M-Attack-V2-Adversarial-Samples
收藏资源简介:
M-Attack-V2对抗样本数据集包含由M-Attack-V2方法生成的对抗图像样本,源自论文《通过细粒度细节目标推进黑盒LVLM攻击前沿》。该数据集主要用于评估大型视觉语言模型(LVLM)在黑盒设置下的对抗鲁棒性。数据集包含两种配置:epsilon_8(扰动预算ε=8/255)和epsilon_16(扰动预算ε=16/255),每种配置各包含100张224×224 RGB格式的PNG图像,共计200张对抗样本。所有样本均基于NIPS 2017对抗竞赛数据集生成,采用L∞范数约束扰动。该数据集适用于图像分类、视觉问答等任务的对抗攻击研究,特别适合用于多模态模型和黑盒攻击场景的基准测试。
The M-Attack-V2 Adversarial Sample Dataset consists of adversarial image samples generated via the M-Attack-V2 method, originating from the paper *Advancing the Frontier of Black-box LVLM Attacks via Fine-grained Detail-targeted Objectives*. This dataset is primarily used to evaluate the adversarial robustness of Large Vision-Language Models (LVLMs) under black-box settings. The dataset includes two configurations: epsilon_8 (perturbation budget ε=8/255) and epsilon_16 (perturbation budget ε=16/255). Each configuration contains 100 PNG-format RGB images with a resolution of 224×224, resulting in a total of 200 adversarial samples. All samples are generated based on the NIPS 2017 Adversarial Competition Dataset, with perturbations constrained by the L∞ norm. This dataset is applicable to adversarial attack research for tasks such as image classification and visual question answering, and is particularly suitable for benchmarking multi-modal models and black-box attack scenarios.



