qsf - quick spam filter

Filtering: qsf [-snrAtav] [-d DB] [-g DB]

[-L LVL] [-S SUBJ] [-Q NUM] [-X NUM]

Training: qsf -T SPAM NONSPAM [MAXROUNDS] [-d DB]

Retraining: qsf -[m|M] [-d DB] [-w WEIGHT] [-aN]

Database: qsf -[p|D|R|O] [-d DB]

Database merge: qsf -E OTHERDB [-d DB]

Allowlist query: qsf -e EMAIL [-m|-M|-t] [-d DB] [-g DB]

Help: qsf -[h|l|V]

qsf reads a single email on standard input, and by default outputs it on standard output. If the email is determined to be spam, an additional header ("X-Spam: YES") will be added, and optionally the subject line can have "[SPAM]" prepended to it.

qsf is intended to be used in a procmail(1) recipe, in a ruleset such as this:

:0 wf | qsf -ra :0 H: * X-Spam: YES $HOME/mail/spam

For more examples, including sample procmail(1) recipes, see the EXAMPLES section below.

Before qsf can be used properly, it needs to be trained. A good way to train qsf is to collect a copy of all your email into two folders - one for spam, and one for non-spam. Once you have done this, you can use the training function, like this:

qsf -aT spam-folder non-spam-folder

This will generate a database that can be used by qsf to guess whether email received in the future is spam or not. Note that this initial training run may take a long time, but you should only need to do it once.

To mark a single message as spam, pipe it to qsf with the --mark-spam or -m ("mark as spam") option. This will update the database accordingly and discard the email.

To mark a single message as non-spam, pipe it to qsf with the --mark-nonspam or -M ("mark as non-spam") option. Again, this will discard the email.

If a message has been mis-tagged, simply send it to qsf as the opposite type, i.e. if it has been mistakenly tagged as spam, pipe it into qsf --mark-nonspam --weight=2 to add it to the non-spam side of the database with double the usual weighting.

The qsf options are listed below.

-d, --database [TYPE:]FILE

Use FILE as the spam/non-spam database. The default is to use /var/lib/qsfdb and, if that is not available or is read-only, $HOME/.qsfdb. This option can also be useful if there is a system-wide database but you do not want to use it - specifying your own here will override the default.

If you prefix the filename with a TYPE, of the form btree:$HOME/.qsfdb, then this will specify what kind of database FILE is, such as btree, gdbm, sqlite and so on. Check the output of qsf -V to see which database backends are available. The default is to auto-detect the type, or, if the file does not already exist, use btree. Note that TYPE is not case-sensitive.

-g, --global [TYPE:]FILE

Use FILE as the default global database, instead of /var/lib/qsfdb. If you also specify a database with -d, then this "global" database will be used in read-only mode in conjunction with the read-write database specified with -d. The -g option can be used a second time to specify a third database, which will also be used in read-only mode. Again, the filename can optionally be prefixed with a TYPE which specifies the database type.

-s, --subject

Rewrite the Subject line of any email that turns out to be spam, adding "[SPAM]" to the start of the line.

-S, --subject-marker SUBJECT

Instead of adding "[SPAM]", add SUBJECT to the Subject line of any email that turns out to be spam. Implies -s.

-n, --no-header

Do not add an X-Spam header to messages.

-r, --add-rating

Insert an additional header X-Spam-Rating which is a rating of the "spamminess" of a message from 0 to 100; 90 and above are counted as spam, anything under 90 is not considered spam. If combined with -t, then the rating (0-100) will be output, on its own, on standard output.

-A, --asterisk

Insert an additional header X-Spam-Level which will contain between 0 and 20 asterisks (*), depending on the spam rating.

-t, --test

Instead of passing the message out on standard output, output nothing, and exit 0 if the message is not spam, or exit 1 if the message is spam. If combined with -r, then the spam rating will be output on standard output.

-a, --allowlist

Enable the allow-list. This causes the email address given in the message's "From:" header to be checked against a list; if it matches, then the message is always treated as non-spam, regardless of what the token database says. When specified with a retraining flag, -a -m (mark as spam) will remove that address from the allow-list as well as marking the message as spam, and -a -M (mark as non-spam) will add that address to the allow-list as well as marking the message as non-spam. The idea is that you add all of your friends to the allow-list, and then none of their messages ever get marked as spam.

