Effective Voice Fuzzing Method for Finding Vulnerabilities in AI Speech Recognition Devices

Abstract

Various biometric technologies are applied to secure authentication, and use Internet of things and smart devices; among them, speech recognition is used as an interface for artificial intelligence (AI) devices as it can conveniently operate and control AI devices through interactive commands. Voice interface, developed with AI technology, is an innovative method to actively control devices. However, the system's vulnerabilities are exploited by malicious attempts on the system, such as dolphin and adversarial attacks on speech recognition devices, and frequent unintended errors occur, such as malfunctioning in response to unauthorized voice signals. In this study, we attempt to effectively find vulnerabilities in these AI speech recognition devices using fuzzing techniques that are generally used to find vulnerabilities in software. In addition, for effective testing, an optimization method for voice fuzzing is proposed to reduce testing time from test coverage perspective. Experimental results demonstrate that the optimization scheme can improve the testing time by approximately 29% by adopting a two-stage fuzzing method.