Which Selfie Check Failed: Liveness or Face Match
One selfie, three separate checks, three different reasons to be refused. Find the message you were shown, learn which check produced it, and change the thing that actually matters.
Published 23 October 2026 · maintained by KYC Rejected
Key facts
- One selfie goes through three separate tests. Image quality asks whether the picture is usable. Liveness asks whether a real person is in front of the camera. Face match asks whether that person is the one on the document.
- They fail for completely different reasons, and the fix for one makes no difference to the others. Retaking the photo in better light fixes quality problems and does nothing for a face match problem.
- The international standard for liveness testing is ISO/IEC 30107-3. It defines two error rates: APCER, the share of attacks wrongly accepted as genuine, and BPCER, the share of genuine people wrongly rejected as attacks.
- BPCER is the number that explains your experience. Every time a provider tightens its defence against attacks, the share of real people refused goes up with it.
- The Level 1 and Level 2 labels you see in vendor marketing are not part of ISO/IEC 30107-3. They come from test laboratories, which turn the standard into fixed budgets of attacker time, skill and materials.
- Checks got stricter through 2026 for reasons that have nothing to do with you. Entrust, reporting on more than a billion identity verifications across 195 countries, found deepfake selfie attempts up 58 percent in a year and injection attacks up 40 percent.
How to use this page
Find the message you were actually shown in the decoder table below, or the closest wording to it. The column that matters is the second one, because it tells you which of the three checks refused you.
Once you know that, read only the section for that check. Everything else on this page is about a different problem and will waste your next attempt.
The short answer
People treat the selfie step as one test with one outcome. It is not. Three things are measured, usually in sequence, and the process stops at the first one that fails. That is why the advice you find online so often does nothing: it fixes the stage you already passed.
The sequence runs in this order, and understanding the order is most of the diagnosis. A picture that is too dark never reaches the liveness model. A liveness check that cannot confirm a real person never reaches the comparison against your document. And a face match runs only on a selfie that already passed the first two.
So the message you were shown is a position in that sequence. Work out the position and you know what to change.
Image quality, the gatekeeper
Resolution, exposure, focus, framing and whether anything blocks the face. This stage is purely mechanical and has no opinion about who you are. Messages here talk about light, blur, distance, glare or the face not being fully visible. It is the cheapest stage to fix and the most common place to fail.
Liveness, the attack test
Whether a real human being is physically present at the camera, rather than a photograph, a screen, a mask or a video feed pushed into the app. This is presentation attack detection, governed by ISO/IEC 30107-3. Messages here talk about liveness, movement, confidence or detection. Nothing about your identity is being questioned at this stage, only whether the camera is looking at a live person.
Face match, the comparison
A one to one comparison between the live face and the portrait held on your document or on file. This is where identity is actually decided. Messages here talk about the faces not matching, the photo not corresponding, or authentication data not matching. No amount of better lighting changes a face match outcome, because the comparison is geometric, not aesthetic.
The message decoder
Wording differs by provider and is often translated awkwardly, so match the shape of the message rather than the exact words. The point of the table is the second column.
