The threat of Hardware Trojans looms large on safety-critical systems. A Design-For-Trust technique to mitigate this threat without significant loss in performance is to implement these systems as a Heterogeneous Secure System $-HSS$. An HSS is built using an array of trustworthy home-grown cores and untrusted but fast third-party cores in a way that prevents unverified results from third-party cores reaching IO peripherals and devices. In this work, we propose to use the unverified results to initiate a speculative execution of subsequent layers of a Neural Network (NN) application on trustworthy cores. Our experiments on six popular NN applications show that on an average, the secure execution on an HSS is slower than the corresponding untrusted execution by up to 6.26% as compared to the slowdown of 80.89% experienced by a conventional trustworthy system.
Paper
Full text
SANNA: Secure Acceleration of Neural Network Applications
Semantic Scholar · Computer Science · 2023
Abstract
The threat of Hardware Trojans looms large on safety-critical systems. A Design-For-Trust technique to mitigate this threat without significant loss in performance is to implement these systems as a Heterogeneous Secure System $-HSS$. An HSS is built using an array of trustworthy home-grown cores and untrusted but fast third-party cores in a way that prevents unverified results from third-party cores reaching IO peripherals and devices. In this work, we propose to use the unverified results to initiate a speculative execution of subsequent layers of a Neural Network (NN) application on trustworthy cores. Our experiments on six popular NN applications show that on an average, the secure execution on an HSS is slower than the corresponding untrusted execution by up to 6.26% as compared to the slowdown of 80.89% experienced by a conventional trustworthy system.