Passive vs active liveness: security, speed, and user effort

Two identical grey plaster mannequin heads, one facing forward and one turned, with an orange arc over the turned one

Passive vs active liveness: what each mode catches, what one US test measured on speed and user effort, and what Southeast Asian regulators ask.

In remote electronic Know Your Customer (eKYC) onboarding, liveness checks come in two modes: active liveness asks the user to act, and passive liveness judges an image already captured. This guide sets them against one US government test, the standards and labs that test them, and Southeast Asian rules. None of the sources cited here measures how users in Southeast Asia respond, and none of the regulator texts picks a mode.

What passive and active liveness actually do

ISO/IEC 30107-1 defines liveness detection as analysis to “determine whether a biometric sample is being captured from a living subject present at the point of capture”. The standard calls it “a subset of presentation attack detection methods”, and presentation attack detection (PAD) covers attacks made at the capture device.

Active liveness detection asks the user to do something the camera can check, such as turning the head or blinking. Each such cue asks for a response that a fake must reproduce, which is why iBeta’s testers tailor attacks on active systems to supply the requested movement, smile, or blink.

Passive liveness detection gives the user no task: the system judges an image or video that has already been captured. The Monetary Authority of Singapore (MAS) discusses liveness detection in an information paper, “Cyber Risks Associated with Deepfakes” (MAS/TCRS/2025/06, 18 September 2025). It says institutions should implement liveness detection “to detect manipulated videos through motion analysis, texture analysis, thermal imaging, 3D Depth analysis and behavioural analysis”. It adds that they could also prompt the user to act. MAS does not call these signals passive and does not define them.

ISO/IEC 30107-1 gives its own examples of liveness evidence. Anatomical characteristics include the “absorption of illumination by the skin and blood”, and involuntary reactions include “the reaction of the iris to light and heart activity (pulse)”. Voluntary reactions include a presentation “in response to a directive cue”. For passive software, the US National Institute of Standards and Technology’s NIST IR 8491 (2023) found that a number of algorithms did better on video than on still images, a gain it calls “highly algorithm-dependent”.

Both labels are industry terms, not terms defined in ISO/IEC 30107-1. Either mode can run at onboarding or later, as part of biometric authentication.

What each mode catches and misses in presentation attack detection tests

What the tests found

The Remote Identity Validation Rally (RIVR) was run for the US Department of Homeland Security’s Science and Technology Directorate (DHS S&T) at the Maryland Test Facility. In late 2025, 645 US volunteers ran one transaction per phone on each of 6 active products, and 12 passive products processed selfies the same volunteers had captured with an app. Results were published in February–March 2026. The test measured different things for each mode, which the speed section explains.

Results varied greatly between products. One of the 6 active and one of the 12 passive systems met both the security and convenience thresholds. The test’s own conclusion: “Low-effort presentation attacks can be effective.”

NIST IR 8491 covered passive, software-based detectors only: 82 algorithms from 45 developers. Accuracy “varied widely” across algorithms, use cases, and attack types, with “only a small percentage of algorithms evaluated” showing notable detection capability. NIST did not test challenge-response methods, so it gives no data on active liveness.

What they did not cover

Neither mode, as tested in these sources, covers injection, where media is fed in place of the camera. ISO/IEC 30107 covers only attacks at the capture device, NIST IR 8491 is “silent on … injection attacks”, and the DHS S&T test left injection out. The FIDO Alliance’s requirements say challenge-response “may help prevent attacks, particularly those less sophisticated”, and require vendors only to document their injection protections.

How long a liveness check takes: speed and user effort in one US test

In the DHS S&T test, the passive capture app ran on three phone models; it allowed two attempts to pass automated quality checks and offered manual capture on the third.

MeasureActive (6 products)Passive (12 products)
Average time (worst case per product)18.1–23.1 seconds per transaction, whole user interaction0.3–7.4 seconds of machine processing only
Positive satisfaction88.1%–96.3% of volunteersNot measured
Genuine users wrongly rejected (median)4.3%1.7%
Worst-case attack acceptance, by product0%–65%1.7%–100%

The two modes were measured differently, so these figures cannot be compared mode against mode. Active time covers the volunteer’s whole interaction, while passive time is machine processing without the volunteer’s capture time. Satisfaction was measured for active systems only.

