For machine learning aficionados, a surprising tale of how anti-fraud measures discovered a Chinese car fraud racket that no-one was looking for:
Had they trained AdWords into anti-car prejudice? Was the model simply broken?The patterns of fraud in the areas that the model was trained to look at (counterfeit goods and phishing) turned out to be very similar to fraud in other areas like car-theft-serving-innocuous-order. Scammers and thieves tend to operate very similar models in a wide range of marketplaces, and the machine learning model was able to generalise sufficiently to detect fraud in an area where no-one was really looking for it.
The answer turned out to be even stranger. They were real cars, but they weren't really for sale. Scammers were taking pictures of cars on the street, and when a hapless customer showed up a few days later offering money, they'd steal the car and hand it over.
Baker and his team weren’t looking for cars or car thieves. But the algorithm saw a pattern of quick buys from new accounts, tied together with larger and more subtle patterns, and deduced something was up.
Buried in the main story, but to my mind equally significant, is the reason why frauds are so prevalent in China:
According to Li, the larger problem is the Chinese financial system, which requires every bank-to-bank transaction to be routed through the central government’s banking authority. As a result, anti-fraud measures are usually slower than criminals. Stopping a payment could take as long as three days, by which time the money is usually unrecoverable.Turns out that centralisation of bank transactions really slows things down - heck, why is this surprising? The Chinese government has no real interest in speeding up its inspection of transactions; thus, scammers can rely (indeed, base their fraud model) on slow bank transactions. It's harder to do this in the West because banks have been competing to speed up transactions, giving the scammers a smaller window in which to conduct their frauds.