-L, --level, --threshold LEVEL

Change the spam scoring threshold level which must be reached before an email is classified as spam. The default is 90.

-Q, --min-tokens NUM

Only give a score if more than NUM tokens are found in the message - otherwise the message is assumed to be non-spam, and it is not modified in any way. The default is 0. This option might be useful if you find that very short messages are being frequently miscategorised.

-e, --email, --email-only EMAIL

Query or update the allow-list entry for the email address EMAIL. With no other options, this will simply output "YES" if EMAIL is in the allow-list, or "NO" if it is not. With -t, it will not output anything, but will exit 0 (success) if EMAIL is in the allow-list, or 1 (failure) if it is not. With the -m (mark-spam) option, any previous allow-list entry for EMAIL will be removed. Finally, with the -M (mark-nonspam) option, EMAIL will be added to the allow-list if it is not already on it.

If EMAIL is just the word MSG on its own, then an email will be read from standard input, and the email address given in the "From:" header will be used.

Using -e automatically switches on -a.

-v, --verbose

Add extra X-QSF-Info headers to any filtered email, containing error messages and so on if applicable. Specify -v more than once to increase verbosity.

-T, --train SPAM NONSPAM [MAXROUNDS]

Train the database using the two mbox folders SPAM and NONSPAM, by testing each message in each folder and updating the database each time a message is miscategorised. This is done several times, and may take a while to run. Specify the -a (allow-list) flag to add every sender in the NONSPAM folder to your allow-list as a side-effect of the training process. If MAXROUNDS is specified, training will end after this number of rounds if the results are still not good enough. The default is a maximum of 200 rounds.

-m, --mark-spam

Instead of passing the message out on standard output, mark its contents as spam and update the database accordingly. If the allow-list (-a) is enabled, the message's "From:" address is removed from the allow-list.

-M, --mark-nonspam

Instead of passing the message out on standard output, mark its contents as non-spam and update the database accordingly. If the allow-list (-a) is enabled, the message's "From:" address is added to the allow-list (see the -a option above).

-w, --weight WEIGHT

When marking as spam or non-spam, update the database with a weighting of WEIGHT per token instead of the default of 1. Useful when correcting mistakes, eg a message that has been mistakenly detected as spam should be marked as non-spam using a weighting of 2, i.e. double the usual weighting, to counteract the error.

-N, --no-autoprune

When marking as spam or nonspam, never automatically prune the database. Usually the database is pruned after every 500 marks; if you would rather --prune manually, use -N to disable automatic pruning.

-p, --prune

Remove redundant entries from the database and clean it up a little. This is automatically done after several calls to --mark-spam or --mark-nonspam, and during training with --train if the training takes a large number of rounds, so it should rarely be necessary to use --prune manually unless you are using -N / --no-autoprune.

-X, --prune-max NUM

When the database is being pruned, no more than NUM entries will be considered for removal. This is to prevent CPU and memory resources being taken over. The default is 100,000 but in some circumstances (if you find that pruning takes too long) this option may be used to reduce it to a more manageable number.

-D, --dump [FILE]

Dump the contents of the database as a platform-independent text file, suitable for archival, transfer to another machine, and so on. The data is output on stdout or into the given FILE.

-R, --restore [FILE]

Rebuild the database from scratch from the text file on stdin. If a FILE is given, data is read from there instead of from stdin.

-O, --tokens

Instead of filtering, output a list of the tokens found in the message read from standard input, along with the number of times each token was found. This is only useful if you want to use qsf as a general tokeniser for use with another filtering package.

-E, --merge OTHERDB

Merge the OTHERDB database into the current database. This can be useful if you want to take one user's mailbox and merge it into the system-wide one, for instance (this would be done by, as root, doing qsf -d /var/lib/qsfdb -E /home/user/.qsfdb and then removing /home/user/.qsfdb).

-B, --benchmark SPAM NONSPAM [MAXROUNDS]

Benchmark the training process using the two mbox folders SPAM and NONSPAM. A temporary database is created and trained using the first 75% of the messages in each folder, and then the entire contents of each folder is tested to see how many false positives and false negatives occur. Some timing information is also displayed.

This can be used to decide which backend is best on your system. Use -d to select a backend, eg qsf -B spam nonspam -d GDBM - this will create a temporary database which is removed afterwards.

