From the Public Safety edition of October 29, 2026
Peer-reviewed University of Pennsylvania experiment finds facial recognition error rates shift sharply with blur, low resolution and lighting in police-style images, with persistent race and gender disparities
University of Pennsylvania (Department of Criminology; Wharton Statistics and Data Science; School of Engineering) · Policing / facial recognition accuracy and forensic reliability · United States
- Publisher
- PLOS One
- Original publication
- PLOS One, Oct. 21, 2026
- Source retrieved
- 2026-10-30
- Event date
- 2026-10-21
What happened
Cuellar, To and Mehrotra generated 50,000 synthetic faces with StyleGAN3, labelled them with FairFace, and built race-and-gender-balanced samples (a 167-face control database and 167 treated probes, 83 target-present and 84 target-absent). They degraded probes by contrast, brightness, motion blur, pose and resolution and ran 1:n identification with DeepFace using ArcFace. Low resolution, high blur and contrast, and low brightness substantially increased the false negative identification rate (FNIR), with reduced resolution the largest effect. At the selected operating point the false positive identification rate (FPIR) was substantially higher than FNIR and highest near baseline image quality, generally falling as degradation increased. Disparities persisted: FPIR was often higher for Black individuals, especially Black women, while the group with the highest FNIR varied by degradation type. The authors argue FRT should not be summarised with one accuracy figure and should be evaluated on the low-quality images used in investigations.
Why it matters
Gives police agencies, prosecutors and courts a peer-reviewed basis for asking for image-quality-specific error rates before relying on a facial recognition lead, especially as courts begin to require disclosure of FRT use.
Evidence and measured results
Opened the PLOS One article and read the methods, results, discussion and limitations sections in full. Peer-reviewed journal article; data are deposited at Harvard Dataverse (not opened).
Limitations and uncertainty
Synthetic faces and controlled single-factor degradations, which may not capture real surveillance imagery; one open-source pipeline (DeepFace/ArcFace), not commercial police systems; small 167-face gallery; FairFace labels and simplified race categories; FFHQ-derived faces skewed white before balancing. U.S. academic study.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-30; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Police and prosecutors rely on facial recognition leads from degraded footage without knowing how error rates change with image quality.
- Stakeholders
- Police FRT or real-time crime centre units
- Prosecutors and defence counsel
- Courts
- Civil rights oversight bodies
- Discovery questions
- Which product, threshold and gallery size does the agency use?
- Are probe images logged with quality information?
- How are FRT leads disclosed in court?
- Value hypothesis
- Condition-specific validation can reduce wrongful leads and strengthen court-defensible records.
- Potential engagement
- Validation protocol and image-quality gate design for the FRT workflow.
- Unsupported claims
- Error rates for any commercial police FRT product
- That quality scoring eliminates bias
Pre-sales engineering
Role takeaway
- Fit
- Fits agencies already using FRT for investigative leads.
- Architecture and integration
- Image-quality scoring before search; logging of score, threshold and candidate list with the case record.
- Prerequisites
- FRT search logs
- Sample of resolved cases
- Image-quality metrics
- Constraints
- Synthetic faces
- Single open-source pipeline
- Small gallery
- Single-factor degradations
- Security
- Search logs contain biometric data and must stay in CJIS-compliant systems with access controls.
- Proof of value
- Retrospective scoring of past probes against resolved outcomes.
Delivery
Role takeaway
- Work
- Add quality scoring, condition logging and supervisor review to the FRT workflow.
- Dependencies
- Vendor export of search logs
- Prosecutor disclosure requirements
- Policy approval
- Ownership
- Agency FRT unit with the prosecutor’s office
- Skills and adoption
- Analyst training on how image quality shifts false positive and false negative risk.
- Governance checkpoints
- Policy update approval
- Periodic audit of low-quality searches
- Disclosure review with prosecutors
- Acceptance criteria
- All searches logged with quality scores
- Review threshold documented
- Disclosure-ready records
- Risks
- False confidence in quality scores
- Vendor resistance to log export
- Over-generalising study numbers
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Suggests attaching image-quality measures (resolution, blur, brightness, contrast, pose) to each probe and reporting condition-specific error rates rather than a single vendor benchmark figure.
Governance
Who approves, reviews and stays accountable for outcomes?
Supports policies that treat FRT results as investigative leads only and that require validation on agency-like image conditions before use or court reliance.
Security and privacy
What data, permissions and controls need testing?
False positives, higher for some demographic groups, translate into risk of wrongful suspicion; the authors cite at least seven confirmed misidentification cases.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Human reviewers of candidate lists need training on how image quality changes the error profile.
Procurement
What should contracts, pricing and exit terms secure?
Agencies can require vendors to provide error rates under degraded-image conditions and by demographic group, not only NIST-style high-quality benchmarks.
Operating model
Which teams own the service once it runs?
Moves validation from one-time vendor selection to per-image judgement about whether a probe is fit for search.
Publication history
- 2026-10-29Public Safety · Issue 537 resources
Stable resource ID: penn-plos-one-facial-recognition-degraded-police-images-error-rates-20261021