Physical Backdoor Dataset

本文介绍了一种利用深度生成模型自动合成物理后门数据集的框架。该数据集用于训练或评估物理后门模型，通过三个自动模块：建议合适的物理触发器、生成中毒候选样本（通过合成新样本或编辑现有干净样本）以及最终精炼最合理的样本，有效地降低了创建物理后门数据集的复杂性。该数据集的应用领域主要集中在深度神经网络的安全性研究，特别是在物理世界中实现后门攻击的研究。

This paper presents a framework for automatically synthesizing physical backdoor datasets using deep generative models. This dataset is designed for training or evaluating physical backdoor models. It effectively reduces the complexity of creating physical backdoor datasets through three automated modules: proposing suitable physical triggers, generating poisoned candidate samples (either by synthesizing new samples or editing existing clean samples), and finally refining the most plausible samples. The application scope of this dataset primarily centers on the security research of deep neural networks, particularly studies on backdoor attacks implemented in the physical world.