自然对抗性数据集

自然对抗性数据集是为评估学习型医疗网络物理系统（CPS）的鲁棒性而创建的。该数据集由宾夕法尼亚大学等机构的研究人员开发，包含九个案例研究，其中包括六个与临床相关的案例。数据集通过使用概率标签从自动化弱监督标签中获得，这些标签结合了嘈杂且易于获取的标签启发式方法。通过这些标签，方法对抗性地排序输入数据，并使用此排序构建一系列逐渐对抗性的数据集。数据集的应用领域是评估医疗CPS中学习组件的鲁棒性，特别是在处理自然发生的对抗性示例时。

The Natural Adversarial Dataset was developed to evaluate the robustness of learning-enabled medical cyber-physical systems (CPS). Constructed by researchers from institutions including the University of Pennsylvania, this dataset contains nine case studies, six of which are clinically relevant. The dataset is derived from automated weakly-supervised labels, which are generated using probabilistic labels that combine noisy but easily accessible label heuristics. Leveraging these labels, methods adversarially rank input data and build a series of progressively adversarial datasets based on this ranking. The application domain of this dataset is to evaluate the robustness of learning components in medical CPS, particularly when handling naturally occurring adversarial examples.