DGPDL: Domain-Guided Prompt Distribution Learning for Generalizable Face Anti-Spoofing

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Overfitting to source domain signals hinders face anti-spoofing generalization.
• Current methods improving source domain diversity offer limited benefits for unseen target domains.
• Domain gaps cause understanding bias in face anti-spoofing models.

Purpose of the Study:

• To propose a novel Domain-Guided Prompt Distribution Learning (DGPDL) method.
• To alleviate understanding bias and improve domain generalization in face anti-spoofing.
• To leverage Vision-Language Models for a unified representation of domain signals.

Main Methods:

• Developed a learnable Domain-Specific Distribution (DSD) to connect various domain elements.
• Constructed optimal Domain-Specific Prompts (DSPs) using Prompt Assemble Attention (PAA) based on style statistics.
• Utilized assembled DSPs in both vision and language branches of Vision-Language Models.

Main Results:

• DGPDL effectively reduces reliance on specific domain appearances by representing domain signals as prompts.
• The model dynamically adapts to unseen target domains without retraining.
• Achieved state-of-the-art performance on several cross-domain face anti-spoofing benchmarks.

Conclusions:

• DGPDL offers a robust solution for face anti-spoofing domain generalization.
• Uniform prompt representation of domain signals ensures applicability to target domains.
• The proposed method significantly improves the robustness and adaptability of face anti-spoofing systems.