CISE Seminar: April 26, 2019 – James Kapinski, Toyota Research Institute, North America

Cyber-physical systems (CPSs) are used in many mission-critical applications, such as airplanes and medical devices, and so these systems are required to work correctly under all conditions. New CPS designs include artificial intelligence (AI)-enabled components, such as deep neural networks (DNNs), which enable them to perform tasks like autonomous driving; however, verifying correct behavior of AI-related aspects of CPS designs is difficult. New development and analysis techniques are needed to increase confidence in the reliability of AI-enabled CPS.

This talk presents new approaches to address these challenges, including a method that provides a formal proof of correctness for the system by learning candidate safety certificates from simulation traces, which are then confirmed using automated reasoning tools. The talk additionally touches on an extension of this approach to perform training of the AI-enabled components in a manner that is guaranteed to satisfy safety properties. Also, the talk will present a falsification-based approach, which uses best-effort methods to automatically identify system behaviors that fail to satisfy requirements. The talk concludes with a discussion of open problems and next steps for test and verification of AI-enabled CPS.