deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Injection attacks bypass liveness checks by feeding KYC apps a stream that skips the camera

Remote onboarding liveness checks certify against printouts and masks, not synthetic streams injected via virtual cameras. Attack volume is rising fast and regulators have begun requiring anti-injection controls.

Injection attacks bypass liveness checks by feeding KYC apps a stream that skips the camera

The camera that never saw the face

Remote identity verification has a blind spot: a liveness check can only analyse the pixels that reach it, and attackers increasingly arrange for those pixels to arrive without ever passing through a real lens. That is the central argument of a technical post on dev.to aimed at teams building remote onboarding and know-your-customer flows.

Two attacks that keep getting confused

The post draws a firm line between two attack classes that are routinely conflated. A presentation attack places a physical object in front of the lens — a printed photo, a screen playing a video, a silicone mask, a 3D head model — and the camera faithfully records what it sees. Conventional liveness defences are built for exactly this: skin texture, micro-expressions, scene depth and lighting changes betray a printout, glare and moiré give away a screen, and unnatural motion exposes a mask.

An injection attack skips the lens entirely. The attacker hands the application a synthetic stream through a virtual camera, an emulator or a rooted phone. Because modern generators produce a face that blinks, reacts to light and follows a turn-your-head command in real time, every signal the liveness algorithm depends on is present — it has just been manufactured.

The certification gap follows directly from this distinction. According to the post, PAD certification under ISO/IEC 30107-3 covers presentation attacks only, by definition, so a fully certified system can remain blind to injection.

Growth in attacks, weakness in detection

The post cites industry figures showing injections growing fastest: iProov reported a 1,151% year-on-year rise in injection attacks on iOS in the second half of 2025, while Entrust reported 40% year-on-year growth and found a deepfake behind one in five biometric fraud attempts.

On the defensive side, the post points to an October 2025 benchmark from Purdue University that tested detectors against real deepfakes pulled from social media rather than lab datasets. The best academic and government systems reached at most 74.78% AUC on images and 73.67% on video, generalised poorly to new data and were vulnerable to simple manipulations, especially on video. The post adds caveats: the study covered political clips, not onboarding verification, which is a different domain, and paid commercial detectors beat free ones while still failing to carry their results over to real-world footage.

The post also debunks a figure circulating on Russian-language sites, the claim that 67% of verification systems are vulnerable to synthetic faces. That number appears nowhere in the Purdue work; it traces back to a vendor's sponsored article, and the study itself concerns deepfake detectors rather than verification systems.

A regulator that already requires anti-injection controls

Kazakhstan, where 80% of bank accounts are opened online according to Kursiv Media as cited in the post, has moved to close the gap. A joint resolution by the ARDFM and the National Bank, No. 39186, in force since 12 July 2026, explicitly requires biometric authentication software to implement self-integrity control, emulator detection and protection against substitution of the built-in camera.

The backdrop includes a January 2025 demonstration at the Digital Almaty forum, where the head of the National Bank's National Payment Corporation bypassed biometric identification with a printed face and a mask, followed by a string of fraud campaigns: fake regulator officials, a deepfake of a Financial Monitoring Agency spokesperson, a fabricated bank-CEO endorsement of an investment platform, and a June 2026 case in which one victim lost US$1,200. The post separates these from the machine-facing threat: most documented cases, like Arup's roughly US$25.6 million loss in Hong Kong in January 2024, are social engineering against humans rather than defeats of automated checks. No publicly documented case in Kazakhstan has yet shown a deepfake passing bank biometrics and ending in a conviction — though cybercrime damage reached KZT 16.4 billion in 2025 according to the Interior Ministry's cybercrime department, and deepfake incidents are not counted as a separate line in official statistics, so absence of evidence is not evidence of absence.

Why it matters

For anyone building KYC or remote onboarding, the practical takeaway is that a PAD certificate is necessary but not sufficient: it says nothing about whether the video stream actually originated at a physical camera. Vendor evaluations should cover stream-integrity controls — emulator and virtual-camera detection, device attestation, self-integrity checks — rather than detection accuracy alone. The Purdue numbers suggest content analysis on its own is a fragile defence, and Kazakhstan's resolution offers an early template for the explicit anti-injection requirements other regulators may soon copy.

  • #biometrics
  • #kyc
  • #security
  • #deepfakes
  • #identity-verification

Related posts