A Fast Reliability Analysis of Image Segmentation Neural Networks Exploiting Statistical Fault Injections

The reliability of hardware running deep neural networks (DNNs) is becoming the object of multiple research works. Fault injections (FIs) are one of the most used solutions to determine the reliability of DNN models. However, defining how many faults to inject in the model is not a trivial task. An exhaustive FI campaign requires injecting, in modern DNNs, billions or trillions of parameters. On the other hand, random FI campaigns do not offer a practical measure of the accuracy of the result. A different approach is to perform a statistical FI: the number of faults to inject is decided based on the number of possible faults and by fixing an error margin and a confidence level on the measured output metric. While the statistical approach offers the best of both worlds, it requires a proper setup to guarantee its statistically significance. In this work, a study on the statistical fault injection procedure on an image segmentation neural network is proposed. In particular, the study compares results from a random FI campaign and an improperly-defined statistical FI campaign, and shows how they fail at highlighting some of the critical aspects of U-Net, a state-of-the-art DNN used for image segmentation. The proposed approach, by injecting only the 0.07% of all the possible faults, accurately measures both the criticality of each layer and of the parameters' bit with an error margin of 1% and a confidence level of 99%.