OEMs & platforms NeuroVerify inside your product
Attackers stopped pointing at your camera. They plug into it.
Virtual cameras, emulators and real-time face swaps feed synthetic faces into identity, onboarding and calling products behind the sensor, where liveness never looks. NeuroVerify embeds in your product through native SDKs, backend SDKs, REST or MCP and checks the face, video or voice itself, so your policy and your interface make the call.
The problem, in numbers
The attack moved behind the sensor. Standards and devices are following it.
- 2,665%
rise in native virtual camera attacks on remote identity verification. iProov links it partly to virtual-camera apps getting into mainstream app stores.
iProov Threat Intelligence Report, 27 Feb 2025 · vendor data - 40%
annual rise in injection attacks. Virtual camera injection is the most common kind, often paired with device emulation.
Entrust 2026 Identity Fraud Report, via Infosecurity Magazine, 18 Nov 2025 · vendor data - SHALL
NIST now requires identity providers to check that media comes from a genuine sensor and to analyse all submitted media for manipulation or forgery.
NIST SP 800-63A-4, Jul 2025 - Every photo
Pixel 10 is the first phone lineup to attach Content Credentials to every Pixel Camera photo, rated Assurance Level 2 under the C2PA Conformance Program.
Google, 10 Sep 2025
Where deepfakes hit platforms
Your customers inherit whatever your product lets through.
Each workflow below trusts media it cannot inspect by eye. The fake is aimed at that trust, and the loss lands on the bank, insurer or employer that bought your product.
- Identity verification platforms
Selfie and liveness in your SDK
A face-swap tool renders the victim’s face in real time and a virtual camera hands it to your SDK as if it were the phone camera. The face blinks and turns on cue, so liveness passes.
NeuroVerify on the selfie or video your SDK capturedRead more - eKYC API providers
Face images arriving over your API
Your customer’s app does the capture, so you never see the device. A generated or replayed face reaches your endpoint with no sign of where it came from.
NeuroVerify in your backend, on every face imageRead more - Video KYC software
Agent-led video verification
The applicant joins the call through a face-swap feed. The agent judges a face on a screen, which is exactly what the swap is built to fool.
NeuroVerify on frames and recorded segmentsRead more - Video-conferencing and contact-centre platforms
The participant or caller you host
A cloned voice asks for a reset; a swapped face sits in a meeting tile. Your customers assume the platform shows them who is really there.
NeuroCall (Beta) for live voice · NeuroVerify on recorded audio and videoRead more - Device and camera makers
Media your hardware captures or receives
Content Credentials prove where a signed capture came from. Most media people send and receive carries none, and a stripped credential looks the same as one that never existed.
NeuroVerify through the Android, iOS and Flutter SDKsRead more - Your trust and fraud team
Disputes and escalations
A customer contests a rejection, or an auditor asks why a fake passed. A score alone does not answer either question.
NeuroDetect forensic workspace and reportRead more
Why your checks pass it
Every control in your stack vouches for something other than the face.
Liveness vouches for presence, attestation for the device, face match for likeness, provenance for origin. An injected face swap can satisfy all four.
Liveness / PAD
Is a live person in front of the sensor?
Anything that never reaches the sensor. ISO/IEC 30107-3 covers attacks at the capture device only; an injected stream skips it.
Device and app integrity
Is this a genuine, unmodified device and app?
It vouches for the device, not the face. Emulator and virtual-camera detection is a race against tools built to evade it.
1:1 face match
Does the selfie match the ID photo?
A face swap built from the victim’s own photos. It matches by design.
Content Credentials (C2PA)
Who signed this capture, and what edits are recorded?
Media with no credentials, which is most of it. Missing provenance proves nothing either way.
Agent on a video call
Does this person look and sound right?
Real-time swaps and cloned voices made to look and sound right.
NeuroVerify
Is this face, video or voice real?
One risk score from six signals, including Face Manipulation Detection and AI Generative Model Analysis. It reads the media itself, however it arrived.
NeuroDetect
What exactly was changed in this file?
Region, frame and audio-segment findings with a forensic report for disputes and audits.
For the product rows, the last column is what the check adds.
Two streams, one check
Your SDK cannot tell the streams apart. The pixels can.
A virtual camera delivers frames through the same interface as the real one. The difference is in the face: blending seams, generator traces, pixel-level edits.
- Phone camera
- Your SDK
- Liveness: pass
- Face-swap tool
- Virtual camera
- Your SDK
- Liveness: pass
- Visual Safety Screening
- Live Presence Verification
- Face Manipulation Detection
- AI Generative Model Analysis
- AI Pixel-Level Manipulation
- AI Content Compliance
Standards and guidance
Buyers will ask which attacks your product was tested against.
These documents define the questions. Neural Defend claims no certification or conformance under them; meeting them is your programme’s decision.
NIST SP 800-63A-4
July 2025Identity providers SHALL use controls that raise confidence media comes from a genuine sensor, such as detecting a virtual camera, emulator or jailbroken device, and SHALL analyse all submitted media for manipulation or forgery. They SHOULD look for signatures of generative AI and deepfake tools.
Platforms selling into US identity programmes will be asked how they meet both halves: the device and the media.
Read the sourceCEN/TS 18099:2024
Valid from Dec 2024A European Technical Specification on biometric data injection attack detection: definitions, injection attack instruments, detection guidance and test plans. Presentation attacks are left to ISO/IEC 30107.
It gives buyers and test labs a shared vocabulary for injection, separate from liveness testing.
