Liveness Detection and KYC Deepfake Fraud: A Defense Guide for Banks and Fintechs

---

Deepfakes now account for roughly 7 percent of all fraud attempts that identity-verification vendor Sumsub screened in 2024, a fourfold jump from 2023, per Sumsub's 2024 Identity Fraud Report. For banks and fintechs, the onboarding selfie is no longer proof that a real person is applying. Liveness detection was built to close that gap, and attackers have already found ways around it.

What is liveness detection? Liveness detection verifies that a real, present human is completing identity verification, blocking photos, replays, and masks. Modern deepfake KYC fraud increasingly bypasses the camera itself through injection attacks, so effective programs layer liveness detection with AI deepfake analysis of the submitted media.

1. Liveness checks

   Require active liveness, not just a static selfie.

2. Injection defense

   Detect virtual cameras and replayed or injected video.

3. Synthetic-media scan

   Run uploaded documents and selfies through a deepfake detector.

4. Layered review

   Escalate edge cases to human review.

What Is Liveness Detection?

Liveness detection is a biometric control that confirms the face in front of the camera belongs to a live human being present at capture, not a photo, a video replay, a 3D mask, or a synthetic rendering. It sits inside the Know Your Customer (KYC) and anti-money-laundering (AML) onboarding flow, usually paired with a document-to-selfie face match, so a bank can tie a verified government ID to the person holding the phone.

The control exists to defeat presentation attacks: any artifact shown to the camera to impersonate someone else. Printed photos, replayed videos on a second screen, silicone masks, and paper cutouts are the classic examples. The international yardstick for measuring how well a system resists these is ISO/IEC 30107-3:2023, which defines testing and reporting methods for presentation attack detection (PAD). Independent labs such as iBeta certify vendors against it, and a Level 1 or Level 2 PAD certification is the first credential a fraud team should ask any liveness vendor to produce.

Liveness was a strong answer to the problem it was designed for. The trouble is that deepfakes changed the problem.

Active vs Passive Liveness Detection

Liveness systems come in two architectures, and the choice shapes both your fraud resistance and your conversion rate. Active liveness asks the user to perform an action, such as blinking, turning their head, or following a moving dot. Passive liveness analyzes a single selfie or short capture in the background, with no instructions, often using texture, depth, and reflectance cues the user never notices.

Neither architecture is inherently safer against deepfakes. Active liveness frustrates simple photo attacks but can be satisfied by a deepfake that animates on command, and the prompts can leak to attackers who script responses. Passive liveness protects the onboarding funnel by removing steps, which is why most high-volume fintechs prefer it, but a convincing synthetic face can pass a passive check just as it passes an active one. The deciding factor is not active versus passive. It is whether the image reaching the model came through a real lens at all.

How Deepfakes Break Traditional KYC

Every camera-based liveness check rests on one assumption: that the pixels it scores were captured by the device's real camera, of a real scene, in real time. Deepfake KYC fraud attacks that assumption from two directions.

Presentation attacks: screens, prints, and masks

Presentation attacks are the original threat, and they happen at the sensor. A fraudster holds up a printed photo, replays a video of the victim on a phone screen, or wears a fabricated mask. Modern PAD models, especially passive ones that read texture and depth, stop most of these. This is the attack liveness handles well, and it is the success story vendors lead with. It is also no longer where the money is being lost.

Injection attacks: virtual cameras and emulators that skip the lens

Injection attacks bypass the camera entirely. Instead of showing something to the lens, the attacker feeds a prepared video stream straight into the verification app, using a virtual camera driver, an emulator, or a hooked mobile app. The liveness model receives a clean, frontal, well-lit deepfake that never passed through any real sensor, so the texture and reflectance cues that catch printed photos simply are not there to fail. This is the failure mode IDV vendors gloss over, because liveness alone cannot see it.

The scale is no longer theoretical. Identity-verification firm iProov reported that injection attacks targeting iOS surged 1,151 percent in the second half of 2025, contributing to a 741 percent annual increase, as covered by Biometric Update (2026). Earlier iProov research documented a 704 percent rise in face-swap attacks across the second half of 2023. The techniques have industrialized into repeatable kits sold to non-experts.

Deepfake selfies and synthetic document photos

The media itself keeps getting cheaper and better. A face-swap or fully synthetic selfie can satisfy a document-to-face match when the document photo is also tampered or AI-generated. Synthetic identity fraud, where an attacker stitches a real document number to a fabricated face, compounds the problem: there may be no genuine victim to flag, and the application can clear sanctions and credit checks. Liveness confirms presence; it does not confirm authenticity of the pixels.

KYC Fraud by the Numbers

The trend lines all point the same way. Treat each figure as the publishing vendor's own measurement, because fraud statistics are marketing artifacts as much as research, and weigh them accordingly.

• Deepfakes made up about 7 percent of fraud attempts in 2024, a 4x increase year over year, per Sumsub's 2024 Identity Fraud Report.
• Global deepfake incidents surged tenfold from 2022 to 2023, per Sumsub research (2023).
• Biometric injection attacks on iOS rose 1,151 percent in H2 2025, per iProov via Biometric Update (2026).
• The U.S. Treasury's FinCEN issued a 2024 alert on deepfake media used to open accounts and defeat verification, summarized by iProov.

For the full cross-industry dataset, see our deepfake statistics roundup.