As arithmetic on this cohort only: at the median, roughly 96 in 100 genuine single attempts with active systems, and 98 in 100 with passive ones, were not wrongly flagged as attacks. This is not a completion rate. The test did not measure retries or drop-off, passive images had already passed automated quality checks, and the medians are across system and phone combinations in a US sample.

Active results varied more across demographic groups, with the widest gap for volunteers aged 46 and over, and results also varied by smartphone model.

What Southeast Asian regulators ask of liveness

Regulator and instrumentModalityWhat it says about liveness
State Bank of Vietnam (SBV): Circulars 50/2024 and 77/2025Must (“phải”)Online confirmation, first or new-device login: certified PAD, 3-minute match, ≤10-failure lock
SBV: Circular 17/2024Must (“phải”)Account opening: biometrics matched to chip ID; no liveness wording
Bank Negara Malaysia (BNM): e-KYC policy documentMay; shall for Appendix 4 accountsLiveness module; minimum PAD test scenarios
Bank of Thailand (BOT): 2023 biometric technology guidelineShouldAttempt limit, growing delay, another method (footnote 7)
MAS: information paperShould; couldLiveness signals; “could also prompt” users to act
Bangko Sentral ng Pilipinas (BSP): Circular No. 1213 (2025)MustNo liveness wording
Otoritas Jasa Keuangan (OJK): POJK 8/2023, 21/2023Must (“wajib”)Biometric factor; no liveness wording

Vietnam’s rules are the most detailed. Circular 50/2024/TT-NHNN, Article 11(5), sets the face-matching rules for confirming online banking transactions. Article 8(6) applies them when an individual customer first logs in to a Mobile Banking app, or logs in on a new device, where biometric checks are required. Face matching must be able to detect presentation attacks with photos, video, and 3D masks, based on international standards such as NIST SP 800-63B, ISO 30107, or FIDO. Article 11(5) also sets an accuracy floor: a false rejection rate below 5% at a false acceptance rate below 0.01%, or equivalent, under the FIDO requirements, on at least 10,000 samples.

The match must finish within 3 minutes (“tối đa 03 phút”), and biometric confirmation must lock after the institution’s limit of consecutive failures, no more than 10. Circular 50 already required the PAD solution to be certified by a FIDO-recognized lab. Since 1 March 2026, Circular 77/2025/TT-NHNN also accepts an accredited certification body. It sets the bar in these words: “đáp ứng tiêu chuẩn ISO 30107 cấp độ 2 (Level 2) hoặc tương đương” (meeting ISO 30107 Level 2 or equivalent). ISO/IEC 30107-3 sets the test method and defines no numbered levels such as “Level 2”. Its clause on “levels of evaluation” names which part of the system is tested: the PAD subsystem, data capture, or the full system. Labs and programs such as iBeta and FIDO set the numbered levels.

Circular 17/2024/TT-NHNN, Article 16, requires a bank opening an account electronically to match the customer’s biometrics against the chip ID card, or through the VNeID electronic identity account. Neither Circular 17 nor its 2025 amendment, Circular 25/2025, sets a liveness or PAD requirement of its own.

Circular 50 names NIST SP 800-63B among its standards. Its current version, NIST SP 800-63B-4 of July 2025, is a US federal guideline. It says biometric systems “SHALL implement PAD for facial recognition” and should show an impostor attack presentation accept rate (IAPAR) below 0.07. After 5 consecutive failures, or 10 with PAD, it requires a delay of at least 30 seconds before each further attempt. After 50 (100 with PAD), it requires disabling biometric authentication and offering another factor. It picks no mode.

BNM’s e-KYC policy document of 15 April 2024 lists liveness detection as one of three e-KYC modules an institution “may” use (G 8.10). For certain current, savings, and investment accounts opened through eKYC, its Appendix 4 makes all three mandatory. Appendix 3 sets minimum test scenarios, with “DeepFake” tests “if possible”, and says the PAD test “may be done in conformance to ISO/IEC 30107-3”. Test samples “shall be representative of and adequately reflect the demographics of an FI’s customers”.

