Slashdot Mirror
Turing Tests to Stop Spam
cexy writes "The Register has a story about how Hotmail and Yahoo! are using Carnegie Mellon developed captcha technology (completely automated public Turing tests to tell computers and humans apart) to stop spammers from automating signups for accounts from which they can send spam. These guys are using captcha too, but to stop incoming spam."
I would rather Yahoo stop spam from getting to my mail acocunt before they concentrate on stopping people from signing up automatically. I'm one of the few people who actually pay for Yahoo "additional" services. I thought I would get better anti-spam support. Not so far. I literally have 10 to 20 an hour and I can't block anymore because Yahoo only allows 100 addressed to be blocked. And considering the smammers are using 12374614187641874@optinmail.com along with other numerous addresses, it's impossible to block the majority of them. Hell I would even be happy if they would start allowing people to block entire domains. That would be a good first step.
My sig of choice is Marlboro
These Turing tests do not stop spam. They discourage spammers from using bogus Hotmail etc accounts to originate spam from. They do this by making it incrementally more expensive to create the accounts; rather than using a bot to create an account a second you have to use a human to create accounts by the minute. So 60 times the effort.
But I don't think that translates into 60 times the cost. The Turing tests are interesting but I don't think that the creation of the accounts ever was a bottleneck in the process in sending spam. You could get a high school kid to create all the accounts you would need for a month in about an hour, and pay him in pr0n.
If the truth were known, Hotmail and Yahoo are just trying to decrease server loads. I bet that when bots create accounts they create hundreds or thousands more than are used, which take up server resources during creation and later as the accounts eat up storage. With Turing tests it is more likely that not too many will be laying around waiting to be used.
=^..^= all your rodent are belong to us
It's no secret, at least it shouldn't be, that Micro$oft is making money selling your hotmail address (yet then they spam you with advertisements for their spam-blocking software)...
Instead of just experimenting by setting up a Hotmail account, has anybody ever tried the other way around? That is, pose as an advertiser and approach Hotmail about e-mailing their users?
Bayesian techniques depend on predicting which elements (usually, which words) are likely to indicate spam, and which are likely to indicate non-spam messages. This can vary highly from user to user, and so it should be done on a per-user basis.
For instance, I am a security administrator and receive a lot of legitimate mail about "antivirus software", and very little legitimate mail about "teenage lesbians." However, my girlfriend's crush, who is an activist lesbian, may well receive a lot of legitimate mail about "teenage lesbians" and only spam about "antivirus software." If we are on the same ISP, then it would be erroneous behavior for my reporting "teenage lesbians" as spam and "antivirus software" as nonspam to throw her spam-filtering out of whack, or vice versa. And yet it is a potential privacy violation for the ISP to be gathering statistics on which one of us gets virus bulletins, and which one is the lesbian.
(Moreover, there also isn't yet any standard mechanism for users to report spamminess or nonspamminess back to normal IMAP or POP mail hosts -- and Bayesian algorithms require sampling both spam and non-spam mail, not just spam reported to an abuse address.)
The filtering mechanisms that should be implemented on the server are general ones -- ones that do not rely on deep inspection into the content of the message. I don't really want ISPs to gather stats on common keywords in users' incoming mail -- do you? It is one thing to examine structural elements of the message, such as the IP address which sent it, or the presence of normal headers; or to statelessly scan the message for static patterns, such as virus signatures or "DISCOUNT HERBAL VIAGRA !!!" It would be quite another thing to gather the kind of data that Bayesian filters involve, for every user on a large end-user system.
From my understanding, the use of image recognition in the captcha test would make it nearly impossible for blind people to pass the test.