The Software That Stares at Goats? [Pentagon Data Mining]

The Software That Stares at Goats? Pentagon Building System For Massive E-Mail Privacy Violation

... Earlier this month the Pentagon announced a new effort to build a system aimed at allowing it to scan billions of communications in order to detect "anomalies" in people's behavior that will predict who is about to snap and turn into a homicidal maniac - or, perhaps, leak damaging documents to a reporter.

[See DARPA-BAA-11-04: Anamoly Detection at Multiple Scales (ADAMS) ]

Citing the case of Maj. Nidal Hasan, the Army psychiatrist charged with killing 13 people in Fort Hood, Texas, the Pentagon's Defense Advanced Research Projects Agency (DARPA) wants to try to identify, before they happen, "malevolent actions" by insiders within the military.

The new project is called ADAMS, for Anomaly Detection at Multiple Scales, and anyone who remembers the battles over the Bush Administration's "Total Information Awareness" (TIA) program may be experiencing a major flashback right about now. TIA, also a DARPA project, was based on a vision of pulling together as much information as possible about as many people as possible into an "ultra-large-scale" database, making that information available to government officials, and sorting through it to try to identify terrorists. Eventually shut down by Congress, it was probably the closest thing to a truly comprehensive monitor-everyone "Big Brother" program that has ever been seriously contemplated in the United States. And many of the problems with TIA are equally present with this ADAMS project.

For one thing, the idea is naive and misguided and it won't work. Statistical data mining has been found to be of limited use in some areas, such as in detecting credit card fraud. But as experts have said, data mining is not good at predicting highly unusual events, because it does not have a large body of examples it can use as a basis for identifying patterns. In fact, there are no patterns with some things. As my colleague Mike German often points out - and he used to work undercover on anti-terrorism cases for the FBI - empirical studies show that there is no such thing as a reliable profile that will predict violent behavior. Incidents in which people turn into homicidal maniacs and begin shooting up their offices are extremely rare and each one has unique origins in the individual psychology, circumstances and life history of the perpetrator. It would probably take a treatment of the length and sensitivity of War and Peace to truly get at the factors that cause an individual to turn into a Maj. Hasan or any other workplace shooter.

Read more.

Comment

GregoryPS
I think the author of this article is making a wrong assumption that the goal of data mining tools is to fully automatically identify the potential threats. In this case, they are very unlikely to work.

However, it is much more realistic to use data mining tools to focus on likely threats, which will still be subject to human review. This is how data mining successfully works in CRM. For example, DM tools do not identify customers that will certainly churn, but only a subset of customers more likely to churn.

So data mining can certainly be useful to identify a subset of threats worth analyzing further. In addition, social network analysis methods can also be used, and not only statistically-based patterns