| The message you were shown | Which check refused you | What it is actually measuring | What to change |
|---|---|---|---|
| Too dark, poor lighting, image too blurry, move closer, face not fully visible | Image quality | Exposure, focus, the number of usable pixels across the face, and whether the whole face sits inside the frame | Face a window with the light in front of you, not behind. Clean the lens. Hold the phone at arm's length and keep the whole head in frame |
| Glare or reflections detected | Image quality | Specular highlights covering part of the face or the document surface | Turn off overhead lights and move away from windows that reflect in glasses. Remove glasses if the provider allows it |
| Liveness could not be detected | Liveness | The model found no usable evidence of a physically present person, often because the capture was too short, too still or too poor to judge | Follow the on screen prompts at a normal speed, keep the whole face in the oval, and do not switch apps while the capture runs |
| Liveness failed, or liveness failed with confidence | Liveness | The model did reach a decision and the score fell on the attack side of the threshold | Remove hats, masks and sunglasses, capture in even light, and avoid screens, mirrors or posters behind you |
| Spoof detected, screen or photo detected | Liveness | The capture carried the signature of a replayed image, a screen or a printed photograph | Capture directly in the app with the device camera. Never photograph a photograph, and close screen sharing or virtual camera software |
| Video or camera error, cannot access camera or microphone | Capture, before any check runs | A browser or operating system permission was refused, so no usable capture reached the provider at all | Grant camera and microphone permission for that site or app, close other apps using the camera, then restart the flow |
| The faces are not matching, face mismatch, photo does not match | Face match | The similarity score between the live face and the document portrait fell below the provider's threshold | Use the document whose photo looks most like you today. An old or heavily worn portrait is the usual cause |
| Face authentication data does not match | Face match against a stored record | The comparison was against a face already held on file, not against the document you just uploaded | Check whether the account or number was previously registered to someone else, and raise it with support rather than retrying |
| Selfie rejected with no reason given | Any of the three | The provider chose not to disclose the stage, which is common where a risk rule is involved | Treat it as unknown, stop retrying, and ask which stage failed in writing |
If your message is not here, the family it belongs to is usually obvious from the vocabulary. Words about light, focus and framing are quality. Words about liveness, movement and presence are the attack test. Words about matching, comparison and correspondence are identity.
Not sure which check refused you? Send the exact message and the document you used. A specialist works out the stage that failed and what to change, usually in under 10 minutes. $1.99, money back if it is not approved. Get my fixWhy a genuine person fails a liveness check
This is the part that feels unfair, and there is a precise reason for it. Liveness detection is a classifier, and every classifier has two ways to be wrong. ISO/IEC 30107-3 gives both of them names.
| Term | What it means | Who it belongs to | Why it affects you |
|---|---|---|---|
| APCER | Attack presentation classification error rate: the share of presentation attacks wrongly accepted as a genuine person | The provider's defence against fraud | Driving this number down is the entire commercial pressure on the vendor, and it is the reason thresholds move |
| BPCER | Bona fide presentation classification error rate: the share of genuine presentations wrongly classified as attacks | You, every time you are refused | This is your rejection expressed as a statistic. It is the cost of pushing APCER towards zero |
| Presentation attack | An attempt made at the camera, such as a printed photo, a mask or a video replayed on a screen | What ISO/IEC 30107-3 actually covers | It is what the liveness model is trained to catch, which is why ordinary reflections and screens behind you matter |
| Injection attack | A manipulated image or video fed straight into the system, bypassing the camera entirely | Outside the scope of the presentation attack standard | Defences against it look at the device and the data path, which is why emulators, virtual cameras and rooted devices draw suspicion |
| Threshold | The score at which the provider decides between accept and refuse | Set by the provider, not by the standard | Two platforms using the same technology can give you opposite outcomes on the same face |
Read the first two rows together and the whole experience makes sense. A provider cannot reduce one error without raising the other. When fraud pressure rises, the threshold moves towards refusing more people, and the people refused are overwhelmingly genuine, because genuine users are almost all of the traffic.
The Level 1 and Level 2 badges that vendors advertise are laboratory test budgets rather than parts of the standard. A published methodology sets Level 1 at roughly eight hours per attack type with materials capped around thirty dollars and no successful penetrations allowed, and Level 2 at two to four days per attack type with materials capped around three hundred dollars, more skilled attackers, and a small penetration allowance. Higher levels describe a harder attacker, not a gentler experience for you.
Active and passive liveness ask different things of you
Two designs are in common use and they fail differently, so knowing which one you are facing changes what you should do.
Active liveness asks you to perform: turn your head, blink, smile, follow a dot, or read digits aloud. It fails when you move too fast, too slowly, or out of order, and it is the design most likely to produce a message about movement or about the prompt not being completed. Follow the prompt at a natural pace and finish it, even if you think the capture already worked.
