Machine Learning Approaches to Detect Safeguards-Relevant Activities and Spoofing in Facility Operations Data

Nuclear facility operators collect many types of operations data to support process control, nuclear safety, nuclear security, and nuclear material control and accountancy. Such data have long been of interest to the international safeguards community to support verification activities, although joint-use of operator data remains limited, partly due to the difficulty of assuring that data from operator instruments are reliable and have not been tampered with. Approaches to assure the integrity of such data could therefore be of interest for safeguards. This paper examines the utility of using machine learning (ML) approaches to detect illicit tampering in multimodal nuclear facility data. Using example nuclear facility data, the paper discusses the development of an algorithm for detecting operational signatures of interest. ML is then used to characterize how difficult it would be to conceal these signatures through spoofing without detection, assuming a subset of the operator instruments are authenticated. Finally, the paper discusses benefits, challenges, and implementation issues that would face these approaches if they are considered for IAEA deployment.