Frank Tore Johansen wrote to [EMAIL PROTECTED]:

Hi, I've been running Spamassassin 2.63 since it came out, but lately more
and more spam seems to slip by its tests.  I have the following local
rules:

99_FVGT_Tripwire.cf  chickenpox.cf   nov2rules.cf      weedsonly.cf
backhair.cf          evilnumbers.cf  oct03_headers.cf
bigevil.cf           local.cf        oct03_rules.cf

evilnumbers.cf, bigevil.cf, backhair.cf and chickenpox.cf is updated
nightly.  I personally get around 520 spam pr day, and with a required
hits of 3, an average of 4 gets through every day.  A colleague of mine
uses a required hits of 5, and the last few day around 40 of 300 spams has
gotten through spamassassin for him.

We both regularly train our bayesian filters with all the spam that gets
through.

Basically, I'm looking for more tuning tips.  Is there any other great
ruleset that I should try out?  How low do you dare set your
required_hits? (Yes, I have whitelisted most common important emails, but
not all).  I haven't tried SURBL yet, could this help greatly?

Our required hits are set at 7.0. On a corpus from a few days ago containing 2300s/2800h, we caught 100.000% of the spam, and let through 100.000% of the ham. How?

0. Diligent corpus maintainership
1. SURBL [ http://www.surbl.org/ ] (easy to install and enable, and YES,
   it helps greatly!)
2. DCC, Razor2, Pyzor (these are easy to install and enable, and DCC
   especially hits a lot of spam)
3. Religious Bayes training
4. About ~500 carefully chosen, tested, and manually rescored SARE rules
   (we only choose the most conservative ones that we see good results
   with in our own mass-check; tripwire hits all the time, but, in a
   corpus containing about ~10K spams, all of its rules combined weren't
   enough to trip *anything* over the threshold).
5. About ~400 local rules that we developed for our *own* corpus. Many
   of them are ham rules that subtract points for mail characteristics
   we see regularly that we've never seen spammers use. We use our own
   eval rules to carefully validate certain mail properties.
6. Manual whitelisting of a *very* small number (8, currently) of
   "problem" senders who always seem to send spammy looking messages.
   Most of them wouldn't trip the threshold, but they were getting
   awfully close. :-)
7. Manual "tweaking" of some of the stock rule scores for 3.0, due to
   scores which, for our own mass-check, were out of line. The scare
   quotes around "tweaking" are there because, in some cases, rule
   scores changed by an order of magnitude.
   (DCC_CHECK went from 0.2 to 2.1 :-)
8. Diligent reading of this list

I think the proof is in the pudding:

Overall Count:  4466 Max:  99.292  Min: -28.806  Mean:  20.284
Spam    Count:  2289 Max:  99.292  Min:   9.230  Mean:  52.145
Ham     Count:  2177 Max:   4.955  Min: -28.806  Mean: -13.216

Interestingly, the hammiest spam scored 9.230/7.0, and the spammiest ham
scored 4.955/7.0, so we've been able to widen the gap between ham and
spam considerably. The percentile ranges aren't shown here (that's going
to be in version 1.4 of my utility :-), but something like 97% of ham
scores < 0.0, and 99.2% of spam scores > 10.0.

These numbers don't include manually whitelisted emails or their scores.
(Otherwise the min would be < -100.0). Our highest scoring rule is ~7.0
points, and that is roughly something like, "if the Bayes score is
between 70-100%, any of the checksum rules hit, and the timezone is far
away, roast 'em". This "roasts" about 25% of our spam, and if it ever
hits ham, they had it coming. :-) It helps with autolearning, greatly
reducing the burden on admins to train the classifier.

These results are relatively normal for us. We don't get 100% all the
time, but the catch rate has been consistently above 99.9% (better than
our human classifiers can do!) for weeks. (1/1000 misses). The only FPs
I've seen have been things that *I* probably wouldn't want to read
anyway, but somebody always subscribes to these things. :-) Maintaining
a "conservative" site-wide filter for this many different users is like
herding cats. I guess we're doing something right, though. :-)

After much whining from our users about having to download hundreds of
tagged spams after a long weekend, we finally introduced a "quarantine"
system where they tell us a score, and we deliver anything above that
score to our own quarantine for admin review. Even though our tagging
threshold is 7.0, one of our users said, "just delete everything of mine
over 6.0.  That should be safe".

The spam fighting effort averages about 15 hours/week for me currently,
which isn't much, considering the amount of junk our users never have to
look at and delete. If it weren't for this, I'd probably be spending at
least 15 hours/week looking for new work, because all of my customers
left... or maybe the more loyal ones would fall on their swords, and
then I'd have their blood on my hands. :-)

None of this would be possible without all the *great* people in this
community who make this stuff work!

- Ryan

--
  Ryan Thompson <[EMAIL PROTECTED]>

  SaskNow Technologies - http://www.sasknow.com
  901-1st Avenue North - Saskatoon, SK - S7K 1Y4

        Tel: 306-664-3600   Fax: 306-244-7037   Saskatoon
  Toll-Free: 877-727-5669     (877-SASKNOW)     North America

Reply via email to