Evaluation of the Impact of Random Computing Hardware Faults on the Performance of Convolutional Neural Networks

Artificial Intelligence (AI) rapidly spreads across high-tech industries and enters almost every safety-critical area such as automotive, aerospace, and medical industries. However, like any other software, AI-based applications are prone to random hardware faults such as a random bit flip in CPU, RAM, or network. Therefore, it is essential to understand how various hardware faults affect the performance and accuracy of AI applications. This paper provides a general description and particular conceptual and implementational features of our recently introduced Fault Injection (FI) framework InjectTF2. InjectTF2 is developed using the TensorFlow 2 API and allows the user to specify fault parameters and perform layer-wise fault injection into the TensorFlow 2 neural networks. It enables the automated injection of random bitflips. The paper describes the software architecture of the framework. The framework is open source and freely available on the GitHub. The application of InjectTF2 is demonstrated with extensive fault injection experiments on a Convolutional Neural Network (CNN) trained using the GTSRB dataset. The experiments' results show how random bitflips in the outputs of the CNNs layers affect the classification accuracy. Such results support not only numerical analysis of reliability and safety characteristics but also help to identify the most critical CNN layers for more robust and fault-tolerant design.