Clay Davis wrote:
Over the past several months I have been saving the spam that slips
through to my users accounts to train my bayes with. I notice that
lately almost all of it has (what I am assuming to be) an attempt to
poison my bayes (a bunch of valid words put together in a nonsensical
paragraph) at the bottom of it.
How much should I worry about this type of spam and how it will affect
my bayes db? Work arounds? Advice?
Thanks, gang.
Clay
Hi, Clay. Without getting into the math behind it, Bayes poisoning is almost impossible. I have been training
my Bayes DB with everything I consider "spam", wether it has a "poison" section or not. I'm almost
always seeing a BAYES_99 result on these "poisoned" emails. Why? Because the key tokens that make it spam
are repeated; the "poison" text is not.
I use a combination of auto-training and hand-correction with my DB. I only
"correct" if the answer is not a BAYES_99. Don't sweat the "poison", Bayes is
almost immune to Iocane, etc.
--
--Michel Vaillancourt
Wolfstar Systems
www.wolfstar.ca