The exception to this is the MySQL backend, where a full database specification must be given (-d MySQL:database=db;host=localhost;...) and the database table given will not be wiped beforehand or dropped afterwards.

As with -T, if MAXROUNDS is specified, training will never be done for more than this number of rounds; the default is 200.

-h, --help

Print a usage message on standard output and exit successfully.

-l, --license

Print details of the program's license on standard output and exit successfully.

-V, --version

Print version information, including a list of available database backends, on standard output and exit successfully.

/var/lib/qsfdb

The default (system-wide) spam database. If you wish to install qsf system-wide, this should be read-only to everyone; there should be one user with write access who can update the spam database with qsf --mark-spam and qsf --mark-non-spam when necessary.

/var/lib/qsfdb2

A second, read-only, system-wide database. This can be useful when installing qsf system-wide and using third-party spam databases; the first global database can be updated with system-specific changes, and this second database can be periodically updated when the third-party spam database is updated.

$HOME/.qsfdb

The default spam database for per-user data. Users without write access to the system-wide database will have their data written here, and the two databases will be read together. The per-user database will be given a weighting equivalent to 10 times the weighting of the global database.

Currently, you cannot use qsf to check for spam while the database is being updated. This means that while an update is in progress, all email is passed through as non-spam.

There is an upper size limit of 512Kb on incoming email; anything larger than this is just passed through as non-spam, to avoid tying up machine resources.

To filter all of your mail through qsf, with the allow-list enabled and the "spam rating" header being added, add this to your .procmailrc file:

:0 wf | qsf -ra

If you want qsf to add "[SPAM]" to the subject line of any messages it thinks are spam, do this instead:

:0 wf | qsf -sra

To automatically mark any email sent to spambox@yourdomain.com as spam (this is the "naive" version):

:0 H * ^To:.*spambox@yourdomain.com | qsf -am

To do the same, but cleverly, so that only email to spambox@yourdomain.com which qsf does NOT already classify as spam gets marked as spam in the database (this stops the database getting too heavily weighted):

# If sent to spambox@yourdomain.com: :0 * ^To:.*spambox@yourdomain.com { :0 wf | qsf -a

# The above two lines can be skipped if you've # already piped the message through qsf.

# If the qsf database says it's not spam, # mark it as spam! :0 H * ^X-Spam: NO | qsf -am }

Remove the -a option in the above examples if you don't want to use the allow-list.

A more complicated filtering example - this will only run qsf on messages which don't have a subject line saying "your <something> is on fire" and which don't have a sender address ending in "@foobar.com", meaning that messages with that subject line OR that sender address will NEVER be marked as spam, no matter what:

:0 wf * ! ^Subject: Your .* is on fire * ! ^From: .*@foobar.com | qsf -ra

For more on procmail(1) recipes, see the procmailrc(5) and procmailex(5) manual pages.

A couple of macros to add to your .muttrc file, if you use mutt(1) as a mail user agent:

# Press F5 to mark a message as spam and delete it macro index <f5> "<pipe-message>qsf -am\n<delete-message>" macro pager <f5> "<pipe-message>qsf -am\n<delete-message>" # Press F9 to mark a message as non-spam macro index <f9> "<pipe-message>qsf -aM\n" macro pager <f9> "<pipe-message>qsf -aM\n"

Again, remove the -a option in the above examples if you don't want to use the allow-list.

Note, however, that the above macros won't work when operating on multiple tagged messages. For that, you'd need something like this:

macro index <f5> ":set pipe_split\n<tag-prefix><pipe-message>qsf -am\n<tag-prefix><delete-message>\n:unset pipe_split\n"

If you use qmail(7), then to get procmail working with it you will need to put a line containing just DEFAULT=./Maildir/ at the top of your ~/.procmailrc file, so that procmail delivers to your Maildir folder instead of trying to deliver to /var/spool/mail/$USER, and you will need to put this in your ~/.qmail file:

| preline procmail

This will cause all your mail to be delivered via procmail instead of being delivered directly into your mail directory.

See the qmail(7) documentation for more about mail delivery with qmail.

If you use postfix(1), you can set up a system-wide mail filter by creating a user account for the purpose of filtering mail, populating that account's .qsfdb, and then creating a shell script, to run as that user, which runs qsf on stdin and passes stdout to sendmail(8).

