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 perturbations demonstrate. We investigate the impact of visual adversarial attacks on current NLP systems on character-, word-, and sentencelevel tasks, showing that both neural and nonneural 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.
Original language | English |
---|---|
Title of host publication | Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) |
Place of Publication | Minneapolis, Minnesota |
Publisher | ACL |
Pages | 1634–1647 |
Number of pages | 14 |
Volume | N19-1 |
DOIs | |
Publication status | Published - 7 Jun 2019 |
Event | 2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics - Hyatt Regency, 1300 Nicollet Mall, Minneapolis, MN, Minneapolis, United States Duration: 2 Jun 2019 → 7 Jun 2019 https://naacl2019.org/ |
Conference
Conference | 2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics |
---|---|
Abbreviated title | NAACL 2019 |
Country/Territory | United States |
City | Minneapolis |
Period | 2/06/19 → 7/06/19 |
Internet address |