Re: IMPORTANT: your message to html-tidy

[Craig Sanders <cas@taz.net.au>]

On Wed, Sep 10, 2003 at 01:06:01AM +0300, Riku Voipio wrote:
> That Pauls theory. This is my theory, bayesian will lead to spam
> like THIS:
> 
> -cut-
> From: John Doe <john.doe@mail>
> To: me <ansa@kos.to>
> Subject: FYI
> 
> Hi, FYI, I just updated my website, what do you think about it?
> <URL:http://www....>
> -cut-

so?  my anti-spam methods cope with this.

any spam that makes it through my filters, whether it gets tagged by
spamassassin or not is used to construct new anti-spam rules.

i routinely add any domains and ip addresses found in spams to both my postfix
access lists and my spamassassin rules.  the first spam to mention an ip
address or domain may get through....the second one wont.

because of all my custom rules, the average score for spam identified by SA on
my systems is about 35.

the high scores also help SA's bayesian filter to learn to recognise spams like
this because they have a large number of spam words without ANY ham words to
lower the bayes score.


the spams that are most likely to get through both SA and my access lists (i.e.
false negative) are nigerian 419 spams, because they vary the phrasing so much.
no problem, i routinely add phrases and names taken from those spams to my
spamassassin rules too.  again, the first copy may get through....subsequent
copies won't.  again, this also increases the spam score and feeds back into
SA's bayesian learning.



craig