Mikhail Shkorin, Machine Learning Engineer, R&D, Verigram.
This is a story about how the team struggled with a seemingly simple task: on a frame, are the person's eyes closed or open, that is, is the person asleep or not. They came back to the task three times, each time about six months later. The talk covers: why this counts as an attack on a verification system, what they did to make their data clean and usable, how they used synthetic data, and what you can take away if you work with dirty data or synthetics, or train lightweight specialized models. The author previously talked about how they defend against video-substitution attacks (there are talks by Verigram on YouTube); they have almost described their whole biometrics system bit by bit already.
Google or a one-time email code. No password.