Silent Face Anti-Spoofing Detection Algorithm Based on Meta-learning and Attention Mechanism

In the field of facial identity authentication, silent face anti-spoofing detection, as a key technology, to prevent fake photos and videos, is receiving widespread attention. To address the issue of low detection accuracy in existing algorithms that similar tasks have not appear in training, silent face anti-spoofing detection algorithm based on meta-learning and attention mechanism is proposed. First, by leveraging meta-learning to quickly adapt to shot samples, the proposed algorithm uses MAML as the main network to capture the commonalities between training tasks, allowing it to quickly adjust parameters in new scenarios or new attack types to achieve efficient detection. Then, in order to focus the model on extracting and analyzing the key features of facial regions, an attention mechanism is introduced in MAML to reduce the interference from the facial background region. Finally, a Fourier spectrum aided supervision branch is designed to improve the accuracy of classification by using the difference of the Fourier spectrum of real/fake face in the frequency domain. Experimental results show that the proposed algorithm achieves good accuracy on public NUAA dataset, with AUC and F1 scores of 0.99 and 0.90, respectively.