Guus Bosman

Today I read 'Uncovering Spoken Phrases in Encrypted Voice over IP Conversations', a very interesting article from the December 2010 issue of ACM Transaction on Information and System Security. (Read the full PDF version here).

The paper details a gap in the security of VBR compressed encrypted VoIP streams. The authors had earlier found that it is possible to determine the language that is spoken on such a VoIP call, based on packet lengths. Now they have expanded their research and show that it's possible to detect entire spoken phrases during a VoIP call. On average, their method achieved recall of 50% and precision of 51% for a wide variety of phrases spoken by a diverse collection of speakers (some phrases are easier to detect than others; the recall various from 0% to 98%, depending on length of the phrase and the speaker).

In other words: they can detect fairly well if a certain phrase is being used in a conversation, even though the VoIP conversation is encrypted!

Fundamentally, this is possible because VoIP packets are compressed using variable bit-rate compression and not typically "padded". Longer phonemes (such as vowels) correspond with longer packets, shorter phonemes (such as fricatives like 's', 'sh' or 'th') use shorter packets -- using sophisticated statistical analysis they can detect whole phrases.

A solution would be to add padding to VoIP packets, but that increases the bandwidth that is needed. Not only does padding increase the bandwidth because of padding itself, but it also negates a big benefit of VBR compression when dealing with quiet periods in a conversation, when one party is listening to another.

A fun read, quite accessible.