Evaluating Resilience Of Deep Learning Models

Deep learning applications have become a valuable tool to solve complex problems in many critical areas. It is important to provide reliability on the outputs of those applications, even if failures occur during execution. In this paper, we present a reliability evaluation of three deep learning models. We use an ImageNet dataset and a homebrew fault injector to make all the tests. The results show there is a difference in failure sensitivity among the models. Also, there are models that despite an increase in the failure rate can keep the resulting error values low.