Read the sourceISO/IEC 30107-3:2023
2023Methods and reporting for testing presentation attack detection. The attacks it considers happen at the capture device during presentation; any other attack is out of scope.
A PAD test report says nothing about injected or swapped media. Read the scope before you read the score.
Read the sourceC2PA Conformance Program
In use from 2025Sets security requirements for products that sign Content Credentials and assigns assurance levels. Google reports its Pixel Camera app at Assurance Level 2.
Provenance at capture is becoming a device feature. Detection covers the media that arrives without it.
Read the source
Integration
One call inside the step you already run.
Pick the surface
A native SDK in your app, a backend SDK, REST, or the MCP server for agent workflows.
Send what you capture
The face image, video or audio your product already collects. No new step for the user.
Read one score
A unified risk score with signal-level evidence, back to your service.
Decide in your product
Your thresholds, your screens, your wording. Your customers see your product.
SDK-first integration
One API.
Every stack.
Native SDKs for Android, iOS and Flutter inside your app; Python, TypeScript, Go and Java on your servers; REST for everything else and an MCP server for agent workflows. Two lines to integrate; live in under 24 hours.
Sign in for API access$ npm install @neuraldefend/sdk
import { NeuroVerifyClient } from "@neuraldefend/sdk"
const result = await new NeuroVerifyClient().detectImage("selfie.jpg")Mobile SDKs
Native capture apps- KAndroid (Kotlin)
- SiOS (Swift)
- FFlutter
Backend SDKs
Server-side checks- PyPython
- TSTypeScript / Node.js
- GoGo
- JJava
REST API
For everything elseAlready have capture? Send the image to POST /detect/image and get every signal back in one JSON response.
AI agents
MCP serverGive approved AI assistants bounded image and video authenticity tools, with API keys kept outside prompts.
Two lines to integrate<200 ms p95Live in under 24 hours
Deployment
Run it where your tenants’ data must stay.
The managed API is the quickest start. Private, container, on-premises and air-gapped deployments are scoped with your environment during evaluation.
Managed cloud API
Neural Defend runs the service. The fastest route to launch.
Private cloud / VPC
A dedicated deployment in your network and region.
Secure container
A portable runtime inside your approved environment.
On-premises
Inside your own data centre, next to your platform.
Air-gapped
Fully isolated, with no external network path.
Embedding, attribution, resale, pricing and volume terms are agreed with Neural Defend. Talk to us about terms. ISO/IEC 27001 · 27701 · 42001 · SOC 2 Type II.
Knowledge centre
Read before you embed.
Sources
Read the evidence yourself.
Third-party publications. Vendor reports are labelled as vendor data. None of them evaluates or endorses Neural Defend.
- iProovAnnual Identity Verification Threat Intelligence Report
27 Feb 2025. Native virtual camera attacks up 2,665%, partly through mainstream app stores; face swaps up 300% on 2023; nearly 24,000 users selling attack tools. Vendor data.
- Infosecurity MagazineGenAI and deepfakes drive digital forgeries and biometric fraud
18 Nov 2025. Reports Entrust’s 2026 Identity Fraud Report: injection attacks up 40% a year, virtual camera injection most common, deepfakes a fifth of biometric fraud attempts. Vendor data.
- NISTSP 800-63A-4, Identity Proofing and Enrollment
July 2025. Digital injection prevention and forged media detection controls for remote identity proofing.
- CENCEN/TS 18099:2024, Biometric data injection attack detection
Technical Specification, valid from 16 Dec 2024. Definitions, instruments, detection guidance and test plans for injection attacks.
- ISO/IECISO/IEC 30107-3:2023, Presentation attack detection testing
Scope limited to attacks at the biometric capture device during presentation.
- GooglePixel and Android bring C2PA Content Credentials to images
10 Sep 2025. Content Credentials on every Pixel Camera photo on Pixel 10; Assurance Level 2 under the C2PA Conformance Program.
Questions
What platform and product teams ask first.
Compare the top 5 deepfake detection APIsHow virtual camera injection is detectedWill our customers see Neural Defend, or our product?
Your product. NeuroVerify returns a score and signals to your service; the screens, wording and decisions are yours. Attribution, resale, pricing and volume terms are agreed with Neural Defend directly. Talk to us about terms
Does this replace our liveness or PAD vendor?
No. Liveness answers whether a person is at the sensor. NeuroVerify asks whether the face, video or voice was manipulated or generated. An injected face swap can pass the first question; the second is where it shows.
Does NeuroVerify detect virtual cameras and emulators?
It analyses the media, not the device. Keep your device-integrity and injection controls; NeuroVerify inspects what they let through. NIST SP 800-63A-4 asks for both.
We are multi-tenant. Can each customer set its own policy?
The response is a score with signal-level evidence. Thresholds and actions live in your policy layer, so you set them per tenant in your own code.
How fast is it, and where does the media go?
Eligible image requests return in under 200 ms at p95 under defined test conditions. The managed API is the quickest start; private cloud, container, on-premises and air-gapped deployments are scoped during evaluation. Neural Defend holds ISO/IEC 27001, 27701 and 42001 certifications and a SOC 2 Type II report. How media is handled
Does NeuroVerify check ID documents?
No. It checks face images, video and audio. Keep your document verification in place and add NeuroVerify on the face and voice steps.
Test it inside your product
Run your own attack set through it before your customers do.
Bring captures from your SDK, including injected and face-swapped samples your red team has made. See the score, the signals and the latency on your own media, then decide where the call sits.