Dr. Ahmed Imteaj Awarded NSF NAIRR Pilot Grant to Advance Safer AI Systems

A research team led by Ahmed Imteaj, Ph.D., has received a grant through the National Science Foundation’s National Artificial Intelligence Research Resource (NAIRR) Pilot program for the project, “ARMOR-HAL: Unified Defense for Adversarial Robustness and Hallucination Mitigation in Large Vision-Language Models.” 

Imteaj, an I-SENSE faculty fellow, is director of FAU’s SPEED Lab and assistant professor in the Department of Electrical Engineering and Computer Science in the FAU College of Engineering and Computer Science. 

As part of the award, Purdue Anvil AI will provide 5,000 NVIDIA H100 Service Units — equivalent to 10,000 GPU hours — representing approximately $50,000 in advanced computing resources. This allocation enables the team to train and rigorously test advanced AI systems at a broader scale.

Large vision language models are artificial intelligence systems that can interpret visual information while simultaneously understanding written language. These models support technologies such as visual conversational assistants, medical image analysis tools, autonomous systems and decision support platforms. Although these systems have advanced rapidly, their deployment in real-world environments remains limited due to two major challenges: jailbreak attacks, which occur when users manipulate instructions to make an artificial intelligence system ignore its built-in safety protections, and hallucination, which happens when an artificial intelligence system produces answers that sound convincing but are not supported by the visual or textual information it receives.

“Current approaches tend to address jailbreak attack and hallucination challenges separately, which can create trade-offs between safety, reasoning ability and factual accuracy,” Imteaj said. “ARMOR-HAL introduces a unified defense framework that strengthens both adversarial robustness and visual grounding simultaneously.”

The awarded computing resources will enable the team to run large-scale stress tests, simulate adversarial attacks and retrain models with enhanced protection. This intensive testing process helps identify weaknesses before deployment in real-world environments. Ultimately, the project supports the development of AI systems that can be trusted in high-stakes applications, reinforcing national priorities in AI security, public safety and technological leadership.

“ARMOR-HAL is designed to help AI systems resist deception and remain honest about what they truly see by embedding safety parameters directly into their reasoning processes, guiding them to rely on what is actually visible in an image and fine-tuning how they prioritize important information,” Imteaj said. “Together, these improvements make the models harder to manipulate and less likely to generate inaccurate or ungrounded responses, ensuring they remain reliable even under stress or attack.”