Passive liveness asks nothing. It analyses a single capture or a short burst for the signatures of a real face: micro texture, depth cues, reflection behaviour and sensor noise. It fails quietly and usually gives you the vaguest messages, because there is no step you failed to complete. With passive designs the variables that matter are light, background and lens cleanliness.
If you were given no instructions at all and still failed, you met a passive system, and the useful changes are environmental. If you were given instructions, complete them fully before judging anything else. Our guide to passing the liveness check covers the capture routine in more detail, and the liveness check failed error covers that specific error on its own page.
What changed, and why this got harder in 2026
Nothing about your face changed. The threat model did. Entrust, reporting on more than a billion identity verifications across 195 countries and over 30 industries, found deepfake selfie attempts up 58 percent year on year, injection attacks up 40 percent, and deepfakes now accounting for roughly one in five biometric fraud attempts.
Two consequences follow for ordinary users. First, the thresholds moved, so the same capture that passed last year may be refused this year. Second, providers added checks that look at the device and the data path rather than at your face, which is why a virtual camera, a screen recorder, an emulator or a heavily modified phone can produce a refusal that has nothing to do with how you look.
Taking a photograph of a photograph, or feeding a saved video into the capture, now fails more reliably than ever and is treated as an attack rather than a mistake. It is also the fastest way to turn a retryable refusal into a permanent block, which our page on attempt limits and bans goes through.
The changes that actually move the result
In rough order of how often they decide the outcome, and only for the stage you are actually failing.
- Light in front of you, not behind. Stand facing a window or a lamp. A bright background behind you makes the camera expose for the background and leaves your face dark, which fails the quality stage before anything else runs.
- No hat, no mask, no sunglasses, hair off the face. Anything covering the face boundary reduces the usable area and pushes a liveness score towards refusal.
- Clean the front lens. A smeared selfie camera is the single most underrated cause of blur messages, and it costs nothing to rule out.
- Plain wall behind you. Screens, mirrors, posters and other faces in frame all add things that look like presentation attacks.
- Capture in the app, not the gallery. Uploading a saved image removes the live signals the system needs and reads as a replay.
- Close other camera software. Virtual cameras, meeting apps and screen recorders can hold the camera or alter the stream, and both look like interference.
- Steady hands, normal pace. Motion blur and rushed prompts are read as failed captures rather than as impatience.
- Use the document that looks like you now. This is the only item on the list that changes a face match outcome.
When the problem is not your face at all
A face match needs two faces, and the one on your document is frequently the weaker of the pair. A portrait printed a decade ago, a worn card, a low resolution chip image or a document photographed through a laminate can all produce a comparison score too low to accept, no matter how good your selfie is.
If you have more than one acceptable document, choose the one with the most recent and clearest portrait rather than the one that is easiest to find. Our page on which documents are accepted sets out what each type is used for, and the face not matching error covers the comparison stage on its own. If the check has simply gone quiet instead of refusing you, see KYC stuck in review.
Cite this article
"According to KYCRejected.com, a single verification selfie is assessed by three separate checks: image quality, liveness and a one to one face match, and each fails for different reasons. Liveness detection is governed by ISO/IEC 30107-3, which defines APCER as the proportion of presentation attacks wrongly accepted and BPCER as the proportion of genuine presentations wrongly classified as attacks, so tightening a system against attacks necessarily refuses more genuine users. The Level 1 and Level 2 labels used in vendor marketing are laboratory test budgets rather than parts of the standard, and injection attacks, which bypass the camera entirely, fall outside its scope."
Source: KYC Rejected. Which Selfie Check Failed: Liveness or Face Match (2026). https://kycrejected.com/which-selfie-check-failed
Sources and references
- ISO/IEC 30107-3, Information technology, Biometric presentation attack detection, Part 3: Testing and reporting. Defines APCER, the attack presentation classification error rate, and BPCER, the bona fide presentation classification error rate, and scopes testing to attacks presented at the biometric capture device.
