Text Classification Attack Benchmark (TCAB)

Text Classification Attack Benchmark (TCAB)是由加州大学欧文分校、加州大学圣地亚哥分校和俄勒冈大学共同创建的大型文本分类攻击基准数据集。该数据集包含超过150万次成功的对抗性攻击实例，这些攻击针对三种分类器，这些分类器是在六个情感分析和滥用检测的源数据集上训练的。TCAB的创建过程是自动化的，便于随着新文本攻击和分类器的发展而扩展。数据集主要用于研究攻击识别，包括攻击检测、攻击标签、攻击定位和攻击特征化，旨在提高文本分类器对抗对抗性攻击的鲁棒性。

Text Classification Attack Benchmark (TCAB) is a large-scale text classification attack benchmark dataset co-created by the University of California, Irvine, the University of California, San Diego, and the University of Oregon. This dataset contains over 1.5 million successful adversarial attack instances targeting three classifiers trained on six source datasets for sentiment analysis and abuse detection. The construction of TCAB follows an automated pipeline, enabling convenient scalability to accommodate the emergence of new text attacks and classifiers. It is primarily used for research on attack recognition, including attack detection, attack labeling, attack localization, and attack characterization, with the goal of enhancing the adversarial robustness of text classifiers.