An Emulation Platform for Evaluating the Reliability of Deep Neural Networks
Corrado De Sio, Sarah Azimi, Luca Sterpone · 2020
In recent years, Deep Neural Networks have been increasingly adopted by a wide range of applications characterized by high-reliability requirements, such as aerospace and automotive. In this paper, we propose an FPGA-based platform for emulating faults in the architecture of DNNs. The approach exploits the reconfigurability of FPGAs to mimic faults affecting the hardware implementing DNNs. The platform allows the emulation of various kinds of fault models enabling the possibility to adapt to different types, devices, and architectures. In this work, a fault injection campaign has been performed on a convolutional layer of AlexNet, demonstrating the feasibility of the platform. Furthermore, the errors induced in the layer are analyzed with respect to the impact on the whole network inference classification.