Range Restriction to Harden CNNs Against Hardware Faults: A Broad Empirical Analysis

Due to the increasing use of Deep Learning in mission/safety-critical application contexts, in the recent past several techniques have been designed to harden the system to guarantee its correct behavior even in presence of faults affecting the hardware. Often, such new techniques are evaluated on a reduced set of Convolutional Neural Network (CNN) models and/or data sets, such that their generality and robustness could actually be limited. This paper presents a broad and systematic experimental evaluation of a state-of-the-art range restriction technique presented in literature, i) by applying it to a large set of CNNs, implementing different functional tasks, and ii) by using multiple datasets. The obtained results demonstrate that the effectiveness of the technique highly depends on the complexity of the considered task; in particular, classification CNNs benefit the most, while regression, image segmentation, and object detection are subject to different levels of benefits.