Seminar: Automated Black-Box Testing of Deep Learning Systems

Abstract

    Major breakthroughs in Deep Learning (DL) algorithms have brought huge success in many application domains such as image recognition, security, and autonomous driving. However, despite tremendous progress, DL-based systems often exhibit erroneous behavior that can lead to potentially critical errors. This raises the question of how we can efficiently test DL-based systems, especially in the context of safety-critical systems, to ensure their correctness and compliance with safety requirements. However, testing techniques for traditional software systems cannot be directly applied to DL-based systems due to the significant differences in terms of internal architecture and programming paradigms. Moreover, most existing works on testing DLbased systems require the access to the internals of DL models, which is often not feasible in many practical contexts.

This talk will focus on the challenges of testing DL-based systems and our automated black-box testing solutions for such systems.

Bio

Dr. Manel Abdellatif is an institutional researcher at École de Technologie Supérieure. Prior to her current position, she was a postdoctoral fellow at the School of EECS, University of Ottawa. She received her PhD in Computer Science from Polytechnique Montreal and her master’s degree in Information Technology from École de Technologie Supérieure. Her research interests include testing machine learning-based systems, service computing, and empirical software engineering. She received several best paper awards and her research works have been published in top venues such as TSE, JSS, ICSOC, and ASE.