Abstract: It is difficult for the current arms control verification technology to draw accurate conclusions about the authenticity of nuclear weapons without detecting sensitive information. By combining the physical mask encryption technology and the K-nearest neighbor algorithm, this work proposes a verification system that is capable of independently encrypting and identifying nuclear weapon identity information. A radiation fingerprint collection device based on neutron fission reaction is built using Geant4 and a database is constructed by building samples under a variety of cheating scenarios. Furthermore, this study employs a KNN algorithm for establishing a machine learning model for identifying unknown items, and measures the robustness and security of the verification system. The results show that when the sample isotopic abundance changes from weapons-grade uranium to lower-level enriched uranium(the abundance of ²³⁵U changes from 96% to 70% and below) or when the sample geometry changes slightly, the system provides excellent discrimination against these cheating scenarios. Using intelligent algorithms, the verification method facilitates independent certification of nuclear weapons, improving efficiency and reducing the risk of manual tampering and prying. Additionally, with physical mask encryption technology, sensitive information is not measured from the beginning to the end, reducing the risk of cheating through software backdoors and other means. On the basis of protecting the sensitive information of the inspected item, nuclear arms control verification technology based on physical encryption and the K-nearest neighbor algorithm will be able to identify its authenticity with high accuracy and efficiency.