[Standards] [Members] 33C3 talk on Signal and current XMPP issue in providing a similar UX

Evgeny Khramtsov xramtsov at gmail.com
Mon Jan 2 10:12:31 UTC 2017


Thu, 29 Dec 2016 16:48:19 -0600
Sam Whited <sam at samwhited.com> wrote:

> I very
> *very* rarely have spam get through my email filters (which is not
> proprietary or centralized), and don't see why individual services
> that use XMPP (server operators) couldn't also implement the exact
> same spam filters on their XMPP servers.

The problem is that spam/ham classification problem for SPIM is a bit
different from those for email because IM messages are short. There
is a research paper showing that Bayesian/Support-vector-machine/etc
filters would have a lot of false positives. Instead, some neural
network is introduced in the research paper which is capable of
detecting a "dialog" characteristic in the IM message. However,
neural networks are very slow to train and thus require the "big data".

Another problem, and I'm constantly repeating here, is that when you
detect a spam message on the server, what would you do with it? Block
silently? I'm not aware of any email server which acts that badly, e.g.
by simply throwing away a "suspicious" message. This is a problem if we
implement heuristics filters with a lot of false positives.


More information about the Standards mailing list