Pop Goes the Verifier: Phoneme-Targeted Data Poisoning of Voice Liveness Detection

Authors: Eshaq Jamdar, Amith Kamath Belman

Published: 2026-10-31 16:35:38+00:00

Comment: Accepted at IEEE International Joint Conference on Biometrics (IJCB) 2026

AI Summary

This paper introduces "Pop Goes the Verifier," a stealthy phoneme-targeted data poisoning attack against voice liveness detection systems like VoicePop. By using FastSpeech 2 with a modified HiFi-GAN vocoder, the attack injects localized pop bursts at plosive phoneme locations and adds Gaussian noise, aiming for greater stealth against signal-level inspection compared to previous, more easily detectable broad-based attacks.

Abstract

Many Automated Speaker Verification Systems (ASVs) commonly employ liveliness detection techniques, such as VoicePop, which identify genuine speech by analyzing low- frequency pop bursts from plosive phonemes using Gammatone Frequency Cepstral Coefficients (GFCC) and an SVM classifier. While effective against traditional replay attacks, such systems are vulnerable to data poisoning techniques. Prior work shows that broadly applied SyntheticPop attacks can collapse model to accuracy to as low as 14.01% using only 20% poisoned samples, but the resulting artifacts are easily detectable when visualizing the waveform energies. This paper introduces Pop Goes the Verifier, a stealthier phoneme injection poisoning attack. By leveraging FastSpeech 2 (FS2) with a modified HiFi-GAN vocoder, we inject localized pop bursts at plosive phoneme locations and overlay subtle Gaussian noise to reduce the visibility of injected artifacts in waveform and spectrogram analysis. Using the ASVSpoof 2019 as the dataset, we observe that the approach achieves a moderate poisoning attack while providing greater resistance to signal-level inspection than the original SyntheticPop attack.


Key findings
The Pop Goes the Verifier attack achieved a moderate poisoning effect, reducing the target VoicePop system's accuracy to 25.66%, which is less severe than SyntheticPop (14.01%) but more stealthy. It demonstrates greater resistance to signal-level inspection through waveform and spectrogram analysis due to localized injection and noise. However, Bhattacharyya score analysis indicates that it does not improve statistical similarity to human speech, achieving a lower score than vanilla FS2 and SyntheticPop.
Approach
The attack modifies the HiFi-GAN vocoder within the FastSpeech 2 pipeline to inject localized pop bursts at phoneme-aligned plosive locations. These bursts are generated with linguistically-grounded parameters (varying timing, duration, amplitude) and combined with subtle Gaussian noise to camouflage the injected artifacts during waveform and spectrogram analysis. The poisoned samples are then used to train the target VoicePop ASV system.
Datasets
ASVSpoof 2019, LJSpeech (for FastSpeech 2 training)
Model(s)
VoicePop (target system using Gammatone Frequency Cepstral Coefficients (GFCC) and SVM classifier), FastSpeech 2 (FS2), HiFi-GAN vocoder, OpenAI's Whisper
Author countries
United States