Homomorphic inference of deep neural networks for zero-knowledge verification of nuclear warheads


Gabriel V. Turturica and Violeta Iancu [ https://www.nature.com/articles/s41598-023-34679-7 ]

In the context of nuclear warhead verification, this work proposes an algorithm that combines for the first time, nuclear resonance fluorescence, machine learning and homomorphic inference to achieve zero-knowledge verification. Two elements ensure the security of the protocol, the implementation of the template-based protocol at the architecture level, using Siamese networks, and the use of homomorphic encryption at inference time. The results of this work are presented in two parts: an extensive 2D example demonstrating the capability of Siamese neural networks to address the warhead verification problem and a single-point analysis highlighting the advantages and current limitations of homomorphic inference.

Validation data point. (a) Schematic representation of the Black Sea warhead. The warhead is composed of concentric spheres composed of HEU, HMX, and WGPu. (b) Monte Carlo simulation results for the Black Sea warhead. The image is obtained by integrating the energy histogram for all eight detectors. (c) Normalized energy histogram summed over the eight detectors for the pixel annotated in the middle panel.

Siamese network architecture for the 2D analysis.