No offence, bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

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

Résumé

We introduce a simple yet efficient sentencelevel 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.
Fichier principal
Vignette du fichier
_ACL_Template__Copy____cyberseq-4-1.pdf (130.9 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04420371 , version 1 (26-01-2024)

Identifiants

Citer

Sergey Berezin, Reza Farahbakhsh, Noel Crespi. No offence, bert - I insult only humans! Multiple addressees sentence-level attack on toxicity detection neural networks. The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), Dec 2023, Resorts World Convention Centre, Singapore. pp.2362-2369, ⟨10.18653/v1/2023.findings-emnlp.155⟩. ⟨hal-04420371⟩
8 Consultations
5 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More