Towards Protecting On-Device Machine Learning with RISC-V based Multi-Enclave TEE

Abstract

On-device machine learning is a trending paradigm that empowers the artificial intelligence of various smart devices, including IoT, mobile, and robotics, etc. On the other hand, this emerging paradigm has brought new security challenges that traditional system security techniques cannot harness. In this paper, we explored the possibility of using multi-enclave Trusted Execution Environments to address these security challenges. We first identified the challenges and threats that on-device machine learning systems are facing. Then, we presented our experimental results of using RISC-V-based multi-enclave TEE to secure on-device machine learning system, demonstrating a promising performance advantage. Finally, we discussed the technical directions for the threats that cannot be completely addressed by the TEE.