“Biometrics are our most unique physical (and behavioral) features that can be practically sensed by devices and interpreted by computers so that they may be used as proxies of our physical selves in the digital realm. In this way we can bond digital data to our identity with permanency, consistency, and unambiguity, and retrieve that data using computers in a rapid and automated fashion.”
Much is made about the breadth of biometric modalities, and indeed research into new, exotic biometrics (ear, gait, odor, etc.) is compelling. But the modalities that are field-proven in large-scale deployments are fingerprint, face, iris, and voice. Behavioral biometrics such as keystroke dynamics are also gaining in performance and adoption. These happen to be the biometric modalities that, today, best meet our tests for uniqueness, permanence, and consistency while also being conducive to capture using sensing devices in an ergonomically and economically practical way. Proprietary techniques that have also been deployed include vascular (palm, finger vein), and hand geometry.
Biometrics are not deterministic but rather probabilistic in nature, so it follows that:
- the more data we have in a biometric sample (or set of samples), the more likely that it is unique,
- there is always some likelihood that two different individuals will generate very similar or equivalent biometric samples, and
- there is always some likelihood of false match or false non-match (Type I or Type II error) results from a biometric comparison.
Some biometric modalities are less permanent over time than others, and some are more difficult to present and capture consistently. Some are more prone to environmental conditions that lead to a low signal-to-noise in the sample, which in turn leads to lower performance.
There is no perfect biometric modality; each has advantages and disadvantages for a given use case. For example, perhaps the most differentiating feature of fingerprints as a modality is that they leave behind evidence at a crime scene as “latents” (e.g. fingerprints on a glass). Irises are perhaps the most consistent, information-dense, “barcode-like” of the modalities. Facial images stand out because they are the biometric modality that humans excel at comparing, and so we can integrate complementary human- and machine-based recognition. Additionally, facial images are abundant in the digital realm, and also can be collected covertly from a distance. Voice is notable for being behavioral as well as physical, and thus the samples available from a given individual are abundant.
Even when our biometric samples are unique, permanent, consistent, and physically bonded to us, the sensors and algorithms we have devised to acquire and analyze them are imperfect. Sensors introduce optical and electrical distortion. Information is lost as sample data is converted from analog to digital form, and then again when the digital signal is compressed. Sampling rates (spatial resolution in the digital domain) significantly impact the quality of biometric samples. The algorithms designed to extract computer-matchable “templates” from a sample vary dramatically in precision and performance, as do algorithms and systems used by computers to rapidly assess their similarity. Machines are good at very fast, reasonably accurate, automated signal processing and template comparison, but they lack a human’s ability to visually perceive, analyze and characterize the similarity of two samples. Nevertheless, our physical selves provide many features that are well-suited for biometric comparison and search, and advances in modern sensing and computing technologies continue to improve the ability of a machine to perform biometric identification extremely quickly and accurately.
1 “Proprietary” here means that the capture and matching software and capture hardware peripheral are inextricably interdependent.