spam classification breaker

[Robin Becker]

This article at the BBC reports on what appears to be a genetic
algorithm or random search method for finding words that apparently fool
bayesian classifiers every time.

http://news.bbc.co.uk/1/hi/technology/3458457.stm

The author apparently had to include html reporting into the emails to
allow his mail client to report back automatically.

Of course if he'd used python the whole process of email generation and
classification could have been done in a single process and would
probably allow easier generation of the magic words.

Why Berkshire, Marriot etc should be allowed through is pretty strange

 

[Skip Montanaro]

I noticed immediately that the author of the article used the term "ham" to
refer to mail which was not spam. Even if SpamBayes dies an ignominious
death in the future at the hands of some ruthless spammers, that will be our
lasting legacy.

Mr. Graham-Cumming could have avoided the overhead of sending himself 10,000
mails by simply selecting words from his archived public presence on the
net: web pages, Usenet posts or archived mailing list posts associated with
his email address. I suspect his genetic algorithm would have been all but
unnecessary. (Google for "John Graham-Cumming" for example.)

This doesn't have to be a tedious process either. In the course of normal
scumbag email harvesting, all the crawler has to do is select a few
non-trivial words from the harvested page and associate them with the email
address(es) on that page. After seeing the same email address a few times
they would have a decent collection of hammy words for use in the "random
words" block of later spam.

Also, unlike the statement the author made:

And, he said, this would have to be repeated for every person a spammer
wanted to reach because they would all have a different list of key
words.

this wouldn't have to be done for all email addresses. Anything which
increases the likelihood that a spam is opened will be seen as an
improvement for the spammer. There's obviously no need for them to get a
100% open rate on spam. If that was the case, they'd already all be out of
business.

These research types. They always do things in the hardest way possible...

Skip

 

[John Graham-Cumming]

Mr. Graham-Cumming could have avoided the overhead of sending himself 10,000
mails by simply selecting words from his archived public presence on the
net: web pages, Usenet posts or archived mailing list posts associated with
his email address. I suspect his genetic algorithm would have been all but
unnecessary. (Google for "John Graham-Cumming" for example.)

This doesn't have to be a tedious process either. In the course of normal
scumbag email harvesting, all the crawler has to do is select a few
non-trivial words from the harvested page and associate them with the email
address(es) on that page. After seeing the same email address a few times
they would have a decent collection of hammy words for use in the "random
words" block of later spam.

Yes, and I've tested this and its possible to find hammy words this
way too, although it wasn't as effective as the technique I pointed
out, nevertheless it is practical and in my experiments I looked at
the uncommon words found in the locus of my email address and around
40% were pure ham!

Another way would be to spider the web page associated with the domain
in the email address. e.g. to attack my address spider www.jgc.org.

All of this indicates that it should be possible to attack Bayesian
filters with a variety of techniques that rely on the fact that they
are naive (i.e. they'll accept a hammy word no matter where it
appears).

John.

 

[Robin Becker]

.....

tomatically.

If I'm a spammer trying to get my pitches seen by you, and you're using a
personal Bayesian classifier, then I need to load my pitches with words that
are very hammy to you. If I don't have access to your personal training
data (if I do, I already own your machine ...), then I need to *deduce*
what's hammy to you. One way to do that is, as John Graham-Cumming noted
here, is for me to send you thousands of messages with different piles of
words, and note which ones did and didn't get caught by your filter. Then
I load my sales pitches with words from the ones that your filter didn't
reject, and avoid words from ones your filter did reject. In order to do
that, I have to know which messages you did and didn't look at. That's the
purpose of the HTML "web bug"/"web beacon"s in the thousands of test
messages. (If your email client renders HTML pages, including fetching
images off the net, a spammer can know when you've rendered their message,
by, e.g., embedding your email address as a parameter in a URL that fetches
a .jpg to display.)

..... are you asserting that spammers don't have access to the pdf that
users are filtering? Each filter may be unique, but they can be biassed.

 

[Robin Becker]

[Robin Becker]

.... are you asserting that spammers don't have access to the pdf that
users are filtering?

Sorry, I couldn't make sense of that question.

Each filter may be unique, but they can be biassed. --

.....OK I guess I'm trying to get at the following hand waving argument.
Since most people agree about what is ham or spam there must be a
general recognizer for each. My question is then, is whether it's
possible to define a camouflage mechanism that turns ham into spam or
vice versa. Most people reading a newspaper article would classify it as
spam. If I insert a short ad v ert into the middle the quick
scan process is gone, but I might be able if everything is
set up correctly to get a forbidden word
set into the text in plain si g ht even
though it's specifically fo r bidden by your
all singing and dancing B a yesian analyser. It is well known
that word/space runs are very distracting which is why printers
have long tried to eliminate them.

I don't believe a small cost will kill all spam; every day I get large
amounts of paper adverts, flyers, business cards etc etc. These have
real cost, but presumably are sufficiently market oriented that they pay
for themselves. Putting a cost on email will just reduce the volume of
spam.