一种针对大语言模型键值缓存的安全性检测方法及系统
By constructing a token tree and combining iterative generation of the shadow large language model with evaluation of the first character generation delay, the problems of detection accuracy and robustness in the shared key-value cache environment of large language models are solved, achieving more efficient privacy protection and system stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-17
AI Technical Summary
In multi-tenant environments with large language models, existing technologies struggle to effectively detect side-channel risks introduced by key-value cache sharing mechanisms, especially in long sequences and complex network environments. This results in low detection accuracy and poor robustness, hindering improvements in privacy protection and system robustness.
By constructing an initial token tree and iteratively generating it using the shadow large language model and the target large language model, and combining the prediction confidence of candidate tokens and the initial character generation delay to assess security, a security assessment mechanism for multiple candidate paths is formed. This reduces the dependence on the temporal characteristics of network traffic and selects the path with the highest security score for matching degree assessment.
It significantly improves the accuracy and robustness of key-value cache security detection in long text and complex network environments, enhances privacy protection in multi-tenant shared environments, and reduces sensitivity to network noise.
Smart Images
Figure CN122153968B_ABST