logo logo

Basic Terminology

Biometric Capture and Extraction process

Biometric sample

For biometrics we need a method of capturing and storing information for later comparison. Biometric sample is a general term for an image of a biometric trait = an information obtained from a biometric capture device, either directly or after processing.

Most common way for storing physiological biometric data is to capture an image of given biometric. This can be a fingerprint ink print or a face photo. In the past, when automatic matching was not an option, these images were used by people for manual comparison. Still today, photographs on ID cards are used to verify an identity of a persons at border control. We use the term biometric sample as a general term for an image of a biometric trait.

Biometric template

Biometric samples cannot usually be used directly for automated comparison. To enable automated comparison, we need to create so called biometric template. The SmartFace allows to automatize such a workflow.

Biometric template, or simply a template, is a set of stored biometric features. To create a template, we usually need to convert a biometric sample into a binary representation. The template is usually small in terms of computer memory use. This allows for quick processing, which is a hallmark of automated biometric authentication. The template provides an expressive representation of the biometric trait.

The process of obtaining all the necessary information and storing a biometric template is also called an enrollment.

Biometric capture

The biometric is presented to the sensor by the person requesting access or a person we want to identify. A camera may capture a face or iris, a sensor may capture a fingerprint, a microphone may capture a voice; in each case, the raw biometric information is acquired. This process is called a biometric capture. The sample is usually captured in a form of a 2D image. In the SmartFace we call this process detection.

Capture process is impacted by environmental condition. For example, capture of the face or iris image may be adversely affected by lighting and background conditions.

Biometric extraction

After capturing a biometric sample, we can proceed with extracting a feature set from the acquired data using feature extraction module - a computer program extracts a pattern of distinguishing features from the digital image. This process is called biometric extraction. In the SmartFace we call this process extraction.

Biometric comparison

The enrollment by itself is useless unless we plan on using this data to resolve some use case or to assist with already existing enrollment process. We need to perform a feature comparison to determine the degree of similarity between 2 or more templates. In the SmartFace Platform an example for such a biometric comparison is the face matching

For comparison, we need at least two templates. To maintain clarity in the process of comparison, we use different terms for the two templates that are being compared:

During comparison operation the matching module compares the feature set extracted at the time - probe template with the reference template using a matching algorithm.

A biometric comparison is an estimation, calculation or measurement of similarity or dissimilarity between 2 templates. Whether these data are coupled with personal information is irrelevant for the comparison process. The whole process is not about names or identity, but tries to decide, if there is a relation between two submitted sets of features. Comparison of probe template a previously stored reference template does not decide if the two templates are from the same person, it just produces a level of similarity.

Probe template

We use the term probe for a biometric sample, that is obtained at a time of comparison, or for a biometric template, that is being verified or identified.

As people change over time, probe samples and templates will most probably also differ over time. You can easily imagine the changes of face via the process of aging. The sample can also differ due to different environment such as lighting or due to behavior of the users, such as angle between the sensor and the camera. Even if you capture two samples of the same modality from the same person right after another, it is most probable that the samples, and therefore the templates also, will at least slightly differ. Probe sample or template is being verified or identified. In SmartFace we call such image Probe Image.

Reference template

We need a sample, or a template already stored in the past that we can use for comparison.

These sample and template are usually called reference sample and reference template. Reference templates can be stored on different platforms, such as on a passport or ID card presented by the person being verified, in a local or online database or just in device memory.

Remember, that higher quality of reference templates leads to more accurate biometric system. In SmartFace we call such image Reference Image.

Comparison score - Matching score

Level of similarity can be used to make final decision about the source of the templates. This similarity or distance is usually represented by a number. The number is called a comparison or matching score.

It indicates the degree of similarity between two biometric feature sets. The score is than passed to the decision subsystem. Usually a higher quality of data leads to higher scores for matches/mate pairs and lower scores for no-matches/non-mate pairs.

