Multi-modal domain adaptation safety evaluation method and system based on efficient parameter fine-tuning

By employing a multimodal domain-adaptive security assessment method with efficient parameter fine-tuning, this method addresses issues such as data privacy leakage, insufficient cross-modal understanding, and resource consumption in the security assessment of large models in sensitive domains. It achieves secure and accurate multimodal data assessment and generates logically rigorous expert-level assessment results.

CN122174277APending Publication Date: 2026-06-09NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NAT INNOVATION INST OF DEFENSE TECH PLA ACAD OF MILITARY SCI
Filing Date
2026-05-11
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

In the security assessment of large models in sensitive fields such as finance and healthcare, there are problems such as data privacy leakage risks, cross-modal semantic gaps, huge resource consumption and difficulty in aligning expert standards. Existing technologies are unable to achieve efficient, secure and accurate multimodal data assessment.

Method used

A multimodal domain-adaptive security assessment method with efficient parameter fine-tuning is adopted. Through an efficient transfer learning framework with feature out-of-domain and data retention, intra-domain self-supervised feature encoding, cross-attention reinforcement mechanism, and improved GRPO expert alignment, cross-modal feature alignment and security assessment are achieved.

Benefits of technology

It achieves physical-level data non-discovery, eliminates the risk of sensitive data leakage, improves recall rate in multimodal scenarios, significantly reduces computing power costs, and generates logically rigorous and conclusive expert-level evaluation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122174277A_ABST
    Figure CN122174277A_ABST
Patent Text Reader

Abstract

This invention discloses a multimodal domain-adaptive security assessment method and system based on efficient parameter fine-tuning, relating to the field of artificial intelligence security assessment technology. The method includes the following steps: S1. Inputting the data to be identified into a security assessment model; S2. The security assessment model applies data augmentation based on data and task characteristics to the input data to be identified, inputting the augmented data into a domain feature encoder to obtain a representation, and mapping the representation to a contrastive learning space to capture core features; S3. The data output from step S2 is processed through a cross-attention feature alignment module, and after parameter fine-tuning, enters the GRPO reinforcement learning stage to generate a security assessment conclusion; S4. Outputting a safe or dangerous security assessment conclusion, and outputting the inference process. This invention completely eliminates the risk of sensitive data leakage, improves recall in multimodal scenarios, and significantly reduces computational costs.
Need to check novelty before this filing date? Find Prior Art