NDSS 2025 – ASGARD

ASGARD is a virtualization-based TEE solution that protects on-device deep neural networks by securely extending TEE boundaries to include SoC-integrated accelerators via secure I/O passthrough, aggressively debloating the trusted computing base, and coalescing TEE-to-REE exits to minimize overhead. Implemented on an RK3588S Android platform with a Rockchip NPU without modifying proprietary vendor software, ASGARD aims to maintain compatibility and achieve model protection with a minimal TCB size and negligible inference latency.