Doing this requires some knowledge of postfix configuration and care needs to be taken to avoid mail loops. One qsf user's full HOWTO is included in the doc/ directory with this package.

A feature called the "allow-list" can be switched on by specifying the --allowlist or -a option. This causes messages' "From:" addresses to be checked against a list of people you have said to allow all messages from, and if a message's "From:" address is in the list, it is never marked as spam. This means you can add all your friends to an "allow-list" and qsf will then never mis-file their messages - a quick way to do this is to use -a with -T (train); everyone in your non-spam folder who has sent you an email will be added to the allow-list automatically during training.

You can manually add and remove addresses to and from the allow-list using the -e (email) option. For instance, to add foo@bar.com to the allow-list, do this:

qsf -e foo@bar.com -M

To remove bad@nasty.com from the allow-list, do this:

qsf -e bad@nasty.com -m

And to see whether someone@somewhere.com is in the allow-list or not, just do this:

qsf -e someone@somewhere.com

In general, you probably always want to enable the allow-list, so always specify the -a option when using qsf. This will automatically maintain the allow-list based on what you classify as spam or non-spam.

The only times you might want to turn it off are when people on your allow-list are prone to getting viruses or if a virus is causing email to be sent to you that is pretending to be from someone on your allow-list.

Because the database format is platform-specific, it is a good idea to periodically dump the database to a text file using qsf -D so that, if necessary, it can be transferred to another machine and restored with qsf -R later on.

Also note that since the actual contents of email messages are never stored in the database (see TECHNICAL DETAILS), you can safely share your qsf database with friends - simply dump your database to a file, like this:

qsf -D > your-database-dump.txt

Once you have sent your-database-dump.txt to another person, they can do this:

qsf -R < your-database-dump.txt

They will then have an identical database to yours.

When a message is passed to qsf, any attachments are decoded, all HTML elements are removed, and the message text is then broken up into "tokens", where a "token" is a single word or URL. Each token is hashed using the MD5 algorithm (see below for why), and that hash is then used to look up each token in the qsf database.

For full details of which parts of an email (headers, body, attachments, etc) are used to calculate the spam rating, see the TOKENISATION section below.

Within the database, each token has two numbers associated with it: the number of times that token has been seen in spam, and the number of times it has been seen in non-spam. These two numbers, along with the total number of spam and non-spam messages seen, are then used to give a "spamminess" value for that particular token. This "spamminess" value ranges from "definitely not spammy" at one end of the scale, through "neutral" in the middle, up to "definitely spammy" at the other end.

Once a "spamminess" value has been calculated for all of the tokens in the message, a summary calculation is made to give an overall "is this spam?" probability rating for the message. If the overall probability is 0.9 or above, the message is flagged as spam.

In addition to the probability test is the "allow-list". If enabled (with the -a option), the whole probability check is skipped if the sender of the message is listed in the allow-list, and the message is not marked as spam.

When training the database, a message is split up into tokens as described above, and then the numbers in the database for each token are simply added to: if you tell qsf that a message is spam, it adds one to the "number of times seen in spam" counter for each token, and if you tell it a message is not spam, it adds one to the "number of times seen in non-spam" counter for each token. If you specify a weight, with -w, then the number you specify is added instead of one.

To stop the database growing uncontrollably, it is automatically pruned after every 500th update (or, in training mode, after every 15 rounds). The pruning step removes any tokens with very small message counts, and scales all counts down if they start getting too large.

Finally, the reason MD5 hashes were used is privacy. If the actual tokens from the messages, and the actual email addresses in the allow-list, were stored, you could not share a single qsf database between multiple users because bits of everyone's messages would be in the database - things like emailed passwords, keywords relating to personal gossip, and so on. So a hash is stored instead. A hash is a "one-way" function; it is easy to turn a token into a hash but very hard (some might say impossible) to turn a hash back into the token that created it. This means that you end up with a database with no personal information in it.

When a message is broken up into tokens, various parts of the message are treated in different ways.

First, all header fields are discarded, except for the important ones: From, Sender, To, Reply-To, and Subject.

Next, any MIME-encoded attachments are decoded. Any attachments whose MIME type starts with "text/" (i.e. HTML and text) are tokenised, after having any HTML tags stripped. Any non-textual attachments are replaced with their MD5 hash (such that two identical attachments will have the same hash), and that hash is then used as a token.

