Lexa: A Liveness Detection Enabled Voice Assistant
Rodolfo Rodriguez, Yanyan Li · 2023
Virtual voice assistants are becoming more prevalent in the modern household or office environment, with the amount of features and account privileges they have rising steadily as well, this along with their lack of traditional authentication methods makes them susceptible to potential attack. Smart assistants are targets of replayed audio attacks, a type of computer intrusion that focuses around playing back an either synthetic or pre-recorded audio command to a smart voice assistant. This paper proposes a liveness detection system for voice activated smart assistants. Our approach is novel in its use of spectrogram images to treat this as a image classification problem. We implement transfer learning techniques to change the domain of a popular image classification model MobileNetV2, to be a binary classifier for replayed voice samples. We implemented an Alexa enabled voice assistant, Lexa, on a raspberry pi by integrating Alexa Voice Service (AVS) and our proposed liveness detection system. A real world evaluation on our proposed Lexa has shown that it is capable of protecting a smart assistant from replayed audio attacks with high accuracy.