TensorFI+: A Scalable Fault Injection Framework for Modern Deep Learning Neural Networks

Abstract

Deep Neural Networks (DNNs) are widely deployed in various applications such as autonomous vehicles, healthcare, space applications. TensorFlow is the most popular framework for developing DNN models. After the release of TensorFlow 2, a software-level fault injector named TensorFI is developed for TensorFlow 2 models, which is limited to inject faults only in sequential models. However, most popular DNN models today are non-sequential. In this paper, we are the first to propose TensorFI+, an extension to TensorFI to support for non-sequential models so that developers can assess resiliency of any DNN model developed with TensorFlow 2. For the evaluation, we conduct a large-scale fault injection experiment on 30 sequential and non-sequential models with three popularly used classification datasets. We observe that our tool can inject faults in any layer for any sequential or non-sequential DNN model, and fault-injected inference incurs only 7.62 x overhead compared to fault-free inference.