face recognition

Face recognition identifies or verifies individuals by matching facial features against enrolled identities using deep learning embeddings. Pipeline: (1) detection (locate faces—MTCNN, RetinaFace), (2) alignment (normalize pose, rotation—landmark-based affine transform), (3) embedding extraction (map face to compact vector—ArcFace, CosFace, FaceNet), (4) matching (cosine similarity or L2 distance against gallery). Training: metric learning with angular margin losses (ArcFace: additive angular margin on softmax) on large-scale datasets (MS1M, WebFace). Accuracy: >99.8% on LFW benchmark, but performance varies significantly across demographics. Applications: device unlock (1:1 verification), surveillance (1:N identification), access control, and photo organization. Privacy concerns: mass surveillance, consent, data protection (GDPR, BIPA), and function creep. Bias: higher error rates for darker skin tones and women documented in multiple studies (Buolamwini, Gender Shades). Mitigations: balanced training data, fairness constraints, and regulatory frameworks. Liveness detection (anti-spoofing) prevents attacks using photos, videos, or 3D masks.