Facial recognition that never phones home

What it does

FaceAISDK_Android bundles face detection, recognition, liveness anti-spoofing, and both 1:1 and 1:N (or M:N) face search into a single offline Android SDK. It targets Android 5 through 16, works with system cameras or UVC USB cameras, and keeps all inference local — no cloud round-trips, no stored face data leaving the device.

The interesting bit

The SDK treats “offline” as a hard constraint, not a fallback mode. That makes it viable for environments with spotty connectivity (rural access control, embedded kiosks) or strict data-sovereignty requirements (finance, healthcare-adjacent apps). The liveness detection covers both passive and active checks — blink, smile, nod, shake — which matters if you’re trying to stop someone from unlocking a door with a printed photo.

Key highlights

• Fully on-device: no network required, no face data uploaded or stored externally
• Supports both system cameras and UVC-protocol USB cameras for custom hardware
• Liveness detection includes silent (passive) and action-based (active) modes
• 1:1 verification for auth scenarios; 1:N/M:N search for identification against a local face database
• Cross-platform SDK family: iOS, Flutter, React Native, and uniApp wrappers exist
• Distributed via Maven Central with modular demo app showing verify, search, and face-enrollment flows

Caveats

• The README notes a hardware requirement: cameras need “wide dynamic value >105Db” for acceptable results, which may exclude cheap or poorly lit setups
• Documentation is primarily in Chinese; English developers will need to lean on code samples and PDF references
• The project appears to be a vendor SDK with a demo wrapper rather than an open-core library — the actual FaceSDKLib module is a black box

Verdict

Worth evaluating if you’re building access-control, attendance, or device-auth apps in China or similar privacy-regulated markets where cloud biometrics are a non-starter. Skip it if you need transparent model weights, heavy customization of recognition algorithms, or first-class English documentation.