A large model security boundary mapping method and system
By constructing an automated attack and defense environment and a security boundary map, the problems of lag and high cost in the security assessment of large language models are solved, and dynamic optimization of the security posture of large models and systemic risk management are realized.
CN122114056APending Publication Date: 2026-05-29INSPUR SOFTWARE TECH CO LTD
View PDF 0 Cites 0 Cited by
Patent Information
- Application Number
- CN202610025297.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-09
- Publication Date
- 2026-05-29
Smart Images
Figure CN122114056A_ABST
Abstract
The present application relates to the technical field of large language model, in particular to a large model security boundary mapping method and system. The large model security boundary mapping method first completes the construction and movement of the adversarial continuous learning loop, and finally realizes the directional model reinforcement and loop iteration based on the atlas feedback. The large model security boundary mapping method and system realize the active, continuous and interpretable management of the security situation of the large model, can automatically induce the massive scattered adversarial samples into the systematic defect categories understandable by humans, quickly locate the high-risk areas, improve the security analysis efficiency, and greatly enhance the confidence and feasibility of deploying large models in the field of high reliability.
Need to check novelty before this filing date? Find Prior Art