Fault Injectors for TensorFlow: Evaluation of the Impact of Random Hardware Faults on Deep CNNs

Today, Deep Learning (DL) enhances almost every industrial sector, including\nsafety-critical areas. The next generation of safety standards will define\nappropriate verification techniques for DL-based applications and propose\nadequate fault tolerance mechanisms. DL-based applications, like any other\nsoftware, are susceptible to common random hardware faults such as bit flips,\nwhich occur in RAM and CPU registers. Such faults can lead to silent data\ncorruption. Therefore, it is crucial to develop methods and tools that help to\nevaluate how DL components operate under the presence of such faults. In this\npaper, we introduce two new Fault Injection (FI) frameworks InjectTF and\nInjectTF2 for TensorFlow 1 and TensorFlow 2, respectively. Both frameworks are\navailable on GitHub and allow the configurable injection of random faults into\nNeural Networks (NN). In order to demonstrate the feasibility of the\nframeworks, we also present the results of FI experiments conducted on four\nVGG-based Convolutional NNs using two image sets. The results demonstrate how\nrandom bit flips in the output of particular mathematical operations and layers\nof NNs affect the classification accuracy. These results help to identify the\nmost critical operations and layers, compare the reliability characteristics of\nfunctionally similar NNs, and introduce selective fault tolerance mechanisms.\n