TL;DR: NIST tested 89 pre-pandemic face-recognition algorithms on masked faces in 2020 and found that "some of the algorithms failed 30 to 50 percent of the time" when matching the same person wearing a mask, while "some of the same algorithms with high false non-match rates (FNMR) with masks had results of less than .01 FNMR without a mask" [1]. The DHS Science and Technology Directorate's 2020 Biometric Technology Rally, with "582 diverse test volunteers representing 60 countries" and "60 facial recognition configurations", measured median identification rates of "about 93 percent" without masks and "about 77 percent" with masks, with the best system "correctly identifying individuals about 96 percent of the time" while masked [2]. Mask shape, color, and coverage matter: NIST found that "Masks that are more round, such as many N95 masks, had fewer false non-match rates, as did light blue masks" and "Masks that covered more of the face, and masks that were black, had higher false match rates" [1]. The 2020 tests covered pre-pandemic algorithms; "newer face recognition algorithms that have been calibrated during the COVID-19 pandemic to account for mask wearing were not tested" in that report [1]. Federal use has grown even as the technical picture has shifted: GAO found "about 60,000 searches" by seven DHS and DOJ law-enforcement agencies through 2022, with most agencies using the systems "without requiring staff take facial recognition training" [3].
Why a Mask Changes the Math
A face-recognition algorithm builds a representation of the face from a gallery image, then measures how close a probe image is to that representation. The nose, mouth, chin, and cheek texture carry a large share of the signal. A standard surgical mask covers the nose, the mouth, and most of the chin, leaving the algorithm with the upper face, the eyes, and a small strip of cheek. A gait, voice, or skin pattern cannot substitute on a still image. The probe has fewer distinguishing features to compare against the gallery, and the matching engine produces less confident scores.
The drop is not linear. It depends on which face is in the gallery, which mask is in the probe, and how the algorithm was trained. Some algorithms were built on bare faces and have no learned representation for the lower face occluded by fabric. Other algorithms were updated after 2020 to specifically handle masked probes, and they have done measurably better. The headline accuracy number from any one vendor is therefore less informative than the conditions under which it was measured.
NISTIR 8311: The 2020 Mask Test
NIST Internal Report 8311, published in July 2020 as "Ongoing FRVT Part 6A: Face recognition accuracy with face masks using pre-COVID-19 algorithms", tested 89 one-to-one face-recognition algorithms on a dataset that paired unmasked gallery photos with masked probe photos [1]. The result that mattered most: on a typical operating point where the false match rate was held to about 0.001, false non-match rates for some algorithms climbed above 0.30 and "some of the algorithms failed 30 to 50 percent of the time" when matching the same person wearing a mask [1]. NISTIR 8311 noted in its summary that "some of the same algorithms with high false non-match rates (FNMR) with masks had results of less than .01 FNMR without a mask" [1]. The same algorithms that could pick a face out of millions unmasked were missing one out of three matches once a mask was in the probe.
NIST's World Privacy Forum summary of the same report adds the qualitative findings. "Masks that are more round, such as many N95 masks, had fewer false non-match rates, as did light blue masks" and "Masks that covered more of the face, and masks that were black, had higher false match rates" [1]. A black mask that covers more of the cheeks costs the algorithm more distinguishing signal than a light blue surgical mask that sits higher on the nose and exposes more skin. The same NIST text notes that "newer face recognition algorithms that have been calibrated during the COVID-19 pandemic to account for mask wearing were not tested" in NISTIR 8311 [1]. The 2020 paper was a snapshot of pre-pandemic systems, not of the algorithms that came after.
NISTIR 8311 also flagged an asymmetry. False non-match rates are the "did the system miss a real match" number; false match rates are the "did the system flag two different people as the same person" number. Both got worse with masks, but the report's quoted finding is that mask color and shape affected FNMR more than FMR. The headline takeaway for someone reading the 2020 paper in 2026 is that the algorithms were a moving target and the worst-performing ones were the ones that had not been re-trained.