The Layered Defense: Liveness Plus Deepfake Detection

If the camera can lie, no single control is enough. The programs holding up against injection attacks treat onboarding as a pipeline with independent checks at each stage, so a deepfake that defeats one layer still has to survive the next.

Layer 1: Liveness at capture

Keep a PAD-certified liveness vendor at the point of capture. It remains your best defense against the high-volume presentation attacks, printed photos, and screen replays that still make up a large share of attempts. Verify the certification level and ask specifically about the vendor's injection-attack and virtual-camera defenses, which are a separate problem from classic PAD.

Layer 2: Media forensics on every submitted frame

This is the layer that addresses injection. Independent of where the media came from, analyze the submitted selfie, video, or document image for the statistical fingerprints of AI generation and manipulation. Because media forensics scores the pixels rather than trusting the capture path, it catches synthetic frames that were injected past the lens and never triggered a sensor-level liveness flag. This is where a deepfake detection API fits.

Our API analyzes video, image, and audio for AI-generated and manipulated content with 95 percent accuracy, returning a verdict of Authentic, Likely Synthetic, or Inconclusive with a TrustScore from 0 to 100. It detects output from current image and video generators such as Midjourney, DALL-E, Stable Diffusion, Sora, Runway, Pika, and Kling. Important honesty note for evaluators: we are a post-hoc media-forensics layer, not a liveness or onboarding SDK. We do not run the camera-side capture or the active-prompt challenge. You keep your liveness vendor and add our analysis as the independent check on the media that vendor passes through.

Layer 3: Document and metadata consistency checks

Cross-check the document against the selfie, the stated identity against external data, and the file metadata against the claimed device and capture path. Inconsistent metadata, impossible timestamps, or emulator signatures are strong injection-attack tells that complement pixel-level forensics.

Where an API fits in your onboarding pipeline

Insert media forensics between capture and decision: the user submits, liveness scores presence, the deepfake detection API scores the media itself, and your decision engine combines both signals before approving, declining, or routing to manual review. Files are deleted from primary storage within 60 seconds of analysis unless you opt into retention, which matters when the media is regulated identity data. For data residency and confidentiality, that default purge is often the difference between a usable vendor and a compliance blocker.

CTA: See how the DeepfakeDetector.ai deepfake detection API adds media forensics to your onboarding flow. Plans from $49 to $599 per month, or create a free account and start with 50 detections.

Evaluating Deepfake Detection for KYC: 7 Questions to Ask Vendors

Liveness vendors and media-forensics vendors solve different halves of the problem. When you assess the forensics layer, these seven questions separate marketing from capability.

1. Injection-attack media. How does accuracy hold on clean, frontal, injected deepfakes that never passed a real sensor, versus easy presentation attacks?
2. Latency. What is per-request scoring time at your onboarding volume? Sub-second keeps the funnel intact.
3. Modality coverage. Image only, or image plus video and audio? Voice onboarding and video KYC need all three.
4. Testing and standards. What independent benchmarks or certifications back the accuracy claim, and is the figure from production or a curated test set?
5. Data retention. Where is identity data stored, for how long, and can you enforce immediate purge for residency and confidentiality?
6. Integration effort. REST API, SDK, or both? What is realistic time to a working pilot inside your decision engine?
7. Pricing model. Per-detection, tiered, or seat-based, and does it scale to your monthly application volume?

FAQ

What is the difference between liveness detection and deepfake detection? Liveness detection checks that a real, present person is at the camera during capture. Deepfake detection analyzes the media file itself for signs of AI generation or manipulation, regardless of how it reached the system. Injection attacks, which feed synthetic video past the camera, defeat liveness alone and require the media-forensics check.

Can deepfakes pass liveness checks? Yes. Documented injection attacks bypass the camera entirely and feed a clean deepfake to the liveness model, which has no sensor-level cues left to fail on. iProov reported triple-digit annual growth in such attacks. Layered controls, liveness plus media forensics plus metadata checks, are what close the gap.

What is KYC fraud? KYC fraud is the use of false, stolen, or synthetic identities to pass a financial institution's onboarding and identity-verification checks. Increasingly it relies on deepfake selfies, AI-generated documents, and synthetic identities that stitch real credentials to fabricated faces.

Is liveness detection required by regulation? Most AML rules do not name liveness detection or deepfake detection specifically. Regulators expect effective customer identification and risk-based controls. The FATF Guidance on Digital Identity (2020) sets out supervisory expectations for reliable, independent digital ID, and U.S. FinCEN has alerted institutions to deepfake-enabled account fraud. Treat these as expectations to meet, not a checklist mandate.

How is deepfake detection integrated into onboarding? Through an API placed between capture and the approval decision. The user submits their selfie or document, liveness scores presence, the deepfake detection API scores the submitted media, and your decision engine combines both before approving or routing to review. Sub-second scoring keeps onboarding conversion intact.

Conclusion: Assume the Camera Can Lie

Liveness detection still earns its place at the front of every KYC flow, but the assumption it rests on, that the camera saw what it scored, no longer holds. Injection attacks feed clean deepfakes straight past the lens, and that is where camera-based liveness goes blind. The fix is architectural: keep PAD-certified liveness at capture, add independent media forensics on every submitted frame, and reconcile documents and metadata before you decide. Score the pixels, not just the presence. If your onboarding pipeline trusts the camera, build the layer that does not.

CTA: Add media forensics to your KYC flow with the DeepfakeDetector.ai deepfake detection API, available from the $49 Starter plan, or create a free account to test it with 50 detections.

Related reading