@inproceedings{creo-pudasaini-2025-silverspeak,
title = "{S}ilver{S}peak: Evading {AI}-Generated Text Detectors using Homoglyphs",
author = "Creo, Aldan and
Pudasaini, Shushanta",
editor = "Alam, Firoj and
Nakov, Preslav and
Habash, Nizar and
Gurevych, Iryna and
Chowdhury, Shammur and
Shelmanov, Artem and
Wang, Yuxia and
Artemova, Ekaterina and
Kutlu, Mucahid and
Mikros, George",
booktitle = "Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect)",
month = jan,
year = "2025",
address = "Abu Dhabi, UAE",
publisher = "International Conference on Computational Linguistics",
url = "https://aclanthology.org/2025.genaidetect-1.1/",
pages = "1--46",
abstract = "The advent of Large Language Models (LLMs) has enabled the generation of text that increasingly exhibits human-like characteristics. As the detection of such content is of significant importance, substantial research has been conducted with the objective of developing reliable AI-generated text detectors. These detectors have demonstrated promising results on test data, but recent research has revealed that they can be circumvented by employing different techniques. In this paper, we present homoglyph-based attacks ({\textquoteleft}A' {\textrightarrow} Cyrillic {\textquoteleft}{\CYRA}') as a means of circumventing existing detectors. We conduct a comprehensive evaluation to assess the effectiveness of these attacks on seven detectors, including ArguGPT, Binoculars, DetectGPT, Fast-DetectGPT, Ghostbuster, OpenAI`s detector, and watermarking techniques, on five different datasets. Our findings demonstrate that homoglyph-based attacks can effectively circumvent state-of-the-art detectors, leading them to classify all texts as either AI-generated or human-written (decreasing the average Matthews Correlation Coefficient from 0.64 to -0.01). Through further examination, we extract the technical justification underlying the success of the attacks, which varies across detectors. Finally, we discuss the implications of these findings and potential defenses against such attacks."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="creo-pudasaini-2025-silverspeak">
<titleInfo>
<title>SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aldan</namePart>
<namePart type="family">Creo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shushanta</namePart>
<namePart type="family">Pudasaini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-01</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Firoj</namePart>
<namePart type="family">Alam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Preslav</namePart>
<namePart type="family">Nakov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nizar</namePart>
<namePart type="family">Habash</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Iryna</namePart>
<namePart type="family">Gurevych</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shammur</namePart>
<namePart type="family">Chowdhury</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Artem</namePart>
<namePart type="family">Shelmanov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuxia</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Artemova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mucahid</namePart>
<namePart type="family">Kutlu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">George</namePart>
<namePart type="family">Mikros</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>International Conference on Computational Linguistics</publisher>
<place>
<placeTerm type="text">Abu Dhabi, UAE</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>The advent of Large Language Models (LLMs) has enabled the generation of text that increasingly exhibits human-like characteristics. As the detection of such content is of significant importance, substantial research has been conducted with the objective of developing reliable AI-generated text detectors. These detectors have demonstrated promising results on test data, but recent research has revealed that they can be circumvented by employing different techniques. In this paper, we present homoglyph-based attacks (‘A’ → Cyrillic ‘\CYRA’) as a means of circumventing existing detectors. We conduct a comprehensive evaluation to assess the effectiveness of these attacks on seven detectors, including ArguGPT, Binoculars, DetectGPT, Fast-DetectGPT, Ghostbuster, OpenAI‘s detector, and watermarking techniques, on five different datasets. Our findings demonstrate that homoglyph-based attacks can effectively circumvent state-of-the-art detectors, leading them to classify all texts as either AI-generated or human-written (decreasing the average Matthews Correlation Coefficient from 0.64 to -0.01). Through further examination, we extract the technical justification underlying the success of the attacks, which varies across detectors. Finally, we discuss the implications of these findings and potential defenses against such attacks.</abstract>
<identifier type="citekey">creo-pudasaini-2025-silverspeak</identifier>
<location>
<url>https://aclanthology.org/2025.genaidetect-1.1/</url>
</location>
<part>
<date>2025-01</date>
<extent unit="page">
<start>1</start>
<end>46</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs
%A Creo, Aldan
%A Pudasaini, Shushanta
%Y Alam, Firoj
%Y Nakov, Preslav
%Y Habash, Nizar
%Y Gurevych, Iryna
%Y Chowdhury, Shammur
%Y Shelmanov, Artem
%Y Wang, Yuxia
%Y Artemova, Ekaterina
%Y Kutlu, Mucahid
%Y Mikros, George
%S Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect)
%D 2025
%8 January
%I International Conference on Computational Linguistics
%C Abu Dhabi, UAE
%F creo-pudasaini-2025-silverspeak
%X The advent of Large Language Models (LLMs) has enabled the generation of text that increasingly exhibits human-like characteristics. As the detection of such content is of significant importance, substantial research has been conducted with the objective of developing reliable AI-generated text detectors. These detectors have demonstrated promising results on test data, but recent research has revealed that they can be circumvented by employing different techniques. In this paper, we present homoglyph-based attacks (‘A’ → Cyrillic ‘\CYRA’) as a means of circumventing existing detectors. We conduct a comprehensive evaluation to assess the effectiveness of these attacks on seven detectors, including ArguGPT, Binoculars, DetectGPT, Fast-DetectGPT, Ghostbuster, OpenAI‘s detector, and watermarking techniques, on five different datasets. Our findings demonstrate that homoglyph-based attacks can effectively circumvent state-of-the-art detectors, leading them to classify all texts as either AI-generated or human-written (decreasing the average Matthews Correlation Coefficient from 0.64 to -0.01). Through further examination, we extract the technical justification underlying the success of the attacks, which varies across detectors. Finally, we discuss the implications of these findings and potential defenses against such attacks.
%U https://aclanthology.org/2025.genaidetect-1.1/
%P 1-46
Markdown (Informal)
[SilverSpeak: Evading AI-Generated Text Detectors using Homoglyphs](https://aclanthology.org/2025.genaidetect-1.1/) (Creo & Pudasaini, GenAIDetect 2025)
ACL