Proceedings of the
The 33rd European Safety and Reliability Conference (ESREL 2023)
3 – 8 September 2023, Southampton, UK
SynthiCAD: Generation of Industrial Image Data Sets for Resilience Evaluation of Safety-Critical Classifiers
IAS, University of Stuttgart, Germany.
ABSTRACT
Due to their versatility, Deep Neural Networks are becoming increasingly relevant for the industrial domain. However, there are still challenges hindering their application, such as the lack of high-quality training data and suitable methods for assessing their robustness to internal computing hardware faults in safety-critical applications. To address these challenges, this paper introduces (i) a new data generation tool SynthiCAD for creating customisable image training data, along with an open-source industrial data set for classification generated by SynthiCAD. In addition, (ii) we categorized and compared existing approaches to fault injection and evaluated software-based fault injection using a VGG19 model trained on our new data set. Our findings show that software-based fault injection is a fast and scalable way to assess the reliability of DNNs under the presence of faults.
Keywords: Fault injection, Robustness, Resilience, DNN, Synthetic data, Computer vision.
