[Spambayes] interesting paper: _Exploiting Machine Learning to Subvert Your Spam Filter_ (fwd)

[Justin Mason]

Hi folks -- in case you haven't yet seen this:


  'This paper shows how an adversary can exploit statistical machine
  learning, as used in the SpamBayes spam filter, to render it useless--even
  if the adversary's access is limited to only 1% of the training messages.
  We further demonstrate a new class of focused attacks that successfully
  prevent victims from receiving specific email messages. Finally, we
  introduce two new types of defenses against these attacks.'

  http://www.usenix.org/event/leet08/tech/full_papers/nelson/nelson_html/


Basically, measuring the effects of loading spams with huge dictionaries
in order to increase false positive frequencies, once the mail has been
trained on.

Would be interested to hear what people think -- personally:

- 1. this is very similar to
  http://www.cs.dal.ca/research/techreports/2004/CS-2004-06.shtml , and I
  haven't seen spammers using those attacks in the intervening 4 years.

- 2. I wonder how big the messages have to be, in order to affect training
  in a relatively small number of messages.  Maybe limiting the number of
  tokens trained on per message, might help.


It might be worthwhile implementing the described "RONI" scheme
to avoid the less targeted form of the issue anyway.

--j.