On Wed, 2006-10-25 at 01:33 +0200, Michael Vasiliev wrote:
> On Monday October 23 2006 20:28, Gil Freund wrote:
> > Hi,
> >
> > I am using a Amavis+SA+CLAM for mail filtering (debian sarge packages).
> > Recently I am being hit by a lot of image spam. Bayesian filtering and 
> > RBL's are not enough.
> 
> These mails are terrible, while my personal accounts get a few dozen each, 
> Linux-IL posting address gets a hundred per day minimum. If we only relied on 
> SA's checking, this list would look much worse than LKML. Most of these 
> emails are a crazy mix of sentences on random topics, probably web-crawler 
> generated, plus a gif image with a nasty ad of some kind (I never care enough 
> to mimedecode them).

The problem is that the image has some random noise added to it, so you
can't even get your bayesian filter to recognize the mime data. This
will also, IMO, in the long run defeat OCRs in the same way that
CAPTCHAs defeat OCRs.

> SA's Bayesian filtering is terrible, I don't know why. (All right, I admit I 
> am too lazy and illiterate at adult-level mathematics to figure out why). I'm 
> running SA and bogofilter in parallel, clearly bogofilter is at least twenty 
> times more accurate. 

I'm using bogofilter in conjunction with a large set of RBLs, and while
I do see a surge in the number of small-caps and viagra image spam, only
rarely there is a false negative - about 4-5 times a week.

--
Oded
::..
If you lost your left arm, your right arm would be left.



=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]

Reply via email to