The Crux of Voice (In)Security: A Brain Study of Speaker Legitimacy Detection

Abstract

A new generation of scams has emerged that uses voice impersonation to obtain sensitive information, eavesdrop over voice calls and extort money from unsuspecting human users. Research demonstrates that users are fallible to voice impersonation attacks that exploit the current advancement in speech synthesis. In this paper, we set out to elicit a deeper understanding of such human-centered "voice hacking" based on a neuro-scientific methodology (thereby corroborating and expanding the traditional behavioral-only approach in significant ways). Specifically, we investigate the neural underpinnings of voice security through functional near-infrared spectroscopy (fNIRS), a cutting-edge neuroimaging technique, that captures neural signals in both temporal and spatial domains. We design and conduct an fNIRS study to pursue a thorough investigation of users' mental processing related to speaker legitimacy detectionwhether a voice sample is rendered by a target speaker, a different other human speaker or a synthesizer mimicking the speaker. We analyze the neural activity associated within this task as well as the brain areas that may control such activity.

Human Brains Can't Detect Fake News: A Neuro-Cognitive Study of Textual Disinformation Susceptibility

“…Their results show that there is a significant increase in several areas of the brain when people view fake websites vs. real websites even though users' accuracy in identifying the legitimacy of the site was close to 50%. The study of [32] investigated users' processing of real vs. fake voices in the context of voice impersonation attacks, with similar findings as to the above studies. Stanford History Education Group conducted a study [18] in which the participants (school/college students) were asked to find the differences in some media contents, e.g., news and ads, identify real-fake Facebook accounts, test the reliability of Facebook posts (whether a web site is trusted).…”

"Hello, It's Me": Deep Learning-based Speech Synthesis Attacks in the Real World

“…The first classical synthesis attack measurement paper [57] uses a traditional survey format and finds that users correctly distinguish between real and Festvoxsynthesized voices (imitating the real speaker) about 50% of the time, regardless of their familiarity with the real speaker. A followup study to this [60] uses the same data and survey format but includes fNIRS brain scanning technology to measure participants' neural activity. They find no statistically significant differences in neural activity when real or synthetic speakers are played.…”

“…Having demonstrated that DNN-synthesized speech can easily fool machines (e.g., real-world SR systems), we now move to evaluate their impact on humans. Different from prior work that uses surveys to measure human perception of speech synthesized by classical (non-DNN) tools [57,60], we assess the susceptibility of humans to DNN-synthesized speech in different interactive settings. For this we conduct two user studies, covering both static survey and "trusted" interaction settings.…”

“…Prior work has established the viability of speech synthesis attacks against prior generations of synthesis tools and speaker recognition systems [28,45,56,57]. Similarly, prior work assessing human vulnerability to speech synthesis attacks evaluates now-outdated systems in limited settings [57,60].…”