BRENNEN

Industries / Finance and business services / Staffing and recruiting / Background checks

Identity

Confirming that an applicant is the person named, from fingerprints, a Social Security number, an ID document or a selfie, and finding the past addresses that decide where records are searched.

5

Milestones achieved, from 1924 to 2023.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1924

    A national fingerprint file

    On 1 July 1924 the US Bureau of Investigation set up a Division of Identification to hold the criminal records collected at the federal penitentiary in Leavenworth, Kansas. The file began with 810,188 records and passed 3.5 million by mid-1932, by the FBI's count. Fingerprints became the national way to tie a person to a criminal record whatever name they gave.

    archives.govfbi.gov

    Credit

    US Bureau of Investigation, now the FBI

  2. 1999

    Fingerprint checks in hours

    IAFIS, the FBI's computerized fingerprint system, went live on 28 July 1999 at a cost of $640 million. It had targets of 2 hours for a criminal search and 24 hours for a civil one, such as a check for a job or license, and in 2009 the FBI reported averages of 8 minutes and 2 to 3 hours. Its successor, Next Generation Identification (NGI), replaced it on 7 September 2014 and, by the FBI's figures, raised machine matching accuracy from 92% to more than 99%.

    fbi.govle.fbi.gov

    Credit

    FBI Criminal Justice Information Services (CJIS) Division

  3. 2009

    Number checks by consent

    Since fiscal year 2009 the US Social Security Administration (SSA) has let businesses that hold a person's signed consent check that a name, birth date and Social Security number match its records, through a service called CBSV. Use fell from 3.1 million checks in fiscal 2021 to 2.1 million in 2022, and SSA raised the fee from $1.00 to $2.25 a check from October 2023, giving falling volume as the main reason. A faster electronic version required by a 2018 law, eCBSV, is open only to financial institutions and the firms that serve them. SSA expected 307 million checks from its first 10 users; in fiscal 2024 it expected 52 million from 20, and $25.5 million of the $66.3 million spent on it by then had not been recovered from fees.

    Lessonan industry's forecast of demand is not a commitment, and a service sized to it can end up priced for users who never arrive.

    oig.ssa.govgovinfo.govlaw.cornell.edugovinfo.govgovinfo.govgovinfo.gov

    Credit

    Social Security Administration; US Congress (Public Law 115-174)

  4. 2022

    Borrowed identities apply remotely

    On 28 June 2022 the FBI's Internet Crime Complaint Center (IC3) warned of more complaints about people using deepfakes and stolen personal details to apply for remote jobs, many in information technology. It reported that pre-employment background checks had found that the details some applicants gave belonged to other people.

    Lessonthe records check caught what the remote interview missed, because the details led back to someone else.

    ic3.gov

    Credit

    FBI Internet Crime Complaint Center

  5. 2023

    Outside labs test selfie checks

    Presentation attack detection is software that tells a live face from a photo, mask or screen held up to the camera. In September 2023 NIST, the US standards agency, published tests of 82 such algorithms from 45 developers against nine kinds of attack, and none caught every kind. In May 2024 the FIDO Alliance, an industry standards body, launched a certification for face verification in which independent labs run at least 10,000 tests on a product and compare its errors by skin tone, age and gender; it had begun certifying ID document checks in October 2022.

    Lessona product sheet that says "liveness" tells a buyer little until an outside lab has attacked the product.

    nist.govfidoalliance.orgfidoalliance.org

    Credit

    NIST (Mei Ngan, Patrick Grother, Austin Hom); FIDO Alliance, with the test labs Ingenium Biometrics and TÜV Informationstechnik

Watch for

  1. Screeners add deepfake checks

    Checkr, a US screening company, launched identity verification in March 2026, and in October 2026 added deepfake detection and matching of a selfie to an ID to a hiring fraud suite that also checks liveness and documents. In its own trial, 36% of incoming résumés were flagged for review, and 50% for software and AI engineering jobs; these were flags, not confirmed fraud.

