一种多模态大模型对抗安全检测系统及方法
By designing a multimodal large-scale adversarial security detection system, generating multimodal collaborative adversarial samples, the system solves the security detection problem of multimodal large-scale models in non-white-box scenarios, and achieves comprehensiveness and scalability of adversarial security assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-07-16
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies lack effective detection schemes for adversarial security detection of multimodal large models, especially in non-white-box scenarios. Furthermore, existing methods fail to fully utilize the synergistic effects between multimodal data, resulting in incomplete adversarial security detection.
A multimodal large-scale adversarial security detection system was designed. By combining an image and text adversarial sample generation module with a multimodal task testing module, multimodal collaborative adversarial samples are generated. Security assessment can be performed by only knowing the structural information of the front-end multimodal encoder.
It realizes multimodal large model adversarial security detection in non-white-box scenarios, can comprehensively evaluate the security performance of the model, and supports the generation of adversarial examples for various multimodal data, with good scalability.
Smart Images
Figure CN118916833B_ABST