· System sensitivity Many systems, such as the fingerprint-recognition devices we tested for this issue, may be tuned to do less strict checking at the expense of opening the system. Administrators have to balance false acceptances versus false rejects, the possibility of fraud versus user convenience.

· Calibration Early fingerprint-recognition systems required careful calibration by a trained expert. And some systems today still require periodic adjustment to assure correct reading. Likewise, voice-recognition systems can require considerable user training, especially when used at more secure sensitivity settings.

· User population Some populations have difficulty using biometric devices. People with light ridge definition in their fingers may have difficulty using fingerprint-recognition systems. Those who work with abrasive substances--construction workers or even people who handle large volumes of paper--can have their ridges worn down. There also are substantial physical differences based on age, gender and ethnicity.

· Environment Users with excessively dry, wet or dirty hands have experienced problems with finger- and palm-recognition systems. People wearing gloves generally can't use these systems; however, the ultrasound-based systems have had limited success detecting prints through thin latex gloves. Face recognition may not be practical in industrial and medical environments where hoods or masks are required.

· Comparison engine location If verification is not performed as part of a network that stores and retrieves biometric data, it will be necessary to enroll each new user at every potential location. In a Windows NT environment, for example, client-only verification is much more limited than domain-based verification, which gives users access at any properly equipped workstation.

·Security of data channels The security of the connectivity between the device and host, as well as the host and any back-end verification engine, is crucial to avoid wire snooping and playback attacks. Many devices that use standard cameras, microphones and other equipment should encrypt or sign packets on the wire. In a Windows NT domain environment, communications between client and server PCs also should be encrypted. Just because it replaces the Windows NT GINA (Graphical Identification and Authentication) logon system doesn't mean that authentication data is protected on its way back to the Primary Domain Controller.

· Host computer power Most low-cost devices rely on the processing power of the host and often require a Pentium processor and at least 32 MB of RAM.

· Storage The biometric template--the digital representation of what the reader detects--should be encrypted where it's stored, and protected storage locations such as smartcards can improve overall security. The size of the template may be a factor. Most fingerprint and iris templates require between 256 bytes and 1 KB per user, though some systems need up to 8 KB. Face-recognition systems can require up to 3.5 KB per user--too large for some smartcards. Architectures based on a separate database engine that runs in tandem to the operating system may require additional host resource and administration.

· Export restrictions If you plan to send these systems overseas, then you must consider export restrictions that fall under the U.S. State Department's national security and criminal control statutes.

Choosing a Strategy Biometrics fits into two broad categories: the physiological, which uses a physical characteristic, and the behavioral, which matches some consistently performed action. Products based on physiological measurements are by far the most common--with finger, palm, hand geometry, face, iris and retina recognition being the most mature technologies. Researchers have proposed many other approaches, including ear geometry, vein structure and even body odor. Within the behavioral camp, voice and handwriting recognition are the most prominent, with acoustic handwriting recognition--literally listening to the way you sign documents--under development.

Iris scanning is the most secure of the "productized" authentication approaches available. These products are too expensive for general use--prices run in the four- to five-digit range per device--but that's about to change. IriScan, the major patent holder for iris-recognition systems, is preparing products that will drop the per-device cost by a factor of 10 or more, dipping down into the $500 range.