Government & public sector NeuroVerify + NeuroDetect
A minister’s voice, a pensioner’s face, a piece of evidence. All three can now be generated.
Public services decide on media nobody can inspect in person: a selfie video that proves a pensioner is alive, a voice note from a senior official, a clip submitted as evidence. NeuroVerify checks the face, video or voice at enrolment and proof of life, and NeuroDetect gives investigators a forensic workspace that runs on-premises or air-gapped.
The problem, in numbers
Auditors count the losses. Regulators and the FBI describe the new method.
- $233bn–$521bn
estimated direct fraud losses to the US federal government each year, fiscal 2018 to 2022. A first government-wide estimate.
GAO-24-105833, 16 Apr 2024 - $100bn–$135bn
estimated fraud in US pandemic unemployment insurance, 11–15% of benefits paid. Not a deepfake figure: it shows what remote benefit claims are worth to attackers.
GAO-23-106696, 12 Sep 2023 - $6,000,000
FCC penalty for robocalls carrying a deepfake of the US President’s voice, sent two days before the 2024 New Hampshire primary under a spoofed caller ID.
FCC 24-104, 26 Sep 2024 - Since 2023
actors have impersonated state, White House and Cabinet-level officials and members of Congress with AI-generated voice messages, reaching their family and acquaintances.
FBI IC3, I-121925-PSA, 19 Dec 2025
Where deepfakes hit government
Six decisions that rest on a face, a voice or a file.
Each one replaced a counter visit with a camera or a phone. That made services reachable for everyone, including the person pretending to be someone else.
Digital identity enrolment
Citizen accounts, digital ID wallets, tax portalsA virtual camera injects a face-swapped stream into the selfie step. The face was built from the stolen document photo, so the face match passes and the face still blinks on cue.
Benefits and pension proof of life
Annual life certificates, remote re-verificationPayments continue as long as a face appears on camera once a year. Old video of the beneficiary, or a photo animated into a talking head, keeps them flowing after the person has died or moved away.
Licensing and permits
Driving licence renewal, permits, professional registersA renewal accepts a new photo and a selfie. A generated face that resembles the holder slides an impostor onto a credential that other agencies will later trust.
Law-enforcement evidence intake
Complaints, tip lines, citizen uploads, case filesAn edited clip or a generated voice recording arrives as evidence. Hashing proves it has not changed since intake. It cannot prove it was real when it arrived.
Election and official communications
Election offices, press offices, emergency alertsA cloned voice of a candidate or official tells the public something false, from a number that appears legitimate. The press office has hours to establish what is real.
Impersonation of senior officials
Staff, counterparts, families of officialsA voice note in a minister’s voice builds rapport, then moves the conversation to another messaging app. The FBI describes the next step as a link that harvests credentials.
Why existing controls miss it
Every control checks the claim against a record. None checks whether the face or voice is real.
Government identity controls were designed for forged documents and borrowed passwords. A generated face carries the right document, the right data and a perfect custody record.
Document verification
Is the ID genuine, and does its photo match the face?
A face generated from that same document photo. The better the fake, the better the match.
Liveness / presentation attack detection
Is a physical person in front of the camera?
Video injected after the camera. NIST SP 800-63-4 now treats injection attacks and forged media as their own threat.
Register data matching
Do the details agree with civil, tax or benefit records?
Anything the records do not yet know: a death abroad, an unregistered move, a stolen identity with clean data.
Caller ID or known sender
Does the number or account belong to the official?
Spoofed numbers, as in New Hampshire, and “new number” messages on encrypted apps.
Recognising the voice
Does this sound like the minister?
A clone trained on years of public speeches. Officials are the easiest people to clone.
Chain of custody
Has the file changed since we received it?
A file faked before it was received. Its custody record is perfect.
NeuroVerify
Is this face, video or voice real?
One risk score from six signals, including live presence, face manipulation and AI generative model analysis, at enrolment or proof of life.
NeuroDetect
Was this file edited or generated?
Findings down to regions, frames and audio segments, and a forensic report that records model version and uncertainty.
For the product rows, the last column is what the check adds.
One intake queue
Each item gets the check that fits it, and a route a person owns.
Faces and voices from a citizen go to NeuroVerify. Files that will be relied on go to NeuroDetect. A flag never closes a case; it moves the case to a stronger check.
- Pension proof of lifeSelfie video · mobile appNeuroVerify
Live presence failed · face manipulation flagged
Hold payment · video call with an officer
- Voice note from a “Deputy Minister”Audio · messaging appNeuroDetect
Synthetic speech in two audio segments
Do not reply · call back on the directory number
- Complaint evidencePhone video · citizen uploadNeuroDetect
No edit or generation findings
Case file · forensic report attached
Illustrative queue. Findings and routes show how results are used, not real cases.
Regulation and guidance
What standards bodies and regulators say.
Plain summaries of public texts. They describe the threat and the obligations; they say nothing about Neural Defend, and using a product does not by itself meet any of them.
- United StatesFinal, Jul 2025NIST SP 800-63-4 Digital Identity Guidelines
The US federal digital identity guidelines add controls for injection attacks and forged media, naming deepfakes, to identity proofing.
- United States8 Feb 2024FCC Declaratory Ruling on AI-generated voices
AI-generated voices are “artificial” under the Telephone Consumer Protection Act. Robocalls that use them need the called party’s prior consent, unless an exemption applies.
- United StatesSep 2023NSA, FBI and CISA: Contextualizing Deepfake Threats to Organizations
Names impersonation of leaders and financial officers, and fraudulent communications that open access to networks, among the most substantial deepfake threats.
