Every remote verification flow asks a stranger to hold a document next to their face and perform for the camera. The control that guards that moment, liveness detection, was built to stop someone holding up a photograph or a screen. The attacks that beat video KYC today do not present anything to the camera at all. They replace the camera. This page follows that distinction through the anatomy of an attack, the way fraud industrialized it, and the checks that answer it, the same path DuckDuckGoose's reporting walks, checked here against the 15 KYC tools in our own directory.

The threat liveness was not built to stop

A presentation attack happens in front of the sensor: a printed face, a replayed video, a mask. An injection attack happens behind it: a virtual camera, an emulator, or a hooked app feed pipes synthetic video directly into the verification stream, and the liveness model dutifully analyzes a person who is not there. Certification has not caught up with the distinction. In our directory, 18 of the 25 tools that name any benchmark name iBeta presentation-attack (PAD) certification under ISO/IEC 30107-3, a standard that tests what is presented to a device. It says nothing about whether the feed itself is genuine. A flow can be fully certified and fully open to injection at the same time.

The anatomy of an attack

The ingredients are commodity. A harvested face (scraped photos, a breached selfie database, or a short social video) feeds a real-time face-swap that maps it onto an operator or a fully synthetic performer. A document is generated or stolen to match. The swapped stream is injected through a virtual camera into a browser session or an emulated phone, the liveness challenge is performed by the swap in real time, blink, turn, smile, and the account opens in a name that belongs to someone else. Gig platforms meet the same attack at a second door: driver and courier accounts are verified once, then rented or resold, which is why the defense guide treats re-verification as a layer of its own.

When fraud becomes a product

The DeepfakeDetect app interface, whose detector layer scores captured media for synthesis artifacts inside a verification flow
The certification count is from the DeepfakeDetect directory: 18 of the 25 tools that name any benchmark name iBeta presentation-attack certification, which tests presentation attacks, not injected feeds.

The operator above is increasingly a customer, not a craftsman. Verification-bypass services sell the whole pipeline: matched document and face sets, injection tooling, and coaching through the challenge, priced per successful account. Fake-ID production has the same shape, generated documents sold in volume with the selfie to match. And organized account-opening rings run the play across banks, wallets and gig platforms in sequence, because a flow defeated once is documented and repeated until it is fixed. The patterns matter more than any single incident count: each is a service with suppliers, which means the defense has to raise the cost of the service, not catch one talented attacker.

What actually works

Feed integrity before face analysis. Attest the device and the camera path first; a detector downstream of an injected feed is grading the attacker's homework. Injection detection as its own layer. Signals of virtual cameras, emulators and tampered streams are detectable, but only by checks designed for them, not by the liveness model. A detector inside the flow. A deepfake model scoring the captured media for synthesis artifacts, the layer our own DeepfakeDetect provides for image, video, audio and text, catches what the swap leaves behind even when the performance is perfect. Document forensics beside the face check. The document and the face arrive through the same pipe; checking only one invites the matched-set attack. Re-verification after onboarding. Account renting defeats any one-time check; periodic and event-triggered re-checks are what make a rented account a wasting asset.

The question every verification flow should answer

Before the next audit or the next incident: if an attacker never shows a real camera anything, which control in your flow notices? If the honest answer is none, the liveness certificate on the wall is answering a question from the previous generation of attacks. The buying side of fixing it, vendors, pricing and evaluation, is in the KYC buyer's guide; the layered design is in the five-layer guide.