Biometric Authentication: How It Works and Where Banks Use It

Biometric authentication from the standards: enrollment, 1:1 comparison, thresholds and error rates, NIST’s limits, and what Malaysia’s RMiT expects of banks.
When a banking app asks an existing customer for a face or a fingerprint to log in or approve a payment, it runs biometric authentication. It compares a fresh sample with the reference stored at enrollment, and a set threshold decides whether the result counts as a match. A match then serves as one factor toward approving the login or payment.
Terms here come from the ISO/IEC biometrics vocabulary. For biometrics as a factor, this guide follows the US National Institute of Standards and Technology (NIST), whose guidelines cover US federal systems. The one regulator discussed is Malaysia’s central bank, Bank Negara Malaysia (BNM). The guide is written for bank fraud, security, and compliance teams and for product managers.
What is biometric authentication?
Biometric authentication is the automated comparison of a person’s biological or behavioral characteristic with a stored reference, used as a factor to confirm who they claim to be. ISO/IEC 2382-37:2022, the international biometrics vocabulary, calls the underlying technology biometric recognition: the “automated recognition of individuals based on their biological and behavioural characteristics.”
In NIST’s guidelines on authentication, SP 800-63B-4, a physical authenticator serves as “something you have,” while the biometric match serves as “something you are.” Something you know is the third factor type. Our guide to digital identity explains how the factor types combine.
ISO/IEC 2382-37 notes that “Use of the term ‘authentication’ as a substitute for biometric verification is deprecated,” applies the same note to biometric identification, and prefers “biometric recognition” for the umbrella. NIST and the market still say “biometric authentication” when a biometric comparison serves as a factor at login or payment approval. This guide follows that usage.
Biometric verification vs identification
ISO defines biometric verification as the “process of confirming a biometric claim through comparison.” At a login, the claim is that the person at the sensor is one specific enrolled person, so the new sample is compared with that person’s stored reference. ISO calls this a one-to-one comparison, often written 1:1. Checking a selfie against the face enrolled for an account is an example of face verification.
Biometric identification starts with no named person. ISO defines it as the “process of searching against a biometric enrolment database to find and return the biometric reference identifier(s) attributable to a single individual.” This is a one-to-many search, often written 1:N.
| Aspect | Biometric verification (1:1) | Biometric identification (1:N) |
|---|---|---|
| Question | Is this the claimed, enrolled person? | Which enrolled person, if any, is this? |
| Compared with | One person’s stored reference | References of many enrolled people |
| Result | Claim confirmed or not confirmed | Candidate list or comparison decision |
| Example use | Login or payment approval | Duplicate enrollment check |
At login, the 1:1 comparison checks the claim of the person opening the app. Before a payment or another high-risk action, it can serve as one factor in a fresh authentication, often called step-up. Onboarding is a different process: identity proofing establishes who a new customer is, and ISO’s duplicate enrollment check, a 1:N search, may run at that stage.
Types of biometric authentication (modalities)
ISO and NIST both give open lists of biometric characteristics, with no fixed number and no ranking. Biological characteristics on both lists include the face, fingerprints, the iris, the retina, the palm, and vein patterns. ISO’s examples also include handwritten signature dynamics, and NIST’s glossary entry “biometrics” lists more behavioral characteristics, such as keystroke cadence, typing speed, screen pressure, and mouse or phone movements. The same entry lists voice prints, and NIST’s US federal guidelines state that “Biometric comparison based on voice SHALL NOT be used.” When a mobile banking app relies on the phone’s own biometric function, the phone reads the face or fingerprint through its camera or fingerprint reader.
How biometric authentication works
ISO calls the whole set-up a biometric system, a “system for the purpose of the biometric recognition of individuals.” It notes that such a system “will contain both biometric and non-biometric components.” The process runs in five steps; the first four are the biometric part.
- Enroll. The sensor captures the customer’s characteristic as a biometric sample, a representation “prior to biometric feature extraction.” The system stores a reference, which ISO allows to be samples, templates, or models. A template is a “set of stored biometric features comparable directly to a biometric probe,” and a stored image, even enhanced or compressed, “is not a biometric template.”
- Capture a probe. At each login or payment approval, the sensor captures a new sample. ISO calls the sample or feature set sent for comparison a probe.
- Compare. The probe is compared with the stored reference, and the comparison produces a score. ISO’s definition of comparison covers similarity or dissimilarity, and its note on the comparison score (37.03.27) warns that “A higher score does not necessarily mean more similar.”
- Decide. A decision policy that includes a threshold turns the score into a comparison decision: a match or a non-match. ISO adds that a decision of “undetermined” may sometimes be given.
- Use the result as a factor. In NIST’s model, a biometric comparison can act as an “activation factor” that unlocks a multi-factor authenticator, and that authenticator then serves as the physical factor.
Step 3 can run on the customer’s device or at a central verifier run by the service. NIST’s guidelines state: “Since the potential for attacks on a larger scale is greater at central verifiers, comparison SHOULD be performed locally.” If comparison is central, the sensor or endpoint “SHALL be authenticated before capturing the biometric sample,” and an authenticated protected channel to the verifier “SHALL be established.”
