Spammers catching on

[Matt Armstrong]

Zack Brown <zbrown at tumblerings.org> writes:

> On Mon, Dec 16, 2002 at 08:09:02PM -0500, Suzanne Skinner wrote:
>> Here's another case of sneaky, bayesian-filter-evading spammers. This one,
>> alas, was successful: it scored around 0.5 with Robinson-Fisher. Fortunately,
>> my SpamAssassin backup (threshold 10.0) caught it with a whopping 29 point
>> score.
>
> It's good that spammers are exploring these strategies so
> early. They give us good 'food' for thought...
>
> I suspect the next great leap forward in anti-spam techniques will
> involve neural nets. After all, that's their big strength -
> categorizing things with a sophistication that human-readable
> algorithms (like bayesian filters) have trouble with.
>
> I imagine the front end of such a tool would be almost identical to
> bogofilter today - new email could be used to train the system
> further, and false positives/negatives could be corrected by hand.
> But instead of the limited system of tokenizing and forming stats,
> the whole email would be sent through the neural system and
> processed according to whatever structures that system chose to
> recognize.

It might also be possible to have pseudo-tokens in the "word"
database.  E.g. .html-font-same-foreground-as-background could be in
the database with a good and SPAM count.  Then it could count against
messages having that property.

The novelty over SpamAssassin is that the filter would still be
training itself.