Abstract
In view of the limitations of current cross-domain face liveness detection models in generalization ability and deep feature representation, this paper proposes a high-performance adaptive fusion network for face anti-spoofing detection. This method innovatively introduces the face depth map fusion mechanism, combines the ResNet-18 backbone network to extract common features in multiple source domains, strengthens the distinguishable feature capture between real faces and spoofing attacks. Meanwhile, a content feature extraction architecture of "dynamic convolution + bottleneck attention module" is designed. It is combined with adaptive instance normalization and central difference convolution for collaborative optimization, breaking through the bottleneck of insufficient representation of depth details in traditional feature extraction. Finally, through adversarial training of domain discriminators, dual alignment of multi-source domain features and categories is achieved, effectively alleviating the problem of cross-domain data distribution differences. A large number of training and test results conducted on the four benchmark datasets of OULU-NPU, MSU-MFSD, CASIA-FASD and ReReplay Attack show that the performance of the proposed method is significantly better than that of the existing algorithms, providing a more innovative technical path for cross-domain face liviality detection.