How Face Recognition Works — and Why Storing Embeddings Beats Storing Faces

Face recognition involves four stages: detecting faces, aligning them, converting them into numerical embeddings, and matching those embeddings against stored ones using a similarity threshold. The key privacy insight is that embeddings — vectors of 128 to 512 numbers — can verify identity without revealing what a person looks like, making them far safer to store than raw images. In the FaceVision system described, detection and embedding run entirely in the browser using ONNX Runtime Web, so facial images never reach the server. This design simplifies consent, reduces breach risk, and makes deletion straightforward — a single database row. Engineers are also cautioned to test for bias across skin tones and age groups, use liveness detection to block photo spoofing, and set thresholds based on their own user data rather than generic benchmarks.