Henrik Krohns wrote:
On Thu, Mar 22, 2007 at 09:55:07AM -0700, Marc Perkel wrote:
Maybe I'm doing something wrong but with the various methods of bayes
poisoning going on I've found that bayes is just lowering the score of
spam and causing more spam to get through.
So is there actually any real proof that Bayes poisoning works? I've yet to
find any evidence. All the cases have been admins/users messing it up
themselves.
In point of fact, my own experience is that poisoning attempts make no difference
at all. Because the number of poison tokens in an established database is so small, they
don't change anything. However the incidence of other spam-positive keys "tips the
hand".
I use auto-learning. Always have. It has NEVER been a problem; if I
get an FP or FN, I resubmit those mails for retraining to the DB.
I've even gone so far as to take a Spam mail that was visually more than 80%
"poison", copy the poison out, put it around another spam mail and mail it to
myself from a dummy account. Result? Bayes_99. Took the same poision, wrapped it
around a legitimate mail and sent it to myself. Result? Bayes_00. You can't keep a
good Bays down; auto-learned or not.
And I'm a little guy; 5000 messages a day ... 10000 if the lists I host
are busy. Its not like I have a massive bayes DB to work against. The Big
Boys should be even more accurate just by raw weight of statistical incidence.
Bayes Poison is fiction; its not even good fiction.
--
--Michel Vaillancourt
Wolfstar Systems
www.wolfstar.ca
