Poisoned Acoustics: Targeted Data Poisoning Attacks on Acoustic Vehicle Classification at Sub-1% Corruption Rates Harrison Dahme ∗

2026

Read Poisoned Acoustics: Targeted Data Poisoning Attacks on Acoustic Vehicle Classification at Sub-1% Corruption Rates Harrison Dahme ∗ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-22258.txt.

Description

Poisoning a training set sounds like it needs many corrupted labels, but Harrison Dahme shows a whisper can suffice. He attacks acoustic vehicle classification on a city-intersection dataset of roughly ten thousand clips across six vehicle classes. A compact convolutional network trained on log-mel spectrograms learns the task well, yet a tiny label-flipping attack turns it against one target class.

Flipping just half a percent of labels makes nearly all target-class test samples misclassify as the wrong vehicle, while overall accuracy shows zero detectable change. Dahme proves the stealth is structural: the worst accuracy drop from a full targeted attack is capped by the minority class share, which for rare trucks sits below training noise. Aggregate accuracy monitoring, he concludes, is provably insufficient as a defense.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE19 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-22258

Browse the full library · About the Tower of Babel