back

Systematic Analysis of Cyber-Physical Systems with Machine Learning Components

  • Date October 8, 2018
  • Hour 11.30 am
  • Room GSSI Main Lecture Hall
  • Speaker Tommaso Dreossi (University of California Berkeley)
  • Area Computer Science

ABSTRACT

Cyber-physical systems are starting to include sophisticated machine learning components. Recent efforts have shown that learning algorithms, that aim to generalize from examples, can produce inconsistent output under small adversarial perturbations. This raises the question: Can the output from learning components lead to a failure of the entire CPS? In this talk, we will present a framework for the systematic analysis of Cyber-Physical Systems with Machine Learning components (CPSML). The talk is structured in three main parts that roughly constitute our framework: 1) The specification, i.e., how the correctness of the system is formalized; 2) The testing algorithm, i.e., how meaningful inputs that lead the system to violate a specification are generated; 3) The environment, i.e., how the environment in which our CPSML operates is modeled. As a case study, we will focus on self-driving vehicles.


References:

- Data augmentation with counterexamples: https://arxiv.org/abs/1805.06962 
- Scenic: a language to generate simulation scenes: https://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-8.html