23 Comments

Summary:

Sift Science says it can help you finger people on your website who are likely to create fraudulent accounts, post fake reviews or do other dastardly deeds. The startup’s service, now in private beta, uses machine learning to help ID the bad guys.

The problem of online fraud, fake reviews and sock puppetry is only going to get worse, according to recent research. But there are ways to identify likely perpetrators and that’s what Sift Science aims to do.

Sift Science team.

The 8-person San Francisco startup uses machine learning to analyze user interaction with web sites and create a digital profile of who will likely perpetrate online fraud, said company co-founder Brandon Ballinger, an ex-Google software engineer.

Companies can use the service — built on Hadoop, HBase, Avro and MongoDB —  by adding some Javascript code to their sites and then using JSON (JavaScript Object Notation) APIs “to track transactions, bans, chargebacks, or custom event types,” according to the company.

Here are some early findings based on the private beta of the service:

1: Fraudsters tend to be nightowls. Most fake accounts are created late at night local time: 3:00 a.m is apparently the witching hour.

2: Bad guys stick with old technology.  People using Chrome on Windows XP are four times more likely to create a fake ID than the average user. (Firefox users are 50 percent more likely than average to create a faux account.)

On the other side of the same coin:

3: Fakers don’t update.  An account created on Safari running on Mac OS X is about 30 percent less likely to be fake. Those running IE9 on Windows 7 are 33 percent less likely than average to be fake.

4: Yahoo email is big. Folks with Yahoo.com email accounts are five times more likely to create a fake account than someone using Gmail.com or Comcast.com addresses.

5: Geography is key. Most traffic coming from Nigeria is fraudulent but also goes through a proxy to disguise its point of origin.

“Based on user actions, we build a model of what a normal user would do on a site versus what a fradulent user would do. We look at the time of account creation, the sequence of pages viewed. If they’re browsing around, they’re probably normal. If they set up an account and jump straight to a transaction, probably not,” Ballinger told me by phone. But then again, they’re tricky. Sift Science found that someone who opens an account, then waits an hour before transacting is 7 times more likely to be fraudulent than the average user.

The proess is similar to Google Analytics in that Sift Science creates a history of user events and comes up with a score for each user that rates the likelihood that he or she is involved in fraud, he said.

Sift Science is heavy on former Googlers:  6 employees are ex-Google engineers. Jason Tan was former CTO of BuzzLabs and Fred Sadaghiani was CTO of Teachstreet.

  1. not to burst your buble, but are you seriously have no clue or just for the startup badge?

    Share
    1. Are you using Chrome on Windows XP?

      Share
  2. This is a great example of a good (big) data product. It’s tracking very specific activity e.g. conversions or signups and the characteristics which call out those which later turn out to be false/fraudulent in order to do some advance screening. This is a very close connection between the start and end of the flow which makes it easier (not not easy) than the much more generic “big data” claims that are trendy right now.

    Share
  3. All well and good. I the bottom line value in impacting efficiencies – but perhaps quite marginally. It will be a commercial success though for sure because of the strength of the team.
    All you have to do these days is plugin some ML to achieve the insights of accepted wisdoms and you’ve got a hight tech start up. I expect similar results could be achieved without ML – by smart application of simple heuristics. After all the problem space here is not complex.
    Still, this is probably a good entry point to the market and will open up other opportunities as they go. Industry begets opportunity.
    Its at least refreshing to hear about a start up thats a bit above the TechCrunch dross. Good luck to them.

    Share
  4. what a fake “technology”

    Share
  5. “Most traffic coming from Nigeria is fraudulent”

    Can you or Sift Science prove that patently absurd and frankly insulting claim?

    Share
    1. I think what they mean is that most traffic that “appears” to come from Nigeria is fraudulent…
      Yet the only ever royal family member to contact me was from Nigeria! So… You can’t really blame their statement..

      Share
    2. Dominic Amann Friday, October 5, 2012

      Hard any one individual to prove, but I can state that 100% if the traffic in my inbox from Nigeria is fraudulent. Would you like me to forward you the contents of my junk box?

      Share
  6. What total prove did you have to justify that Most Fake Traffic comes from Nigeria.

    Fake traffics can come from anywhere in the globe, due to proxy functionality. Erase that perception you are trying to inject.

    Share
  7. This reminds me of nefrology (I believe it was called) in which cranum measurements were taken as predictors of criminal nature. In the end all you will have are inter-historically (because things change) and internally contradicting statistics that prove both that chrome users are the best and the worst computer-frauds in the world.

    Share
    1. That would be “Phrenology”. Some of Terry Pratchett’s books also featured the opposite, retrophrenology, where the practitioner would change the shape of the patient’s cranium in order to influence whether their criminal tendencies (or willingness to use Chrome or XP). Each retrophrenologist came equipped with a big mallet… I suppose the lart would be a similar technology.

      Share
  8. “Most traffic coming from Nigeria is fraudulent”

    Do you have evidence of your claim? I think is pure predudice.

    Share
  9. Hi, speaking for my company, Sock Puppets R Us, we have taken a keen interest in this post. Your article is a very good start, but would you please provide another more detailed list all the steps we should be taking to appear more real?

    Share
    1. Was thinking the same thing socky. Seurity companies rarely give the bad guys a roadmap to the was they get screened out

      Share
  10. This is nothing but bs, fraudsters will adapt quickly. Simply setting my user agent to Safari running on Mac, clicking a few pages and not logging in from Nigeria will defeat your screening procedure? How long do you think it takes for them to catch on?

    Share

Comments have been disabled for this post