How does facial recognition technology actually work?

The answer

Modern facial recognition systems map 68 or more landmark points on a face - such as the distance between eyes, jawline contour, and nose bridge width - creating a unique mathematical 'faceprint.' Deep learning algorithms convert this into a numerical vector (embedding) that can be compared against a database in milliseconds. Current systems achieve over 99% accuracy under ideal conditions, but performance drops significantly with poor lighting, angled faces, or darker skin tones due to training data biases.