Security and privacy in speaker recognition systems

Thesis

Voice interfaces continue to grow in popularity, with standalone systems being deployed in our homes, smart assistants being added to our phones and smart watches, and voice based software being added to phone call systems. As part of this trend voice interfaces are also deploying personalised functionality, with the unique features of an individual's voice being used as a biometric to guard access to this.

In this thesis we examine the security and privacy aspects of these systems, with a particular focus on remotely accessed speaker recognition. First, we evaluate the susceptibility of speaker recognition systems to attacks by developing an attack method that allows an adversary to impersonate a chosen user. We demonstrate the capabilities of this attack across several different systems, showing that an adversary can still perform this method even when restricted to audio data of limited quantity and quality.

Having demonstrated the vulnerability of speaker recognition to attacks, we investigate methods to enhance the privacy of those using such systems. We first propose and evaluate a method of voice anonymisation, which removes identifying information but maintains the prosody of the speech. We follow this by developing an alternate method, which allows the user of a remote speech system to protect their own voice from capture, by replacing it with an alternate voice when they speak and removing all voice information from the final audio. We show that this method can be used to maintain the identity of the user over time, and is more resilient to attacks than the status quo for users with exposed voice traits.

Finally, we explore a new method for collecting biometric datasets. We design, implement and evaluate a system for collecting these datasets remotely. This allows researchers to scale their dataset collection more effectively, allowing for the creation of larger and more useful biometric datasets in the future.

Authors: Turner, H

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: privacy; voice; speaker; speech; security; impersonation; mechanical turk
• Subjects: Computer security; Automatic speech recognition