A Lightweight Memory Protection Scheme with Criticality-Aware Encryption and Embedded MAC for Secure DNN Accelerators

As deep neural network (DNN) accelerators have been used for various mission-critical applications, ensuring the security of these accelerators has become a pivotal concern. This paper proposes a lightweight and efficient memory protection scheme for secure DNN accelerators. We leverage the data characteristics of DNNs to reduce the overhead of memory protection while providing robust security features. Specifically, we propose a criticality-aware bit position encryption (CBPE) method and an embedded message authentication code (EMAC) method to lower the overhead of memory encryption and integrity verification. Experimental results demonstrate that the proposed memory protection scheme can defend against most memory-related attacks with very small overhead in performance ($\sim 1 \%$), area (${\lt}1 \%$), and power (${\lt}2 \%$) even when deployed on a small energy-efficient DNN accelerator.