In addition to single-word tokens from textual message parts, qsf adds doubled-up tokens so that word pairs get added to the database. This makes the database a bit bigger (although the automatic pruning tends to take care of that) but makes matching more exact.

As well as using the textual content of email to detect spam, qsf also uses special filters which create "pseudo-tokens" based on various rules. This means that specific patterns, not just individual words, can be used to determine whether a message is spam or not.

For example, if a message contains lots of words with multiple consonants, like "ashjkbnxcsdjh", then each time a word like that is seen the special token ".GIBBERISH-CONSONANTS." is added to the list of tokens found in the message. If it turns out that most messages with words that trigger this filter rule are spam, then other messages with gibberish consonant strings will be more likely to be flagged as spam.

Currently the special filters are:

GTUBE

Flags any message containing the string XJS*C4JDBQADN1.NSBN3*2IDNEN*GTUBE-STANDARD-ANTI-UBE-TEST-EMAIL*C.34X as spam - useful for testing that your qsf installation is working.

ATTACH-SCR

ATTACH-PIF

ATTACH-EXE

ATTACH-VBS

ATTACH-VBA

ATTACH-LNK

ATTACH-COM

ATTACH-BAT

Adds a token for every attachment whose filename ends in ".scr", ".pif", ".exe", ".vbs", ".vba", ".lnk", ".com", and ".bat" respectively (these are often viruses).

ATTACH-GIF

ATTACH-JPG

ATTACH-PNG

Adds a token for every attachment whose filename ends in ".gif", ".jpg" or ".jpeg", and ".png" respectively.

ATTACH-DOC

ATTACH-XLS

ATTACH-PDF

Adds a token for every attachment whose filename ends in ".doc", ".xls", or ".pdf" respectively (these tend to indicate a non-spam email).

SINGLE-IMAGE

Adds a token if the message contains exactly one attached image.

MULTIPLE-IMAGES

Adds a token if the message contains more than one attached image.

GIBBERISH-CONSONANTS

Adds a token for every word found that has multiple consonants in a row, as described above. Spam often contains strings of gibberish.

GIBBERISH-VOWELS

Adds a token for every word found that has multiple vowels in a row, eg "aeaiaiaeeio".

GIBBERISH-FROMCONS

Like GIBBERISH-CONSONANTS, but only for the "From:" address on its own.

GIBBERISH-FROMVOWL

Like GIBBERISH-VOWELS, but only for the "From:" address on its own.

GIBBERISH-BADSTART

Adds a token for every word that starts with a bad character such as %.

GIBBERISH-HYPHENS

Adds a token for every word with more than three hyphens or underscores in it.

GIBBERISH-LONGWORDS

Adds a token for every word with over 30 characters in it (but less than 60).

HTML-COMMENTS-IN-WORDS

Adds a token for every HTML comment found in the middle of a word. Spam often contains HTML inside words, like this: w<!--dsgfhsdgjgh-->ord

HTML-EXTERNAL-IMG

Adds a token for every HTML <img> (image) tag found that contains :// (i.e. it refers to an external image).

HTML-FONT

Adds a token for every HTML <font> tag found.

HTML-IP-IN-URLS

Adds a token for every URL found containing an IP address.

HTML-INT-IN-URL

Adds a token for every URL found containing an integer in its hostname.

HTML-URLENCODED-URL

Adds a token for every URL found containing a % sign in its hostname.

Normally, filters will just cause a token to be added, and these tokens are processed by the normal weighting algorithm. However the GTUBE filter will immediately flag any matching message as spam, bypassing the token matching.

The inbuilt binary tree database backend ("btree") will not necessarily provide the best performance, but is provided because using it requires no external libraries.

If, when qsf was compiled, the correct libraries were available, then it will be possible to use qsf with alternative database backends. To find out which backends you have available, run qsf -V (capital V) and read the second line of output. To see how well a backend performs, collect some spam and non-spam and use qsf -d BACKEND -B SPAM NONSPAM (see the entry for -B above).

Some people find that they get the best performance out of the gdbm backend; this is a library that is widely available on many systems.

To efficiently share a qsf database across multiple machines, you may find the MySQL backend useful. However, using it is a little more complicated.

