IDENTITY VERIFICATION + KYC

Know your customer — even when face, voice, and ID are AI-generated.

Contact usGet started
VOICE FRAUD

TTS clones bypass biometric checks

Open-source models generate studio-quality voice at near-zero cost. Standard biometric tools weren't trained on this generation of TTS.

FACE SWAP

Face-swap defeats liveness detection

Real-time face-swap runs in-browser. Liveness checks relying on movement or blink are no longer a sufficient signal.

DOCUMENT FORGERY

AI-generated IDs pass OCR checks

Inpainting produces ID photos with correct fonts, lighting, and metadata. OCR checks evaluate the text — not whether the image is synthetic.

Trusted by
mOdels shipping daily

A convincing synthetic identity used to take skill and money. Not anymore.

Open-source TTS clones voices in seconds. Face-swap runs on consumer hardware. Generative models produce ID photos that clear verification checks. These aren't sophisticated attacks — they're free tools on a laptop, getting cheaper with every model release.
voice
Biometric checks weren't trained on this generation of TTS.
Open-source models generate audio indistinguishable from authentic speech at zero marginal cost per sample.
face
Liveness checks that rely on blink or motion are no longer enough.
Real-time face-swap runs in-browser. The attacker doesn't need a photo — they need a webcam.
documents
Inpainting produces IDs with correct fonts, lighting, and metadata.
OCR checks evaluate the text. They don't evaluate whether the image is AI-generated.
THE PROBLEM

AI identity fraud, detection gaps, and how Resemble AI closes them

KYC was built on one assumption: matching face, voice, and document means a real person. Generative models now produce all three — as a coordinated matched set — for zero marginal cost. Here's how Resemble detects them across every vertical.

id verification + kyc/aml
Banking + fintech
Attack vectors:
Synthetic voice
AI face swap
Fabricated ID documents
Coordinated identity packages
Modalities
Audio
Video
Image
Products
Resemble Detect
Resemble Intelligence
Without detection
A single pipeline produces a matching face, voice, and document in under a minute — all three pass their respective checks independently
Synthetic identities clear liveness, biometric, and OCR simultaneously — no single check sees the coordinated attack
Standard KYC tools were evaluated on human fraud — not built to detect AI-generated input at zero marginal cost
How Resemble AI detects it
Multimodal analysis across voice, face, and document in a single API call — cross-modal consistency itself flags coordinated synthetic identities
Resemble Intelligence explains the verdict in human-readable form — abnormalities, confidence reasoning, and audit trail for compliance review
On-premise deployment — PII never leaves your environment, meeting DORA, HIPAA, and data residency requirements
ID FRAUD • ACCOUNT TAKEOVER
Telecom + marketplace
Attack vectors
Cloned voice verification
Deepfake selfie
Synthetic ID documents
Modalities
Audio
Image
Products
Resemble Detect
Resemble Intelligence
Without detection
Voice clones defeat voice biometric authentication at SIM registration and account creation — indistinguishable from the account holder
Synthetic selfies and deepfake faces pass profile photo and liveness verification on marketplace platforms
AI-generated IDs with correct fonts, metadata, and lighting pass document review and OCR checks
How Resemble AI detects it
Audio analysis detects TTS artifacts and voice-clone signatures — works on recorded samples for async verification flows
Image synthesis analysis catches deepfake selfies — checks the image itself for generation artifacts, not just metadata
Binary verdict per document — watermark present or not. More defensible than a probabilistic score for account suspension and appeals
Compliance • regulatory identity
Government + regulated
Attack vectors
Synthetic identity packages
AI-gnenerated biometrics
Deepfake video verification
Modalities
Audio
Video
Products
Resemble Detect
Resemble Intelligence
Without detection
AI-generated biometrics defeat remote government ID verification — benefits enrollment, passport renewal, and tax registration are all at risk
Deepfake video in notarized or witnessed remote verification bypasses human review
Explainability gap — standard confidence scores don't satisfy regulatory requirements for rejected applications
How Resemble AI detects it
Air-gapped and on-premise deployment — citizen data and PII stay within national or agency infrastructure
Video + audio + document analysis in a single call — covers every modality used in remote government verification
Resemble Intelligence produces a human-readable audit trail per verdict — structured for regulatory review and appeals processes
how resemble detect works

A detection score and a full explanation. One API call.

