CEN/TS 18099: What Injection Attack Testing Proves

CEN/TS 18099 is the European technical specification for biometric injection attack detection. It defines how to test whether a system can spot synthetic data inserted directly into its pipeline, bypassing the camera entirely. Published in 2024, it grades resistance across three attack levels and is now referenced in national identity-proofing requirements. Germany’s Federal Office for […]

CEN/TS 18099 is the European technical specification for biometric injection attack detection. It defines how to test whether a system can spot synthetic data inserted directly into its pipeline, bypassing the camera entirely. Published in 2024, it grades resistance across three attack levels and is now referenced in national identity-proofing requirements.

Germany’s Federal Office for Information Security put a date on it. Its TR-03147 evaluation guidance requires that biometric injection attack detection be tested by an accredited laboratory to CEN/TS 18099, at level substantial for baseline identity proofing and level high for extended identity proofing, before the end of 2026.

For anyone selling or buying remote identity verification that touches the European market, that deadline is close.

Most providers are not ready for the question behind it. Ask a provider whether their liveness check handles deepfakes and you’ll usually get a yes, followed by a presentation attack certificate that does not answer the question.

And the number who could sit this exam today, let alone pass it, is very small. The specification is two years old. Accredited laboratories only began offering the testing recently, and the systems that have cleared it at Substantial level or above worldwide still number in the handful.

We’re proud to be one of them.

What a biometric injection attack is

An injection attack replaces the biometric feed before it ever reaches the application. No artifact is held up to a lens. The attacker works further upstream, through a virtual camera, an emulator, a hooked system function or manipulated network traffic, and the system receives a video stream that looks exactly like a genuine capture.

The Age Check Certification Scheme, which runs accredited testing against the specification, draws the line plainly in its 2026 service documentation: presentation attack detection covers physical spoofing such as masks and printed photos, while injection attack detection covers attacks that enter through virtual webcams, APIs and manipulated signal inputs.

The camera never sees the attack

A presentation attack requires something physical in front of a sensor. A printed photograph, a tablet replaying a video, a silicone mask. Liveness detection is the control built for exactly that case.

The capture device records a real scene containing a fake object, and detection works by finding the evidence of that object: screen moire, print texture, flat depth, unnatural reflection.

An injection attack removes the scene. No photograph, no tablet, no mask, nothing in front of the lens at all, so there are no such signals to find. The camera is still running. It just isn’t the thing feeding the application.

The detection problem moves from analyzing an image to establishing where the image came from.

Why injection scales and presentation does not

Physical presence caps a presentation attack. Someone has to build the artifact, hold it up, and sit through the session. One attacker, one attempt at a time.

Injection removes that ceiling. An attack delivered through software can be scripted and run in parallel across thousands of sessions with no marginal effort per attempt. iProov’s 2025 threat intelligence reports a 740% rise in injection attacks on iOS over the year, which is the kind of growth curve automation produces and manual attacks do not.

Where deepfakes enter the pipeline

Think of the deepfake as the thing being smuggled in, and injection as how it gets through the door.

Keeping those apart matters commercially, because they are sold as one thing and tested as two. A generated face has to reach the verification system somehow, and the route it takes is what CEN/TS 18099 measures.

A provider can detect synthetic faces well and still accept any feed that arrives. Another can reject tampered feeds while having no view on whether the face inside a genuine feed was generated. Buying one and assuming the other is the most common gap we see in remote onboarding stacks.

!Diagram of injection points in a biometric capture pipeline

What CEN/TS 18099 specifies

CEN/TS 18099 gives injection attack testing a shared vocabulary and a repeatable method, which is what was missing while every player described its own defenses in its own terms. The specification was published by CEN, the European Committee for Standardization, in 2024.

Two terms carry most of the weight, and a test report is unreadable without them.

Injection Attack Method and Injection Attack Instrument

The Injection Attack Method, or IAM, is how an attack is delivered. Exploiting a software library, manipulating network traffic, running the app in an emulator, hooking a system function to intercept and replace the biometric data.

The Injection Attack Instrument, or IAI, is what gets delivered. The synthetic face, the deepfake video, the manipulated image sequence the attacker wants accepted as genuine.

