Facial recognition uses an image or video to create a mathematical representation of facial features and compare it with enrolled templates. A result is normally a similarity score evaluated against a threshold, not an absolute statement that two people are the same.
Detection, verification, and identification
Face detection locates a face in an image. It does not identify the person.
Verification (one-to-one) asks whether a captured face matches one claimed identity, such as a device owner.
Identification (one-to-many) searches a gallery for possible matches and is more likely to require human review and strong governance.
Face analysis may estimate attributes or quality but should not be confused with identity matching.
How the matching pipeline works
Capture: a camera obtains an image or video under particular lighting, angle, resolution, and compression.
Quality and alignment: software finds the face and normalizes pose, scale, and landmarks.
Template creation: an algorithm converts features into a numeric representation.
Comparison: the probe template is compared with an enrolled template or a gallery.
Decision: policy applies a threshold and may require another factor or human review.
A template is not an ordinary photograph, but it is sensitive biometric data and may support recognition or correlation. It needs protection, retention limits, and a deletion process.
Accuracy and decision thresholds
A false match incorrectly links different people. A false non-match fails to recognize the enrolled person. Tightening a threshold usually reduces false matches while increasing false non-matches. The appropriate balance depends on whether the system unlocks a personal device, admits someone to a building, deduplicates records, or generates an investigative lead.
Performance can change with image quality, aging, pose, expression, lighting, camera, occlusion, demographic group, gallery size, and operational workflow. A laboratory accuracy figure does not guarantee performance in a different population or environment.
Security threats
Presentation attack: a photo, screen, mask, replay, or synthetic media is presented to a sensor.
Injection attack: manipulated media bypasses the physical sensor and enters the processing pipeline.
Morphing: one image is blended from multiple people to support identity fraud.
Template or account theft: enrollment records, administrative access, or recovery paths are compromised.
Poisoned enrollment: the wrong person or poor evidence becomes the trusted reference.
NIST discusses presentation attacks and face morphing in its current overview of facial-recognition security.
Privacy and governance questions
What is the specific purpose and legal basis, and is a less intrusive method adequate?
Were people clearly informed, and can they use a practical alternative where required?
Are original images, templates, candidate lists, audit records, and derived attributes retained only as needed?
Which provider, employee, or agency can search, export, or share the data?
How are accuracy, demographic effects, appeals, human review, deletion, and misuse audited?
Security controls
combine high-risk verification with another independent factor and secure recovery;
use tested presentation-attack detection and protect the sensor-to-matcher path;
encrypt templates, separate identifiers, restrict administrator access, and log searches and exports;
test the intended population and environment at documented thresholds;
treat identification results as leads requiring corroboration, not automatic adverse decisions;
maintain incident response and notification procedures for biometric compromise.
Facial recognition FAQ
Can a face be changed like a password?
Not easily. That makes template protection, revocable representations, and alternative factors especially important.
Does liveness detection stop every spoof?
No. It reduces selected presentation attacks and must be tested against the sensor, attack types, accessibility needs, and operating environment.