The DHS 2020 Biometric Technology Rally
The DHS Science and Technology Directorate ran the third annual Biometric Technology Rally at the Maryland Test Facility in fall 2020. AFCEA's January 5, 2021 report on the rally, drawing on DHS's own press materials, describes the test population as "582 diverse test volunteers representing 60 countries" run through "60 facial recognition configurations" over "10 days of human testing" [2]. The DHS summary, as quoted in the AFCEA piece, is that "Without masks, median system performance demonstrated a about 93 percent identification rate, with the best-performing system correctly identifying individuals about 100 percent of the time. With masks, median system performance demonstrated a about 77 percent identification rate, with the best-performing system correctly identifying individuals about 96 percent of the time" [2].
The rally's key practical finding was a separation between top and median. The top system was at 96 percent masked, almost matching its own unmasked performance. The median dropped by 16 percentage points, from 93 percent to 77 percent. The quote attributed to DHS's Arun Vemury is that "this isn't a perfect 100 percent solution, but it may reduce risks for many travelers" [2]. The mask dataset at the rally was diverse: "the ability of biometric acquisition systems and matching algorithms to reliably collect and match images of individuals wearing a diverse array of face masks" [2]. The 16-point gap between top and median is the spread a vendor's marketing has to be read against.
What Has Changed Since 2020
Two things changed between the NISTIR 8311 test and subsequent evaluations: more algorithms were trained on masked faces, and the rally datasets grew. NIST's follow-up evaluations on the VISABORDER dataset, hosted on the FRTE Face Mask page, are updated as vendors submit new algorithms [4]. Vendor press releases summarized in trade coverage reported False Non-Match Rates below 1 percent on masked probes for several leading vendors in 2023 and 2024 testing rounds. NIST's published FRTE Face Mask page is the canonical place to read the current leaderboard rather than relying on the headline numbers below.
The improvement reflects the same engineering pattern seen elsewhere in computer vision. A category of failure is exposed by a public dataset, vendors add masked-face training data to their models, the new generation of algorithms performs better on the masked-face test, and the public test becomes a snapshot of the previous generation. NIST runs the test continuously for this reason, and the published FRTE Face Mask page is updated as vendors submit new algorithms [4]. The numbers on the leaderboard should not be treated as the accuracy the average user will encounter at an airport or a stadium today. Those use vendor-chosen algorithms on systems the vendor continues to update, and the published report is one static measurement per submission.
How Federal Agencies Are Using It Anyway
The Government Accountability Office report GAO-23-105607, published September 12, 2023, examined face-recognition use across seven DHS and DOJ law-enforcement agencies. The report's headline number, drawn from agency disclosures, was that the agencies had together run "about 60,000 searches" through their face-recognition services when they did not have training requirements in place [3]. The same report found that "All seven agencies initially used these services without requiring staff take facial recognition training" [3]. For the FBI specifically, GAO reported that "only 10 staff completed facial recognition training of 196 staff that accessed the service" [3]. The FBI did not finalize its Facial Recognition Technology Use Policy Directive until December 2023 [3]. ICE's Homeland Security Investigations arm "did not have a process to periodically monitor whether staff were taking required training" [3].
What that tells the reader is that the federal use is wide, the controls are thin, and the technical accuracy question is downstream of the policy question. A 2020 pre-pandemic algorithm at 30 to 50 percent false non-match rate is being used today in an environment where 60,000 of those searches happened before training was even required. The 2024 NIST leaderboard shows vendor accuracy on masked faces has improved since 2020. Whether the federal systems being used have been updated to match that leaderboard is a separate question, and one GAO did not directly answer in 2023.
What the Mask Itself Does to the Number
The NISTIR 8311 mask-color and mask-shape findings are the most actionable for a person making a one-time decision about what to wear. The combination that cost the algorithm the least distinguishing signal was a round, light blue mask that covered a smaller share of the cheeks. N95-style respirators produced fewer false non-matches than flat surgical masks, which NIST attributes to the curved shape exposing more of the upper cheek. Black masks and any mask pulled down to cover more of the cheeks produced the worst numbers on both FNMR and FMR.
