AMAZE: Recognizing Speakers with Amazon's Echo Dot Device

Given the growing popularity of voice assistants such as Amazon's Alexa/Echo product, this work investigates a potential privacy concern with this type of product. We collect the encrypted TCP traffic moving from the Echo Dot to the Alexa Voice Service (AVS), and then use machine learning techniques to determine who, of a finite set of speakers, is speaking to the Amazon Alexa/Echo product. We achieve close to 80% speaker identification accuracy with two speakers and approximately 30% speaker identification accuracy with 12 speakers. Both results are statistically significant when compared to random guessing. In this work we discuss the privacy implications of encrypted speaker identification using the Amazon Alexa/Echo product. We also present our speaker identification techniques, as well as an analysis of our results. Since this work investigating privacy concerns is in its initial stages, we also outline potential future work.