Resemble Detect evaluates documents, selfie, and voice in a single call. Resemble Intelligence layers a human-readable explanation on top — so your compliance team understands the verdict, not just the number.
Detects AI-generated documents, not just altered ones
Identifies synthetic audio across TTS and voice-clone types
Catches real-time face-swap in video verification flows
Intelligence: abnormalities, confidence reasoning, audit trail
On-premise — PII never leaves your environment
View on GitHub
THE MATH IS clear

Not running detection means your onboarding is open to synthetic identities

Estimates are conservative. Actual exposure is likely higher.
GET A FRAUD EXPOSURE ESTIMATE
Banking
Daily onboarding
200 applications
Synthetic ID rate
1%
Average fraud per account
$15,000
Annual exposure
$10,950,000
Detection cost
$18,250
Net protected
$10,931,750
Fintech
Daily onboarding
1,000 applications
Synthetic ID rate
2%
Average fraud per account
$3,500
Annual exposure
$25,550,000
Detection cost
$36,500
Net protected
$25,513,500
Marketplace
Daily onboarding
5,000 applications
Synthetic ID rate
0.5%
Average fraud per account
$800
Annual exposure
$7,300,000
Detection cost
$91,250
Net protected
$7,208,750
ONE PLATFORM. EVERY identity attack.

Detect synthetic media across every verification modality

Voice fraud
ATTACK TYPE

Synthetic voice in verification

TTS audio submitted for voice biometric verification — indistinguishable from authentic speech to standard tools.

DETECTION METHOD

Audio multimodal analysis

Scores for TTS generation artifacts, voice-clone signatures, and absence of natural speech disfluencies.

OUTPUT

Confidence score + audit log

Per-sample verdict with confidence score, flagged for review or auto-blocked above threshold.

Face swap
ATTACK TYPE

Real-time face-swap in video verification

Live face-swap via webcam defeats video liveness checks — bypasses motion, blink, and depth-based signals.

DETECTION METHOD

Video + image analysis

Frame-by-frame analysis for deepfake artifacts, including temporal consistency checks across the video sequence.

OUTPUT

Per-frame confidence + temporal flag

Confidence score with temporal consistency signal. Anomalous sequences flagged with frame-level evidence.

Document forgery
ATTACK TYPE

AI-generated identity documents

Synthetic or inpainted IDs with correct fonts, lighting, and metadata that pass OCR and resolution checks.

DETECTION METHOD

Image synthesis analysis

Checks for synthesis artifacts in the image itself — not just metadata — catching patterns OCR cannot see.

OUTPUT

Confidence score + tamper evidence

Per-document verdict with tamper evidence, structured for compliance and audit review workflows.

BUILT FOR onboarding WORKFLOWS

Runs where your identity data lives

Real-time at the verification step or async batch across existing queues. All output is audit-ready for compliance and dispute review. Identity data never leaves your environment.
Cloud API or on-premise deployment
Real-time (<200ms) and async batch
Image, video, and audio
Confidence score + audit log per submission
Resemble Intelligence explanation per verdict
Custom models for your claim type and data
Frequently asked questions
Can deepfakes bypass KYC verification systems?
Yes. Current generative models produce synthetic faces, voices, and documents that pass standard liveness, biometric, and OCR checks. Resemble Detect is trained on these attack vectors including coordinated multi-modal identity packages.
How do you detect AI-generated identity documents?
Resemble Detect analyzes the image itself for synthesis artifacts not just metadata catching inpainted and generated documents that pass OCR and resolution checks.
What's the difference between Resemble Detect and Resemble Intelligence?
Resemble Detect returns a mathematical confidence score per modality. Resemble Intelligence available with Detect produces a full human-readable explanation of the verdict for compliance and legal review.
Is on-premise deployment available?
Yes. Resemble Detect supports cloud, on-premise, and air-gapped deployment. Identity data and PII never leave your environment, meeting DORA, HIPAA, and financial services data residency requirements.
How does voice liveness detection work in remote KYC?
The API analyzes submitted audio for TTS generation artifacts and voice-clone signatures. Works on recorded samples — fits both synchronous and asynchronous onboarding flows.
What accuracy does Resemble Detect achieve?
97.9% pooled accuracy across image, video, and audio attack vectors. Custom models trained on your onboarding data and attack profile can improve on this baseline.
Get complete generative AI security
Join thousands of developers and enterprises securing with Resemble AI
Contact us