With a little practice, almost anyone can learn to imitate a signature. But only the most highly skilled forger can rip it off just the right way, with the same variations in speed, the same order of crossing the t's and dotting the i's.

With that in mind, Cornell University engineers have developed a signature-verification system that can reject most forgeries while still allowing for the minor variations in a person's real signature. The system runs on ordinary desktop computers and could be used for point-of-sale applications.

"We collect 49 features of a signature," explained Toby Berger, the J. Preston Levis Professor of Electrical Engineering at Cornell, who developed the system along with graduate students Luan Ling Lee, Erez Aviczer and Yi-Jen Chiu. "Some are static features that you could measure from the end result, like the shape of letters, or the maximum distance between the highest and lowest points. Others are dynamic features that must be measured as the person is signing, like the maximum forward velocity and where and when in the signature it occurs."

Many signature-verification systems depend on having a person sign at the same speed every time, Berger noted, but this one allows for speed variations. "We noticed in the lab that sometimes people signed fast and sometimes slow," Berger said. "It might depend on whether or not you were wearing your overcoat."

The system requires that the person sign on a graphics pad, or on a piece of paper placed over the pad. The pad reports the position of the pen about once every five to 10 milliseconds.

"We reject a real signature about 1 in 200 to 1 in 500 times, depending on the conditions under which it is signed," Berger said, "while rejecting up to 93 percent of forgeries. That's what you want in a point-of-sale system: you want to accept genuine signatures almost all the time. If you fail to reject a phony signature sometimes, that's not so serious; without a system like this you have no ability to reject at all."

The rejection rate for false signatures falls to around 50 percent for a "timed forgery" done by a skilled forger. But this requires that the forger watch as you write your signature -- perhaps several times -- and then imitate the timing as well as the shape.

On the other hand, a perfect tracing won't fool the system at all. "It's easy to reject tracing," Berger said. "You lock up your hand and lose dynamics in your wrist." If this system became common, he added, people might no longer have to sign the backS of their credit cards, and forgers would have nothing to copy.

Berger said the system could also be set up to reject a higher percentage of false signatures at the expense of making it more likely to reject real ones, for high-security applications.

Over 10,000 signatures, including many deliberate forgeries, were collected in Berger's laboratory at Cornell from about 105 volunteers. It's believed to be the largest database of dynamic signatures collected anywhere so far. "Most people signed about 50 times," Berger recalled. "One person signed 1,000 times."

Some volunteers were asked to write a few signatures as fast as they could. The system has a mode in which the verification decision doesn't depend on the overall speed of writing. In this mode it measures the time at which some features occur relative to the overall time taken to sign.

The collected forgeries were of three types: "simple," where the forger knew only how to spell the signature, "statically skilled," where the forger was allowed to view finished genuine signatures and to practice writing copies of them, and "timed," where a statically skilled forger was coached until his signing duration matched that of the genuine signature.

The signature database was used to develop a set of "average features." "A lot of the features we used have been used over the years by people trying to recognize handwriting, such as the developers of the Apple Newton," Berger reported. "Others were ones we chose based on our experience in collecting signatures."

"By knowing what most people do with each of these features, we can figure out which features of someone's signature are particularly unique," he explained. "You could decide that your system is going to look at the same features for everyone, or you could use individualized sets of features. You're not going to store all the features for everyone at the point of sale."

The system, for which Cornell has applied for a patent, is not "computationally intensive" compared to other such systems, Berger said. A laboratory prototype of it running on a desktop computer with a 486 microchip processor returns results in less than one-and-a-half seconds. A commercial system with a custom-designed chip could run considerably faster, he said.

Berger teaches courses in communications, information theory, probability and communication networks in Cornell's School of Electrical Engineering. Lee is now a professor of electrical engineering at the State University of Campinas in Brazil. Aviczer studied at Cornell on leave from AT&T Bell Laboratories. Chiu is conducting doctoral research on video compression under Berger's direction.

The work was funded in part by the Brazilian Science Agency, the State University of Campinas and the National Science Foundation.