To use the MySQL backend you will need to create a table with the fields key1, key2, token, value1, value2 and value3. The token, value1, value2, and value3 fields must be VARCHAR(6), BIGINT or INT, and BIGINT or INT respectively, and indexing on the token field is a good idea. The key1 and key2 fields can be anything, but they must be present.

For example:

USE mydatabase; CREATE TABLE qsfdb ( key1 BIGINT UNSIGNED NOT NULL, key2 BIGINT UNSIGNED NOT NULL, token VARCHAR(6) DEFAULT '' NOT NULL, value1 INT UNSIGNED NOT NULL, value2 INT UNSIGNED NOT NULL, value3 INT UNSIGNED NOT NULL, PRIMARY KEY (key1,key2,token), KEY (key1), KEY (key2), KEY (token) );

The key1 and key2 fields allow you to have multiple qsf databases in one table, by specifying different key1 and key2 values on invocation.

Instead of specifying a database file with the --database / -d option, you must specify either a specification string as described below, or the name of a file containing such a string on its first line.

The specification string is as follows:

database=DATABASE;host=HOST;port=PORT; user=USER;pass=PASS;table=TABLE; key1=KEY1;key2=KEY2

This string must be all on one line, with no spaces.

DATABASE

is the name of the MySQL database.

HOST

is the hostname of the database server (eg "localhost").

PORT

is the TCP port to connect on (eg 3306).

USER

is the username to connect with.

PASS

is the password to connect with.

TABLE

is the database table to use. If a table with this name does not exist when qsf is called in update or training mode, then it will be created if permissions allow this to be done.

KEY1

is the value to use for the key1 field.

KEY2

is the value to use for the key2 field.

Since command lines can be seen in the process list, it is probably best to specify a filename (eg qsf -d mysql:qsfdb.spec) and put the specification string inside that file.

If you have problems with qsf, please check the list below; if this does not help, go to the qsf home page and investigate the mailing lists, or email the author.

Nothing is being marked as spam.

First, use the -r option to switch on the X-Spam-Rating header, and check that this header appears in email passed through qsf. If it does not, then it is likely that qsf is not being run at all - check your configuration of procmail(1) or its equivalent.

If you are seeing X-Spam-Rating headers, and different emails have different scores, then you may simply need to retrain your database a little more. Take more spam email and pass it to qsf -m.

If you are seeing X-Spam-Rating headers but they all give the same spam rating, then the most likely reason is that qsf is not reading any database. Make sure that whatever is processing the email has read permissions on /var/lib/qsfdb and/or ~/.qsfdb - and make sure that, if you are using ~/.qsfdb, what your database creator thought was ~ ($HOME) is the same as it is for whatever is processing the email.

Retraining sometimes takes a very long time.

With the obtree backend or 2-column MySQL or SQLite tables, every 500th retrain (-m or -M), the database is pruned. On some systems this may take some time, and during this time the database is locked (except when using the MySQL or SQLite backends). If you constantly do a lot of retraining and want to avoid this, then use the -N option to suppress auto-pruning, and then have a cron(8) job or something run a manual prune (qsf -p) every now and again.

Running qsf from procmail fails with an error.

If you can run qsf from the command line, but in your procmail log file you get errors about "qsf: cannot execute binary file", then contact your system administrator for help. It may be that incoming email is handled by a different server to the one you normally shell into, and either they are of a different architecture or operating system, or the mail server is not permitted to execute user-owned binaries.

The following people have contributed suggestions, comments, patches, and testing:

Tom Parker <http://www.bits.bris.ac.uk/palfrey/>

Dr Kelly A. Parker

Vesselin Mladenov <http://www.antipodes.bg/>

Glyn Faulkner

Mark Reynolds

Sam Roberts

Scott Allen

Karsten Kankowski

M. Kolbl

Micha Holzmann

Jef Poskanzer <http://www.acme.com/jef/>

Clemens Fischer <http://ino-waiting.gmxhome.de/>

Nelson A. de Oliveira

The author:

Andrew Wood <andrew.wood@ivarch.com>

http://www.ivarch.com/

Project home page:

http://www.ivarch.com/programs/qsf/

If you find any bugs, please contact the author, either by email or by using the contact form on the web site.

procmail(1), procmailrc(5), procmailex(5)

Someone has written a guide to using qsf with KMail that can be found at:

http://www.softwaredesign.co.uk/Information.SpamFilters.html

This is free software, distributed under the ARTISTIC license.