For the inverse case, a reader who wants the algorithm to fail: the worst-case mask from NIST's 2020 testing is a dark fabric mask pulled low to cover more of the cheeks. The best-case evasion mask is the opposite of a visible ID, since the same features the algorithm needs are the ones you want to obscure. Anti-facial-recognition accessories, the dazzle-pattern glasses and reflective eyewear that other guides cover, sit on top of the mask and degrade the upper-face signal the algorithm falls back on once the lower face is gone. The combination is what the algorithms of 2024 to 2026 have to defeat, not the mask alone.
What This Means for You
The bottom line is that mask effectiveness against face recognition has been improving against you since 2020. NIST's 2020 test caught a generation of pre-pandemic algorithms in their worst state, and vendor performance on masked faces has improved substantially since. Vendor press releases for the leading 2023 and 2024 NIST submissions reported False Non-Match Rates below 1 percent on masked probes; the published NIST leaderboard is the authoritative reference for current accuracy [4]. The 96 percent reported by DHS for the top performer in the 2020 rally is a 2020 number, not a 2026 one, and current best-in-class systems have improved past it.
What a mask still does: it removes a large portion of the lower face from the probe, which is the single most distinctive feature in a bare-face comparison. The eye region, brow ridge, and skin texture above the mask still match, but with much less signal than the same comparison on a bare face. The matching engine produces less confident scores, and less confident scores are exactly the threshold a privacy-resisting subject benefits from. None of this is a guarantee, and none of it substitutes for the structural defenses covered elsewhere: the warrant requirements, the algorithmic transparency rules, and the legislative bans that are the actual fix.
The Honest Takeaway
NISTIR 8311, published in July 2020, documented the moment face recognition ran into mask-wearing and lost. Pre-pandemic algorithms missed one out of three masked matches at the strictest operating point; some failed the same test half the time [1]. The DHS 2020 Biometric Technology Rally at the Maryland Test Facility, with 582 volunteers from 60 countries and 60 different facial recognition configurations, measured median identification rates of about 93 percent without masks and about 77 percent with masks; the top system still hit about 96 percent masked [2]. Since then, vendors have retrained on masked-face data and NIST's ongoing leaderboard on the VISABORDER dataset shows multiple vendors below 1 percent false non-match rate on masked faces [4]. The federal use of face recognition grew in parallel, with GAO reporting about 60,000 searches by seven DHS and DOJ law-enforcement agencies through 2022, the majority of which happened before training was required [3]. Mask shape and color still matter, with round, light blue masks costing the algorithm less signal than dark, full-coverage masks [1]. None of this changes the fact that a mask removes a large portion of the lower face from the probe, and a less confident match is exactly what a privacy-resisting subject benefits from.
Sources
- World Privacy Forum, summarizing NISTIR 8311 (NIST Ongoing FRVT Part 6A: Face recognition accuracy with face masks using pre-COVID-19 algorithms, July 2020): 89 algorithms tested, 30 to 50 percent failure rate, mask shape and color findings, pre-pandemic algorithm scope
- AFCEA Signal Media, DHS Test Shows Biometrics Work With Face Masks (January 5, 2021; 2020 Biometric Technology Rally at Maryland Test Facility, 582 volunteers, 60 countries, 60 facial recognition configurations, 93 percent unmasked and 77 percent with-masks median, 96 percent with-masks top)
- Government Accountability Office, GAO-23-105607, Facial Recognition Services: Federal Law Enforcement Use (September 12, 2023; seven DHS and DOJ agencies, about 60,000 searches, training gaps, FBI December 2023 policy directive)
- NIST, Face Recognition Technology Evaluation (FRTE) Face Mask Effects (ongoing benchmark on VISABORDER dataset; 2024 leaderboard for masked 1:1 verification)
Published: October 18, 2026