TCAB: A Large-Scale Text Classification Attack Benchmark

TL;DR

TCAB is a comprehensive large-scale dataset for analyzing, detecting, and understanding adversarial attacks on text classifiers, facilitating research in attack detection, localization, and characterization.

Contribution

Introduces TCAB, a large, automated dataset of 1.5 million adversarial attack instances on text classifiers, enabling diverse attack analysis tasks.

Findings

Contains 1.5 million attack instances across multiple classifiers and datasets

Includes human-labeled subset for semantic preservation analysis

Supports attack detection, localization, and target labeling tasks

Abstract

We introduce the Text Classification Attack Benchmark (TCAB), a dataset for analyzing, understanding, detecting, and labeling adversarial attacks against text classifiers. TCAB includes 1.5 million attack instances, generated by twelve adversarial attacks targeting three classifiers trained on six source datasets for sentiment analysis and abuse detection in English. Unlike standard text classification, text attacks must be understood in the context of the target classifier that is being attacked, and thus features of the target classifier are important as well. TCAB includes all attack instances that are successful in flipping the predicted label; a subset of the attacks are also labeled by human annotators to determine how frequently the primary semantics are preserved. The process of generating attacks is automated, so that TCAB can easily be extended to incorporate new text attacks and better classifiers as they are developed. In addition to the primary tasks of detecting and labeling attacks, TCAB can also be used for attack localization, attack target labeling, and attack characterization. TCAB code and dataset are available at https://react-nlp.github.io/tcab/.