Momentum Contrast for Representative Face Presentation Attack Detection

With the widespread usage of using face recognition systems, they became vulnerable to presentation attacks encountered by attackers. To tackle this issue, face presentation attack detection (PAD) methods are implemented. However, these methods have several shortcomings including the generalizability of unknown attacks. This thesis targets two main problems that face PAD methods. The first problem that this work target to solve is databases annotation problems. Annotating databases with labels is time-consuming, to solve this problem, a representative learning model (MoCo framework in this thesis) is used as it focuses on unsupervised learning databases. The second problem that this work target is the insufficient PAD data. Most PAD databases are manually collected especially presentation attack samples, thus they are labor-intensive and small-scale. This thesis target this problem by training the model on a face recognition database such as CASIA-Web database which is a very large-scale public facial recognition database, not a PAD database, which is collected randomly in the wild where images are diverse from illumination, sensors, identity. This work proves that using face recognition databases to learn face representation, can be adapted to be used in detecting presentation attacks and the model can benefit from using extra existing face recognition data besides the model becomes more familiar with diverse setups and illuminations within face images. Finally, the classification model suggested by the state-of-art MoCo, is extended by applying pseudo labeling to it, which improved the general results.