Identification by physical characteristics is as old as humanity. Recognizing people by their voices or appearance, and impersonating people by assuming their appearance, was widely known in classical times. Efforts to find physical characteristics that uniquely identify people include the Bertillion cranial maps, fingerprints, and DNA sampling. Using such a feature to identify people for a computer would ideally eliminate errors in authentication.

Biometrics is the automated measurement of biological or behavioral features that identify a person. When a user is given an account, the system administration takes a set of measurements that identify that user to an acceptable degree of error. Whenever the user accesses the system, the biometric authentication mechanism verifies the identity. This is considerably easier than identifying the user because no searching is required. A comparison to the known data for the claimed user's identity will either verify or reject the claim. Common characteristics are fingerprints, voice characteristics, eyes, facial features, and keystroke dynamics.

Fingerprints

Fingerprints can be scanned optically, but the cameras needed are bulky. A capacitative technique uses the differences in electrical charges of the whorls on the finger to detect those parts of the finger touching a chip and those raised. The data is converted into a graph in which ridges are represented by vertices and vertices corresponding to adjacent ridges are connected. Each vertex has a number approximating the length of the corresponding ridge. At this point, determining matches becomes a problem of graph matching. This problem is similar to the classical graph isomorphism problem, but because of imprecision in measurements, the graph generated from the fingerprint may have different numbers of edges and vertices. Thus, the matching algorithm is an approximation.

Voices

Authentication by voice, also called speaker verification or speaker recognition, involves recognition of a speaker's voice characteristics or verbal information verification. The former uses statistical techniques to test the hypothesis that the speaker's identity is as claimed. The system is first trained on fixed pass-phrases or phonemes that can be combined. To authenticate, either the speaker says the pass-phrase or repeats a word (or set of words) composed of the learned phonemes. Verbal information verification deals with the contents of utterances. The system asks a set of questions such as "What is your mother's maiden name?" and "In which city were you born?" It then checks that the answers spoken are the same as the answers recorded in its database. The key difference is that speaker verification techniques are speaker-dependent, but verbal information verification techniques are speaker-independent, relying only on the content of the answers.

Eyes

Authentication by eye characteristics uses the iris and the retina. Patterns within the iris are unique for each person. Hence, one verification approach is to compare the patterns statistically and ask whether the differences are random.. A second approach is to correlate the images using statistical tests to see if they match. Retinal scans rely on the uniqueness of the patterns made by blood vessels at the back of the eye. This requires a laser beaming onto the retina, which is highly intrusive. This method is typically used only in the most secure facilities.

Faces

Face recognition consists of several steps. First, the face is located. If the user places her face in a predetermined position (for example, by resting her chin on a support), the problem becomes somewhat easier. However, facial features such as hair and glasses may make the recognition harder. Techniques for doing this include the use of neural networks and templates. The resulting image is then compared with the relevant image in the database. The correlation is affected by the differences in the lighting between the current image and the reference image, by distortion, by "noise," and by the view of the face. The correlation mechanism must be "trained." Several different methods of correlation have been used, with varying degrees of success . An alternative approach is to focus on the facial features such as the distance between the nose and the chin, and the angle of the line drawn from one to the other.

Keystrokes

Keystroke dynamics requires a signature based on keystroke intervals, keystroke pressure, keystroke duration, and where the key is struck (on the edge or in the middle). This signature is believed to be unique in the same way that written signatures are unique. Keystroke recognition can be both static and dynamic. Static recognition is done once, at authentication time, and usually involves typing of a fixed or known string. Once authentication has been completed, an attacker can capture the connection (or take over the terminal) without detection. Dynamic recognition is done throughout the session, so the aforementioned attack is not feasible. However, the signature must be chosen so that variations within an individual's session do not cause the authentication to fail. For example, keystroke intervals may vary widely, and the dynamic recognition mechanism must take this into account. The statistics gathered from a user's typing are then run through statistical tests (which may discard some data as invalid, depending on the technique used) that account for acceptable variance in the data.

Combinations

Several researchers have combined some of the techniques decribed above to improve the accuracy of biometric authentication. Three scientists combined voice sounds and lip motion with the facial image. Scientists describe a "supervisor module" for melding voice and face recognition with a success rate of 99.5%. The results indicate that a higher degree of accuracy can be attained than when only a single characteristic is used.

Caution

Because biometrics measures characteristics of the individual, people are tempted to believe that attackers cannot pose as authorized users on systems that use biometrics. Two assumptions underlie this belief. The first is that the biometric device is accurate in the environment in which it is used. For example, if a fingerprint scanner is under observation, having it scan a mask of another person's finger would be detected. But if it is not under observation, such a trick might not be detected and the unauthorized user might gain access. The second assumption is that the transmission from the biometric device to the computer's analysis process is tamperproof. Otherwise, one could record a legitimate authentication and replay it later to gain access.