NAS-FAS: Static-Dynamic Central Difference Network Search for Face Anti-Spoofing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face anti-spoofing (FAS) is crucial for secure face recognition.
• Current FAS methods often rely on hand-designed networks, limiting performance.
• Existing neural architecture search (NAS) primarily focuses on classification tasks, not specialized domains like FAS.

Purpose of the Study:

• To develop the first FAS method utilizing neural architecture search (NAS).
• To address the challenges of network generalization across different acquisition conditions and spoofing attack types in FAS.
• To create task-aware neural architectures optimized for face anti-spoofing.

Main Methods:

• Proposed NAS-FAS, a novel FAS method based on neural architecture search (NAS).
• Developed a specialized search space incorporating central difference convolution and pooling operators.
• Utilized a static-dynamic representation for spatio-temporal discrepancy mining.
• Introduced Domain/Type-aware Meta-NAS for robust cross-domain and cross-type knowledge transfer.
• Released the CASIA-SURF 3DMask dataset for evaluating NAS transferability.

Main Results:

• NAS-FAS achieved state-of-the-art performance on nine FAS benchmark datasets.
• Demonstrated superior performance across four different testing protocols.
• Validated the effectiveness of the proposed search space and meta-learning approach.
• Showcased strong generalization capabilities for cross-dataset and cross-type evaluations.

Conclusions:

• NAS-FAS represents a significant advancement in face anti-spoofing by leveraging NAS.
• The proposed method and dataset facilitate more robust and generalizable FAS systems.
• This work highlights the potential of NAS for specialized biometric security tasks.