High Quality Audio Adversarial Examples Without Using Psychoacoustics

Wei Zong, Yang-Wai Chow, Willy Susilo

Abstract: In the automatic speech recognition (ASR) domain, most, if not all, current audio AEs are generated by applying perturbations to input audio. Adversaries either constrain norm of the perturbations or hide perturbations below the hearing threshold based on psychoacoustics. These two approaches have their respective problems: norm-constrained perturbations will introduce noticeable noise while hiding perturbations below the hearing threshold can be prevented by deliberately removing inaudible components from audio. In this paper, we present a novel method of generating targeted audio AEs. The perceptual quality of our audio AEs are significantly better compared to audio AEs generated by applying norm-constrained perturbations. Furthermore, unlike approaches that rely on psychoacoustics to hide perturbations below the hearing threshold, we show that our audio AEs can still be successfully generated even when inaudible components are removed from audio.