Accuracy terms: false matches, false non-matches, and the threshold
Each comparison rests on a measurement from the sensor, and NIST notes that this measurement “is subject to noise and presentation variations that require setting an acceptance threshold.” Two kinds of error follow. In ISO’s terms, a false match is a match decision for a probe and a reference that come from different people. A false non-match is a non-match decision for a probe and a reference from the same person and the same characteristic.
ISO counts these errors at two levels, single comparisons and whole transactions, and defines a verification transaction as “one or more verification attempts resulting in resolution of a biometric claim.”
| Term | Level | What it measures |
|---|---|---|
| False match rate (FMR) | Comparison | Share of different-person comparisons wrongly decided as a match |
| False non-match rate (FNMR) | Comparison | Share of same-person comparisons wrongly decided as a non-match |
| False accept rate (FAR) | Transaction | Share of transactions with false claims wrongly accepted |
| False reject rate (FRR) | Transaction | Share of verification transactions with true claims wrongly rejected |
ISO defines a threshold as a “numerical value (or set of values) at which a decision boundary exists,” and both FMR and FNMR depend on it. Setting it is a trade-off. A stricter threshold lets fewer impostors through and turns away more genuine users, and a looser one does the reverse. NIST observes that “The chosen threshold usually emphasizes a low FMR to maximize security since false non-matches can often be mitigated by repeating the measurement or using an alternative authentication method.”
Presentation attacks and liveness detection
A presentation attack targets the sensor. ISO/IEC 2382-37 defines it as a presentation to the capture subsystem “with the goal of interfering with the operation of the biometric system.” Presentation attack detection (PAD) is the “automated discrimination between bona-fide presentations and biometric presentation attacks,” and ISO adds that “PAD cannot infer the biometric capture subject’s intent.” NIST’s guidelines describe liveness detection as a subset of PAD methods that checks whether a sample comes from a living subject present at the point of capture.
ISO/IEC 30107-1:2023 is the framework standard for PAD terms. Its scope clause excludes the standardization of specific detection methods and limits the attacks considered to those “at the capture device during the presentation and collection of the biometric characteristics.” The standard’s introduction also notes that PAD mechanisms “are subject to errors, both false positive and false negative.”
Injection attacks, which bypass the capture device, fall outside that scope. NIST’s guidelines address them separately: the verifier “SHOULD determine the performance and integrity of the sensor and its associated endpoint.” Specific attack techniques and types of liveness checks are not described here.
What NIST’s guidelines require for biometrics
SP 800-63B-4 became final in July 2025 and supersedes SP 800-63B. These US federal guidelines focus on people who interact with US government information systems over networks. The base volume of the suite, SP 800-63-4, says private organizations “may consider the use of these standards where appropriate.” In NIST’s text, SHALL marks a requirement, SHOULD a recommendation, and MAY a permitted option.
The starting point is that “A biometric characteristic is not recognized as an authenticator by itself.” Section 3.2.3 and its subsections then set these rules:
- “Biometrics SHALL only be used as part of multi-factor authentication with a physical authenticator (i.e., ‘something you have’).”
- “The biometric characteristic SHALL be presented and compared for each authentication operation.”
- “An alternative non-biometric authentication option SHALL always be provided to the subscriber.”
- “Biometric data SHALL be treated and secured as sensitive personal information.”
- The system “SHALL implement PAD for facial recognition” and “SHOULD implement PAD for iris and fingerprint modalities.”
- “Biometric comparison based on voice SHALL NOT be used.”
- Presenting a biometric to activate an authenticator “SHALL be a separate operation from unlocking the host device (e.g., smartphone).” The factor that unlocks the device “MAY be used in the authentication operation.” Agencies MAY lower this requirement for authenticators managed by or for the credential service provider (CSP) that have short agency-determined inactivity timeouts and biometric systems meeting the requirements above.
NIST gives its reasons for supporting “only a limited use of biometrics for authentication.” The false match rate alone “does not account for active impersonation attacks.” Biometric comparison “is probabilistic, whereas the other authentication factors are deterministic,” and template protection schemes that allow a characteristic to be revoked have limited availability. “Biometric characteristics do not constitute secrets”: “A facial image can be obtained by taking a picture,” and latent fingerprints can be lifted from objects someone touches. The use of biometric characteristics, especially when stored for central verification, also “introduces new privacy concerns.”
A match shows only that the comparison score, under the decision policy, fell on the match side of the threshold. Whether the sample came from a live person through a trusted sensor is a separate question, answered by PAD and endpoint checks.
Where banks use biometric authentication: Malaysia’s RMiT
This section reads biometric authentication in banking through one regulator’s text. BNM’s Risk Management in Technology (RMiT) policy document has been in effect since 28 November 2025, except where it states otherwise; the current version was issued on 25 September 2026. It applies to the institutions listed on its cover, from licensed banks, Islamic banks, insurers, and takaful operators to approved e-money issuers, operators of designated payment systems, registered merchant acquirers, and intermediary remittance institutions. Paragraphs marked “S” are standards that “must be complied with,” and non-compliance may result in enforcement action. Paragraphs marked “G” are guidance “encouraged to be adopted.”
