Domain Adaptation Based on ResADDA Model for Face Anti-Spoofing Detection

Different datasets have more apparent differences due to lighting, background and image quality issues, which makes the generalization problem of face anti-spoofing detection more prominent. A domain adaptive method for face spoofing detection based on ResADDA model is proposed, which adopts the ResNet34 network to extract deep convolutional features, and draws on the GAN network idea to use adversarial training by alternately optimizing the domain discriminator and feature encoder, adjusting the parameters of the target domain feature encoder and reducing the difference of feature distribution between the target domain and the source domain to improve the detection ability of the model on the target domain. Crossover experiments on the publicly available dataset CASIA-FASD and Replay-Attack are conducted to verify the effectiveness of the ResADDA model which is superior to other methods.