PXL Vision PXL Deepfake Detector
Get API access Request a demo
AI-generated fictional person: a man in a dark t-shirt looking calmly into the camera

PXL Deepfake Detector

This person wants to open an account.

He isn't real. He's a deepfake. The PXL Deepfake Detector catches him at onboarding.

Request a demo

This person does not exist.

Synthetic Face98.7 % confidenceFlagged by the PXL Deepfake Detector

Scroll to run the detector

Anyone can make deepfakes now, at scale.

Not too long ago this took a skilled operator and a workstation. Today it takes a phone and a subscription.

+704 %face-swap attacks in remote identity verification, H1 to H2 2023iProov 2024 Threat Intelligence Report
+58 %deepfaked selfies in 2025Entrust 2026 Identity Fraud Report
5 minbetween detected deepfake attempts in 2024Entrust 2025 Identity Fraud Report
2 to 4xyear-on-year growth in deepfake attempts across PXL Vision's own onboarding trafficPXL Vision, 2026

PXL Deepfake Detector

You cannot see it.
We can.

A trained human reviewer can sometimes spot basic Deepfakes. Our Deepfake Detector sees far more than a human and also catches the attacks with no visible tell at all.

Request a demo Get API access

Customers who trust PXL Vision

The record

What we can show.

99.5 %+genuine selfies pass through
99 %+detection rate on our benchmark
60+AI generation tools covered
1.8 M+attack samples benchmarked

  • Catches what it has never seen

    Built to generalise to generators it was never trained on, which matters when a new tool launches every month.

  • Rapid retraining

    New generators are added continuously. Detection is an arms race and the cadence is the weapon.

  • Developed with Idiap

    Built in collaboration with Idiap Research Institute on peer-reviewed biometric research, not on a scraped media corpus.

  • Aligned to eIDAS 2.0

    ETSI TS 119 461 and CEN/TS 18099, the standards your auditors will be asking about.

  • Native to the PXL platform

    One switch for existing customers. REST API or SDK for everyone else. Flagged cases can route to manual review.

  • Hosted in Switzerland and the EU

    Processed and stored in data centres in Switzerland and the EU, under Swiss data protection law and the GDPR. What is kept, and for how long, is set in your data processing agreement.

Four ways a face can lie.

Four fictional persons, four real manipulations. Made in an afternoon with tools anyone can subscribe to.

AI-generated fictional person: a young man with curly hair, the original selfie before the face swap AI-generated fictional person: face swap, the same head and hair with an older man's face swapped in
UnmanipulatedFace swap

Case 01 · Face swap

One person's face on another person's head.

What a human reviewer might notice: a seam at the hairline and the jaw, and a face thirty years older than the hair.

Drag the handle to compare

AI-generated fictional person: face synthesis, a man with a perfectly symmetric face and flawless skin, generated from nothing
Move over the face to magnify

Case 02 · Face synthesis

A face that never existed, generated synthetically.

What a human reviewer might notice: mirror symmetry, skin without a single pore, a smile too even to be a smile.

Move over the face to magnify

AI-generated fictional person: a young woman with dark hair, the original selfie before the age edit AI-generated fictional person: face editing, the same woman aged thirty years
Age 28ScrollAge 58

Case 03 · Face editing

The same person, thirty years later. A real face altered to age 30 years.

What a human reviewer might notice: texture that changes around the eyes and the mouth, and nowhere else.

Scroll to age the face

Case 04 · Face reenactment

A real face image, turned into a puppeteered character with expressions it never made.

What a human reviewer might notice: a mouth and eyes that move out of sync with the head.

AI-generated fictional person: a woman inside an onboarding camera frame, face-swap manipulation, the scan sweeping her face before the verdict Synthetic Verify your identity

PXL Deepfake Detector at work

Every selfie, checked inside your flow.

The user takes a selfie as usual. The verdict arrives with the confidence and the suspected manipulation type, before the damage happens.

Verdict
Synthetic
Confidence
98.7 %
Manipulation
Face swap
Likely generator beta
Diffusion face-swap family

The PXL Deepfake Detector, built for onboarding and verification selfies.

An onboarding selfie is a controlled scene: one face, camera-facing, front-lit, captured live. General-purpose media detectors are built for anything, so they drift on this job. The PXL Deepfake Detector is tuned for that one scene.

It reads the captured image itself, so it makes no difference whether a manipulated face arrived through the phone camera, a virtual camera or an injected stream. Injection attack detection runs as its own layer in PXL Ident. And it checks the selfie, not the portrait printed on the document: morphing attacks on document photos are a separate check.

Where it fits

  1. ProblemA synthetic identity arrives as a plausible document and a face reenacted or generated to match it. A real person sits in front of the camera, so liveness passes.

    DamageAn account that belongs to no real customer. It turns into a mule account, draws credit that is never repaid and leaves a KYC file the regulator can pick apart.

    SolutionThe detector checks the selfie frame inside the onboarding flow, next to liveness and document verification, and a synthetic verdict blocks or routes the case before the account exists. For EU providers this is the step that eIDAS 2.0 and ETSI TS 119 461 expect remote identity proofing to defend.

  2. ProblemA stolen or bought document paired with a swapped face is enough to activate, replace or port a SIM.

    DamageThe SIM is the doorway to one-time codes for banking and email. One fraudulent activation turns into account takeovers elsewhere, and the telco carries the disputes and the financed handset that is never paid off.

    SolutionThe detector checks the selfie at activation, replacement and porting. Telcos on PXL Ident switch it on without integration work.

