Silent-Hidden-Voice Attack on Speech Recognition System

Abstract

In this paper, we propose a method for creating a hidden voice that is perceived as silence by a human. The proposed method creates a silent hidden voice that is mistakenly classified as a target phrase by the target model; it does this by configuring the loss function so that the probability of classification into the target phrase by the target model is highest. In an experimental evaluation using the Mozilla Common Voice dataset as the test data source and TensorFlow as the machine learning library, the proposed method created a silent hidden voice that had a 100% attack success rate for a target phrase on a target model while minimizing the average distortion to 187.81.