Robustness and Sensitivity Assessment of Deep Neural Networks

Nowadays, Deep Neural Networks (DNNs) are employed in many fields to perform different tasks, such as autonomous driving, always listening, augmented reality, computer vision and many others. Some of the listed tasks, as the autonomous driving, are safety-critical. Aware of that, ensuring the reliability of this technology is very important since, an error of assessment could endanger human lives. DNNs are known to be intrinsically error resilient since, their topology exploits data redundancy. Therefore, even if an error occurs the DNNs might be still capable to correctly predict the result. However, this may change whenever the DNNs are implemented in Hardware (HW) for the reason that, the target HW have limited processing resources in which multiple neurons are mapped and potentially be corruptible by a single fault. The aim of this master thesis is to study novels Hardware Description Level (HDL) fault injectors able to reduce the expensive time cost of the state-of-art HDL fault injectors. The attention was drawn onto permanent and transient fault models. Indeed, the thesis will provide a pipelined fault injector for permanent fault and an optimized fault injector for transient fault. Both fault injectors were tested in fault injections campaign to asses their effectiveness. Such frameworks make feasible the reliability analysis at the HW level of systems based on neural network, reducing the required simulation time of at least 60%.