面向多模态大模型的时序对抗样本安全评测方法及系统
By introducing temporal consistency loss and optical flow guidance mode, the adversarial perturbation generation method is optimized, which solves the problem of insufficient temporal consistency and robustness in the temporal security evaluation of multimodal large models, and realizes efficient and comprehensive evaluation of model security in dynamic environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU ANQUAN DIGITAL INTELLIGENCE TECH CO LTD
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods for evaluating the temporal security of multimodal large models are insufficient in terms of temporal consistency and robustness assessment. They lack time-dimensional indicators for time-series data, have low efficiency in generating adversarial data, and cannot comprehensively assess the security of models in dynamic environments.
We introduce temporal consistency loss and optical flow guidance mode, and generate objective function by defining temporal consistency loss Ltemp and loss function Ladv. We optimize adversarial perturbation to generate temporal adversarial examples, and evaluate the model security by calculating perturbation imperceptibility and temporal robustness indices.
The generated adversarial examples are closer to real-world attack scenarios, enhancing the real threat and effectiveness of security assessments. They can efficiently generate high-quality temporal adversarial examples and provide multi-dimensional security assessment results.
Smart Images
Figure CN122153920B_ABST