With Eyes Wide Shut: Detecting Sleeping People and Other Remote Verification Attacks Without Consent

Mikhail Shkorin23 May 2026

Problem

Remote identity checks (KYC for loans, betting, government services) decide from a phone camera, in milliseconds and without an operator, whether a live and correct person is in the frame. A sleeping person defeats this: the camera is real, the face is alive and matches the database, so liveness and face matching both pass. Someone nearby, such as a roommate or relative, can then take out a loan in the sleeper's name. Texture-based presentation attack detection cannot help, because the face is genuine.

Example from the article: in a dormitory, barracks, hostel or rented flat, someone opens the lending app, points the camera at a sleeping neighbor and applies for a loan.

Solution in brief

Add a separate lightweight single-frame model that answers one question: are the eyes closed. It is trained on cleaned real frames plus synthetic closed and open eyes, and evaluated by the false attack rate at a fixed 5% false reject rate. A second, independent face-occlusion detector covers cases where drawn-on eyes fool the first model.

Questions this case answers
  • How to detect whether a person is coerced or not attentive during remote identity verification from a single frame
  • How to use synthetic images for one class without teaching the model to detect the generator
  • How to choose a decision threshold when false rejections for honest users are capped
  • How to detect a sleeping or coerced person during remote face verification from a single frame

Tools and methods

  • Eye-strip classifier — MobileNetV1 on a crop of the eyes region, binary open / closed.
  • FacePoke — open face editor that closes and opens eyes via facial keypoints, used for synthetic data.
  • FAR @ FRR = 0.05 — target metric: attacks passed at a cap on blocked honest users.
  • Face occlusion detector — second MobileNet, 7 classes, trained on separate data.
Continue reading — it's free

Google or a one-time email code. No password.