Re: [Evolution] Spamassassin 2.50 ...

[Philippe Chartier <chartiep cybersciences com>]

Le jeu 20/02/2003 à 17:36, Tony Earnshaw a écrit :
> ... has been released, 15:50 CET, for those using SA.
>
> For the record: My standard MUA is Evo 1.2.1, secondary is Horde Imp CVS
> - both with uw-IMAP 2002b and Exim smtp 4.12. Yes, Linux/Gnome 1.4.
>
> For the great unwashed, I've been using 2.50-CVS since Sept. last year
> and with Bayes analysis since mid Jan. last. And it's absolutely
> fantastic.
>
> Mind you, my solution's an MTA solution, Exim 4.12 with SA-Exim built in
> (actually a shared *.so library) to the Exim binary.
>
> Although Sendmail, Postfix, Qmail and Procmail solutions are available
> for the unwashed. There are even M$ Windows possibilities.
>
> I have my e-post address everywhere, so spam and shitty MS executables
> are increasing at an exponential rate. My reaction is British English
> "two horns to the devil," US English "middle finger up."
>
> I almost never get to read spam sent to me, nor receive MS viruses.
> They're mostly refused with an smtp 550 at data time, but with a trigger
> higher than the "define-as-spam points." Any spam or viruses that pass
> the smtp trigger, Exim's own filter catches and files/writes to the mail
> spool either for consenting adults or the sysadmin.
>
> 2.50 plusses over 2.43/2.44, according to Justin Mason - one of its
> pappas:
>
> - Bayesian filtering, using a Bayesian-style form of probability-
>   analysis classification.  This uses an algorithm based on the one
>   detailed in Paul Graham's 'A Plan For Spam' paper, along with aspects
>   taken from Graham Robinson's work, and the chi-combining technique
>   developed by the SpamBayes project.
>
> - Auto-learning.  This trains the Bayesian filter automatically, based
>   on the results from traditional SpamAssassin diagnosis.   It uses a
>   set of heuristics and separate thresholds to ensure (as much as is
>   possible) that it trains on guaranteed non-spam and spam. Old, unused
>   tokens are automatically expired.
>
> - much-improved rule set.  A whole new set of rules based on Message-Id
>   analysis is now in place, which accurately detects forged headers from
>   a wide range of spamware.   Many inaccurate rules have been dropped.
>   HTML tests much improved, with a set to detect image-only spam.
>
> - new default format for detected-spam messages; the message is
>   encapsulated as a MIME part, with a preview and the spam report
>   in the main part of the message.
>
> - Score sets.  Based on whether you are using just SpamAssassin rules,
>   adding network tests, and using a trained Bayesian database,
>   SpamAssassin will use a set of scores appropriately to gain the
>   maximum degree of accuracy.
>
> - Italian, Polish, Spanish, French and German rule sets and
>   translations.
>
> - Much improved reliability with spamd.  The problems with signals
>   have been cleared up thanks to a pipe-based child tracking system,
>   and all spamd-hanging bugs reported have proved unreproducable.
>
> - Unicode problems with Red Hat 8 and perl 5.8 fixed.   Works on Perl
>   5.005, 5.6.x, and 5.8.x.
>
> - Taint-safe.  SpamAssassin runs with perl's taint-checking enabled for
>   better security.
>
> - Razor 1 support is now officially deprecated.
>
> - "spamc -c" was not working, fixed.  This fix required increasing the
>   revision of the spamd protocol; only difference is that now more than
>   one protocol header can appear in the reply from spamd.
>
> Best,
>
> Tony

For those who were hesitating to do the jump from 2.43 to 2.50 or to
start using spamassasin...

I've been running Spamassassin 2.50 and, although it's supposed to be
more or less beta, it works very well! I previously try to use a
combination of Spamassassin 2.43 and Bogofilter to filter my flow of
100-200 spams/day. But it was never satisfactory. Never managed to train
bogofilter properly. Bogo always got a lot of false positives. I found
that the new Spamassassin with Bayesian filters is quite able to take of
the job all by itself. If it could just be integrated in a future
version of Evolution... ;-)

-- 
Philippe