Speaker Recognition in Encrypted Voice Streams

摘要

Transmitting voice communication over untrusted networks puts personal information at risk. Although voice streams are typically encrypted to prevent unwanted eavesdropping, additional features of voice communication protocols might still allow eavesdroppers to discover information on the transmitted content and the speaker. We develop a novel approach for unveiling the identity of speakers who participate in encrypted voice communication, solely by eavesdropping on the encrypted traffic. Our approach exploits the concept of voice activity detection (VAD), a widely used technique for reducing the bandwidth consumption of voice traffic. We show that the reduction of traffic caused by VAD techniques creates patterns in the encrypted traffic, which in turn reveal the patterns of pauses in the underlying voice stream. We show that these patterns are speaker-characteristic, and that they are sufficient to undermine the anonymity of the speaker in encrypted voice communication. In an empirical setup with 20 speakers our analysis is able to correctly identify an unknown speaker in about 48% of all cases. Our work extends and generalizes existing work that exploits variable bit-rate encoding for identifying the conversation language and content of encrypted voice streams.