[ILUG] Re: cohosh covert
rcunniff at stp.dias.ie
Wed Apr 14 17:57:30 IST 2004
On Wed, 14 Apr 2004, Bryan O'Donoghue wrote:
> Ronan Cunniffe wrote:
> hmmm, this guy claims a Bayesian algorithm will hit 995 spams per thousand,
> with zero false positives.
I could claim that too...
Bayesian filters aren't magic, they're a piece of (mathematical)
engineering, and have precise and calculable modes of failure. The
spammers have calculated and precisely targetted one of them.
Bayesian filters use past patterns as templates to judge which of N bins
a new e-mail matches, and by adding the new e-mail to the relevant bin,
can increase the resolution of the template patterns.
You simply can't do this if the messages are perfectly random (this being
one of the practical tests of randomness). What you can do is force the
spammers out the main body and into an attachment. Whether this is
progress, I don't know.
More information about the ILUG