Splitting them is what makes a result diagnostic rather than a grade. This is the part most buyers skip, and it’s the part that tells you whether a provider is strong or just lucky. A system can be strong against one delivery method and blind to another while the instrument stays identical, so a single pass or fail tells a buyer much less than the method-by-method breakdown underneath it.

The three levels

Testing runs across three tiers of attack sophistication. The Age Check Certification Scheme publishes the following breakdown for its accredited testing, and Germany’s BSI uses the parallel vocabulary of basic, substantial and high in TR-03147.

LevelAlso calledWhat the attack looks like
Level 1BasicReplayed video, screen recording, browser-based signal replay
Level 2Intermediate, substantialDeepfake injection, synthetic biometric input, signal manipulation
Level 3Advanced, highCustom-crafted injections, API tampering, virtual sensor spoofing

Level 2 is where deepfake attacks on onboarding formally enter the test plan. That is also the level Germany’s BSI sets as the floor for baseline identity proofing, which makes it the practical minimum for most regulated remote onboarding rather than an aspiration.

How attack potential is actually rated

A level means something because of how the specification scores the effort behind an attack. BSI’s TR-03147 evaluation guidance describes the CEN/TS 18099 rating methodology as estimating attack potential across two phases: identification, covering what it takes to discover and first build the attack, and exploitation, covering what it takes to reuse it against other identities.

That second phase is the one buyers underweight. An attack that takes a specialist three weeks to develop and then thirty seconds to replay against any subsequent applicant is not a research curiosity, and the two-phase score is what separates it from one that stays expensive every time.

Does a PAD certificate cover injection attacks?

No. A presentation attack detection certificate says nothing about injection resistance, and the standard behind it says so itself.

ISO/IEC 30107-3:2023, the international standard for presentation attack detection testing, states in its own scope that overall system-level security and vulnerability assessment sit outside what it covers. Presentation attacks are defined at the capture device. An injected feed never reaches a capture device, so no amount of PAD testing exercises it.

Two standards, two attack surfaces

 ISO/IEC 30107-3CEN/TS 18099
Attack surfaceArtifact presented to the capture deviceData inserted into the pipeline
Tests forPrint, replay, mask, partial attacksVirtual cameras, emulators, hooked functions, API tampering
Says nothing aboutInjection attacksPhysical presentation attacks
Issuing bodiesAccredited PAD laboratoriesAccredited IAD laboratories

Neither certificate substitutes for the other, and a remote onboarding flow is exposed to both at once.

Three questions to put to a provider

A provider claiming deepfake detection has not thereby claimed injection resistance. Three questions separate the two:

  1. Has your injection attack detection been tested to CEN/TS 18099 by an accredited laboratory, and at which level?
  2. Which injection attack methods were in scope, and which were excluded?
  3. Which product and which version was tested, and on what date?

A provider that answers the first with a presentation attack certificate has answered a different question.

Who has to evidence this, and by when?

Accredited CEN/TS 18099 testing is already a dated requirement in at least one national scheme, and the population it covers is wider than biometric suppliers. Remote identity proofing is rarely confined to one market, so a requirement written for one jurisdiction tends to reach the suppliers serving it from everywhere else.

The BSI TR-03147 requirement

Germany’s Federal Office for Information Security set the clearest deadline in its TR-03147 evaluation guidance. Identity proofing targeting baseline assurance must have its injection attack detection tested by an accredited laboratory to CEN/TS 18099 level substantial, which is level 2.

Extended assurance requires level high, which is level 3. Both deadlines land at the latest before the end of 2026.

Two consequences follow for anyone selling into that market. A supplier without an accredited result has nothing to offer a customer who needs one, and a buyer who accepts a presentation attack certificate in its place has not met the requirement.

Who this reaches

The Age Check Certification Scheme lists the systems its 2026 injection attack detection testing is built for: facial recognition and age estimation providers, KYC and onboarding platforms, biometric SDK suppliers, identity providers integrating age verification or facial age estimation, and any system that processes biometric data remotely or through an API.

That last category is broad on purpose. A system does not need to be sold as a biometric product to fall inside it.

What the certification produces

Accredited testing against the specification produces artifacts, not a badge. The Age Check Certification Scheme publishes five deliverables for its 2026 testing:

  1. A structured test plan aligned to CEN/TS 18099
  2. An injection simulation report detailing attack vectors and outcomes
  3. Pass or fail results per attack level
  4. A certificate of conformity for systems that meet the criteria
  5. A listing on its certified registry

The simulation report is the document to ask for. The certificate just tells you it passed; the report tells you what it was up against.

