[SpamCop-List] Re: Mail Daemon Spam - what is it?

[Michael Lefevre]

Marjolein Katsma wrote:
> Michael Lefevre (michael.spamcop at michaellefevre.com) wrote in
> news:c24olm$vvi$1 at news.spamcop.net: 
>> The killer point is that bounces tend to include the original
>> content, which may well be an actual spam, and depending on the format
>> of the bounce, it can be tricky to work out where the included message
>> begins.
>
> Um, no, that's not a killer point - on the contrary, it's a way to
> actually recognize a real bounce. *If* that pattern occurs, it's most
> likley a bounce (not that every bounce would have the original content):
> it's one of the things a human uses to recognize a real bounce, and a
> nice example of what could be used to develop a better heuristic to
> recognize actual bounces. 
>
> Actually, you could probably use bayesian technology to "teach" a system
> to recognize real bounces. 

Yes, I was thinking about it in terms of using the existing code.

[snip]
> I don't think it's necessarily a lot of work. Bayesian and other
> pattern-recognition algorithms are already out there; it's more a matter
> of wiring them together.

That's why it is a lot of work though, I think.

There are Bayesian systems out there, but I'm not there are too many that
you could copy and paste into a Perl CGI.

Even if you could do Bayesian, you'd need accurate humans to train it. The
problem we're trying to solve is that the humans are getting things
wrong.

>> Anyway, I'm just pointing out why it's not that easy to fix - I
>> imagine Julian will do something at some point...
>
> I won't say it's easy - but I'm sure it's possible. And at some point,
> something will have to be done - start thinking and testing now! 

I think Julian is still busy trying to solve bigger problems which have
existed for some time, like Spamcop falling for forged headers...

-- 
Michael