Exploiting Privacy-preserving Voice Query in Healthcare-based Voice Assistant System
Thamer Altuwaiyan, Mohammad Reza Hadian, Samuel Rubel, Xiaohui Liang · 2020
Voice Assistant Systems (VAS) such as Amazon Echo and Google Home are becoming a popular technology for medical health systems among patients and caregivers. The VAS devices allow patients and caregivers to interact with them via voice commands. In most cases, the users' private voice data is fully disclosed to the VAS server, which may raise severe privacy concerns, especially in case of medical information which are clearly sensitive. In this paper, we propose a privacy-preserving voice query scheme in the healthcare-based voice assistant system, which enables the users to use voice commands for uploading medical data and later retrieving them. The VAS server in this case has no access to the original voice command or the data stored but it can accurately respond to user's query. Our scheme consists of two voice matching techniques with weak and strong privacy levels, where the former discloses only the voice feature, and not the original voice to the server. The latter further uses an obfuscation function to hide the voice features, thus the data is fully protected. We evaluate the performance of our proposed scheme by conducting experiments on self-generated voice data set, from three different languages, English, Chinese, and Arabic. We prove that our proposed scheme can achieve the privacy preservation of the voice data, and up to 98% accuracy in responding to voice queries.