NSF NAIRR Pilot Award
NAIRR Pilot Award for Privacy-Preserving Medical AI
School of Computing Professor Receives NAIRR Pilot Award from the U.S. National Science Foundation for Privacy-Preserving Multimodal AI in Healthcare
CARBONDALE, Ill. — June 5, 2026
Dr. Alvi Ataur Khalil, Assistant Professor in the School of Computing and director of TITANS lab at SIU, has received a grant through the U.S. National Science Foundation (NSF)'s National Artificial Intelligence Research Resource (NAIRR) Pilot program for the project, "SplitMind: Multimodal AI via Split Learning for Privacy-Critical Visual Intelligence." The award provides a full allocation of 10,000 GPU-hours on the NCSA Delta supercomputer, one of the nation's leading AI computing systems.
Multimodal AI systems combine the ability to interpret visual information, such as X-rays, pathology slides, or retinal scans, with the ability to understand written language. These systems power emerging technologies including medical image analysis tools, clinical decision-support platforms, and visual question-answering assistants that can respond to questions posed by physicians about a patient's images. Although the underlying models have advanced rapidly, their deployment in healthcare and other sensitive domains remains limited due to a fundamental obstacle: modern AI training typically assumes that data from many institutions can be pooled in a central location. In healthcare, this assumption is incompatible with privacy regulations such as HIPAA and GDPR, and with the institutional duty to protect patient information.
SplitMind introduces a U-shaped split learning framework for transformer-based medical visual question answering, paired with a novel mechanism called Question-Guided Token Pruning (QG-TP). In this design, each participating institution keeps its raw medical images and diagnostic labels entirely on its own systems. Only a compressed, task-aware representation of each image, with question-irrelevant visual information removed before transmission, is sent to a shared central server for the heavy computation required by large transformer models.
The awarded computing resources will enable Dr. Khalil’s team to run large-scale transformer training, simulate adversarial privacy attacks such as gradient inversion and membership inference, and evaluate the system's robustness under realistic distributed conditions. This intensive testing process is essential for identifying weaknesses before such systems are deployed in real-world clinical environments. Ultimately, the project supports the development of AI systems that can be trusted in high-stakes healthcare applications, reinforcing national priorities in AI security, privacy, and trustworthy AI.
“SplitMind is designed to let hospitals and research institutions collaborate on AI models that none of them could build alone, without any of them having to send patient images or diagnostic labels outside their own walls," Khalil said. "By making the architecture itself privacy-aware, and by removing visual information that isn't needed for the task before it's ever transmitted, we can offer a stronger and more principled guarantee than approaches that protect data only at the network layer.”
— Dr. Alvi Ataur Khalil
Assistant Professor, School of Computing; Director, TITANS lab
The project is expected to contribute to broader efforts in trustworthy AI, supporting NAIRR Pilot focus areas in AI security, interpretability, and trust. Results will be disseminated through peer-reviewed AI venues and public release of code and benchmarks.
Media Contact: School of Computing Communications • socinfo@siu.edu