The BOT’s Guideline for Biometric Technology Adoption in Financial Services (September 2023) says providers should set procedures to detect impostors at registration, for example using “presentation attack detection or liveness detection technology, and other controls when possible”. Footnote 7 to its recommended practice on attempt limits points to NIST SP 800-63B and Thailand’s Electronic Transactions Development Agency for fewer than 10 failed presentation attack detection attempts. It then recommends a growing delay, refusing authentication, and another method.

BSP Circular No. 1213, Series of 2025 (May 2025), amends IT risk-management rules under the Anti-Financial Account Scamming Act. It contains no liveness wording, and neither do OJK’s POJK 8/2023, on anti-money laundering, counter-terrorism and counter-proliferation financing programs, and POJK 21/2023, on digital services by commercial banks.

What “Level 2” means: ISO 30107, iBeta and FIDO

iBeta is a US lab accredited by NIST’s National Voluntary Laboratory Accreditation Program (NVLAP), with ISO 30107-3 in its scope. It defines its own three test levels. Level 1 uses artifacts costing up to $30 and allows no successful attacks. Level 2 uses artifacts up to $300, such as resin or latex masks, and allows 1%.

Those dollar figures are iBeta’s limits on test materials, and they say nothing about what attacks cost outside the lab. To pass, a product must now reject no more than 15% of genuine users, or 10% for Level 3. iBeta says its result “does not translate to a certification of the vendor product”.

The FIDO Biometrics Requirements (v4.1, January 2025) set two levels for face verification. Level 1 allows a 1% false acceptance rate and a 7% false rejection rate. Level 2 tightens false acceptance to 0.033%, with false rejection at 7% or 5% depending on the reference image. Both levels cap attack acceptance at 7% per attack type and 4% across all types, over 6 Level A and 8 Level B attack types.

FIDO’s face-verification page lists v4.1 as active and does not list v5.0; its biometric-component page also lists v5.0, a file headed “Review Draft” (1 September 2026), and sets v4.1 to sunset on 31 March 2027. Version 4.1 does not prescribe active or passive liveness.

FIDO’s Level 2 and iBeta’s Level 2 are different scales. FIDO’s sets error rates for the face verification system; iBeta’s sets attack effort and cost, plus an allowed success rate.

Our reading: when Vietnam’s circular asks for “ISO 30107 Level 2 or equivalent”, the level comes from a lab or a program such as iBeta or FIDO, since ISO/IEC 30107-3 sets the test method.

How to choose and measure: a checklist

These steps apply whichever mode of liveness detection you run.

In a remote eKYC funnel, the three candidate drop-off points are document capture, the selfie step, and the wait for manual review. The liveness check sits inside the selfie step, so its cost shows up in that step’s completion. The DHS S&T test cannot say which mode your customers will finish faster. Measure this before you change anything, and decide in advance how much completion you will trade for the attack coverage you need.

In our view, a reasonable starting point by step is as follows; no regulator or standard sets it.

StepMode to start withWhen to escalate
Onboarding selfiePassiveActive challenge on low confidence or higher-risk accounts
Transaction confirmationPassiveActive challenge on high-value transfers; Vietnam’s Circular 50 applies
Account recovery or new deviceActiveAnother method after repeated failures; Vietnam’s Circular 50 applies

Each escalation adds the active step’s cost for the cases it covers: in the DHS S&T test that was 18–23 seconds of user time and a median false rejection of 4.3%, on US volunteers and the products tested. This passive-first, escalate-to-active flow is what vendors call hybrid or step-up liveness.

Checklist
Name the attacks you must stop, by market

The DHS S&T test (US) used three attack classes, BNM’s Appendix 3 (Malaysia) sets 2D and 3D masks and “ShallowFake” images, and Circular 50/2024 (Vietnam) names photos, video, and 3D masks.

Ask for the lab report

Check the lab, level, product version, device, and genuine-user rejection limit (iBeta; FIDO). Rates from one lab are not necessarily reproducible by another (ISO/IEC 30107-3).

Ask which mode the report covers

Ask which mode, configuration, and capture flow the lab tested. If it covers a different mode or configuration, treat it as evidence for that setup only.

