#870 Void: Voice liveness detection through a spectrogram analysis of voice commands

Popular mobile devices are now being equipped with voice assistants such as Siri and Google Now to provide new ways to interact with devices using voice. However, due to the open nature of voice channels, adversaries could easily record people’s use of voice commands, and replay them to spoof voice assistants. To defend against such spoofing attacks, we present a lightweight and efficient voice liveness detection system called Void (Voice liveness detection): it exploits different characteristics between human voices and voices replayed through speakers with respect to spectral power patterns analyzed over an audible frequency range to effectively detect voice spoofing attacks. To evaluate the performance of Void, we performed experiments on the two datasets: (1) 229,991 voice samples collected from 120 participants with 15 speakers and (2) 18,016 voice samples in the “ASVspoof 2017” dataset with 42 participants and 26 speakers. For both datasets, Void is capable of achieving accuracy of over 99% and 98% in detecting voice replay attacks with less than 1% and 5.1% equal error rate (EER), respectively. Moreover, we demonstrate that Void is resilient against various forms of adversarial attacks with hidden voice and inaudible voice commands –Void achieves 96% and 93% accuracy in detecting even hidden voice command and inaudible voice command attacks, respectively.