Abstract

Autonomous vehicles (AVs) rely on deep neural networks (DNNs) for critical tasks such as environment perception—identifying traffic signs, pedestrians, and lane markings—and executing control decisions like braking, acceleration, and lane changing. However, DNNs are vulnerable to adversarial attacks, including structured perturbations to inputs and misleading training samples that can degrade performance. This presentation begins with an overview of adversarial training, emphasizing the impact of input sizes on DNNs' vulnerability to cyberattacks. Subsequently, I will share our recent findings that explore the hypothesis that DNNs learn piecewise linear relationships between inputs and outputs. This conjecture is crucial for developing both adversarial attacks and defense strategies in machine learning security. The last part of the presentation will focus on recent work on using error-correcting codes to safeguard DNN-based classifiers.

Bio

Saif Jabari is a Global Network Associate Professor of Civil and Urban Engineering at New York University. His research interests center on theoretical aspects of traffic flow, specifically topics related to modeling uncertainty and emergent phenomena. The applications focus on traffic operations problems, including traffic state estimation and prediction, distributed traffic control, and cybersecurity.

Prior to joining NYUAD, Jabari was a Post-Doctoral Researcher in the Mathematical Sciences and Analytics Department at the IBM T.J. Watson Research Center in Yorktown Heights, NY. Jabari received his Ph.D. in Civil Engineering from the University of Minnesota, Twin Cities in 2012 and his B.Sc. degree in Civil Engineering from in the University of Jordan in 2001.

Related Info