Detecting Audio Attacks on ASR Systems with Dropout Uncertainty

TL;DR

This paper introduces a novel defense mechanism for ASR systems that leverages dropout-induced uncertainty to detect various adversarial audio attacks, demonstrating high accuracy across multiple datasets.

Contribution

The study presents a dropout-based uncertainty method to effectively detect adversarial audio attacks on end-to-end ASR systems, including robust detection against noise-reduction immune attacks.

Findings

High detection accuracy on Mozilla CommonVoice dataset

Effective against optimized perturbation and frequency masking attacks

Robust detection even for noise reduction immune attacks

Abstract

Various adversarial audio attacks have recently been developed to fool automatic speech recognition (ASR) systems. We here propose a defense against such attacks based on the uncertainty introduced by dropout in neural networks. We show that our defense is able to detect attacks created through optimized perturbations and frequency masking on a state-of-the-art end-to-end ASR system. Furthermore, the defense can be made robust against attacks that are immune to noise reduction. We test our defense on Mozilla's CommonVoice dataset, the UrbanSound dataset, and an excerpt of the LibriSpeech dataset, showing that it achieves high detection accuracy in a wide range of scenarios.