Text Processing Like Humans Do: Visually Attacking and Shielding NLP Systems

TL;DR

This paper introduces visual adversarial attacks on NLP systems, demonstrating their vulnerability compared to human robustness, and proposes shielding methods to improve model resilience, though challenges remain.

Contribution

It is the first comprehensive study on visual adversarial attacks in NLP and evaluates multiple defense strategies to enhance robustness.

Findings

Neural and non-neural models are highly sensitive to visual attacks, with performance drops up to 82%.

Shielding methods improve robustness but do not fully recover original performance.

Humans remain robust against visual text modifications, unlike current NLP models.

Abstract

Visual modifications to text are often used to obfuscate offensive comments in social media (e.g., "!d10t") or as a writing style ("1337" in "leet speak"), among other scenarios. We consider this as a new type of adversarial attack in NLP, a setting to which humans are very robust, as our experiments with both simple and more difficult visual input perturbations demonstrate. We then investigate the impact of visual adversarial attacks on current NLP systems on character-, word-, and sentence-level tasks, showing that both neural and non-neural models are, in contrast to humans, extremely sensitive to such attacks, suffering performance decreases of up to 82\%. We then explore three shielding methods---visual character embeddings, adversarial training, and rule-based recovery---which substantially improve the robustness of the models. However, the shielding methods still fall behind performances achieved in non-attack scenarios, which demonstrates the difficulty of dealing with visual attacks.