- European UnionApplies from 2 Aug 2026AI Act, Article 50 transparency obligations
Providers must mark AI-generated output in a machine-readable way. Deployers that publish deepfakes must disclose that the content is artificial.
- European Union2022, updated Jan 2024Europol Innovation Lab: Facing reality?
Lists manipulated electronic evidence among criminal uses of deepfakes and calls a consistent, transparent chain of custody essential.
Deployment and data sovereignty
Evidence stays inside your boundary.
Air-gapped
No network path out. For classified networks and evidence that must never leave the building.
NeuroDetectOn-premises
Runs on hardware you own, inside your accreditation boundary.
NeuroDetectSecure container
Packaged for your own platform, alongside existing case systems.
NeuroDetectPrivate cloud / VPC
In your own cloud account and region, for data residency.
NeuroDetectAPI and SDKs
One call from your enrolment or proof-of-life app. Android, iOS, Flutter, REST and server SDKs.
NeuroVerifyISO/IEC 27001 · 27701 · 42001 · SOC 2 Type II. Compare deployment models · Trust centre
Knowledge centre
Read further, by public-sector workflow.
- Workflow guide
Deepfake risk controls for public benefits
Where a generated face or voice enters a benefits journey, and which checks to place at enrolment, change of details and proof of life.
Read article - Workflow guide
Identity proofing for digital licensing
Renewals and remote applications create credentials other agencies rely on. How to stop an impostor before the card is issued.
Read article - Workflow guide
Synthetic media review at evidence intake
A review step for citizen-submitted video and audio, before a file shapes an investigation.
Read article - Workflow guide
Deepfake response for election communications
Who verifies, who corrects, and how fast, when a fake statement from a candidate or official spreads.
Read article - Verified case
New Hampshire AI voice robocalls
The FCC record on the deepfake robocalls: spoofed caller ID, the penalty, and the carrier settlement.
Read article - Threat analysis
Deepfakes targeting government officials
How voice-message campaigns against officials and their contacts are built, and where they break.
Read article - Deployment
On-premises deepfake detection
What changes when detection runs inside your boundary: updates, hardware, logging and evidence export.
Read article - Evaluation
Top 5 deepfake detection APIs
A side-by-side comparison to bring to a procurement or technical evaluation panel.
Read article
Sources
Read the evidence yourself.
Third-party publications. None of them evaluates or endorses Neural Defend.
- US GAOGAO-24-105833, Fraud Risk Management
16 Apr 2024. Federal government loses an estimated $233bn–$521bn a year to fraud (FY2018–2022).
- US GAOGAO-23-106696, Unemployment Insurance
12 Sep 2023. Pandemic UI fraud likely $100bn–$135bn, 11–15% of benefits paid.
- US FCCForfeiture Order FCC 24-104, Steve Kramer
Adopted 26 Sep 2024. $6,000,000 for spoofed robocalls carrying a deepfake presidential voice before the New Hampshire primary.
- US FCCFCC makes AI-generated voices in robocalls illegal
8 Feb 2024. Declaratory Ruling: AI-generated voices are “artificial” under the TCPA.
- FBI IC3I-121925-PSA, Senior US officials continue to be impersonated
19 Dec 2025. Activity since at least 2023; AI-generated voice messages; officials’ family and acquaintances targeted.
- FBI IC3I-051525-PSA, Senior US officials impersonated
15 May 2025. Text and AI voice messages to current and former officials and their contacts; move to another platform.
- NISTSP 800-63-4 Digital Identity Guidelines
Final Jul 2025. Adds controls for injection attacks and forged media (“deep fakes”).
- NSA, FBI, CISAContextualizing Deepfake Threats to Organizations
Sep 2023. Impersonation of leaders and financial officers among the most substantial threats.
- EuropolFacing reality? Law enforcement and the challenge of deepfakes
2022, updated Jan 2024. Manipulated evidence, and the case for a transparent chain of custody.
- European UnionRegulation (EU) 2024/1689, the AI Act
Article 50 transparency obligations for synthetic content and deepfakes apply from 2 Aug 2026.
Questions
What public-sector buyers ask first.
Compare the top 5 deepfake detection APIsPreserving evidence in synthetic media incidentsCan it run with no internet connection?
NeuroDetect runs on-premises or fully air-gapped, as well as in a secure container, a private cloud or VPC, or as a web app. NeuroVerify is an API called from your enrolment or proof-of-life app; discuss where it runs for your environment during evaluation. Deployment options
Does NeuroVerify check passports or ID cards?
No. NeuroVerify analyses face images, video and audio. Keep your document verification; NeuroVerify answers the question it cannot: is the face presenting that document real?
Can a score refuse a benefit or a licence?
It should not. The result goes to your official or your policy. A flagged proof of life moves to a stronger route, such as a video call with an officer or an in-person check, so a real pensioner is never cut off by a score.
Will the forensic report hold up in court?
The NeuroDetect report records observations, uncertainty, provenance, model and version context, and analyst notes, so another examiner can see how each finding was reached. Admissibility is for your counsel and the court.
Does this make us compliant with NIST SP 800-63-4 or the EU AI Act?
No product makes an agency compliant. NeuroVerify and NeuroDetect supply evidence for the controls those frameworks describe; your assessors decide whether a control is met.
What about live phone calls to officials?
NeuroCall, in Beta, analyses a caller’s voice during the call. Voice notes and recordings already received go to NeuroDetect. About the NeuroCall Beta
Evaluate inside your boundary
Test it on the cases you have already closed.
Bring past proof-of-life videos, voice notes sent to officials and evidence files, including ones you know were fake. Run them in your own environment and see what NeuroVerify and NeuroDetect find, and what the report looks like on file.