Model-Driven Universal Voice Sensing Threat based on Built-in Sensors
Published in Young Scholar Workshop: Sensing (青年论坛：感知专场), CCF HHME 2021, 2021
The Chinese title of this talk is 模型驱动的通用内建传感器语音感知威胁研究.
Abstract: As the prevalence of mobile devices and smart voice techniques, Voice User Interface (VUI) gradually becomes the next generation of human-computer interactions. However, under the bright future of voice interactions, privacy concerns like voice eavesdropping and abusement raise wide awareness. Traditional privacy preserving techniques mainly defend the outside eavesdropping threat, while ignoring the inside leakage threat. Especially the enrichment of built-in sensors, such as accelerometer, megnetometer, gyroscope, etc., brings conveniences for users but introducing considerable security problems. This talk will focus on the voice sensing, analyzing the side-channel leakage threat of mobile speakers. Based on it, I will further present a model-driven method, recovering voice information from undersampled sensing data from built-in sensors, and their relative experimental results.