Cici: ipaai: XAI-driven data forensics for nuclear edge llms: provenance, detection, and recovery

The National Science Foundation Office of Advanced Cyberinfrastructure awarded Kansas State University $899,868 on November 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop an explainable artificial intelligence framework that secures large language models deployed at nuclear facilities.

The project builds interpretability methods that trace model outputs back to specific inputs and internal representations, then uses those signals to detect and remove poisoned or low-quality training data through machine unlearning. Hardware-based fingerprinting techniques confirm model origin and provenance. At inference time, layered defenses screen inputs, flag risky prompts, and validate outputs before recommendations reach operators. The team validates methods using operational data from a research nuclear reactor. Deliverables include openly shared tools, datasets, a fingerprint library, new university courses, and K–12 outreach to build workforce capability in security-aware technical practices.

The award runs through October 31, 2029, with place of performance in Manhattan, Kansas.

Abstract

Nuclear reactors supply a large share of reliable electricity, and reactor operators increasingly turn to large language models to help interpret procedures, sensor logs, and alarm messages in real time. Running these models on-site, close to the reactor, keeps sensitive data in-house and speeds decision-making, but IT also opens new ways for these systems to fail or be attacked. Poisoned or low-quality training data, foundation models that carry hidden backdoors, and manipulated prompts can each push a model toward confident but incorrect recommendations, a serious hazard that could threaten safety or force a costly shutdown. This project builds an explainable artificial intelligence framework that secures the entire workflow of nuclear-domain large language models. The team develops interpretability methods that trace a model's outputs back to specific inputs and internal representations, then uses those signals to detect and remove poisoned or low-quality training data through machine unlearning. To confirm a model's origin, hardware-based fingerprinting techniques are employed. At inference time, layered defenses screen inputs, flag risky prompts, and validate outputs before any recommendation reaches an operator. The methods are validated with real operational data from a research nuclear reactor. By establishing a unified approach to data integrity, model provenance, and output reliability, the project advances the safe deployment of large language models in critical infrastructure, protecting public welfare and national security. Openly shared tools, datasets, and a fingerprint library, along with new university courses and k?12 outreach, will help build a security-aware technical workforce. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are planned for this award.