A high-performance adaptive fusion network for face anti-spoofing detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Current cross-domain face liveness detection models struggle with generalization and deep feature representation.
• Existing methods face limitations in accurately distinguishing real faces from spoofing attacks across diverse datasets.

Purpose of the Study:

• To propose a high-performance adaptive fusion network for face anti-spoofing detection.
• To enhance generalization ability and deep feature representation in cross-domain face liveness detection.

Main Methods:

• Introduced a face depth map fusion mechanism combined with a ResNet-18 backbone for feature extraction.
• Designed a content feature extraction architecture using dynamic convolution and a bottleneck attention module.
• Employed adaptive instance normalization, central difference convolution, and adversarial training for dual alignment of multi-source domain features and categories.

Main Results:

• The proposed method significantly outperforms existing algorithms on four benchmark datasets (OULU-NPU, MSU-MFSD, CASIA-FASD, ReReplay Attack).
• Demonstrated improved performance in cross-domain face liveness detection by effectively alleviating data distribution differences.

Conclusions:

• The adaptive fusion network offers an innovative technical path for cross-domain face liveness detection.
• The method provides a robust solution for face anti-spoofing by overcoming limitations in generalization and feature representation.