One of your users isn't interacting with the system the way he or she usually does. It could be that the user is trying to do something new.

Or it could be that the person working in your user's account isn't your user.

The ability to detect fraud is increasingly becoming a holy grail for enterprises, which not only want to protect their users from fraud and theft but also to be able to tell prospective customers that their systems are more secure than their rivals'. There's just one problem: Accurately detecting fraud in user accounts is really, really hard.

"We recently did a test at an [enterprise] customer site where the customer gave us three months of log data and asked us to find the fraud," says Tom Miltonberger, founder and CEO of Guardian Analytics, a startup that specializes in online fraud detection. "The company we were competing with came up with 500 alerts -- and the user didn't end up finding any real fraud among them."

Such results are typical in today's behavioral analysis and anomaly detection systems, which are designed to help identify potential fraud but often generate far more false positives than useful data, Miltonberger observes.

Enter Guardian Analytics, a startup that offers a technology called "fraud modeling," a combination of user behavioral analysis and artificial intelligence-type reasoning. Guardian Analytics essentially allows companies to create sophisticated patterns of "normal" online interaction with a particular system or account, making it easier to identify anomalous -- and potentially fraudulent -- behavior.

"The thing about today's fraud is that there are so many different ways to hide it," Miltonberger says. "There are proxies, obfuscation, and so many other tools that fraudsters use. And when they penetrate an account, the things that fraudsters do don't look radically different than what a typical user would do -- they're just transferring funds, collecting account data, and so forth.

"As a result, most systems today have trouble detecting fraud just by analyzing usage patterns. Either they miss [the fraud] entirely, because the differences are so subtle, or they turn in so many false alarms that you can't really make any sense of them."

Guardian Analytics, on the other hand, does dynamic account modeling, creating models of each user's behavior in a particular account. Using this model, Guardian Analytics can not only tell what the user has done in the past, but it can also predict what he or she will do in the future -- and recognize activity that falls outside these established norms.

The Guardian Analytics approach is much more accurate than current its counterparts, Miltonberger says. "In the test we did at the customer site, where our competitor came up with 500 alerts, we only came up with five," he states. "Three of the five ended up being fraud cases that the user knew about when they gave us the data."

Fraud modeling will become increasingly useful as attackers adopt hybrid approaches to account theft, Miltonberger says. "We see a lot of that now, where the fraudster uses online methods to scope out the user's account information, but then uses offline approaches, such as wire transfer, to actually do the theft," he says. "When they limit their online activity, you have fewer chances to detect them, so accurate analysis becomes even more important."

For now, Guardian Analytics is focusing most of its attention on applying the fraud modeling technology to banks and other financial institutions, where account fraud is most critical and potentially harmful. Over time, however, this model could be applied to many other environments, such as e-commerce, SAS, sales force automation, and even social networking services.

"One of the beauties of this technology is that it will automatically adapt to recognize the user's behavior, whatever that behavior is," Miltonberger says. "It will recognize the patterns and alert you if there's activity outside those patterns."

The Guardian Analytics technology, which is sold as software that sits alongside the server, is sold on an annual subscription basis that includes frequent updates, much like an antivirus application. It isn't cheap -- prices range from about $70,000 per year for a small bank to more than $1 million for a large institution.

"When you compare that to the cost of other authentication technologies, like tokens, though, this is a pretty attractive alternative," Miltonberger says. "The cost of deploying tokens to your users would end up being much higher -- and a lot more of a hassle in time and administration."

Have a comment on this story? Please click "Discuss" below. If you'd like to contact Dark Reading's editors directly, send us a message.