Where the requirement is heading

ISO/IEC 25456 is in development and will extend injection attack detection requirements beyond Europe. The European Digital Identity Wallet program and high-assurance remote identity proofing are the commercial drivers pulling it forward, and BSI’s own guidance positions CEN/TS 18099 and ISO/IEC 25456 as a catalog the two standards share.

For a provider outside Europe, the sequence matters more than the geography. Requirements of this shape tend to arrive in one market and then propagate.

Where injection resistance actually comes from

Passing an injection attack test is downstream of a design decision. If your system treats its own camera input as trustworthy, it has nothing to detect with: by the time a frame is analyzed, the substitution has already happened.

Resistance comes from establishing provenance at capture rather than inspecting pixels afterwards. Published liveness benchmarks tell you how good the detector is. They do not tell you anything about where the frames it looked at came from.

Device and app attestation

Attestation answers a question that happens before any biometric is examined: is this a genuine build of the application, running on a genuine device, in a state that has not been tampered with. A rooted device, a modified client or an emulator fails that check regardless of what the camera subsequently reports.

Virtual camera and emulator detection

A virtual camera presents itself to the operating system as an ordinary capture device while sourcing its frames from a file or a stream. Passive liveness running on those frames will assess them as it would any others.

Detecting one means examining the capture source itself: which device the frames claim to come from, whether that device behaves like real hardware, and whether the session shows the signatures of an emulated environment.

HyperVerge runs identity verification inside digital onboarding for banks and fintechs, which is where our fraud prevention and deepfake detection actually get tested. In production remote onboarding, the injected sessions we see are rarely exotic. They are ordinary applications running against a virtual camera, and they are caught at the capture layer rather than by looking harder at the face.

What an independent injection evaluation looks like

BixeLab put HyperVerge’s HyperKYC SDK through CEN/TS 18099 at the Substantial level, which is Level 2, between 1 July and 10 September 2026. Four injection methods, 10 attack instruments, face swaps and deepfake videos and avatar reenactment among them.

All 630 injected transactions were rejected. None of them got through.

The evaluators worked the way a motivated attacker does. They rooted a device and concealed the root, decompiled and hooked our SDK to intercept the camera feed, then pushed deepfake and face-swap video in through an OBS virtual camera on a Mac, HDMI capture on Windows and an iOS virtual-camera module.

“BixeLab’s evaluators approached our system the way a well-resourced attacker would.” Hariprasad PS, Head of AI, HyperVerge

Root detection was partially bypassed during the work. The application still never captured the virtual-camera frames, which is the outcome that decides whether an injected session reaches a decision.

Across 300 genuine runs the bona fide error rate was 1.33%, on Android SDK 2.12.1, iOS 1.6.1 and Web 11.0.0.

That result makes HyperVerge the first identity verification provider in India to clear independent injection testing at this level. It also puts us in a group you can still count on your fingers: systems anywhere that have been measured against both this specification and the presentation attack standard, and held up under both.

BixeLab records the injection outcome as an evaluation rather than an accredited certification, which is what the specification supports today. A result this good against 10 attack instruments is still not a promise about the eleventh. Nobody can make that promise, and a provider who does is telling you something about their marketing rather than their engineering.

Keeping it true after the result

A certificate names a product, a version and a date. Shipping software moves past all three.

“BixeLab’s evaluators approached our system the way a well-resourced attacker would.” Hariprasad PS, Head of AI, HyperVerge

Worth asking any provider for a re-test cadence and pinning the tested version into the contract. A buyer who only checks the date on the certificate is still ahead of most.

Why so few systems have passed this

Count the identity verification systems that have been independently evaluated against CEN/TS 18099 at Substantial level or above and you run out of names quickly. That is not because the industry is uninterested. It is because the entry requirements are steep in a way that a certification list never shows.

The specification was published in 2024. Accredited laboratories offering the testing are few and recent, and you can’t just book a slot and hope. An evaluation of this shape exposes architecture, not tuning, so a system that treats its own camera input as trustworthy has nothing to submit.

Ten weeks of laboratory time, in our case, and that was only the stretch with an evaluator attached to it.

