On Sat, 8 Mar 2003, Vernon Schryver wrote:
From: "Hallam-Baker, Phillip" <pbaker(_at_)verisign(_dot_)com>
...
First premise: Filters do not work better than 95%
All the filters proposed thus far are based on technologies that are known
to have severe limitations. That includes the Bayesian filtering approach
for which some have been claiming ridiculous 99.5% success rates with no
failures.
I'm not sure whether "the" numbers is 80%, 90%, 95%, or perhaps even
99%, but I'm sure it's not 100%. Part of the cause for the ridiculous
claims of 99.5% averages for Bayesian filters (but notably not on
official Bayesian web pages) is that someone who receives one or two
legitimate messages per week and 28 spam/day really can see better
than 99% accuracy. People who receive 100's of legitimate messages/day
will have other views.
The accuracy of a bayesian filter very much depends on training, and the
variance in your regular email and its differentiation from spam. On my
personal email, of which including mailing list subscriptions I get about
400 mails a day, about 50-100 of which are spam, bayesian filters *are*
99.9% accurate.
Throw in a mixed user base and that figure drops quite a bit (though not
unreasonably so).
The bad thing is that bayesian filters are thwartable (though no more so
than any other body-based mechanisms are, including hashes).
Matt.
_______________________________________________
Asrg mailing list
Asrg(_at_)ietf(_dot_)org
https://www1.ietf.org/mailman/listinfo/asrg