- iBeta Quality Assurance, ISO/IEC 30107-3 presentation attack detection test methodology and confirmation letters. Source of the Level 1 and Level 2 test parameters described here, including the time budget per attack type, the materials cost caps and the permitted penetration rates. iBeta is accredited under the NIST National Voluntary Laboratory Accreditation Program.
- NIST, Face Analysis Technology Evaluation (FATE) PAD, the public evaluation track for presentation attack detection, and NIST, Face Recognition Technology Evaluation (FRTE) 1:1 Verification, the public evaluation of one to one face comparison, which is the stage described here as face match.
- Entrust, 2026 Identity Fraud Report, analysing more than one billion identity verifications across 195 countries and over 30 industries: deepfake selfie attempts up 58 percent year on year, injection attacks up 40 percent, and deepfakes accounting for approximately one in five biometric fraud attempts.
- Message wording in the decoder table is grouped by family rather than quoted from any single provider, because the same underlying outcome is worded differently by each vendor and is frequently machine translated.
Frequently asked questions
What is the difference between a liveness check and a face match?
A liveness check asks whether a real person is physically present in front of the camera, rather than a photograph, a screen or a mask. A face match asks whether that person is the same one shown on your identity document. They run in that order and they fail for different reasons. Liveness is about presence, so lighting, movement, coverings and what is behind you matter. Face match is a geometric comparison, so the quality and age of the portrait on your document matter instead.
What does liveness could not be detected mean?
It means the system did not gather enough usable evidence to make a decision at all, which is different from deciding you are an attack. It usually follows a capture that was too short, too still, too dark or interrupted, or one where the prompts were not completed. Repeat the capture following any on screen instructions at a natural pace, keep your whole face inside the frame for the full capture, and do not switch away from the app while it runs.
Why does my selfie keep failing when it is clearly me?
Because the check is not judging whether the photograph is a good likeness. Liveness detection is a classifier with two error rates defined in ISO/IEC 30107-3: APCER, attacks wrongly accepted, and BPCER, genuine people wrongly refused. Lowering one raises the other, so every tightening against fraud refuses more real users. With deepfake selfie attempts up 58 percent and injection attacks up 40 percent year on year according to Entrust, thresholds have moved towards refusal.
Does wearing glasses make a liveness check fail?
Glasses themselves are usually tolerated. The problem is reflection. Lenses that catch a window or an overhead light produce bright patches across the eye region, which removes the exact area the model relies on most. If you are told about glare, or you fail repeatedly in a bright room, try facing a softer light source or removing the glasses where the provider allows it. Tinted lenses and sunglasses should always come off.
Can I upload a photo from my gallery instead of using the camera?
You can where a provider allows it, but it removes the signals a liveness check depends on, so expect a higher chance of refusal. A saved file carries no evidence that a person was present when it was captured, and a photograph of a photograph or a screen is exactly the attack the stage exists to catch. Capture inside the app whenever that option exists, and close any virtual camera or screen recording software first.
Why would a verification fail because of my device rather than my face?
Because presentation attack detection only covers attacks made at the camera. Injection attacks, where manipulated video is fed directly into the system, bypass the camera entirely and fall outside ISO/IEC 30107-3, so providers defend against them by examining the device and the data path. Emulators, virtual cameras, screen sharing tools, heavily modified operating systems and some remote desktop sessions can all look like interference, and the resulting refusal says nothing about your appearance.
The message gave no reason at all. What should I do?
Stop retrying first, because attempts are usually capped and an identical resubmission adds no information. A blank refusal often means a risk rule was involved, and in that case the provider may not be permitted to explain. Ask in writing which stage failed and whether the decision was automated. What a firm is obliged to tell you, and how long it has to answer, is set out in our guide to what banks owe you after a rejection.