George,

Thanks again for the stats. These do verify that spammers are obfuscating the Yahoo redirection code and those lines need to stay in the filter as a result. At least I wasn't wasting my time when I came up with that stuff :)

I didn't get too much else out of the results though. Maybe I'll reorder the test types in the OBFUSCATION filter, and I did make a change to what will become the next version of GIBBERISH where I moved the "Words, Acronyms and Stock Market Symbols" section below the "Auto-generated Codes" section, but I don't yet see any need to tweak the files line for line, only section by section because management is important.

Matt



George Kulman wrote:

Matt,

Here are two analyses.  The 11-15 to 11-30 covers the period from when I
implemented your filters until I began using SKIPIFWEIGHT and MAXWEIGHT
which obviously has some effect on the stats.  The 11-15 to 12-21 expands
the prior set to include the additional filters.

There's also the weighting effect to consider.  While I run the OBFUSCATION
and Y!DIRECTED at hold weight (15), I use the GIBBERISH like the COMMENTS
test and accumulate weight per hit.  Since my SKIPIFWEIGHT is set to my
DELETE weight (60), the filters will run until that's reached.

These stats aren't a big deal to produce since its all in a SQL database.

I'll be implementing your new filter versions this coming weekend (with new
names to avoid commingling stats).  I do strip out comments since they
become meaningless as the filter contents are resequenced by my system.

George



-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Matthew Bramble
Sent: Monday, December 22, 2003 10:32 PM
To: [EMAIL PROTECTED]
Subject: Re: [Declude.JunkMail] GIBBERISH 2.0.1, single file filter with END functionality. functionality. functionality. functionality.



George,


I think that logic can get you 95% of the way there with something as convoluted as this, that is run only about 1/3 of the time, and considering that you are only battling for about 2% of the processing power required by this filter alone, which shouldn't be too terribly much. Removing the comment blocks would probably have a bigger effect :) Changing to the new version of the filter should definitely help, though this isn't by far my most weighty filter.

Here's something that I've very curious about though...the Y!DIRECTED filter contains a bunch of BODY searches for obfuscated strings, something that is almost totally redundant with the OBFUSCATION filter. I would be very curious to see how often those lines are hit because they could be dumped for a measurable performance increase. Any chance you want to take a crack at that? I wouldn't be surprised to see them never hit.

Matt



George Kulman wrote:



Matt,

I use LOGLEVEL HIGH for my data collection and analysis

stuff and, as Bill


pointed out, all hits are reflected.

I've started to use SKIPIFWEIGHT. The result of course is

that filters are


bypassed and the statistics are skewed.

For example on Friday 12/19, 15291 emails were processed by

Declude on my


system. Only 4604 were processed by the GIBBERISH filter.

Of these 1328


had a total of 3854 hits.

My quandary now is to decide whether to use the new control

functions of


SKIPIFWEIGHT, MAXWEIGHT and END to reduce processing

overhead or to collect


a full set of evaluation data by letting everything run.

It's truly a


catch-22 situation. If I collect all of the data, then I

gain no benefit,


since all of the processing takes place. If I take advantage of the
analysis data, I reduce my processing workload but


effectively destroy the


validity of the statistical data which is now skewed by my filtering
control.

George





-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Matthew Bramble
Sent: Monday, December 22, 2003 3:17 PM
To: [EMAIL PROTECTED]
Subject: Re: [Declude.JunkMail] GIBBERISH 2.0.1, single file filter with END functionality. functionality.



George,


That's good data to have. I would have to assume that something tagged as gibberish in the main test would be random, and that's

fairly well

indicated by the somewhat tight range of the two character

strings.

Unless you are using a logging feature that I'm not aware

of, you are

only showing the last hit that the filter produces, and

that explains

why the Z strings are mostly bunched at the top. I've got these ordered alphabetically and will probably leave them there for management purposes.

