Accuracy and fairness of facial recognition technology in real-world, low-quality police images: An experiment
- 1Department of Criminology, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America.
- 2Department of Statistics and Data Science, Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America.
- 3University of Pennsylvania, School of Engineering and Applied Science, Philadelphia, Pennsylvania, United States of America.
Abstract
Facial recognition technology (FRT) is increasingly used in criminal investigations, yet most evaluations of its accuracy rely on high-quality images, unlike those often encountered by law enforcement. This study examines how five common forms of image degradation (contrast, brightness, motion blur, pose shift, and resolution) affect FRT accuracy, and how that varies by gender and race. Using synthetic faces generated by StyleGAN3 and labeled with FairFace, we simulate degraded images and evaluate performance using the DeepFace framework with the ArcFace model in 1:n identification tasks. By using simulated faces, we are able to estimate conditional FPIR and FNIR error-rate curves across degradation levels under controlled conditions, providing practical guidance on how database search performance changes as questioned image quality deteriorates. We find that high motion blur and contrast and low resolution and brightness substantially increase FNIR. At the selected operating point, FPIR is substantially higher than FNIR and is highest near baseline image quality, generally decreasing as image degradation increases and demographic disparities persist across degradation conditions. Demographic disparities vary by error type: FPIR is often higher for Black individuals, particularly Black women, whereas the demographic groups with the highest FNIR differ across degradation conditions. These disparities raise concerns about fairness and reliability of the use of FRT. These findings highlight the importance of evaluating algorithmic performance using the low-quality images encountered in criminal investigations to ensure both fairness and validity.
Related
- Shaping metabolite profile, dehydration kinetics, quality retention, and sorption properties of red water lily (Nymphaea x rubra) petals through drying technology
- Long-term impact of the diagnosis on quality of life, social skills, and person recognition strategies in congenital prosopagnosics: A cross-sectional study
- Associations of Word Recognition in Noise with Perceived Hearing Handicap and Quality of Life in Older Adults with Sensorineural Hearing Loss.
- Social inequity aversion and fairness preference in polar bears
- Fairness decision-making in younger and older adults and associated neural indices
BibTeX
@article{cuellar2026accuracy,
title = {{Accuracy and fairness of facial recognition technology in real-world, low-quality police images: An experiment}},
author = {Maria Cuellar and Hon Kiu James To and Arush Mehrotra},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0353793},
url = {https://doi.org/10.1371/journal.pone.0353793},
note = {CC BY},
}
RIS
TY - JOUR
TI - Accuracy and fairness of facial recognition technology in real-world, low-quality police images: An experiment
AU - Maria Cuellar
AU - Hon Kiu James To
AU - Arush Mehrotra
PY - 2026
DA - 2026/10/21
JO - PLOS One
DO - 10.1371/journal.pone.0353793
UR - https://doi.org/10.1371/journal.pone.0353793
AB - Facial recognition technology (FRT) is increasingly used in criminal investigations, yet most evaluations of its accuracy rely on high-quality images, unlike those often encountered by law enforcement. This study examines how five common forms of image degradation (contrast, brightness, motion blur, pose shift, and resolution) affect FRT accuracy, and how that varies by gender and race. Using synthetic faces generated by StyleGAN3 and labeled with FairFace, we simulate degraded images and evaluate performance using the DeepFace framework with the ArcFace model in 1:n identification tasks. By using simulated faces, we are able to estimate conditional FPIR and FNIR error-rate curves across degradation levels under controlled conditions, providing practical guidance on how database search performance changes as questioned image quality deteriorates. We find that high motion blur and contrast and low resolution and brightness substantially increase FNIR. At the selected operating point, FPIR is substantially higher than FNIR and is highest near baseline image quality, generally decreasing as image degradation increases and demographic disparities persist across degradation conditions. Demographic disparities vary by error type: FPIR is often higher for Black individuals, particularly Black women, whereas the demographic groups with the highest FNIR differ across degradation conditions. These disparities raise concerns about fairness and reliability of the use of FRT. These findings highlight the importance of evaluating algorithmic performance using the low-quality images encountered in criminal investigations to ensure both fairness and validity.
SN - 1932-6203
ER -