Revealing Statistical Trends while Maintaining Differential Privacy as Applied to a Reprocessing Facility

According to International Atomic Energy Agency’s (IAEA) STR-393, Development and Implementation Support Programme for Nuclear Verification 2020–2021, the agency has requested the exploration of data analysis methods and tools to strengthen the synthesis and evaluation of information (e.g., nuclear material flow analysis, near real-time accountancy, and process monitoring tools). This request is driven by the large number of declarations received by the IAEA on a daily basis, challenging its ability to detect undeclared activities in a timely manner. Their use of the Palantir platform is an important advance, but it partly addresses this challenge considering the rapidly changing field of artificial intelligence and machine learning, further increasing the volume of declarations and open-source information. This limitation adds an unknown risk to the detection of undeclared activities and could compromise the trust in the IAEA’s capabilities to meet its mandate in support of the NPT. Since the IAEA and member states have limited funding to invest in this area, employing commercial analytical services to assist in this effort could alleviate such data analyses demands. This project explores the application of differential privacy to inject encoded “noise” into a state’s declared data for a fictitious spent fuel reprocessing facility while not compromising the security of the data. This capability becomes even more important as the expansion of nuclear fuel cycle facilities continues worldwide while IAEA resources remain stagnant. Differential privacy is already used on databases containing sensitive information, such as the U.S. Census, allowing powerful private analytical engines to evaluate important statistics without compromising personal information. For future work, this project will be expanded to include differential privacy applications in a fictitious Gas Centrifuge Enrichment Plant (GCEP)