Facial recognition statistics span two different tasks: 1:1 verification, which compares a person with a claimed identity, and 1:N identification, which searches a gallery for candidates. The figures below cover U.S. federal agency use, NIST benchmark conditions and a peer-reviewed comparison of named algorithms, with dates and limitations kept explicit.
Key Facial Recognition Statistics
The most useful headline figures are:
- In fiscal year 2020, 18 of 24 surveyed U.S. federal agencies reported using facial recognition.
- In fiscal year 2020, 16 of 24 surveyed U.S. federal agencies reported digital-access or cybersecurity use.
- In fiscal year 2020, 14 surveyed federal agencies authorized facial recognition to unlock agency-issued smartphones.
- In fiscal year 2020, 10 surveyed federal agencies reported facial-recognition research and development.
- Through fiscal year 2023, 10 surveyed federal agencies reported plans to expand facial-recognition use.
- Between October 2019 and March 2022, seven selected DOJ and DHS agencies used systems owned by other entities.
- Between October 2019 and March 2022, the agencies with available data conducted about 60,000 facial-recognition searches before training requirements were in place.
- By April 2023, two of seven selected agencies had implemented facial-recognition training requirements.
- In September 2023, three of seven selected agencies had facial-recognition-specific civil-rights or civil-liberties policies.
- In a 2021 comparative study, A2017b was trained on about 5.7 million images.
- In the same study, A2019 achieved an AUC of 0.999779 on East Asian faces.
- NIST reported false-positive rates varying by up to a factor of 720 across tested demographic groups in its 2022 demographic analysis.
- In a NIST example, Nigerian women aged 60 and over had a false-match rate of 1 in 35 under the stated algorithm and threshold.
- NIST’s current 1:N identification table listed 682 developer entries on September 3, 2026.
- In NIST’s fixed-candidate investigation condition, each 1:N search returns 50 candidates.
- Since February 14, 2022, about 3% of border images and 7% of kiosk images contained multiple faces handled by the FRTE API.
Contents
- Federal adoption and planned expansion
- Search volume, training and governance
- Verification accuracy across algorithms and populations
- False-positive and false-negative disparities
- 1:N identification, thresholds and scale
- Accuracy trends by image quality and demographic group
Federal Facial Recognition Adoption and Planned Expansion
The U.S. Government Accountability Office surveyed 24 Chief Financial Officers Act agencies about reported facial-recognition use in fiscal year 2020. These are self-reported agency results, not a census of government, household or commercial use.
- GAO’s Facial Recognition Technology: Current and Planned Uses by Federal Agencies found that 18 of 24 surveyed agencies used a facial-recognition system in FY 2020.
- Sixteen agencies reported using facial recognition for digital access or cybersecurity, while 14 authorized it to unlock agency-issued smartphones.
- Two agencies reported testing facial recognition to verify identities of people accessing government websites.
- Six agencies reported using it to generate leads in criminal investigations, and five reported physical-security uses such as access control or watchlist monitoring.
- Ten agencies reported facial-recognition research and development, including varied projects such as masks and image manipulation; this is not the same as deployment.
- Ten agencies reported plans to expand use through FY 2023, a forecast of reported intentions rather than completed expansion.
The use categories can overlap because agencies could report more than one purpose. The investigation and physical-security figures therefore should not be added to estimate a total number of agencies or systems.
Facial Recognition Search Volume, Training and Governance Data
A later GAO review examined seven selected Department of Justice and Department of Homeland Security agencies. Its scope was narrower than the 24-agency FY 2020 survey and covered October 2019 through March 2022 for service use.
| Measure | Result | Period and scope |
|---|---|---|
| Agencies owning facial-recognition technology | 3 of 7 | As reported in 2023; selected DOJ/DHS agencies |
| Agencies using systems owned by others | 7 of 7 | October 2019–March 2022 |
| Nongovernment services used | 4 | October 2019–March 2022 |
| Searches before training requirements | About 60,000 | Agencies with available data, October 2019–March 2022 |
Source: GAO’s Facial Recognition Technology: Federal Law Enforcement Agency Efforts Related to Civil Rights and Training. Ownership excludes systems accessed through other entities, and the search total is approximate because GAO noted data-availability limitations.
All seven agencies initially used the covered services without requiring user training first. By April 2023, two had implemented training requirements; of the five without implemented requirements, two continued using services and three had halted use.
Governance policies were also uneven. In September 2023, three of the seven agencies had facial-recognition-specific civil-rights or civil-liberties policies, while four lacked such policies; policy existence does not measure implementation or effectiveness.
