Maestro: A New Platform for Building Reliable and Robust AI Systems
In April 2020, the National Science Foundation (NSF) issued a Dear Colleagues Letter, announcing its intention to “fund a small number of Early Concept Grants for Exploratory Research (EAGER) to encourage advances in cybersecurity education.” After reading about this opportunity for cybersecurity proposals, Assistant Professor Zhou (Joe) Li in UCI’s Henry Samueli School of Engineering contacted two colleagues in the Donald Bren School of Information and Computer Sciences (ICS) — Computer Science Professors Sameer Singh and Sergio Gago-Masagué. The trio put together a proposal, “Multi-Level Attack and Defense Simulation Environment for Artificial Intelligence Education and Research,” and last August, they were awarded $300,000 to develop and test their proposed simulation environment.
“This project will provide a platform — Maestro — that will help advance both AI cybersecurity research and education, while also allowing implementation of active learning strategies through gamification in AI project-based courses,” says Gago-Masagué. The team has spent the past year developing Maestro, which is now 80% complete, and they plan to test it during winter and spring 2022 in the course CS 175: Project in Artificial Intelligence.
“CS 175: Projects in AI presents a unique opportunity to implement these ideas,” says Singh, who has taught the course, which lets small groups of students develop AI and machine learning algorithms for use in Minecraft. “As AI and ML evolve quickly, this course lets us adapt to the changing demands, giving students a hands-on experience with the latest ML techniques that may not be covered in other courses.” The team is now redesigning CS 175 to use and assess Maestro.
Specifically, Maestro will help students learn how to build reliable and robust AI systems by letting them test their systems against specific attacks. By including a default library of potential attacks and defense strategies, Maestro will increase awareness of potential system vulnerabilities in a variety of domains. “Maestro will also facilitate the testing of customized attacks and defense strategies and provide valuable feedback,” says Gago-Masagué. “For example, Maestro can provide a gamification learning environment where users/students can join the attack or defense team to exploit or defend a given AI system.”
The goal is to better prepare organizations to address the growing number of cybersecurity threats amid the rapid evolution and expansion of AI and its applications. “Some groups and companies focus on specific applications based on their domains of interest, but the effort is costly and most of them require multidisciplinary teams working on strategies difficult to standardize to other domains,” explains Gago-Masagué. “We are missing an entity that can set standards and priorities in the AI cybersecurity domain. We are also missing a good educational strategy to make the future workforce aware of the risks and strategies in AI cybersecurity.”
As evidenced by the recent string of cybersecurity attacks in the U.S., the need for better understanding the threats and building appropriate defenses is great. This team aims to meet this need using an approach that, as outlined in their project abstract, applies “synergistic efforts in AI, cybersecurity and education that will produce significant research and societal impacts.”
— Shani Murray