Jarvis1111/RobustVLGuard
收藏官方服务：
资源简介：
RobustVLGuard是一个多模态安全数据集，包含三个子集：对齐的安全数据、不齐的安全数据和与安全无关的通用指令遵循数据。这个数据集旨在通过高斯噪声增强和图像文本对齐/不齐的微调来训练和提高视觉语言模型对扰动型对抗攻击的鲁棒性。
RobustVLGuard is a multimodal safety dataset containing three subsets: aligned safety data, misaligned safety data, and safety-agnostic general instruction-following data. This dataset is designed to train and enhance the robustness of Vision-Language Models against perturbation-based adversarial attacks through Gaussian noise augmentation and fine-tuning under image-text alignment/misalignment.
提供机构：
Jarvis1111
