Voice Command Fingerprinting with Locality Sensitive Hashes

摘要: Smart home speakers are deployed in millions of homes around the world. These enable users to interact with other IoT devices household and provide voice assistance such as telling weather reminding appointments. Although smart facilitate many aspects our life, security privacy concerns should be analyzed addressed. In this paper, we show that an attacker sniffing network traffic can infer commands compromise users. Specifically, propose a method utilizes fingerprint without need for extracting features machine learning algorithms. We evaluated proposed on traces 100 different speakers. Our approach correctly infers 42% while models 22% 34%. also effectiveness padding recommended preventing command fingerprinting observed accuracy drops down 15% methods ranges from 6% padding.