Slashdot Mirror
Machine Learning Susses Out Social-Network Fraud
CowboyRobot writes "Machine learning techniques can be used to detect fraud and spies on social networks based on certain features, such as the number of followers and the number of devices used to access the network. Certain characteristics of social-network accounts have a high correlation with fraud and can be used to differentiate between real and fake accounts, a researcher presenting at the SOURCE Boston Conference said this week. Using machine learning techniques, Vicente Diaz, a senior security analyst with security software firm Kaspersky Lab, found that seven characteristics of Twitter profiles could identify fraudulent accounts 91% of the time. The number of devices from which a user accesses the service, the ratio of followers to people following an account, the average number of tweets to each person, and the number of tweets to an unknown receiver are all features that correlate strongly to fraudulent accounts, he says."
In related news, social network machine learning fraud bots get algorithm update based on current fraud detection algorithms.
It's not necessarily friends directly posting crap on your page. A lot of fraud/spam on Facebook comes from these pages set up specifically to attract followers so the page can be sold for huge advertising bucks. They'll post exploitative pictures of injured animals, maimed soldiers, etc. with captions like "1 SHARE = 1 RESPECT". No matter how often I've warned my friends against forwarding this stuff, they'll do exactly what they are told because they don't want to be accused of not caring about puppies or war heroes or orphans or Jesus or whatever.
The end result is, no matter how hard I try to avoid it and how careful I am to restrict my account only to friends and colleagues I personally know, I still get spam from these phony accounts plastered all over my news feed.
"So I would be a fraud if I had a facebook account."
Precisely. There are several things wrong with trying to actually use this in the real world.
(1) 91% is not nearly good enough. Period.
(2) Even if it were 99.9% accurate, it would still not be good enough. Because it runs into the base rate fallacy.
(3) Similar but not related to the base rate fallacy, is that a statistical correlation between datasets of millions says nothing about an individual account.