Abstract
Targeted adversarial attacks against Automatic Speech Recognition (ASR) are thought to require white-box access to the targeted model to be effective, which mitigates the threat that they pose. We show that the recent line of Transformer ASR models pretrained with Self-Supervised Learning (SSL) are much more at risk: adversarial examples generated against them are transferable, making these models vulnerable to targeted, zero-knowledge attacks. We release an adversarial dataset that partially fools most publicly released SSL-pretrained ASR models (Wav2Vec2, HuBERT, WavLM, etc). With low-level additive noise achieving a 30dB Signal-Noise Ratio, we can force these models to predict our target sentences with up to 80% accuracy, instead of their original transcription. With an ablation study, we show that Self-Supervised pretraining is the main cause of that vulnerability. We also propose an explanation for that curious phenomenon, which increases the threat posed by adversarial attacks on state-of-the-art ASR models.
Abstract (translated)
URL
https://arxiv.org/abs/2209.13523
