A Vulnerability Test Method for Speech Recognition Systems Based on Frequency Signal Processing

摘要

With the development of information technology and the popularity of smart devices, voice recognition system, which is installed in many devices such as smartphones and automatic control equipment, is playing an increasingly important role in people's lives. Its appearance has greatly facilitated people's lives. However, their potential safety issues should also receive sufficient attention. Human perception of speech signals is affected by prior knowledge and context. It is natural for a people to unconsciously make up the obscure part of a voice signals he heard. But it doesn't work this way for a voice recognition software. In this condition, wrong results which are far from normal people's recognition will be returned from the software. That case rose above shows that voice recognition systems have potential vulnerability, which makes it possible that designed dangerous commands may be executed without operator's conscious. This paper firstly proposes a method to process voice signal and produce a lot of processed signals which are similar in human auditory but different in actual. Then, we propose a strategy to select processed voice signals aiming at a target speech recognition system. Finally, a way to test the vulnerability of voice recognition system is proposed in passing. Experimental results revealed that the above strategies effectively changed the output texts from speech recognition systems while the input voices were auditory changed little. This result confirmed the vulnerability of voice recognition system.