A Reliability Analysis of a Deep Neural Network

Deep Learning, and in particular its implementation using Convolutional Neural Networks (CNNs), is currently one of the most intensively and widely used predictive models for safety-critical applications like autonomous driving assistance on pedestrian, objects and structures recognition. Today, ensuring the reliability of these innovations is becoming very important since they involve human lives. One of the peculiarities of the CNNs is the inherent resilience to errors due to the iterative nature of the learning process. In this work we present a methodology to evaluate the impact of permanent faults affecting CNN exploited for automotive applications. Such a characterization is performed through a fault injection enviroment built upon on the darknet open source DNN framework. Results are shown about fault injection campaigns where permanent faults are affecting the connection weights in the LeNet and Yolo; the behavior of the corrupted CNN is classified according to the criticality of the introduced deviation.