Abstract: With the further development of nuclear disarmament and nuclear safeguards policies, future nuclear arms control agreements may limit all tactical nuclear warheads in the arsenal, as well as deployed and planned strategic nuclear warheads. Although there are corresponding military control verification methods now, the shielding methods for sensitive information are still not perfect. The existing methods include indirect verification of nuclear warhead vehicles, attribute measurement, and template matching. The zero knowledge protocol’s protection of sensitive information and its acceptable level of confidence make it an important means of potentially changing nuclear arms control verification, belonging to a type of template matching method. However, because zero knowledge protocols currently rely on templates for statistical matching verification, relying on templates can lead to sensitive information leakage and verification limitations, only verifying nuclear weapons of the same design model as the template. Finding ways to replace templates has become an urgent problem to be solved. Innovative research on zero knowledge protocols based on deep learning algorithms was proposed, which combines the discipline of artificial intelligence machine learning. The traditional method of template matching was replaced by the method of artificial intelligence deep learning. On the basis of relevant declaration data, a massive sample library was established through Monte Carlo method. After machine learning, artificial intelligence programs can identify whether the results of random interactive verification match their declaration situation according to the zero knowledge protocol without the need for matching templates. This new method innovatively combines artificial intelligence machine learning, zero knowledge protocols in cryptography, Monte Carlo simulation, and nuclear security verification. In order to demonstrate its protection against sensitive information, the Monte Carlo simulation of this study used multiple nuclear warhead design configurations, and even used nuclear material components of the same quality and abundance as the nuclear warhead. The results show that various design configurations, including nuclear material components with similar quality and abundance and sufficient to undergo chain reactions, can be identified as true warheads by deep learning programs, verifying their good sensitive information protection. In order to demonstrate the accuracy of its discrimination, the Monte Carlo simulation of this study used various missing abundance of nuclear warhead raw materials, the use of lead to replace fissionable materials, and even the use of low concentration nuclear material loose parts as non-nuclear warheads to reflect the situation of nuclear weapon nuclear material loss and replacement. The results show that various situations and even nuclear material components with similar quality and abundance and sufficient to undergo chain reactions can be identified as true warheads by deep learning programs, verifying its good discrimination accuracy. This research technology can not only be applied in the fields of nuclear arms control verification and nuclear safeguards supervision, but also in other fields, with important significance, practical application value, and broad application prospects.