Silent-Face-Anti-Spoofing

Why we included this project

Face liveness is the step that decides whether the person in front of the camera is actually there rather than a printed photo, a screen showing a video, a silicone mask, or a 3D likeness. This project from Minivision publishes the full training pipeline for a silent liveness model, so the user just looks at the camera and never has to follow a scripted set of prompts. The network is a heavily pruned MobileFaceNet variant with an auxiliary Fourier-spectrum supervision branch, cut from 0.224G to 0.081G FLOPs with modest accuracy loss. Measured speeds run from about 19ms on a Kirin 990 phone to 90ms on a low-end Rockchip RK3288 board, which makes on-device deployment realistic. Along with training and test scripts, the repo includes an Android APK and its deployment code, so you can try the detector on a real phone quickly, and the README spells out the limits: results depend on the camera model, the face needs to be roughly upright, and test images must be captured live with a camera rather than downloaded.