[Spambayes] Spammers vetting with Bayesian Filters?

[David McNab]

Hi,

I've been training my SpamBayes setup for 2 weeks now. Initially things
had gotten really peaceful, but I'm now seeing an increasing number of
spams each day that are getting labelled as 'ham'.

>From this, I can only hazard to guess that some of the major spamhausen
are now vetting their messages with Bayesian filters.

Given the revenue that can be 'earned' from sending spam, it would be
well worth any spammer's time and energy to set up several
bayesian-based filters, train them, then use them for vetting messages.

A few iterations of rewording messages and checking bayesian scores, and
the message is ready for sending.

I'm even seeing a few nigerian scam variants making it through as 'ham'.

Does anyone have any thoughts about this?

I switched from text-based challenge/response to Bayesian, because of
concerns that some new correspondents would feel deterred/insulted by
the challenge, or be too technically n00bish to understand what's going
on.

So, where does this leave us in the arms race?
How long will it be before even the best-trained bayesian setups start
getting false positive and false negative rates climbing towards 50%?

Could challenge/response, with increasingly difficult turing tests,
become the norm?

-- 
Cheers
David