[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [AVT] Comments on draft-perkins-avt-srtp-vbr-audio-02
If the average VAD overhang time is known by an interceptor, then it
is still possible to estimate the average talk spurt length (and,
thus, potentially the language). So the overhang statistics themselves
should be randomly varied from call to call.
It would indeed be good to see more quantitative analysis, e.g. about:
- The size of the information leak
- How fast the leak shrinks by adding overhang
- How this leak compares to other leaks such as call duration and IP
addresses.
It would make little sense to close a small leak if bigger ones are
still wide open.
koen.
Quoting Cullen Jennings:
On the advice to overhang the VAD by a random amount, I think I need
more advice to implement this. Clearly if I put a random number from
0 to 1 ms this won't work and if I pick a uniformly distributed
random number between 0 and 1 day, it will more or less disable VAD.
No complaints about the approach to this but I think I need more
advice to be able to implement anything useful.
Is there other research that collaborates the spot-me stuff. I sort
of wonder about it at times because the implications of it are that
the success at picking the language when the system has all the
unencrypted RTP, that the would be better than some of the results I
see. I'm not saying I don't believe the spot-me stuff - I would just
want to look at it much deeper before I decided and I was wondering
what others had found. I often see results where the training data
was too closely correlated with the the data used for the test
results._______________________________________________
Audio/Video Transport Working Group
avt at ietf.org
https://www.ietf.org/mailman/listinfo/avt