Threshold

We need the decision subsystem to perform a decision using some parameters. A threshold regulates the system decision. The system infers that pairs of biometric samples generating scores higher than or equal to the threshold are matches/mate pairs (that is, they are matching and they likely belong to the same person).

Consequently, pairs of biometric samples generating scores lower than threshold are marked no-matches/non-mate pairs that is, they are not mathing and thus they likely belong to different persons. Thresholds allow us to judge whether the sample features match the template based on whether the similarity score exceeds a threshold.

Verification (1:1) and Identification (1:N)

Depending on the application context, a biometric system may either operate in a verification mode or in an identification mode.

Verification (1:1)

In a biometric verification task, an enrolled user claims an identity and the system verifies the authenticity of the claim based on their biometric features. Verification requires the person being verified to lay claim to an identity, so that the system has a binary choice of either accepting or rejecting the person’s claim. The aim is to prevent multiple people from using the same identity or more specific to prevent unauthorized persons to use some else’s identity. Verification is also called one-to-one comparison.

Verification typically involves:

A verification decision can be based on of one or more attempts as dictated by the security policy.

Identification (1:N)

Outside of comparing one template to another to verify a user claim, we can also use a biometric system to determine an identity of a user. In the identification mode, the system recognizes an individual by searching the templates of some or all the users in the database for a match. The system conducts a one-to-many (1:N) comparison to establish an individual’s identity without the subject having to claim it. The 1:N comparison provides the results as a list of possible matching candidates. An identification system must perform many comparisons, therefore it requires more perfomance.

Identification typically involves:

  • biometric capture
  • biometric extraction
  • comparison against some or all templates in the enrolment database, producing a similarity score for each comparison;
  • judgement on whether each matched template is a potential candidate for the user, based on whether the similarity score exceeds a threshold and/or is among the highest number of scores returned

Identification and Verification errors

Both Identification and Verification will either accept or reject the claim. The decision outcome is erroneous if either a false claim is accepted - false accept - or a true claim is rejected - false reject.

If we receive a score above the threshold from a non-mate pair meaning, that the templates were obtained from different people, we encounter a false accept error. On the other hand, if we receive a score below the threshold from a mate pair meaning, that the templates were obtained from the same person, we encounter a false reject error.

False Positive vs False Negative Now that we know what types of errors can occur in a verification scenario we should be able to measure them.

FAR - False Accept Error

The False accept error scenario can occur in any system and we need to decide on how secure our system should be. For example, if we use a four-digit PIN number in our system, there is 1:10 000 chance that someone will guess the PIN at first attempt. We can increase the security by requiring more digits, adding other characters or limiting number of attempts.

In a biometric system security level is defined by a False Accept Rate or FAR, which represents the probability that the systems incorrectly accept an unauthorized person. The False Accept Rate of a biometric system can therefore be defined as the fraction of impostor scores exceeding the threshold. On an algorithm level, FAR is the probability that the system incorrectly declares a successful match between the input pattern and a non-matching pattern in the database.

A very low FAR may be the most important factor in a highly secure access control application, where the primary objective is to not let in any impostors.

FRR - False Reject Error

The False reject error when it comes to a verification system, lead to a convenience issue. While FAR was defined as a fraction of impostor scores above the threshold, measuring a convenience of a biometric system is represented by the False Reject Rate or FRR and may be defined as the fraction of genuine scores falling below the threshold. FRR is the frequency with which a genuine user is not correctly recognized and hence denied access. One of the main contributors to FRR is noisy data.

These two metrics are not independent. Therefore, when setting a threshold, we usually choose our required FAR and then measure system FRR on a specific dataset. After the measurement we can tune the threshold so that it suits our use case the best.

This tradeoff is important, as a system that accepts all the right people, and is convenient, might not reject all the wrong people, and be less secure. To set the desired balance of FAR and FRR, many systems have variable thresholds.

