As a user on the Reddit discussion suggest, adding a liveness detection feature would enable the users of the project to build solutions like attendance marking or face unlock systems.

• Using smile or blink detection may not be reliable, as a video containing spoof smiling/blinking faces may be played in front the camera. Moreover, Mediapipe's Face Detector does not provide smile or blink detection (MLKit does).

• CNN-based models trained on real and spoof images can be used only if they are lightweight (as it needs to execute directly on camera frames) and support execution on TensorFlow Lite.

Projects

• https://github.com/sakethbachu/Face-Liveness-Detection
• https://github.com/minivision-ai/Silent-Face-Anti-Spoofing/tree/master/resources/anti_spoof_models (models are in the PyTorch format, but can be converted easily to the TF format following TFLite conversion)