Natural Backdoor Datasets

自然后门数据集是由芝加哥大学和普林斯顿大学的研究团队创建的，旨在研究物理后门攻击。这些数据集包含真实图像中的物理对象，这些对象被用作触发器，以研究深度学习模型在面对此类攻击时的脆弱性。数据集的创建过程涉及从现有的多标签对象数据集中识别和重新标记潜在的触发器子集，以便用于训练模型进行物理后门攻击。这些数据集的应用领域主要集中在提高对物理后门攻击的认识和防御，从而增强计算机视觉模型的安全性。

The Natural Backdoor Dataset suite was developed by a research team from the University of Chicago and Princeton University to investigate physical backdoor attacks. This suite contains physical objects in real-world images that act as backdoor triggers, aiming to evaluate the vulnerability of deep learning models against such attacks. The development process involves identifying and relabeling potential trigger subsets from existing multi-label object datasets for use in training models to conduct physical backdoor attacks. These datasets are primarily applied to raise awareness and develop defensive measures against physical backdoor attacks, thereby enhancing the security of computer vision models.