Gaze | Facial authentication for Linux (sudo, lock screen, GDM), on-device

We built Gaze (v0.2.1), a facial authentication system for Linux. It hooks into PAM, so it works for sudo, the lock screen, and GDM login. Everything runs on-device. No cloud, no account.

It's been a slow build. v0.1.0 took about 3 months, and v0.2.0 landed roughly a month and a half after that. We'd rather get things right than ship fast.

What it does:

• PAM integration for sudo, polkit, and any PAM-aware auth stack
• Liveness anti-spoofing on by default. A local MiniFASNet-V2 model checks the face crop after a match on the RGB path, and eye-motion analysis handles the IR path.
• Infrared camera support for Windows Hello-style IR cameras, including driving the IR emitter. You can enroll RGB and IR templates and combine them with hybrid policies: require both, either one, or fall back to IR when it's too dark.
• GNOME Shell extension for the lock screen, plus optional face unlock at GDM login
• Works with Hyprland (hyprlock), KDE Plasma, and LXQt alongside GNOME
• A libadwaita desktop app for enrolling, testing, and health checks
• DBus API (com.gundulabs.Gaze) if you want to build on top of it
• Multiple face profiles per user, so you can enroll a default, one with glasses, whatever varies
• gaze refine-face to sharpen recognition in dim light or at odd angles
• Security levels from low to maximum that swap the detector/recognizer models and match threshold, plus a custom level if you want to tune it yourself
• Optional TPM 2.0 template encryption that seals your face templates to the machine, so a stolen disk can't read them
• Multi-user support (gaze add-face work -u alice)
• gaze doctor for a quick health check, gaze uninstall for a clean removal
• Models download on first run and stay local under /var/cache/gaze

One-line install (Debian/Ubuntu/Fedora/Arch):

sh curl -fsSL https://gaze.gundulabs.com/install.sh | sh

One honest caveat: face recognition isn't perfect, so false accepts and rejects happen. Liveness and an IR camera raise the bar, but keep a password as a fallback and don't rely on Gaze as your only factor.

Repo: https://github.com/GunduLabs/gaze Docs: https://gaze.gundulabs.com