On End-to-End White-Box Adversarial Attacks in Music Information Retrieval

Small adversarial perturbations of input data can drastically change the performance of machine learning systems, thereby challenging their validity. We compare several adversarial attacks targeting an instrument classifier, where for the first time in Music Information Retrieval (MIR) the perturbat...

Full description

Bibliographic Details
Main Authors: Katharina Prinz, Arthur Flexer, Gerhard Widmer
Format: Article
Language:English
Published: Ubiquity Press 2021-07-01
Series:Transactions of the International Society for Music Information Retrieval
Subjects:
Online Access:https://transactions.ismir.net/articles/85
_version_ 1819129223970816000
author Katharina Prinz
Arthur Flexer
Gerhard Widmer
author_facet Katharina Prinz
Arthur Flexer
Gerhard Widmer
author_sort Katharina Prinz
collection DOAJ
description Small adversarial perturbations of input data can drastically change the performance of machine learning systems, thereby challenging their validity. We compare several adversarial attacks targeting an instrument classifier, where for the first time in Music Information Retrieval (MIR) the perturbations are computed directly on the waveform. The attacks can reduce the accuracy of the classifier significantly, while at the same time keeping perturbations almost imperceptible. Furthermore, we show the potential of adversarial attacks being a security issue in MIR by artificially boosting playcounts through an attack on a real-world music recommender system.
first_indexed 2024-12-22T08:40:19Z
format Article
id doaj.art-c71f066b116b49218e709cf89a1bf47e
institution Directory Open Access Journal
issn 2514-3298
language English
last_indexed 2024-12-22T08:40:19Z
publishDate 2021-07-01
publisher Ubiquity Press
record_format Article
series Transactions of the International Society for Music Information Retrieval
spelling doaj.art-c71f066b116b49218e709cf89a1bf47e2022-12-21T18:32:15ZengUbiquity PressTransactions of the International Society for Music Information Retrieval2514-32982021-07-014110.5334/tismir.8546On End-to-End White-Box Adversarial Attacks in Music Information RetrievalKatharina Prinz0Arthur Flexer1Gerhard Widmer2Johannes Kepler University LinzJohannes Kepler University LinzJohannes Kepler University LinzSmall adversarial perturbations of input data can drastically change the performance of machine learning systems, thereby challenging their validity. We compare several adversarial attacks targeting an instrument classifier, where for the first time in Music Information Retrieval (MIR) the perturbations are computed directly on the waveform. The attacks can reduce the accuracy of the classifier significantly, while at the same time keeping perturbations almost imperceptible. Furthermore, we show the potential of adversarial attacks being a security issue in MIR by artificially boosting playcounts through an attack on a real-world music recommender system.https://transactions.ismir.net/articles/85adversarial attackinstrument classificationmusic recommendationvalidity
spellingShingle Katharina Prinz
Arthur Flexer
Gerhard Widmer
On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
Transactions of the International Society for Music Information Retrieval
adversarial attack
instrument classification
music recommendation
validity
title On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
title_full On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
title_fullStr On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
title_full_unstemmed On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
title_short On End-to-End White-Box Adversarial Attacks in Music Information Retrieval
title_sort on end to end white box adversarial attacks in music information retrieval
topic adversarial attack
instrument classification
music recommendation
validity
url https://transactions.ismir.net/articles/85
work_keys_str_mv AT katharinaprinz onendtoendwhiteboxadversarialattacksinmusicinformationretrieval
AT arthurflexer onendtoendwhiteboxadversarialattacksinmusicinformationretrieval
AT gerhardwidmer onendtoendwhiteboxadversarialattacksinmusicinformationretrieval