A Generalized Method for Silent Live Face Detection

Facial recognition technology is widely used in the field of public safety and plays an important role. However, with this comes the increasingly serious problem of facial attacks. Existing facial recognition systems are susceptible to attacks from presentation methods such as printed photos, video replays, makeup, and 3D masks. Therefore, in-depth research on live face detection technology is necessary. Currently, there are two mainstream live face detection methods used in industry. One is based on user cooperation to detect expressions or actions for judgment, which has a long verification time and poor user experience, making it unsuitable for large-scale deployment and use. The other relies on additional hardware (such as passive infrared and depth cameras) for multimodal live face detection, which has higher accuracy but requires strict image acquisition conditions and higher equipment costs, making it difficult to promote. This paper proposes a neural network-based silent live face detection method, which only requires the RGB image of the face to be collected to obtain the recognition result directly, without requiring the user to cooperate with any behavioral actions or introduce additional multimodal devices. Our network (referred to as TudouNet) significantly improves both accuracy and generalization ability, while having low computational complexity. Therefore, it can be deployed on resource-and energy-constrained embedded systems, achieving real-time processing.