The appendices cited below are referenced from S 12.1, a standard on security controls for digital services. Under S 12.2, an institution that has not yet put the controls in paragraphs 12.1 to 12.9 in place is required to be ready to show in writing that its alternative measures work at least as well, and to bear the liability for fraud caused by the gap. Appendix 3 sets controls for multi-factor authentication (MFA) and devices. Appendix 11 sets fraud detection standards for financial transactions through digital services for retail customers.
| RMiT paragraph | What it requires (key words) | What it means for biometrics |
|---|---|---|
| App. 3 ¶2(b), (c) | Technology “accurate,” with “a minimal false acceptance rate”; standards “where available” | Biometrics named; RMiT sets no number for the rate |
| App. 3 ¶3(a) | One device per holder by default; designated devices for joint/business; informed opt-out | Fixes which phone’s biometric can approve transactions |
| App. 3 ¶4(a) | “adopt MFA for financial and high-risk non-financial transactions” | A biometric is one factor; MFA needs another type |
| App. 3 ¶5(c) | MFA technology and channels “more secure than unencrypted SMS” | Applies to MFA as a whole; biometrics not singled out |
| App. 3 ¶8 | Own-account financial transactions below RM10,000: MFA optional, “adequate safeguards” | The MFA step, biometric or not, may be skipped there |
| App. 11 ¶2(f), retail | Analytics watch device “biometric credentials” changes where biometrics alone authenticate financial transactions | Covers apps relying on the phone’s biometric function |
Appendix 3, paragraph 2, names biometrics: technology used in identity proofing and authentication methods is required to be “accurate, have a minimal false acceptance rate … across all authentication methods, including biometrics, tokenized processes, and contactless systems.” Under Appendix 11, paragraph 2, an institution’s fraud risk analytics “must include,” at minimum, a list of indicators. Point (f) is “Changes in biometric credentials registered in a device, where biometric technologies are solely used for financial transaction authentication purposes.” It continues: “This includes the scenario where a mobile banking application relies on the biometric verification function of the device, mobile operating system, or a third party application.”
In that scenario, the comparison runs in the phone’s own biometric function, against whatever credentials are registered on the device. A change to those credentials changes who can produce a match, and RMiT treats the change as a fraud indicator.
Our own analysis, in NIST’s terms: when the phone’s biometric approves a payment, it acts as the activation factor, and the bound phone is the “something you have.” Device binding under paragraph 3(a) fixes which phone counts as that factor, and monitoring under point (f) watches who can unlock it, so the two controls work together. None of these provisions makes biometrics compulsory.
Design checklist for a biometric login or payment approval
A dash means none of the provisions cited here addresses the point.
| Decision | NIST SP 800-63B-4 (US federal) | BNM RMiT (Malaysia) |
|---|---|---|
| Where the comparison runs | §3.2.3.3: comparison “SHOULD be performed locally” | — |
| Sensor and endpoint trust | §3.2.3.3: integrity “SHOULD” be checked; if central, sensor “SHALL be authenticated” | App. 3 ¶2(a): “highly resistant to cyber threats affecting customer devices” |
| Presentation attack detection | §3.2.3.2: “SHALL” for face; “SHOULD” for iris, fingerprint | — |
| Physical factor and device binding | §3.2.3: biometrics only with a physical authenticator (“SHALL”) | App. 3 ¶3(a): one device per holder by default, with exceptions |
| Non-biometric option | §3.2.3: alternative “SHALL always be provided” | — |
| After repeated failed attempts | §3.2.3.3: limit failed attempts, delay, then disable; offer another factor if available | — |
| Biometric credential changes | — | App. 11 ¶2(f): biometric changes where biometrics alone authenticate retail financial transactions |
| New device for MFA | — | App. 11 ¶2(a)(iii), retail: deposit-account activity after MFA registered on new device |
Sources checked on 10 November 2026. This article is for informational purposes only and does not constitute legal or regulatory advice. ISO standards and NIST’s US federal guidelines are cited here as technical references. RMiT applies to the Malaysian institutions it lists. Organizations should confirm how these requirements apply to them with the relevant authority or legal counsel.
Biometric authentication answers a narrow question: does this probe come from the person who enrolled this reference? The answer is a probabilistic decision, and a match says nothing about whether the sample came from a live person at a trusted sensor. That is why most of the design work sits around the comparison.
For a bank team, the checklist above comes down to two questions: which device may produce a match, and what happens when a match fails or the device changes. NIST’s guidelines for US federal systems answer several of these questions in their own terms. For the Malaysian institutions it covers, RMiT gives several answers as standards. They include default device binding for transaction authentication and fraud analytics that watch for changes in device biometric credentials where biometrics alone authenticate retail customers’ financial transactions.