The counterbalances though are definitely something that I will use your information for reordering them. I believe I made an attempt to order these in the 2.0 filter version according to what I thought

would be

more common as well as what would be a faster search (BODY searches are slower than other things and will go lower in general,

though a BODY

search for base64 goes at the top because it is fairly common). Because of this and along with the above mentioned issue, the hit stats therefore aren't a perfect indication of what would save the most processing power, but it definitely helps if you just make some assumptions. I hadn't gathered any stats myself on the Auto-generated Codes that I added in about a month or so ago, and it's nice to see that they're getting hit since I was really just brainstorming

about what

types of things might be seen. I might remove some entries

though if

they aren't showing being hit since they are BODY searches and expensive. I'll probably still leave that list of Auto-generated Codes in alphabetical order though for management purposes. This

shouldn't

make a big difference considering that the most common one only gets hit about 1-3% of the time (don't know how common the filter fails a later line which ends up getting logged instead).

If Declude did log every line that hits in a filter, you would see things like GIBBERISH hitting some attachments thousands of

times per

message, and I don't think that's worth the trouble. Data

like this

will make a much bigger impact on performance if you run it against filters where hits can only occur once in a file due to unique data or exact matching. Kami has a bunch of those.

Thanks,

Matt



George Kulman wrote:





Matt,

I thought you might be interested in the attached data which



analyzes the




GIBBERISH and ANTI-GIBBERISH filters by number of hits on my



system from




11/15 through yesterday.

If you're looking for "effectiveness" you should set the entries in
descending order of probability. I use a variation which




looks at date of




most recent hit as well as hit count, although that's more



important with




filters that are being modified on a continual rather that a



fairly static




filter such as these two.

George







-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Matthew Bramble
Sent: Monday, December 22, 2003 9:52 AM
To: [EMAIL PROTECTED]
Subject: [Declude.JunkMail] GIBBERISH 2.0.1, single file filter with END functionality.



I've made some huge leaps forward recently in terms of the




processing



power required to run Declude with the custom filters that I have installed. This was done by way of the SKIPIFWEIGHT

functionality

introduced in the latest beta, but also by way of re-ordering my filters in the Global.cfg file so that the easiest to process custom filters are run first in the hopes of avoiding the need to run more

costly ones.


This new version of GIBBERISH makes use of functionality introduced in the 1.77 beta, however the most recent interim release, 1.77i7, should be used in order to guarantee proper operation (initial versions would always end processing, and effectively disabled the filters). The END functionality removes the need to have ANTI filters since the filter can be stopped before it gets to the main filter matches, and it also presents another opportunity to save on the processing power required to run such things. This also makes use of the MAXWEIGHT functionality to limit the max score as well as end processing once a single hit has been scored. Note that the filter will only log (at the LOW



setting) and



show WARN actions when the filter is tripped and an END was not hit...which is great! No more looking at non-scoring custom filters due to counterbalances :D

Please read through the file and follow these instructions if you already have GIBBERISH installed:

1) Comment out the ANTI-GIBBERISH custom filter in your



Global.cfg




2) Change the score of the GIBBERISH filter to 0 in your Global.cfg.
3) Change the scoring of the filter to match your


system (it is

scored by default for base 10 systems). This can be done
by changing the MAXWEIGHT and Main Filter lines to reflect the multiple of 10 that your system is based on.
4) Change the SKIPIFWEIGHT score to reflect your delete weight, or whatever weight you would like for the filter to
be skipped if the system has already reached it before processing the filter.


The file can be downloaded from the following location:







http://www.mailpure.com/software/decludefilters/gibberish/Gib




berish_v2-0-1.z




ip

Please report any issues with the new filter format. As

soon as bugs

stop being reported, I will move to convert the other dual

file filters

into single file alternatives which make use of the END

functionality.

Until the functionality goes into a full release, I'm going

to continue

to primarily provide the old style filters on my site.

Matt




---
[This E-mail was scanned for viruses by Declude Virus (http://www.declude.com)]

---
This E-mail came from the Declude.JunkMail mailing list.  To
unsubscribe, just send an E-mail to [EMAIL PROTECTED], and
type "unsubscribe Declude.JunkMail".  The archives can be found
at http://www.mail-archive.com.

Reply via email to