziqi Zhang

Overview: Ziqi Zhang presents at IEEE S&P on "No Privacy Left Outside," research examining the insecurity of TEE-based model deployment for on-device LLM protection.

Key Topics: LLMs as expensive sensitive assets requiring protection on edge devices; TEE deployment to prevent white-box model access; model stealing and membership inference attacks against TEE-protected models; and comparative evaluation of TEE partitioning schemes.

Technical Details: The research evaluates prior TEE-based defenses across model stealing and membership inference metrics, finding some schemes enable attacks 4.26x more effective than black-box baselines. Low-privileged database users can exploit partitioning weaknesses.

Takeaways: TEE-based LLM deployment provides insufficient protection against model stealing and membership inference. Model owners must evaluate defense effectiveness against realistic attack scenarios before trusting edge deployment.