/
/
Deepfakes and Injection Attacks: How Remote Identity Verification Is Defeated, and Defended
Deepfakes and Injection Attacks: How Remote Identity Verification Is Defeated, and Defended

Remote identity verification relies on a simple premise: a real, live person is in front of the camera. Deepfakes break that premise. But the deepfake itself is only half the problem.
What decides whether an attack succeeds is how the fake is delivered, and there are two very different methods: presentation attacks and injection attacks.
Liveness detection has long handled the first. The second, injection attacks, bypasses the camera entirely and is now the harder, faster-growing threat. This article explains the difference, how injection attacks are detected, and why the new CEN/TS 18099 standard matters.
The deepfake is the content, not the attack

It helps to separate two things that are often blurred together. A deepfake is synthetic media, a face, a video, or a voice generated or manipulated by AI. The attack is the method used to feed that media into an identity-verification system.
The same deepfake can be delivered in completely different ways, and the defence that stops one does not necessarily stop the other. This also connects deepfakes to synthetic identity fraud, where fabricated identities, not just fabricated faces, are used to open accounts.
Presentation attacks vs injection attacks
There are two ways to get a fake past a camera-based check. Understanding the distinction is the single most important thing for anyone evaluating identity-verification security.

Presentation and injection are two ways to deliver the same deepfake, with different defences.
Presentation attacks
A presentation attack shows a fake artefact to the genuine camera: a printed photo, a mask, or a deepfake video played on a screen held up to the device. Because the attack still goes through the camera, it can be caught by liveness detection, also called presentation attack detection (PAD), which looks for the signs that what the camera sees is not a real, present person.
Injection attacks
An injection attack does not use the camera at all. Instead, the attacker inserts synthetic media directly into the data stream the verification system receives, using virtual cameras, emulators, manipulated mobile apps, or intercepted network traffic and APIs. The system is handed a perfect deepfake as if it had come from a real camera. Because nothing was ever shown to a physical lens, liveness detection alone may see a flawless face and pass it. This is why injection attacks are considered more dangerous, and why they are rising.
Aspect | Presentation attack | Injection attack |
Delivery | A fake is shown to the real camera | Synthetic media is injected into the data stream |
Typical method | Printed photo, mask, or screen replay | Virtual camera, emulator, manipulated app, API or network interception |
Reaches the camera? | Yes | No, it bypasses the camera |
Primary defence | Liveness / presentation attack detection (PAD) | Injection attack detection (IAD) |
Relative threat | Established and well mitigated | Newer, harder to detect, and growing |
Why injection attacks are surging

Two trends are pushing injection attacks up the threat list. First, generative AI has made convincing deepfakes cheap, fast, and accessible, so the quality of the fake is no longer the bottleneck. Second, the tooling to inject that fake, virtual cameras, emulators, and app-manipulation kits, is increasingly packaged and shared, lowering the skill needed to run an attack at scale.
The result is a shift from one-off, physical spoofing toward automated, pipeline-level attacks that can be repeated against many onboarding flows at once.
For an attacker, injection offers a better payoff: a flawless face delivered straight to the system, with no camera artefacts to give it away. That combination, better fakes and easier delivery, is why injection attack detection has moved from a nice-to-have to a core requirement.
Why liveness detection is no longer enough
For years, defeating facial recognition and biometric verification meant fooling the camera, and PAD was the answer. Injection attacks change the threat model: the adversary skips the camera and targets the pipeline.
A system that only does presentation attack detection can be strong against someone holding up a screen and still be blind to a deepfake injected through a virtual camera. Closing that gap needs a second, distinct capability: injection attack detection.
How injection attacks are detected

Injection attack detection (IAD) focuses on the integrity of the capture path rather than only the image. The common techniques include:
Virtual-camera and emulator detection. Identifying when the feed comes from a software camera or emulated device rather than genuine hardware.
Capture and stream integrity. Checking that the media reached the system through a legitimate, unbroken capture path, not injected mid-stream.
Device and signal analysis. Examining metadata, timing, and hardware signals that genuine camera capture produces and injected media often lacks.
Tamper and interception checks. Detecting manipulated apps, hooked APIs, or intercepted network traffic used to insert content.
In practice, robust remote identity verification runs both PAD and IAD together, so that an attacker cannot simply switch delivery method to get through.
Where injection attacks hit hardest
Injection attacks are most attractive where remote onboarding unlocks something valuable and the attack can be automated.
That makes high-assurance, high-volume flows the prime targets: opening bank and payment accounts, onboarding to crypto exchanges, and any process that grants access, credit, or the ability to move money.
The same technique scales to synthetic identity fraud, where one toolkit can spin up many fabricated identities and push each through onboarding with an injected deepfake.
The common thread is that the attacker never has to be physically present, so distance, volume, and repeatability all work in their favour, exactly the opposite of the assumptions behind in-person identity checks.
The standard that changed the conversation: CEN/TS 18099

Until recently, injection attack detection was hard for buyers to assess, because there was no common yardstick. CEN/TS 18099 is a European technical specification that provides a framework for evaluating injection attack detection, giving the market a shared reference for what good looks like.
Alongside requirements such as ETSI TS 119 461 for identity proofing under the EU's trust-services framework, it means a vendor's resistance to injection attacks can now be tested against an external standard rather than asserted in a datasheet.
For regulated firms, this matters at procurement time. A liveness claim is no longer sufficient on its own; the right questions are whether a provider performs both presentation and injection attack detection, and whether that detection has been evaluated against a recognised standard.
What this means for regulated onboarding
As the EU moves toward higher-assurance remote identity under eIDAS 2.0 and the Digital Identity Wallet, the bar for remote identity verification is rising, and so is the sophistication of the attacks against it.
Firms that onboard customers remotely, banks, fintechs, insurers, crypto platforms, need to treat injection attacks as a first-class threat, not a footnote to deepfakes.
The practical takeaways are simple: assume deepfakes will be thrown at your onboarding, require both PAD and IAD, and prefer providers whose detection has been independently evaluated.
Questions to ask an identity verification provider

Because the difference between presentation and injection defence is invisible in a demo, it pays to ask directly:
Do you perform both presentation attack detection and injection attack detection, or only liveness?
Has your injection attack detection been evaluated against CEN/TS 18099 or an equivalent standard?
How do you detect virtual cameras, emulators, and manipulated capture paths?
Do you assess the integrity of the capture pipeline, not just the quality of the image?
How is your detection updated as new injection techniques appear?
A provider that can answer these concretely, and point to external evaluation, is managing the threat. One that falls back on a general liveness claim is not.
How Qoobiss addresses injection attacks

Qoobiss builds presentation and injection attack detection into ONTRACE, its remote identity verification product, so that a deepfake is caught whether it is shown to the camera or injected into the stream.
ONTRACE has been put forward for evaluation of its presentation and injection attack detection under CEN/TS 18099 and ETSI TS 119 461, as part of readiness for eIDAS 2.0, which you can read about in our newsroom.
The point is not that deepfakes can be ignored; it is that they can be delivered two ways, and that a verification system has to be tested against both, to a standard, rather than trusted on a claim.
Frequently asked questions
What is an injection attack in identity verification?
What is the difference between presentation and injection attacks?
Can liveness detection stop deepfakes?
What is CEN/TS 18099?









