Today, AI-based system is developing rapidly, for example, some of Autonomous Vehicles are already providing servics on public roads. However, as we all know, AI components can be easily fooled by adversarial examples and there have also been many failures of CPS system in the real world for which the evidence points to a failure (in part) of AI-based perception. Our goal is to build a platform for comprehensive security analysis of AI-Stack of Autonomous Vehicles, Robot, UAV and other AI-based CPS systems. Specifically, our AI Security Platform can realize CPS perception module simulation-based testing based on adversarial samples and automated fuzzing in virtual testing.

Available Work

• DeVer:Platform includes state-of-the-art adversarial examples, defense and robustness verification method of image-inputed classify models at the digital level.
  
• VerifAuto:Autonomous driving automated fuzzing toolkit for generating scenarios using SCENIC in LGSVL.
  

Ongoing Work

• DeVer extension of target detection model
  
• Extension of DeVer under simulation conditions
  
• VerifAuto's expansion into robotics and UAV testing