The injection evaluation ran from 1 July to 10 September 2026. BixeLab spent it doing what a determined attacker does with unlimited patience and no deadline: rooting a device and hiding the root, pulling our SDK apart, hooking the functions that handle the camera feed, and pushing synthetic video at us from three directions at once.

The attack surface is wider than most teams plan for

Four injection methods and 10 attack instruments went into the test plan. Face swaps, deepfake videos and avatar reenactment, delivered through an OBS virtual camera on a Mac, HDMI capture on Windows, and a virtual-camera module on iOS.

Three delivery routes, three operating systems, and an evaluator free to combine them.

Teams preparing for this tend to build against the attack they have already seen in production, which is almost always one route on one platform, usually the one that showed up in a fraud review last quarter and got someone’s attention. That’s not what arrives.

Three platforms, not one

Android SDK 2.12.1, iOS 1.6.1 and Web 11.0.0 were all in scope. The presentation evaluation covered Android SDK 2.10.0 and iOS 1.6.1 on Samsung Galaxy S25 and iPhone 13 hardware.

Platform coverage is where preparation for a face verification evaluation usually comes apart. Root and emulator detection behave differently on Android than on iOS, and the web has neither to lean on, so a capture-integrity story that holds on mobile can have nothing underneath it in a browser.

Where HyperVerge stands on this

We’re proud of that result, and of the engineering choice underneath it: checking where a frame came from before asking whether the face in it is real. It makes HyperVerge the first identity verification provider in India to clear independent injection testing at the Substantial level. The details are in the announcement.

So when you put the three questions from earlier to your shortlist, put them to us as well. We have the laboratory, the level, the methods in scope and the tested versions ready to share.

FAQs

What is CEN/TS 18099?

 

CEN/TS 18099 is a European technical specification, published by the European Committee for Standardization in 2024, that defines testing methods and requirements for biometric data injection attack detection. It grades a system’s resistance to injected synthetic data across three attack levels and is applied by accredited testing laboratories.

What is the difference between a presentation attack and an injection attack?

 

A presentation attack shows a physical artifact, such as a printed photo or a mask, to the capture device. An injection attack bypasses the device entirely and inserts synthetic data directly into the pipeline through a virtual camera, emulator, hooked function or API. Different surfaces, different standards.

Does ISO/IEC 30107-3 cover injection attacks?

 

No. ISO/IEC 30107-3:2023 governs presentation attack detection, where an artifact is presented to a capture device. Its published scope places overall system-level security and vulnerability assessment outside its remit. A presentation attack certificate therefore carries no information about injection resistance.

What are the three CEN/TS 18099 levels?

 

Level 1, basic, covers replayed video, screen recordings and browser-based signal replay. Level 2, intermediate or substantial, covers deepfake injection, synthetic biometric input and signal manipulation. Level 3, advanced or high, covers custom-crafted injections, API tampering and virtual sensor spoofing.

Who needs CEN/TS 18099 certification?

 

Facial recognition and age estimation providers, KYC and onboarding platforms, biometric SDK suppliers, identity providers integrating age verification, and any system processing biometric data remotely or through an API. Germany’s BSI requires accredited testing for regulated identity proofing before the end of 2026.

Does a deepfake detection claim cover injection attacks?

 

No. Deepfake detection works out whether the media is synthetic. Injection attack detection works out whether the feed carrying it was real. Plenty of players do one of those well and have never tested the other, so ask which attack methods were in scope and who ran the evaluation. Talk to our team about how HyperVerge’s deepfake and injection defenses hold up at the capture layer.

Nupura Ughade

Nupura Ughade

Content Marketing Lead

LinedIn
With a strong background B2B tech marketing, Nupura brings a dynamic blend of creativity and expertise. She enjoys crafting engaging narratives for HyperVerge's global customer onboarding platform.

Related Blogs

CEN/TS 18099: What Injection Attack Testing Proves

CEN/TS 18099 is the European technical specification for biometric injection attack detection....

ISO/IEC 30107-3: What a PAD Certificate Really Says

ISO/IEC 30107-3 is the international standard for testing biometric presentation attack detection....

DPDP Act and Rules: A Compliance Guide for Regulated Entities

The Digital Personal Data Protection Act, 2023 (DPDP Act) is India’s law...