Enrollment Errors/Failures

While there are several types of errors that occur in biometric systems, there are two major classes of errors that relate to the system’s accuracy; comparison errors and decision errors. These error rates can tell us how well given biometric system performs compared to a different one.

Comparison errors are erroneous matches or nonmatches that could be considered machine malfunctions and are represented by comparison score being too low for mate templates and too high for non-mate templates.

Decision errors are impacted by a threshold and candidate count in case of identification and are usually a result of a comparison errors.

Failure To Detect (FTD)

Failure to find a biometric sample.

An enrollment process starts with capturing a biometric sample. There are situations in which the sample is not visible to the sensor. This can lead to Failure to Detect (FTD) error. In case of fingerprints this error occurs when a sample approaches the scanner, but the scanner fails to detect its presence. In the case of the face, the face might not be clearly visible on the camera, might be too small in the image or due to light and other conditions it was not detected. In general, this error occurs when we are unable to locate a biometric sample.

Failure to Capture (FTC)

Failure of the biometric capture process to produce a captured biometric sample of the biometric characteristic of interest, despite the attempt to do so.

For instance, in fingerprint scanning, failure to capture might occur if the sensor fails to obtain a clear or distinct image of the fingerprint due to various reasons such as dirt on the sensor, improper placement of the finger, or technical issues with the scanning device. Similarly, in facial recognition systems, failure to capture might occur if the lighting conditions are poor, there’s an obstruction on the face, or if the camera quality is inadequate to capture facial features accurately.

Such failures can lead to authentication errors, where the system either rejects a valid user or mistakenly accepts an unauthorized individual due to the inability to properly capture and match the biometric data. Manufacturers and developers aim to reduce failure to capture rates to ensure the reliability and effectiveness of biometric systems in various applications, such as security access control, identity verification, or attendance tracking.

Failure to Process (FTP)

Failure to produce template from biometric sample.

If we have already captured a biometric sample, we need to extract a biometric template. Outside of extracting template successfully, there are two other possibilities. After capture the biometric sample is sent to the feature extraction module for processing. If the captured sample is of poor quality, the feature extraction algorithm may fail to extract a usable feature set. Common causes could be diseases or conditions that deteriorate a person’s biometric trait. In some cases, the quality of the biometric template extracted from the biometric sample may be so poor that the user is asked to present the biometric data again. Regardless of the cause – if it is an extractor failure or QA decision - these errors are known as Failure to Process (FTP).

Failure to Acquire (FTA)

Failure to accept for subsequent comparison the output of a biometric capture process, a biometric sample of the biometric characteristic of interest.

Each of the enrollment errors mentioned earlier are important to consider for workflow of biometric enrollment. However, it is often not effective to measure all these errors separately, especially in a benchmark comparing two different biometric systems. The three types of errors (FTD, FTC, and FTP) mentioned in previous sections are often combined into one single measure called the Failure to Acquire. A high FTA rate will affect the throughput of the resulting biometric system and increase user frustration. One way to lower FTA is by lowering QA requirements of the capture and feature extraction modules, such as changing the quality requirements for a succesful enrollment - changing the minimum face size or face image quality. But this will put additional burden on the later modules - such as comparison. The FTA error rate is important measure for enrollment, verification and identification operations.

Failure to Enroll (FTE)

Failure to create and store a biometric enrollment data record for an eligible biometric capture subject, in accordance with a biometric enrollment policy.

As the result of an enrollment operation is a biometric template stored, so that it can be later accessed, overall success of enrollment is often measured only as success rate of template storage. The Failure to Enroll (FTE) rate denotes the proportion of users that cannot be successfully enrolled in a biometric system. Outside of failures mentioned, users, that do not possess given modality, like missing finger or an eye, are also calculated into the FTE. There is a tradeoff between the FTE rate and the error rates of the comparison. If the QA rules are disabled, then templates can be created from poor quality samples, but such noisy templates would result in higher comparison errors.