Adversarial attacks on RNN-based deep learning systems

Abstract

Automatic Speech Recognition (ASR) systems have been growing in prevalence together with the advancement in deep learning. Built within many Intelligent Voice Control (IVC) systems such as Alexa, Siri and Google Assistant, ASR has become an attractive target for adversarial attacks. In this research project, the objective is to create a black-box over-the-air (OTA) attack system that can mutate an audio into its adversarial form with imperceptible difference, such that it will be interpreted as the targeted word by the ASR. In this paper, we demonstrate the feasibility and effectiveness of such an attack system in generating perturbation for the DeepSpeech ASR.