Facebook DeepFace: the social network's face recognition AI

Facebook was one of the first social networks to develop its own artificial intelligence for verifying and recognizing faces. DeepFace, created by Facebook researchers and presented at the CVPR computer vision conference in 2014, achieved 97.35% accuracy on the well-known Labeled Faces in the Wild (LFW) dataset, very close to human-level performance on the same benchmark (97.53%).

DeepFace aligned each face in 3D so that it looked straight at the camera, then passed it to a nine-layer neural network with more than 120 million connection weights, trained on about four million images uploaded by Facebook users.

Facebook used face recognition to suggest whom to tag in photos. From December 2017, its Photo Review feature alerted users when someone uploaded a photo of them and when someone used their photo as a profile picture, a common sign of impersonation. In 2019 these features were brought under a single Face Recognition setting, turned off by default for new users.

Why Facebook switched face recognition off

On November 2, 2021, Meta announced that it was shutting down the Face Recognition system on Facebook and deleting more than a billion people’s facial recognition templates; more than a third of Facebook’s daily active users had opted in. Meta cited growing societal concerns and the lack of clear rules from regulators. The system also cost Meta dearly in court, including a $1.4 billion settlement with Texas in July 2024.

Face recognition at Meta today

Meta now uses face matching for narrower purposes. Since October 2024 it has tested it to detect scam ads that misuse public figures’ faces and to help people regain access to compromised accounts with a video selfie. In July 2026 it launched Facebook Verified, a free badge showing that a profile belongs to a real person, confirmed with a short video selfie that Meta checks against the profile photos. Meta also uses AI for other checks, such as estimating users’ ages.

On the research side, DeepFace was soon overtaken: Google’s FaceNet, published in 2015, achieved 99.63% accuracy on the same dataset.