Deepfake Detection for Public Benefits Claims: Controls Across the Claim Lifecycle

Why Deepfake Detection for Public Benefits Claims Is on Program Integrity Agendas

Benefits programs have already absorbed one wave of large-scale imposter fraud. The US Department of Labor Office of Inspector General, in its oversight summary last updated 5 June 2025, says more than $888 billion in federal and state unemployment insurance was paid for benefit weeks in the pandemic period. Applying the estimated improper payment rate of 21.52 percent, it concludes that at least $191 billion was likely paid improperly. It also cites a 2023 GAO estimate that fraud in the period from April 2020 to May 2023 ran at 11 to 15 percent, or up to $135 billion.

Most of those claims did not use deepfakes. They used stolen or synthetic identities pushed through online forms at a volume states could not check by hand. The OIG notes that 45 percent of states responded by hiring identity verification contractors that used facial recognition. That shift moved the attack surface to the camera, and generative tools now make it cheap to produce a face that matches a stolen ID photo or to replay a genuine claimant's video.

The lesson for program designers is that the face step is now a target in its own right. A control that only compares the selfie to the document portrait will accept a generated face built to match that portrait.

Where Synthetic Faces and Voices Enter a Benefits Claim

A benefits case is not a single event. The same claimant may verify at application, again at each recertification, again to prove they are alive for a pension, and again when they change a bank account or phone number. Each touchpoint has its own attack and its own right control.

Recertification and proof of life deserve particular attention because they repeat for years. India's Jeevan Pramaan service, run by the Ministry of Electronics and Information Technology, lets pensioners authenticate biometrically to generate a Digital Life Certificate that is shared with the disbursing agency, and offers face authentication as one route. Remote face checks like this spare older people a branch visit, and they are also the step a fraudster must beat to keep a deceased person's pension flowing.

Protecting the Payment Destination Change

Changing where money goes is the moment a takeover pays off. An attacker who already holds the claimant's personal data only needs to pass one check to redirect every future payment. That makes the change request a better place for strict verification than the routine weekly or monthly certification.

Treat the request as high risk regardless of how it arrives. Require a fresh face capture or a known-channel confirmation, notify the claimant through the contact details on file before the change takes effect, and hold the first payment to a new account long enough for a genuine claimant to object. If the request arrives by phone and the caller sounds right, the voice alone should not count as proof; voice cloning tools can work from short recordings, and claimants' voices are often available from earlier calls or social media.

Routing Flags to Caseworkers Without Cutting Off Benefits

People who rely on benefits cannot wait weeks for a wrongly flagged claim to clear. Design the review path before switching on detection, and measure it. A manipulation flag should open a case with the capture, the signals that fired, capture conditions and claim history attached. The caseworker can then request another capture, move the claimant to an in-person or assisted route, or escalate to an investigator.

NIST SP 800-63A-4 gives a useful pattern even for programs outside its formal scope. It says credential service providers should offer trusted referees to applicants who fail fraud checks in unattended remote processes, that referees must receive a summary of the fraud failures, and that a redress process must exist for applicants who dispute outcomes. The guideline lists older individuals, persons with disabilities and people without access to computing devices among those who may need that help, which overlaps closely with benefits populations.

Watch the numbers by group. If pensioners using older phones fail the face step at several times the rate of other claimants, that is a testing and device problem to fix, not a fraud signal.

What Detection Will Not Tell a Benefits Agency

Synthetic media analysis tells you whether a capture shows signs of generation, face swapping, replay or editing. It does not establish eligibility, income, residence or whether a claimant is alive beyond the moment of capture. It cannot see a genuine claimant who is being coerced into recertifying for someone else, and it does not check documents.

Error rates also move with the population. Low light, old cameras, compression on slow mobile networks and assistive devices all affect results, so test on media representative of your claimants before setting thresholds. Agencies should keep the decision about benefits with program rules and trained staff, and treat the detector as one input among identity records, payment history and device signals.

How NeuroVerify Supports Benefits Verification

NeuroVerify checks a face image, video or audio for deepfakes and manipulation beyond liveness. A single request returns six signals, including Live Presence Verification, Face Manipulation Detection and AI Generative Model Analysis, combined into one risk score with signal-level evidence. That evidence is what a caseworker needs to decide whether to ask for a new capture or escalate. The agency's policy, not the score, decides the outcome.

NeuroVerify returns results in under 200 ms at p95 on eligible image requests under defined test conditions, so recertification and proof-of-life checks can run inline in a mobile app. Mobile SDKs for Android, iOS and Flutter and backend SDKs for Python, TypeScript, Go and Java shorten integration, and deployment can be managed cloud, private cloud or VPC, secure container, on premises or air-gapped where benefit data must stay in a government environment.

Neural Defend ranked first of 36 systems on the public VendorBench-100 still-image benchmark, with an MCC of 0.876 against 0.646 for the next-best vendor. That benchmark does not cover video, audio or injection attacks, so any pilot should test the agency's own capture channels.