Re: [fetchmail-devel] [Bug #15103] Fetchmail Fails with ASSP (http://assp.sourceforge.net/)

[Frederic M. <fre...@wo...>]

Matthias Andree a écrit :
> Frederic Marchal schrieb:
>
> [in response to BerliOS Bug #15103]
>
>   
>> I solved this issue by running ASSP in test mode (check option "Bayesian 
>> Test Mode") and prefixing the subject with something like [SPAM] (in 
>> option "Prepend Spam Subject"). It means ASSP won't stop mails seen as 
>> bayesian spam but will tag the subject. The mails can then be filtered 
>> out automatically by the final recipient.
>>     
>
> This, and my being bogofilter's co-maintainer (bogofilter also does
> Bayesian filtering), prompted me to have a glimpse at ASSP, and I must say
> that the promises it makes are way too bold. ASSP's self-advertising as the
> best tool aside, the assertion that Bayes filter were to intelligently
> decide is just crap. Bayes filters are making statistical decisions based
> on past training. "Maintenance free" doesn't work for them.
>   
ASSP's implementation of Paul Graham's algorithm is the most inefficient 
I know but it does more than bayesian statistics and that compensate for 
its incapacity at properly classify e-mails. In my case, I have a bunch 
of spamtraps addresses and ASSP learns from the e-mails it nets. Since 
spammers are sending tons of e-mails to every addresses they know, we 
can filter them out by simple comparison between the mails received by 
the spamtraps and the mails received by regular users. So, it does its 
own training even though a Bayes filter is not adequate considering how 
it performs in ASSP.

Beside, ASSP sits on the SMTP port. Therefore, it sees the mails sent by 
local users too and can whitelist the responses, preventing them from 
being filtered out and learning hams in the process. It can also learn 
from SMTP violations the spammers usually do and blacklist the offending 
senders very early and for every user of the server.

I don't believe ASSP is a great piece of software but it is the only one 
I know of that can do that and it does require very little or no action 
from the users it protect (especially no command line to run to learn 
more spam/ham).

>> By the way, I don't advise to let ASSP delete or reject blindly bayesian 
>> spams because, as you noticed, it doesn't take the whole mail into 
>> account and, in general, it does a poor job at filtering spams. It tends 
>> to have a high rate of false positives and you will loose good e-mails 
>> if you don't keep an eye on what it does.
>>     
>
> But then it seems to me that if you're using fetchmail or getmail, that
> integrating bogofilter, spamprobe or qsf in the mail system is more
> efficient than ASSP if it's prone to high resource use.
>
> At least bogofilter and SpamProbe are much faster than SpamAssassin's bayes
> mode. :-)
>   
Thanks for the advertisement :-) but my users are all windows 
(mal-)formatted and it is impossible to get them to run a linux command 
to classify their e-mails...

Frederic