On-device deployment of large language models is increasingly adopted for latency and privacy reasons, but it exposes proprietary model weights to reverse engineering and intellectual property theft. Trusted execution environments (TEEs) are often used to mitigate these risks, yet they are not a complete solution.

The paper introduces BiShield-TEE, a scheme that shields LLM partitions with a TEE and adds weight obfuscation as an extra layer of defense. However, the authors demonstrate that this unilateral obfuscation is insecure, meaning an attacker can recover the original weights despite the protection.

The findings underscore the difficulty of protecting model IP on-device and suggest that stronger, multi-layered defenses are necessary. The full analysis is available on arXiv.