No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks

Sergey Berezin; Reza Farahbakhsh; Noel Crespi

doi:10.18653/v1/2023.findings-emnlp.155

No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks

Sergey Berezin
, Reza Farahbakhsh
, Noel Crespi

Telecom Sudparis

Résultats de recherche: Le chapitre dans un livre, un rapport, une anthologie ou une collection › Contribution à une conférence › Revue par des pairs

Résumé

We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.

langue originale	Anglais
titre	Findings of the Association for Computational Linguistics
Sous-titre	EMNLP 2023
Editeur	Association for Computational Linguistics (ACL)
Pages	2362-2369
Nombre de pages	8
ISBN (Electronique)	9798891760615
Les DOIs	https://doi.org/10.18653/v1/2023.findings-emnlp.155
état	Publié - 1 janv. 2023
Evénement	2023 Findings of the Association for Computational Linguistics: EMNLP 2023 - Hybrid, Singapour Durée: 6 déc. 2023 → 10 déc. 2023

Série de publications

Nom	Findings of the Association for Computational Linguistics: EMNLP 2023

Une conférence

Une conférence	2023 Findings of the Association for Computational Linguistics: EMNLP 2023
Pays/Territoire	Singapour
La ville	Hybrid
période	6/12/23 → 10/12/23

Accès au document

10.18653/v1/2023.findings-emnlp.155

Autres fichiers et liens

Lien vers la publication dans Scopus

Contient cette citation

Berezin, S., Farahbakhsh, R., & Crespi, N. (2023). No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. Dans Findings of the Association for Computational Linguistics: EMNLP 2023 (p. 2362-2369). (Findings of the Association for Computational Linguistics: EMNLP 2023). Association for Computational Linguistics (ACL). https://doi.org/10.18653/v1/2023.findings-emnlp.155

Berezin, Sergey ; Farahbakhsh, Reza ; Crespi, Noel. / No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. Findings of the Association for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics (ACL), 2023. p. 2362-2369 (Findings of the Association for Computational Linguistics: EMNLP 2023).

@inproceedings{468cd5d562204687900fe8148f45dbd3,

title = "No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks",

abstract = "We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.",

author = "Sergey Berezin and Reza Farahbakhsh and Noel Crespi",

note = "Publisher Copyright: {\textcopyright} 2023 Association for Computational Linguistics.; 2023 Findings of the Association for Computational Linguistics: EMNLP 2023 ; Conference date: 06-12-2023 Through 10-12-2023",

year = "2023",

month = jan,

day = "1",

doi = "10.18653/v1/2023.findings-emnlp.155",

language = "English",

series = "Findings of the Association for Computational Linguistics: EMNLP 2023",

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

pages = "2362--2369",

booktitle = "Findings of the Association for Computational Linguistics",

}

Berezin, S, Farahbakhsh, R & Crespi, N 2023, No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. Dans Findings of the Association for Computational Linguistics: EMNLP 2023. Findings of the Association for Computational Linguistics: EMNLP 2023, Association for Computational Linguistics (ACL), p. 2362-2369, 2023 Findings of the Association for Computational Linguistics: EMNLP 2023, Hybrid, Singapour, 6/12/23. https://doi.org/10.18653/v1/2023.findings-emnlp.155

No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. / Berezin, Sergey; Farahbakhsh, Reza; Crespi, Noel.
Findings of the Association for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics (ACL), 2023. p. 2362-2369 (Findings of the Association for Computational Linguistics: EMNLP 2023).

Résultats de recherche: Le chapitre dans un livre, un rapport, une anthologie ou une collection › Contribution à une conférence › Revue par des pairs

TY - GEN

T1 - No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks

AU - Berezin, Sergey

AU - Farahbakhsh, Reza

AU - Crespi, Noel

PY - 2023/1/1

Y1 - 2023/1/1

N2 - We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.

AB - We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.

UR - https://www.scopus.com/pages/publications/85183291122

U2 - 10.18653/v1/2023.findings-emnlp.155

DO - 10.18653/v1/2023.findings-emnlp.155

M3 - Conference contribution

AN - SCOPUS:85183291122

T3 - Findings of the Association for Computational Linguistics: EMNLP 2023

SP - 2362

EP - 2369

BT - Findings of the Association for Computational Linguistics

PB - Association for Computational Linguistics (ACL)

T2 - 2023 Findings of the Association for Computational Linguistics: EMNLP 2023

Y2 - 6 December 2023 through 10 December 2023

ER -

Berezin S, Farahbakhsh R, Crespi N. No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. Dans Findings of the Association for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics (ACL). 2023. p. 2362-2369. (Findings of the Association for Computational Linguistics: EMNLP 2023). doi: 10.18653/v1/2023.findings-emnlp.155

No offence, Bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks

Résumé

Série de publications

Une conférence

Accès au document

Autres fichiers et liens

Empreinte digitale

Contient cette citation