  3. ProblemAccount takeover increasingly starts with a reenacted face, driven from a profile photo or a scraped video, presented at a step-up check: a password reset, a new device, a large transfer.

    DamageAn account that passed every check at onboarding is emptied under the genuine customer's name, and the dispute lands on your desk.

    SolutionThe detector asks whether the face in front of the camera has been manipulated at all. Face matching only asks whether it is the same person. Together they close the gap either one leaves open.

  4. ProblemMarketplaces, gig platforms and dating apps meet synthetic faces at scale: profiles built around a person who never existed.

    DamageReal users get scammed by accounts that carry your verified badge, and a ring can run hundreds of them before anyone notices.

    SolutionThe detector runs at profile verification and again on a schedule. A face later found in a ring can be traced to every account it fronted.

  5. ProblemImpostor employees pass remote identity checks with a candidate's selfie swapped to match a purchased identity.

    DamageSomeone you never vetted holds a company laptop and access to payroll and customer data from day one.

    SolutionA deepfake check at the pre-employment identity step, before background checks are ordered, keeps the manipulated face from ever entering the record.

  6. ProblemInsurers see aged or edited faces reused across policies, and synthetic policyholders who exist only as a document and a face.

    DamageClaims paid to people who do not exist and policies written on invented lives, in a book of business you cannot trust when the audit comes.

    SolutionWhere a claim or a policy change carries a selfie step, the detector flags the manipulated ones and routes them to manual review instead of straight-through processing.

  7. ProblemFraud rings reuse the same synthetic face under new names, and each account passed onboarding on its own.

    DamageA ring keeps operating across dozens of accounts, each one legitimate in isolation, until the losses are added up.

    SolutionRunning the detector over stored selfies of high-risk accounts surfaces the repeats. Through the PXL Fraud Protection Network the pattern can be shared across providers before the next onboarding flow is hit.

Open an account
AI-generated fictional person: an older woman with glasses at the account opening step, her face puppeteered into an expression she never made Synthetic, face reenactment
Document · passedLiveness · passedDeepfake · flagged
Activate your SIM
AI-generated fictional person: a blond woman at the SIM activation step, her selfie carrying another woman's face swapped in Synthetic, face swap
Document · passedLiveness · passedDeepfake · flagged
Confirm it is you
AI-generated fictional person: a dark-haired woman at a password reset step, her selfie edited to age her twenty-five years Synthetic, face editing
Document · passedLiveness · passedDeepfake · flagged
Verify your profile
AI-generated fictional person: a young woman at a platform profile check, a face generated from nothing Synthetic, face synthesis
Document · passedLiveness · passedDeepfake · flagged
Candidate identity check
AI-generated fictional person: a bearded man in glasses at a candidate identity check, his selfie carrying another man's face Synthetic, face swap
Document · passedLiveness · passedDeepfake · flagged
Submit a claim
AI-generated fictional person: a woman with braids at a claim step, her selfie edited to age her thirty years Synthetic, face editing
Document · passedLiveness · passedDeepfake · flagged
Periodic re-check
AI-generated fictional person: a woman in a blue blouse at a periodic re-check, her selfie carrying an older woman's face Synthetic, face swap
Document · passedLiveness · passedDeepfake · flagged

Questions about the PXL Deepfake Detector.

A face image or selfie from an onboarding or monitoring step. It returns a verdict, a confidence score, the manipulation type (face swap, face synthesis, face editing or face reenactment) and, in beta, the family of generator most likely behind it. It checks the selfie, not the portrait printed on the document.

No. It is one layer in the verification flow. It works next to liveness, presentation attack detection and injection attack detection, document verification and NFC chip reading, and it is native to PXL Ident.

Every verdict comes with a confidence score, and the operating point is part of the model release, so there is no threshold for you to tune. A result the detector is not confident about does not have to end the transaction: it can go to manual review, for example PXL Check, instead of a hard block.

It is tuned for the range of conditions seen in real onboarding flows. A poor capture lowers the confidence rather than forcing a verdict, and most flows ask the person for a retake.

In Switzerland and the EU, under Swiss data protection law and the GDPR. What is kept, for how long, and whether it may be used to improve the detector is set in your data processing agreement. How PXL Vision handles personal data in general is described in the privacy policy.

Yes. There is a REST API for any stack and an SDK. PXL Ident customers switch it on without any integration work.

The model is retrained continuously and released on a cadence, and it is built to generalise to generation methods it has not seen. More than 60 generation tools are covered today.

Not yet. Presentation attack detection has a certification standard, ISO/IEC 30107-3, and injection attack detection has CEN/TS 18099, with ISO/IEC 25456 in draft. For deepfake detection itself no standard or certification scheme exists, as the 2026 market report by Biometric Update and Goode Intelligence records. That is why we publish how we measure.

No. Every face on this site is an AI-generated fictional person. No real individual is depicted.

PXL Deepfake Detector, integration and access

One call. One verdict.

Send the selfie, get the verdict, the confidence and the manipulation type. Nothing to train, no threshold to tune.

REST API for any stack
SDK or PXL Cloud
PXL Ident customers: one switch, no integration

Three ways to wire it in

In the same step as liveness and document checks, one decision at the end
As a gate in front of costlier checks and manual review
As a look back over accounts onboarded before it was switched on
Illustrative request and response
POST /v1/deepfake/analyze
image: selfie.jpg

{ "verdict": "synthetic", "confidence": 0.987,
  "manipulation": "face_swap",
  "generator_family": "diffusion_face_swap" }   // beta

Request a demo or API access