Learning to Learn Face-PAD: a lifelong learning approach

A face presentation attack detection (face-PAD) system is in charge of determining whether a face corresponds to a presentation attack or not. The vast majority of proposed solutions consider a static scenario, where models are trained and evaluated in datasets where all types of attacks and conditions are known beforehand. However, in a real-world scenario, the situation is very different. There, for instance, the types of attacks change over time, with new impersonation situations appearing for which little training data is available. In this paper we propose to tackle these problems presenting for the first time a con-tinuallearning framework for PAD. We introduce a continual meta-learning PAD solution that can be trained on new attack scenarios, following the continual few-shot learning paradigm, where the model uses only a small number of training samples. We also provide a thorough experimental evaluation using the GRAD-GPAD benchmark. Our results confirm the benefits of applying a continual meta-learning model to the real-world PAD scenario. Interestingly, the accuracy of our solution, which is continuously trained, where data from new attacks arrive sequentially, is capable of recovering the accuracy achieved by a traditional solution that has all the data from all possible attacks from the beginning. In addition, our experiments show that when these traditional PAD solutions are trained on new attacks, using a standard fine-tuning process, they suffer from catastrophic forgetting while our model does not.