[pLog-svn] r6088 - plog/branches/lifetype-1.2/class/security

[Mark Wu]

Hi Jon:

Actually, the situation won't happened in our scenario.

We don't have the mechanism to tell the next pipeline filter that the
previous one is invalid, and you need to treat the comment as spam and train
it again.

So, even we run again all filters again, the Bayesian filter still treat the
comment  as a"non-spam" comment

But, another scenario may happened:

If the Bayesian filter caught it as a spam, but his auhtimage is wrong. At
that time, the comment already add to database.

That's why some people complain to me the auhtimage fails when comment
marked as spam.

Even so, we still need to run all filters once, not twice.

Mark

> -----Original Message-----
> From: plog-svn-bounces at devel.lifetype.net 
> [mailto:plog-svn-bounces at devel.lifetype.net] On Behalf Of Jon Daley
> Sent: Thursday, November 29, 2007 11:29 PM
> To: LifeType Developer List
> Subject: Re: [pLog-svn] r6088 - 
> plog/branches/lifetype-1.2/class/security
> 
> Ok, so think about this scenario:
> 
> Comment is not caught by bayesian filter, but is caught by auth image.
> 
> Bayesian filter trains as non-spam.
> Auth image catches it as spam.
> 
> The old method would re-run the bayesian filter so it trains 
> the text as spam, so even if the spammer figures out the 
> auth-image code, his text is now trained correctly.
> 
> The new method causes the spam text to be trained as non-spam 
> and left in the database, making the bayesian filter not work as well.
> 
> This is why we need to run the pipeline twice.