Deep learning, audio adversaries, and music content analysis

We present the concept of adversarial audio in the context of deep neural networks (DNNs) for music content analysis. An adversary is an algorithm that makes minor perturbations to an input that cause major repercussions to the system response. In particular, we design an adversary for a DNN that takes as input short-time spectral magnitudes of recorded music and outputs a high-level music descriptor. We demonstrate how this adversary can make the DNN behave in any way with only extremely minor changes to the music recording signal. We show that the adversary cannot be neutralised by a simple filtering of the input. Finally, we discuss adversaries in the broader context of the evaluation of music content analysis systems.