The advent of artificial intelligence in scientific research has brought about transformative changes, yet with these advancements come significant security challenges. Traditional cybersecurity measures often fail to address vulnerabilities in AI models, datasets, and automated systems that are essential for research. The VERITAS project, formally known as the Verified Infrastructure for Trustworthy AI in Science, is tackling these issues head-on by integrating AI Assurance into the scientific research framework.
Led by Anita Nikolich, the VERITAS initiative has received an $896,000 grant from the US National Science Foundation’s Cybersecurity Innovation for Cyberinfrastructure program. This project brings together experts in adversarial AI, research cyberinfrastructure, data science, and workforce development to address AI-specific security challenges in scientific workflows. A major component of VERITAS is the creation of model cards and dataset datasheets that document the origins, development, and limitations of AI models and datasets. This documentation is crucial for ensuring traceability and accountability throughout research workflows.
A key innovation of the project is the introduction of the AI Assurance Engineer role. These professionals will assess emerging AI projects prior to deployment, examining model files for malicious behavior, evaluating software vulnerabilities, and reviewing the autonomy of AI agents. This role is currently being piloted at the National Center for Supercomputing Applications.
Training is another cornerstone of VERITAS, with hands-on modules available through the National Data Platform Education Hub. These modules train students to identify compromised data, inspect models for anomalies, and detect weaknesses in scientific AI pipelines, emphasizing responsible disclosure practices.
Additionally, VERITAS is implementing red-teaming practices within scientific research, a strategy typically used in the industry but seldom seen in research infrastructure. By proactively testing systems for vulnerabilities, the project aims to identify weaknesses before flawed models become integrated into research processes. The initiative aspires to make AI Assurance a standard part of research cyberinfrastructure, ensuring the security and reliability of AI systems in scientific applications.