Face Verification Accuracy Across Algorithms and Populations
A peer-reviewed comparison evaluated four named algorithms—A2011, A2015, A2017b and A2019—on East Asian and Caucasian face groups. The AUC values are threshold-independent results from that study’s test set, not a current market-wide leaderboard.
| Algorithm | Training history | AUC, Caucasian faces | AUC, East Asian faces |
|---|---|---|---|
| A2011 | Not stated in supplied comparison | 0.981614 | 0.977027 |
| A2015 | Not stated in supplied comparison | 0.990328 | 0.973814 |
| A2017b | About 5.7 million images | 0.999721 | 0.999186 |
| A2019 | Close to 1 million images | 0.9997343 | 0.999779 |
Source: Accuracy comparison across face recognition algorithms: Where are we on measuring race bias?. Training-image counts describe algorithm history and are not test-set accuracy measures.
The study found A2017b and A2019 more accurate than A2015 and A2011 across all three GBU partitions. Because the algorithms, datasets and operating conditions are study-specific, the ranking should not be treated as a universal commercial ranking.
AUC also should not be confused with a single deployment accuracy percentage. It summarizes discrimination across thresholds, while operational systems choose a threshold that trades false matches against false non-matches.
Facial Recognition False-Positive and False-Negative Disparities
NIST’s demographic analyses show why a single overall accuracy number can hide materially different outcomes between groups. The benchmark labels can include sex, age and country or region of birth, depending on the dataset; country of birth should not be rewritten as biological race.
- NIST reported within-group false-positive rates varying by up to a factor of 720 across tested demographic groups in its 2022 FRVT demographic analysis.
- The same analysis reported false-negative rates varying by around a factor of 3 across demographic groups.
- In one NIST example, Polish men aged 35–50 had a false-match rate of 1 in 26,000, while Nigerian women aged 60 and over had a rate of 1 in 35 under the stated example algorithm and threshold.
- NIST described the 1-in-35 rate while targeting 1-in-25,000 for Polish men aged 35–50; the report also gives a rounded 1-in-26,000 example elsewhere, so these are not separate trials.
Source: NIST’s FRVT – Face Recognition Vendor Test – Demographic Summaries. The example rates are algorithm- and threshold-specific, not population-wide estimates.
NIST’s 1:1 demographic summary fixes the overall false-match rate at 0.00003 for the reported algorithm comparisons. Those comparisons use same-sex, same-age-group and same-region-of-birth pairs, so the setup is narrower than an arbitrary real-world search.
A separate NIST 2019 summary found false-positive rates often differed by factors of 10 to more than 100 across demographic groups, while false-negative differentials often varied by factors below 3. These are often-range summaries across tested algorithms, and results depend on algorithm, image quality and demographic definition.
1:N Facial Identification Statistics, Thresholds and Search Scale
NIST distinguishes 1:N identification from 1:1 verification. In 1:N identification, one probe is searched against a gallery; false-positive identification rate and false-negative identification rate therefore describe a different task from false-match and false-non-match rates.
- NIST’s FRTE 1:N identification table listed 682 developer entries on a page last updated September 3, 2026; entries are benchmark submissions, not market share.
- FRTE reports false-negative identification rate at a threshold limiting false-positive identification rate to 0.003, with the threshold set separately for each algorithm and image-column condition.
- In the fixed-candidate investigation condition, FRTE returns 50 candidates per search, with human review assumed.
- The demographic FPIR comparison uses a baseline of 0.002, or 1 in 500 searches, for women born in Eastern Europe.
- The same trial flags cells at 20, 40 and 80 times that baseline across eight demographic comparison groups; these are display bands, not three independent population estimates.
- FRTE’s MXOG inequity summary uses eight FPIR values, reflecting the evaluation design rather than a universal demographic taxonomy.
Source: NIST’s Face Recognition Technology Evaluation (FRTE) 1:N Identification. The threshold and candidate settings are benchmark conditions, not universal deployment settings.
Since February 14, 2022, about 3% of border images and 7% of kiosk images contained multiple faces handled by the FRTE API. These proportions describe FRTE image streams, not every border or kiosk system.
Facial Recognition Accuracy Trends by Image Quality and Demographic Group
The peer-reviewed study separated images into Good, Bad and Ugly partitions and examined low false-accept-rate operating points. It plotted performance at FAR 0.0001 and 0.001, analytical choices that do not represent every deployment threshold.
- At FAR 0.0001 in the Good partition, A2015 showed greater accuracy for Caucasian faces than for East Asian faces.
- At FAR 0.0001 in the Bad partition, A2011 and A2015 showed greater verification accuracy for Caucasian faces than for East Asian faces.
- In the Ugly partition, all algorithms except A2011 showed greater accuracy for Caucasian than East Asian faces.
- A2019 and A2017b showed little or no race bias above FAR 0.001, while lower FARs revealed bias in this study.
Source: Accuracy comparison across face recognition algorithms: Where are we on measuring race bias?. These directional results apply to the named algorithms and study data, and are threshold- and dataset-dependent.
The review notes that a common practice is to set false-accept thresholds at 1/1,000, 1/10,000 or lower. That observation describes examples of application practice, not a measured global adoption rate, and it reinforces why image quality and the chosen operating point belong beside any facial recognition accuracy statistic.