Project Grant 2526620

The National Science Foundation's Division of Computer and Network Systems awarded a $410,000 Project Grant to Columbia University on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop security verification methods and tools for pre-trained artificial intelligence models shared through open online repositories. The research team will deliver novel approaches to address three critical vulnerabilities in the machine learning model supply chain: (1) detection and mitigation of model spoofing attacks where adversaries upload malicious models under deceptive names using anomaly detection schemes and architectural signature analysis; (2) securing the model deserialization process through automated, least-privilege mechanisms and safe subset definitions to prevent arbitrary code execution; and (3) additional security controls for model selection, loading, and management that integrate software engineering principles with machine learning techniques.

The project, which extends through September 30, 2030, will advance the security of AI infrastructure supporting the U.S. research enterprise and AI-enabled systems across government and industry. Deliverables include methods and tools enabling users to verify model trustworthiness before integration into scientific workflows and operational systems, alongside educational and workforce development activities including student training, research opportunities, and collaboration among universities, industry, and other stakeholders to strengthen the nation's competitiveness in secure artificial intelligence development and deployment.