What learned representations and influence functions can tell us about adversarial examples

Shakila Mahjabin Tonni; Mark Dras

doi:10.18653/v1/2023.findings-ijcnlp.35

What learned representations and influence functions can tell us about adversarial examples

Shakila Mahjabin Tonni, Mark Dras

School of Computing

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

Abstract

Adversarial examples, deliberately crafted using small perturbations to fool deep neural networks, were first studied in image processing and more recently in NLP. While approaches to detecting adversarial examples in NLP have largely relied on search over input perturbations, image processing has seen a range of techniques that aim to characterise adversarial subspaces over the learned representations.

In this paper, we adapt two such approaches to NLP, one based on nearest neighbors and influence functions and one on Mahalanobis distances. The former in particular produces a state-of-the-art detector when compared against several strong baselines; moreover, the novel use of influence functions provides insight into how the nature of adversarial example subspaces in NLP relate to those in image processing, and also how they differ depending on the kind of NLP task.

Original language	English
Title of host publication	Findings of the Association for Computational Linguistics
Subtitle of host publication	IJCNLP-AACL 2023
Place of Publication	Stroudsburg
Publisher	Association for Computational Linguistics
Pages	392-411
Number of pages	20
ISBN (Electronic)	9798891760189
DOIs	https://doi.org/10.18653/v1/2023.findings-ijcnlp.35
Publication status	Published - 2023
Event	13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: Findings of the Association for Computational Linguistic, IJCNLP-AACL 2023 - Nusa Dua, Bali, Indonesia Duration: 1 Nov 2023 → 4 Nov 2023

Conference

Conference	13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: Findings of the Association for Computational Linguistic, IJCNLP-AACL 2023
Country/Territory	Indonesia
City	Nusa Dua, Bali
Period	1/11/23 → 4/11/23

Access to Document

10.18653/v1/2023.findings-ijcnlp.35Licence: Unspecified

Cite this

@inproceedings{0c31241e5d264c7bb636fbe2fcc9daa5,

title = "What learned representations and influence functions can tell us about adversarial examples",

abstract = "Adversarial examples, deliberately crafted using small perturbations to fool deep neural networks, were first studied in image processing and more recently in NLP. While approaches to detecting adversarial examples in NLP have largely relied on search over input perturbations, image processing has seen a range of techniques that aim to characterise adversarial subspaces over the learned representations. In this paper, we adapt two such approaches to NLP, one based on nearest neighbors and influence functions and one on Mahalanobis distances. The former in particular produces a state-of-the-art detector when compared against several strong baselines; moreover, the novel use of influence functions provides insight into how the nature of adversarial example subspaces in NLP relate to those in image processing, and also how they differ depending on the kind of NLP task.",

author = "Tonni, {Shakila Mahjabin} and Mark Dras",

year = "2023",

doi = "10.18653/v1/2023.findings-ijcnlp.35",

language = "English",

pages = "392--411",

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

publisher = "Association for Computational Linguistics",

note = "13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: Findings of the Association for Computational Linguistic, IJCNLP-AACL 2023 ; Conference date: 01-11-2023 Through 04-11-2023",

}

Tonni, SM & Dras, M 2023, What learned representations and influence functions can tell us about adversarial examples. in Findings of the Association for Computational Linguistics: IJCNLP-AACL 2023. Association for Computational Linguistics, Stroudsburg, pp. 392-411, 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: Findings of the Association for Computational Linguistic, IJCNLP-AACL 2023, Nusa Dua, Bali, Indonesia, 1/11/23. https://doi.org/10.18653/v1/2023.findings-ijcnlp.35

What learned representations and influence functions can tell us about adversarial examples. / Tonni, Shakila Mahjabin; Dras, Mark.
Findings of the Association for Computational Linguistics: IJCNLP-AACL 2023. Stroudsburg: Association for Computational Linguistics, 2023. p. 392-411.

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

TY - GEN

T1 - What learned representations and influence functions can tell us about adversarial examples

AU - Tonni, Shakila Mahjabin

AU - Dras, Mark

PY - 2023

Y1 - 2023

N2 - Adversarial examples, deliberately crafted using small perturbations to fool deep neural networks, were first studied in image processing and more recently in NLP. While approaches to detecting adversarial examples in NLP have largely relied on search over input perturbations, image processing has seen a range of techniques that aim to characterise adversarial subspaces over the learned representations. In this paper, we adapt two such approaches to NLP, one based on nearest neighbors and influence functions and one on Mahalanobis distances. The former in particular produces a state-of-the-art detector when compared against several strong baselines; moreover, the novel use of influence functions provides insight into how the nature of adversarial example subspaces in NLP relate to those in image processing, and also how they differ depending on the kind of NLP task.

AB - Adversarial examples, deliberately crafted using small perturbations to fool deep neural networks, were first studied in image processing and more recently in NLP. While approaches to detecting adversarial examples in NLP have largely relied on search over input perturbations, image processing has seen a range of techniques that aim to characterise adversarial subspaces over the learned representations. In this paper, we adapt two such approaches to NLP, one based on nearest neighbors and influence functions and one on Mahalanobis distances. The former in particular produces a state-of-the-art detector when compared against several strong baselines; moreover, the novel use of influence functions provides insight into how the nature of adversarial example subspaces in NLP relate to those in image processing, and also how they differ depending on the kind of NLP task.

UR - http://www.scopus.com/inward/record.url?scp=85188511295&partnerID=8YFLogxK

U2 - 10.18653/v1/2023.findings-ijcnlp.35

DO - 10.18653/v1/2023.findings-ijcnlp.35

M3 - Conference proceeding contribution

AN - SCOPUS:85188511295

SP - 392

EP - 411

BT - Findings of the Association for Computational Linguistics

PB - Association for Computational Linguistics

CY - Stroudsburg

T2 - 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics: Findings of the Association for Computational Linguistic, IJCNLP-AACL 2023

Y2 - 1 November 2023 through 4 November 2023

ER -

What learned representations and influence functions can tell us about adversarial examples

Abstract

Conference

Access to Document

Other files and links

Fingerprint

Cite this