How the VERITAS Project is Revolutionizing AI Security for Scientists

Post Views: 12

Artificial intelligence models, datasets, and automated systems critical to research are vulnerable to threats that traditional cybersecurity measures fail to identify.

The VERITAS Initiative

The VERITAS (VERified Infrastructure for Trustworthy AI in Science) project seeks to address gaps in AI security by integrating AI Assurance into scientific research infrastructure. Spearheaded by Anita Nikolich, the initiative has secured a $896,000 grant from the US National Science Foundation’s Cybersecurity Innovation for Cyberinfrastructure program. It unites experts in adversarial AI, research cyberinfrastructure, data science, and workforce development to tackle security challenges specific to AI-driven scientific workflows.

Core Components of VERITAS

Standardized Documentation Practices

The project focuses on creating model cards and dataset datasheets to document the origins, development, and limitations of AI models and datasets. These records enable traceability of issues through research workflows, ensuring transparency and accountability.

AI Assurance Engineer Role

A new role, the AI Assurance Engineer, will evaluate emerging AI projects before deployment. Responsibilities include analyzing model files for malicious behavior, assessing software vulnerabilities, and reviewing AI agent autonomy. This role is piloted at the National Center for Supercomputing Applications.

Educational Training Modules

VERITAS develops hands-on training through the National Data Platform Education Hub. Modules teach students to identify compromised data, inspect models for anomalies, and detect weaknesses in scientific AI pipelines. These exercises emphasize responsible disclosure practices.

Red-Teaming in Scientific Research

VERITAS introduces red-teaming practices to scientific research, a method commonly used in industry but rarely applied to research infrastructure. By systematically testing systems for vulnerabilities, the project aims to uncover weaknesses before flawed models or agents become embedded in research processes.

Workforce Development and Education

Students participating in training challenges engage with scientific models, datasets, and infrastructure via the National Data Platform. This experience prepares them for responsible disclosure and fosters a workforce capable of addressing AI security challenges in research.

Conclusion

If successful, VERITAS could establish AI Assurance as a standard element of research cyberinfrastructure. This framework would enable institutions to evaluate the security and reliability of AI systems before integrating them into critical scientific processes.

“Conventional cybersecurity approaches cannot adequately protect AI systems used in research. Poisoned datasets or backdoored models may generate outputs that appear valid but contain subtle errors, evading detection by standard safeguards like firewalls or antivirus software.”