    Achieved whenAn independent lab certifies the checks, for example under FIDO's face verification or document programs, or an audit reports how many flagged applicants turned out to be impostors.

    siliconangle.comidtechwire.com

    Claimed by

    Checkr, March and October 2026

  2. Deepfake attacks, tested by government

    The Science and Technology Directorate of the US Department of Homeland Security (DHS) is running a second Remote Identity Validation Rally (RIVR-2) at its Maryland Test Facility, with a track for detecting deepfake attacks on face checks. Products for the selfie-matching and ID-validation tracks were due on 20 November 2026, and results for attack detection are planned for June 2027.

    Achieved whenThe facility publishes each tested system's deepfake acceptance and false rejection rates.

    mdtf.org

    Claimed by

    DHS Science and Technology Directorate, September 2026

Open

  1. 01

    Inject video into selfie checks

    A public test set of videos of generated faces of people who do not exist, fed into selfie checks through a virtual camera instead of being shown to a real one, with re-encoded and replayed variants, and a harness that reports how often each kind of attack passes open liveness detectors. Real faces come only from an existing dataset whose license allows this use.

    The brief
    Credit · unclaimed
  2. 02

    Test where traces send searches

    A simulation, built entirely from made-up people, of how US screeners choose which county courts to search. Give each person ten years of moves drawn from real migration patterns, made-up court records in some counties, and an address trace (the list of past addresses, drawn from credit files, that a screener uses) with realistic gaps for renters, young people and recent arrivals. Measure how often common rules, such as searching every county on the trace for the past seven years, miss the county that holds the record.

    The brief
    Credit · unclaimed

Open 01 · the brief

Inject video into selfie checks

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set of videos of generated faces of people who do not exist, fed into selfie checks through a virtual camera instead of being shown to a real one, with re-encoded and replayed variants, and a harness that reports how often each kind of attack passes open liveness detectors. Real faces come only from an existing dataset whose license allows this use.

Why it is still open

NIST's 2025 identity guidelines say that presentation attack detection alone is not sufficient against media injected between the camera and the checking software. DHS's deepfake results are not due until June 2027. A December 2025 paper says the literature "offers limited insight into its practical implementation and evaluation", and the nearest public set, a July 2026 benchmark of 2,994 passport and ID face images, says it is not a video or liveness capture set. Europe's specification for detecting injection attacks, CEN/TS 18099:2024, is sold, and no public test data for it was found.

pages.nist.govarxiv.orghuggingface.coevs.ee

Built with what exists

LivePortrait to animate generated faces; v4l2loopback, a virtual camera for Linux; Silent-Face-Anti-Spoofing, an open passive liveness model; DeepfakeBench for open deepfake detectors.

  1. First step

    Generate 20 made-up faces, inject 50 clips through v4l2loopback, and score them with Silent-Face-Anti-Spoofing.

  2. Done looks like

    At least 500 injected clips across at least four kinds of attack, labeled and published with the harness, and pass rates by kind for at least two open detectors that someone else reruns and matches.

Open 02 · the brief

Test where traces send searches

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A simulation, built entirely from made-up people, of how US screeners choose which county courts to search. Give each person ten years of moves drawn from real migration patterns, made-up court records in some counties, and an address trace (the list of past addresses, drawn from credit files, that a screener uses) with realistic gaps for renters, young people and recent arrivals. Measure how often common rules, such as searching every county on the trace for the past seven years, miss the county that holds the record.

Why it is still open

The CFPB's 2019 snapshot of the industry describes screeners that search the courts of each place a person is found to have lived and others that rely on private databases. In a 2024 study of 101 people, Sarah Lageson and Robert Stewart found that 90% of participants had at least one record missing from regulated checks and 92% from unregulated ones. The National Consumer Law Center (NCLC) wrote in 2019 that no industry-wide accuracy study had been done, and this research found no test that separates misses caused by where a search looked from misses caused by what the records held.

library.nclc.orgojp.govnclc.org

Built with what exists

The US Census Bureau's county-to-county migration flows from the American Community Survey; Python with Faker for made-up people; a fixed random seed so results can be repeated; GitHub.

  1. First step

    Build 1,000 people from one state's migration flows and compare the seven-year rule with searching only the current county.

  2. Done looks like

    100,000 made-up people, published code and data, and a table of miss rates for at least four search rules, split by how often the person moved, which a second person reruns with the same seed and matches exactly.