Adversarial Examples for Automatic Speech Recognition: Attacks and Countermeasures

Shengshan Hu, Xingcan Shang, Zhan Qin, Minghui Li, Qian Wang, Cong Wang

Abstract: Speech is a common and effective approach for communication between humans and modern mobile devices such as smartphones or home hubs. The remarkable advances in computing and networking have popularized automatic speech recognition (ASR) systems, which can interpret received speech signals on mobile devices and enable us to remotely control and interact with those devices. Despite promising development, audio adversarial examples, a new kind of attack on advanced ASR systems, are found to be extremely effective in imitating human speech while fooling mobile devices to produce incorrect commands. In this article, we provide a systematic survey of audio adversarial examples in the literature. We first present an overview of the architecture of ASR systems and outline the basic attack philosophy. Followed by a brief introduction of the state-of-the-art solutions to audio adversarial examples, a comprehensive comparison is presented. Finally, after discussing existing countermeasures to defend ASR, we highlight several promising future research directions and challenges on constructing more robust and practical audio adversarial examples.