Efficient Black-Box Adversarial Attacks on Neural Text Detectors

Vitalii Fishchuk; Daniel Braun

Efficient Black-Box Adversarial Attacks on Neural Text Detectors

Vitalii Fishchuk, Daniel Braun

Industrial Engineering & Business Information Systems

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

Abstract

Neural text detectors are models trained to detect whether a given text was generated by a language model or written by a human. In this paper, we investigate three simple and resource-efficient strategies (parameter tweaking, prompt engineering, and character-level mutations) to alter texts generated by GPT-3.5 that are unsuspicious or unnoticeable for humans but cause misclassification by neural text detectors. The results show that especially parameter tweaking and character-level mutations are effective strategies.

Original language	English
Title of host publication	Proceedings of the 6th International Conference on Natural Language and Speech Processing (ICNLSP-2023)
Editors	Mourad Abbas, Abed Alhakim Freihat
Publisher	Association for Computational Linguistics (ACL)
Pages	78-83
Number of pages	6
ISBN (Electronic)	979-8-89176-065-3
Publication status	Published - 1 Dec 2023

Keywords

2024 OA procedure

Access to Document

https://aclanthology.org/2023.icnlsp-1.8Licence: Taverne

Cite this

@inproceedings{c3d5214cffc940b0bfc29d737bf130b5,

title = "Efficient Black-Box Adversarial Attacks on Neural Text Detectors",

abstract = "Neural text detectors are models trained to detect whether a given text was generated by a language model or written by a human. In this paper, we investigate three simple and resource-efficient strategies (parameter tweaking, prompt engineering, and character-level mutations) to alter texts generated by GPT-3.5 that are unsuspicious or unnoticeable for humans but cause misclassification by neural text detectors. The results show that especially parameter tweaking and character-level mutations are effective strategies.",

keywords = "2024 OA procedure",

author = "Vitalii Fishchuk and Daniel Braun",

year = "2023",

month = dec,

day = "1",

language = "English",

pages = "78--83",

editor = "Mourad Abbas and Freihat, {Abed Alhakim}",

booktitle = "Proceedings of the 6th International Conference on Natural Language and Speech Processing (ICNLSP-2023)",

publisher = "Association for Computational Linguistics (ACL)",

address = "United States",

}

Efficient Black-Box Adversarial Attacks on Neural Text Detectors. / Fishchuk, Vitalii; Braun, Daniel.
Proceedings of the 6th International Conference on Natural Language and Speech Processing (ICNLSP-2023). ed. / Mourad Abbas; Abed Alhakim Freihat. Association for Computational Linguistics (ACL), 2023. p. 78-83.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Efficient Black-Box Adversarial Attacks on Neural Text Detectors

AU - Fishchuk, Vitalii

AU - Braun, Daniel

PY - 2023/12/1

Y1 - 2023/12/1

N2 - Neural text detectors are models trained to detect whether a given text was generated by a language model or written by a human. In this paper, we investigate three simple and resource-efficient strategies (parameter tweaking, prompt engineering, and character-level mutations) to alter texts generated by GPT-3.5 that are unsuspicious or unnoticeable for humans but cause misclassification by neural text detectors. The results show that especially parameter tweaking and character-level mutations are effective strategies.

AB - Neural text detectors are models trained to detect whether a given text was generated by a language model or written by a human. In this paper, we investigate three simple and resource-efficient strategies (parameter tweaking, prompt engineering, and character-level mutations) to alter texts generated by GPT-3.5 that are unsuspicious or unnoticeable for humans but cause misclassification by neural text detectors. The results show that especially parameter tweaking and character-level mutations are effective strategies.

KW - 2024 OA procedure

M3 - Conference contribution

SP - 78

EP - 83

BT - Proceedings of the 6th International Conference on Natural Language and Speech Processing (ICNLSP-2023)

A2 - Abbas, Mourad

A2 - Freihat, Abed Alhakim

PB - Association for Computational Linguistics (ACL)

ER -

Efficient Black-Box Adversarial Attacks on Neural Text Detectors

Abstract

Keywords

Access to Document

Fingerprint

Cite this