PhaseFool: Phase-oriented Audio Adversarial Examples via Energy Dissipation

Ziyue Jiang, Yi Ren, Zhou Zhao

Keywords: Audio adversarial examples, audio adversarial attacks, automatic speech recognition

Abstract: Audio adversarial attacks design perturbations onto inputs that lead an automatic speech recognition (ASR) model to predict incorrect outputs. Current audio adversarial attacks optimize perturbations with different constraints (e.g. lp-norm for waveform or the principle of auditory masking for magnitude spectrogram) to achieve their imperceptibility. Since phase is not relevant for speech recognition, the existing audio adversarial attacks neglect the influence of phase spectrogram. In this work, we propose a novel phase-oriented algorithm named PhaseFool that can efficiently construct imperceptible audio adversarial examples with energy dissipation. Specifically, we leverage the spectrogram consistency of short-time Fourier transform (STFT) to adversarially transfer phase perturbations to the adjacent frames of magnitude spectrogram and dissipate the energy that is crucial for ASR systems. Moreover, we propose a weighted loss function to improve the imperceptibility of PhaseFool. Experimental results demonstrate that PhaseFool can inherently generate full-sentence imperceptible audio adversarial examples with the 100% targeted success rate within 500 steps on average (9.24x speed-up over current state-of-the-art imperceptible counterparts), which is verified through a human study. Most importantly, our PhaseFool is the first to exploit the phase-oriented energy dissipation in the audio adversarial examples rather than add perturbations on the audio waveform like most previous works.

One-sentence Summary: We propose a phase-oriented algorithm named PhaseFool that adversarially dissipate the energy that is crucial for ASR systems and efficiently generate imperceptible audio adversarial examples.