Proving Which Spam Filters work Best

pirateninja writes "Dr. Gord Cormack decided to find and prove what the best spam filter is. In his study he looked at the major spam filters (DSPAM, SpamAssassin, etc.) along with those submitted by various academics. The results are quite surprising, with a previously unheard-of spam filter, which uses ideas from various compression algorithms, performing the best overall. He recently presented the results and methodology used in a presentation titled 'Spam Filters, Do they Work? and Can you prove it?'" Note that this is a video of his presentation.

Comments

In my experience...

[vivin]

... the ones which have worked best (for me) are Bayesian Spam Filters (A Plan for Spam, SpamBayes - a free filter) and CRM114 The Controllable Regex Mutilator (Paul Graham mentions it here). I've always had a very high success rate with these.

Re:In my experience...

[Red Alastor]

I like popfile because it's a bayesian filter that sorts into any arbitrary categories you want, not just spam and ham.

http://popfile.sourceforge.net/

Got to go with Brightmail

[saha]

We use Brightmail on our campus and our users love it with its very low false positive and pretty accurate flagging of SPAM. Another campus uses DSPAM and some people are up in arms at the prospect of losing their Brightmail to switch to DSPAM. Personally, DSPAM isn't nearly as good and has flagged many legitamate messages and sent them to the Junk folder.

I also echo a gripe of other posters. Its nice to have a video but 500MB video file it a bit much. A 50KB pie chart or bar graph would have been nice.

Flaw in the test

[lheal]

The spammers actively try to subvert the more popular filters. That gives a lesser-known one a decided advantage, one which will go away as it becomes more popular.

As with most choices like this, factors such as ease of use, speed, and resource efficiency can overshadow selectivity. No system is perfect, so it's perfectly reasonable to go with a system that's pretty good if you already are using it, rather than switching to the latest cool thing.

I have found that using two dissimilar systems in a chain is quite effective.

text versions of the material

[martin-boundary]

For those who don't relish downloading 400MB worth of video (why can't somebody cut out the audio as a standalone MP3?), the material of the talk is also available in text mode.

The official tests of spamfilters were done in last year's TREC conference, you can read the writeup here (or pdf overview).

You can duplicate those tests yourself if you download the evaluation toolkit (GPL). It's a modular system where you can add a mail corpus (either one of the public TREC ones, or you can make your own trivially), and add a spamfilter package (there are 10 or so to download from the web, or create your own as per documentation).

There's also a video talk given at Microsoft research which should cover pretty much the same ground, if text mode is slashdotted :).

There's a new scheduled test towards the end of the year at TREC 2006.

Possible Text Version

[sciop101]

On-line Supervised Spam Filter Evaluation
Gordon Cormack and Thomas Lynam

Full Text, May 29, 2006 - PDF Format

http://plg.uwaterloo.ca/~gvcormac/spamcormack.html /

Re:Possible Text Version

[gvc]

Bogofilter works great. Or SpamAssassin but only if you force-feed it its own judgements. In both cases you have to correct classification errors.

Fidelis Assis (who has now gone solo after having participated in the CRM114 project) shows great results for his recent solo effort: OSBF-lua Bratko's PPM spam filter -- the one that did great at TREC -- is not yet packaged as a drop-in filter. Same for my DMC spam filter.

The actual TREC 2005 tests referred to in TFA are here.

Out of Date and Worthless

[prandal]

This paper's a complete waste of time.

He tested spamassassin 2.3 - that's ancient! I'd imagine the other tools are similarly obsolete.

We currently use SA 3.1.4 with a well-trained Bayes database and Razor, Pyzor, and DCC.

Throw in a few custom rules and a selection of rules from http://www.rulesemporium.com/ and the results are outstanding.

With the new sa-update feature the core rules are updated between point releases, which came in useful this week dealing with the new image spams which seemed to be designed to avoid detection by spamassassin. Thanks Theo.

And the folk on the spamassassin-users mailing list really rock.

Re:Out of Date and Worthless

[gvc]

I assume the paper that you are describing is the 2004 study. The paper described in the talk (which was given 6 months ago or so) described results of the TREC 2005 Spam Track which took place in November 2005. It included a test SpamAssassin 3.x, not 2.3.

TREC 2006 evaluations are now underway.

While it is reasonable to conjecture that spam has changed so as to defeat spam filtering techniques, or will change so as to defeat the PPM technique that did well at TREC, the historical evidence does not support this conjecture. In particular:

• The spam filters tested in 2004 give pretty well exactly the same performance on 2005 and 2006 data.
• New versions of the filters are a little bit better, but not by leaps and bounds, and also get about the same results over the last 2.5 years of data.
• There is no evidence that "Bayesian poisining" is a viable technique for defeating statistical spam filters in anything but a very artifical laboratory environment where the poisoner has access to the recipient's inbox

The subject of the paper -- and the talk -- is primarily about testing methodology and the need for controlled scientific investigation. So I hesitate to endorse the simplistic notion of a "winner" of the TREC evaluation. However the technique that did very well was indeed quite novel, so here's a characterization.

Andrej Bratko used PPM -- a well-known data compression technique to compress ham and spam separately. Well actually he didn't compress them but just build the statistical model necessary to compress them. Then he simply (tentatively) added the unknown message to each model and chose the one that compressed it best. The general technique of using compression has been mentioned here and elsewhere but Bratko used a much stronger compression scheme and was somewhat clever about it.

I later reproduced Bratko's results using DMC -- a compression schem that I invented 20 years ago -- and got some interesting results. We have a journal article in press describing it and also an evaluation paper at CEAS 2006.

Bratko A., Cormack G. V., Filipic B., Lynam T. R. and Zupan B., Spam Filtering Using Statistical Data Compression Models

Torrent

[vivin]

Here is a torrent I made of the xvid file. It should work (I hope).

Dspam floats my boat

[Zzeep]

I receive (no kidding) around 600 spam mails per day, versus approximayely 30 real e-mails. I've been using dspam for over a year now (with very faithful training), and there is maybe 1 false positive every few weeks (less than 1 in 10.000) and every few days a few (usually "new") spam mails get through, which I ofcourse immediately train, to never see those kind again. So I am very very positive about dspam. What I do miss though is something like a good and reliable service (better than the RBL's I know) that can block SMTP clients on the fly (like DSL home users and such) to reduce the immense load on our mailservers (I work for an ISP) caused by all the spam (that also has to go through a virus scanner, clamav).