Instrument the selfie step

Before any change, record four metrics by age band and device:

  • completion: the share of users who finish the step;
  • retries: attempts per user who finishes;
  • time: seconds from the start of the step to the result;
  • false rejections: genuine users failed; reference points, not targets: medians of 4.3% (active) and 1.7% (passive) in the DHS S&T test, iBeta’s 15% cap, FIDO’s 7% or 5%, and Vietnam’s below 5% (Circular 50/2024).
Make one change, then re-test

When you add an escalation from the table above, the gain to check is attack coverage on the escalated cases; the cost to watch is completion and false rejections at the selfie step. We suggest repeating step 4 after the change and comparing the result with your baseline.

Set the failure path in advance

Cap failed attempts, then add delays and offer another method: the BOT guideline recommends a growing delay (footnote 7), and NIST SP 800-63B-4 requires at least 30 seconds per further attempt. For online banking confirmation in Vietnam, lock after no more than 10 consecutive failures (Circular 50/2024).

Plan for false negatives

BNM’s FAQ (question 4) says false-negative customers “should not be immediately dismissed”, so define a remedial route.

In Vietnam, check the certificate

For transaction confirmation and new-device logins, confirm the PAD certification matches Circular 77/2025’s “ISO 30107 Level 2 or equivalent” wording.

This article is general information and does not constitute legal advice. Sources were checked on 12 November 2026. Translations of Vietnamese texts are ours, and the BOT guideline is quoted from its unofficial English translation. Test results apply to the products tested. Institutions should confirm how these texts apply to them with their regulator or legal counsel.

Conclusion

Passive and active liveness are industry labels for two ways of running presentation attack detection. The DHS S&T test measured the two modes differently, and its products varied greatly within each mode. No Southeast Asian regulator text we checked chooses a mode. Vietnam sets the most detailed rules, for confirming online banking transactions, while its account-opening rule adds no liveness requirement of its own.

Taken together, the sources suggest that the product and its test report tell you more than the label does. With no regional user data to go on, the completion cost of either mode is something to measure at your own selfie step, before and after any change.

FAQ

Question author
What is the difference between active and passive liveness?
Active liveness asks the user to do something the camera can check, such as turning the head or blinking. Passive liveness analyzes a captured selfie or video without a task. Both are industry labels; ISO/IEC 30107-1 treats liveness detection as part of presentation attack detection.
NeuroVision expert
Question author
How does passive liveness detection work?
The system examines a captured image or video for signs that it shows a live person, without asking the user to act. ISO/IEC 30107-1's examples of liveness evidence include the skin and blood absorbing light and the iris reacting to light. Singapore's central bank lists signals such as motion, texture, and 3D depth analysis for liveness detection, without defining them. NIST's 2023 test of 82 passive detectors found that accuracy varied widely between algorithms and attack types, and it did not cover injection.
NeuroVision expert
Question author
Is active liveness more secure than passive?
Not as a rule. In a 2025 US government test by DHS S&T, results varied greatly between products of both kinds. One active and one passive system met both the security and convenience thresholds, and low-effort attacks still worked against some systems. The test measured the two modes differently.
NeuroVision expert
Question author
How long does a liveness check take?
In a 2025 US government test by DHS S&T, active checks took about 18 to 23 seconds per transaction on average. Passive systems took 0.3 to 7.4 seconds to process an image, without the user's capture time, so the figures cannot be compared. No source cited here measures this for users in Southeast Asia.
NeuroVision expert
Question author
Do regulators in Southeast Asia require a specific type of liveness?
None of the regulator texts cited here chooses between passive and active. Vietnam requires presentation attack detection certified to "ISO 30107 Level 2 or equivalent" for confirming online banking transactions, and its account-opening rule adds no liveness requirement. ISO/IEC 30107-3 defines no numbered levels; labs and programs such as iBeta and FIDO set them. Malaysia sets minimum test scenarios, Thailand recommends attempt limits, and Singapore says institutions could prompt users to act.
NeuroVision expert
Question author
What does "iBeta Level 2" mean?
It is a test level defined by iBeta, an accredited US lab, under ISO/IEC 30107-3. Level 2 uses artifacts costing up to $300, such as resin or latex masks, and allows at most 1% successful attacks. ISO/IEC 30107-3 defines no numbered levels, FIDO's Level 2 is a different scale, and iBeta says a pass is not a product certification.
NeuroVision expert