Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and Entailment

TL;DR

This paper introduces TextFooler, a robust baseline for generating natural adversarial text to evaluate and improve the robustness of NLP models like BERT, demonstrating superior attack success and preservation of semantics.

Contribution

The paper presents TextFooler, a simple yet strong adversarial attack method for NLP that outperforms existing methods in success rate, preserves semantics, and is computationally efficient.

Findings

TextFooler outperforms state-of-the-art attacks in success rate.

It maintains semantic content and grammaticality.

The attack is computationally efficient, linear in text length.

Abstract

Machine learning algorithms are often vulnerable to adversarial examples that have imperceptible alterations from the original counterparts but can fool the state-of-the-art models. It is helpful to evaluate or even improve the robustness of these models by exposing the maliciously crafted adversarial examples. In this paper, we present TextFooler, a simple but strong baseline to generate natural adversarial text. By applying it to two fundamental natural language tasks, text classification and textual entailment, we successfully attacked three target models, including the powerful pre-trained BERT, and the widely used convolutional and recurrent neural networks. We demonstrate the advantages of this framework in three ways: (1) effective---it outperforms state-of-the-art attacks in terms of success rate and perturbation rate, (2) utility-preserving---it preserves semantic content and grammaticality, and remains correctly classified by humans, and (3) efficient---it generates adversarial text with computational complexity linear to the text length. *The code, pre-trained target models, and test examples are available at https://github.com/jind11/TextFooler.