Math Says Better Passenger Screening Beats Better Profiling

If you do the math, improving the detection of plastic explosives and guns would likely be more effective at preventing and deterring terrorist attacks on planes than figuring out better methods for trying to pinpoint terrorists with watch lists and algorithms, according to a mathematical study to be published in December.

Susan Martonosi, a mathematics professor at Harvey Mudd college, and Arnold Barnett of MIT's Sloan School of Management devised a mathematical model for judging the effectiveness of various security measures at preventing and deterring a terrorist who wanted to sneak a weapon through security. While they lacked data on how well the watchlists, puffer machines and x-ray machines, Martonisi and Barnett developed a mathematical model which showed that improvements in the screening that all passengers go through was more likely to improve airline safety than better passenger-centric solutions, such as the long-delayed Secure Flight program.

The mathematical model offers some preliminary indications about improved checkpoint security. Improving the baseline screening applied to all passengers by a fixed amount might be more beneficial than improving either the terrorist detection probability or the marginal benefit of secondary screening by the same amount. Deterrence might appreciably magnify the benefits of even modest gains in the chance of thwarting a terrorist at the security checkpoint. Not knowing the key parameters, however, we can hardly offer these last thoughts as policy recommendations.

One of the parameters the study couldn't figure out: what expectation of success is necessary for would-be terrorists to act on their plot (in math terms, they call it r).

The deterrence threshold r is the hardest parameter to estimate, given that terrorists will hardly take part in public surveys. Moreover, the threshold presumably varies from group to group. A RAND report argued that fear of ignominious failure does weigh heavily on terrorist groups, so it is unlikely that r is close to zero (Davis and Jenkins 2002). At the same time, many of these groups show too much sophistication to believe that success is guaranteed, so they presumably do not insist on a r close to 100 percent. A distribution for r seems more palatable than a point estimate: Here we approximate r as uniformly distributed between 25 percent and 75 percent.

The full paper will be published in the December 2006 issue of Interfaces, which is published by Institute for Operations Research and the Management Sciences (INFORMS).

Photo: nerd herd