Secure Deep Neural Network Inference on TrustZone

GuardiaNN is a fast and secure on-device neural network (NN) inference framework for TrustZone-enabled devices. It secures the sensitive data of NNs by encrypting the data in main memory and by isolating the execution of the NNs within the secure world. Its key advantage is its rapid neural network execution speed; it minimizes the slow main memory accesses, leverages widely-available cryptographic hardware on the embedded devices, and overlaps the main memory accesses with NN operations. It is designed to be portable, and runs on any mobile/embedded devices implementing ARM TrustZone.

Cortex-A TrustZone, Cryptographic Hardware

ARM Platform with TrustZone, evaluated with STM32MP157DK2

Manufacturers, Embedded system developers

This framework can be ported to ARM Platform with TrustZone including cryptographic hardware and secure SRAM