 |
 |
 |
 |
| |
|||||
|
|
|
Home > News & Events > MITRE Publications > The MITRE Digest > | |||||||||||||||||||
Biometric Systems: Finding a Face in the Crowd October 2003
The distinguishable characteristics of each person—from fingerprints to facial planes—are fueling a new wave in security technology. But much of "biometrics" is still in its infancy, and no system exists that can accurately identify suspected terrorists without their being aware of it. Biometrics systems are automated methods that use a physiological or a behavioral characteristic to either identify a person or authenticate a person's claimed identity. Retina, iris, and fingerprint scans are examples of biometrics that are used to authenticate a claimed identity. These systems are in use now in some airports, prisons, and hospitals to control access. "Face recognition is a popular concept because it's easy to capture in a passive mode," says Paul Lehner, chief engineer in MITRE's Information Technology Division. "Potentially, terrorists can be identified without them even knowing they are being observed, but this kind of system is still more potential than reality." MITRE is conducting Army-funded research to address the need for biometric systems engineering, working in conjunction with the Department of Defense Biometrics Management Office, U.S. Army, and the Intelligence Technology Innovation Center. Specifically, MITRE is developing improved methods for predicting the performance of biometric system designs. Face recognition works by comparing a photographic image (of subjects walking into a building, for example) with a database of stored images. The software programs in most of the systems available today use appearance-based classifiers. Or they attempt to measure some of the nodal points on your face, such as the distance between your eyes, the width of your nose, the distance from eye to mouth, or the length of your jaw line. Algorithms to Templates Algorithms turn the extracted features and measurements into a template, which is then compared with a stored database of reference templates. The output from each template comparison process is a match score. Software applications must establish an appropriate threshold—a value where all scores above are determined to be a positive match, or not a match. Threshold values are sometimes set by default or based on the vendor's recommendation. However, it's always preferable to calibrate the threshold according to the data and according to a trade-off between the impact of the two types of errors—false matches and false non-matches. Tampa, Florida, has experimented with a biometric system on several occasions. In January 2001, the city used it at the Super Bowl by scanning the faces of people coming into the stadium and comparing them with images in a database of digital mug shots. The software flagged some people, but all were false identifications. In Ybor City, a part of Tampa, the police tried using face recognition to catch sex offenders and runaways in the entertainment district. Again, too many false positives made the system unworkable. Such systems have also been tried in airports around the country—with similar results. The problem is that these systems are not yet as accurate in the field as they are in the lab, according to Nick Orlans. Orlans is a principal engineer in the Biometrics Group of MITRE's Applied Information Technology Group and co-author of a recent book on biometrics. "In evaluating the current systems for use, there wasn't any specific test data to suggest that they would or would not perform as advertised," says Orlans. "Vendors claim they have a hot technology, but it's hard to judge how well it works. There's a problem in transitioning test results to fielded systems. The current systems don't address a wide range of environments and integration issues." In our MITRE research project, managed jointly by Orlans and Lehner, we explore two different approaches to improved prediction of biometric system performance. The first uses synthetically generated faces to better understand the performance of face recognition systems. The second uses Bayesian networks to model and predict total system performance that takes into account multiple biometrics and a person's behavior. From Hollywood to MITRE MITRE's synthetic face generation experiments use software from Singular Inversions, Inc., called FaceGen, which is normally used in the entertainment industry to do facial animation. The FaceGen software, combined with a rendering environment, generates alternate images of faces that vary by lighting, camera angle, facial expression, age, backgrounds, and even something as simple as the subject wearing a hat or sunglasses. "If the biometric security system can't respond to these variables," says Orlans, "its accuracy decreases and too many false identifications occur while real suspects pass through undetected." One of the more challenging variables to isolate is the way people age. If a person's photo in the system's database was taken 10 years ago, is it possible to identify the person today? To explore this question, the research team used the age-morphing features of FaceGen to generate a series of synthetically aged images for a set of 50 faces. "We use synthetic data because it allows us to isolate and finely control specific test parameters," says Orlans. "In the case of aging, we can rapidly generate a series of aged images or we can isolate other parameters, say facial hair or expression, in a very controlled way."
A robust face recognition biometrics system must also be able to account for a variety of poses, lighting conditions, and facial positions. A pose experiment was also conducted with synthetic data to isolate and measure the effect of camera angle in one-degree increments. "The experiment demonstrated that if the face was more than 5 degrees off from a full front image, the face recognition system was likely to fail," Orlans says. "Capabilities of commercial products have improved, but there are still significant engineering challenges." How to Catch a Soda Thief The second part of MITRE's research involves the prediction of total system performance. "A single biometric device can often be defeated simply by avoiding that device," says Lehner, "but if the total security system is designed properly, we should be able to identify people who are trying to avoid being identified." An entry security system includes not just the biometric device, but also guards, baggage checks, ID checks, and the standard operating procedures for using these elements. The purpose of the Bayesian network modeling is to represent all of these elements in a single model that predicts the chance that a suspect will defeat the system, as well as false identifications. Bayesian network models are a new class of probability models that support combining known probability information (e.g., performance of a biometric sensor) with expert judgment estimates (e.g., how a suspect might try to avoid detection).
To test the performance prediction model, MITRE set up the Biometrics Game Room. People are encouraged to steal a can of soda from a refrigerator in the Biometrics Game Room without being detected or identified. The model predicts how likely thieves are to succeed. The purpose of this modeling approach is to provide an engineering tool that is useful for designing and evaluating systems. The point of the game is to test the accuracy of performance predictions. "For example, we'll predict that the system is flawed and that 7 out of 10 people will beat it. If 7 out of 10 people actually manage to steal the soda without being detected, we know we have a useful predictive model," says Lehner. "It's the kind of model you can apply to the real world and predict how well a proposed system design will work." The research team is learning some interesting things from the Biometrics Game experiment. Camera location is important, for instance. "We were impressed by how clever people are in defeating a system," says Lehner. "It makes it an interesting challenge to model, but it represents the real world. In our first test, we had eight people go through undetected. Each of them had a different way of beating the system, but they all succeeded because the location of the cameras was obvious and they acted accordingly. So the model now considers both the coverage of the cameras and how obvious the camera locations are. We're making progress on our models as we go through different scenarios." The MITRE team hopes its modeling technique will provide an effective, more structured basis for risk management on a large biometrics system. The team is expanding MITRE's knowledge of the field and collaborating with others to expand knowledge further. For example, the MITRE team has authored two white papers on synthetic-generated data and seeks collaboration with other researchers to further validate the use of synthetic images. The ultimate goal is to use the information to better understand how systems will perform and to take some specific steps toward making them more robust and more predictable. This will help our sponsors choose the most effective systems to meet their needs in the future. —by David Van Cleave Related Information Articles and News |
||||||||||||||||||||
| Page last updated: February 16, 2004 | Top of page |
Solutions That Make a Difference.® |
|
|
