Large model security protection method and device, equipment, storage medium and computer program product
By splitting attack requests into text fragments and generating distributed instruction image sequences, attacks are launched against multimodal large models. This solves the problem of security defense failure in multimodal large language models under multi-image sequence joint reasoning scenarios and improves security protection capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING QIHOOD TECHNOLOGY CO LTD
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-17
AI Technical Summary
Existing multimodal large language models lack effective security mechanisms in multi-image sequence joint reasoning scenarios, resulting in defense failure and an inability to effectively detect and defend against attacks in which malicious intent is dispersed across multiple images.
The key text instructions in the attack request are split into multiple text fragments, and a background image corresponding to each text fragment is generated. A distributed instruction image sequence is generated based on each text fragment and the background image. The multimodal large model is attacked through the distributed instruction image sequence, and security protection is implemented based on the attack results.
By disassembling harmful instructions into text fragments without independent malicious semantics and distributing them into multiple independent images, security vulnerabilities of multimodal large models in multi-image sequence joint reasoning scenarios can be detected, thereby improving the security protection capabilities of multimodal large models.
Smart Images
Figure CN122413441A_ABST