2023-2024 New Frontiers Fellows

Veronica Diaz-Pacheco

North Carolina State University, College of Engineering

Research Topic

Analysis of the Grid’s Vulnerability to Coordinated Cyberattacks

Research Summary

The reliable operation of the U.S. bulk electric power grid is crucial to our safety and economic well-being. Due to ongoing efforts to integrate information, communication, and sensing technologies into our critical infrastructures, we can now think of the grid as a complex web of physical and cyber components working in unison to deliver electricity to our homes. The benefits of these technologies are clear: remote control of operations, real-time availability of data, etc. What is not clearly understood is the extent to which introducing these devices can increase the risk of disruptions to power generation and transmission. Towards this effort, this work will study the resilience of the U.S. bulk electric power grid to cyberattacks. More specifically, using a bi-level optimization model, this work aims to identify the sets of cyberattacks whose impacts result in worse-case scenario losses of service in terms of increases in operating costs and potential for load shedding. The results will help gain insight to improve risk management decision-making and design power systems robust against the growing threat of cyberattacks.

______________________________________________________________________________________________________________________________

Vlas Zyrianov

University of Illinois, Department of Computer Science

Research Topic

Asset-free LiDAR Simulator to Train Robust and Safe Autonomy

Research Summary

Self-driving cars and other autonomous systems are typically trained in simulators before being deployed in the real world. Simulations provide an opportunity to safely collect large amounts of training data, especially of rare or dangerous situations (e.g., a pedestrian crossing in front of a car). However, creating realistic simulators is costly because it requires designing large amounts of detailed 3D models and environments manually. In our paper “Learning To Generate Realistic LiDAR Point Clouds” (published at ECCV 2022), we have shown that generative modeling (using a technique similar to the one used in Stable Diffusion and DALL-E 2) can be applied to 3D LiDAR scans (which are readily available in datasets) to generate scans of 3D environments with realistic layouts and cars. Using the Delta system, I plan to utilize computation resources to train an improved generative model on LiDAR data and to investigate the potential applications that such a model can bring (e.g., training safer and more reliable autonomous vehicles in an asset-free simulator, efficiently storing a highly detailed 3D model of our world, localization tasks, etc.).