VERITAS: Securing AI in Scientific Research

A new initiative called VERITAS aims to improve the security of AI models and datasets used in scientific research. The project addresses vulnerabilities that traditional cybersecurity tools often miss. VERITAS, which stands for VERified Infrastructure for Trustworthy AI in Science, received an $896,000 grant. It is led by Anita Nikolich from the University of Illinois School of Information Sciences. Here's the thing: The project focuses on integrating AI Assurance directly into existing research infrastructure. This means scientists won't need to become cybersecurity experts, and cybersecurity teams won't need to become machine-learning specialists. The plan involves three key efforts. First, standardized documentation, like model cards and dataset datasheets, will trace the origin and purpose of AI components. Second, a new role, the AI Assurance Engineer, will review novel AI projects for security risks. Third, educational challenges will train students to identify compromised data and vulnerabilities in AI pipelines. What's interesting is that VERITAS is also bringing "red-teaming," a method of deliberately probing systems for weaknesses, into scientific research computing. This is a practice common in the tech industry but rarely applied to scientific AI. The bottom line: This project seeks to ensure the integrity and trustworthiness of AI systems that are critical for scientific discovery.