EvolMusic: towards musical adversarial examples for black-box attacks on speech-to-text

ABSTRACTAutomatic Speech Recognition (ASR) has undergone substantial improvements since the incorporation of deep learning. However, the vulnerability of neural networks to imperceptible adversarial perturbations exposes ASR-based devices to potentially serious threats. So far, imperceptibility of audio adversarial examples has been associated with small, or inaudible perturbations. In this paper, we expand the domain of viable audio adversarial examples to include audible, but inconspicuous adversarial perturbations. We present EvolMusic, the first targeted adversarial attack based on musical note-sequences. Our musical perturbations are generated via an adaptive evolutionary approach in a black-box setting. We evaluate our attack against DeepSpeech v0.9.1 using the Fluent Speech Commands dataset.