Answers are grounded in this paper and cite their sources.
Masked-face verification systems apply decision thresholds set on development data, but whether these keep their target false acceptance rate (FAR) after deployment is rarely tested. We transferred thresholds set at 0.1% and 1% FAR on synthetically masked LFW to two external masked-face datasets, RMFDVER and the mixed-provenance COMASK20, using two ArcFace/MS1M-pretrained models and identity-level bootstrapping. At the 1% target, the observed FAR rose to 2.5–2.8% on RMFDVER and 16–21% on COMASK20. FAR inflation also occurred for reconstructed unmasked comparisons, and transfer between the external datasets was either too permissive or too strict. Model replacement and synthetic-mask adaptation improved AUC and true acceptance rate (TAR) but did not restore the nominal FAR, and replacing the detection–alignment path changed AUC substantially while leaving FAR nearly unchanged. Recalibration with 25–100 target identities lowered median FAR, but optimistic independent-pair upper bounds at the median split were ≤1% in only 4 of 12 settings, so near-target results remain uncertain. In an exploratory fixed-holdout analysis, unmasked calibration was conservative, whereas condition-matched calibration recovered TAR but left FAR near the target. Within the tested protocols, the results indicate that recognition accuracy and decision-threshold robustness should be evaluated separately and that deployment requires site-specific threshold validation.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: