Adv-ESC: Adversarial attack datasets for an environmental sound classification

Abstract

In recent years, deep learning techniques have been well explored to accomplish sound classification tasks and in particular, in the classification of environmental sounds. Despite outstanding success in sound classification, various studies illustrate that deep models can be easily tricked by a carefully engineered adversarial input generated by adding a small perturbation to the legitimate input, indistinguishable to human. In this work, we present a comprehensive study investigating the effect of adversarial attacks on the performance of deep models developed to classify environmental sounds. Besides investigating deep model’s performance in the presence of three prevalent adversarial attacks, we are also releasing the first Adversarial environmental sounds (Adv-ESC) dataset, built on two popular benchmark Environmental Sound Classification (ESC) datasets, i.e. ESC-10 and DCASE 2019 Task-1A Challenge datasets. This Adv-ESC dataset can serve as a benchmark to train robust models that are less vulnerable to adversarial attacks. The Adv-ESC dataset can be downloaded from: https://github.com/achyutmani/Adversarial-Attack-ESC-Datasets-Adv-ESC-.