Learning Knowledge-Based Prompts for Robust 3D Mask Presentation Attack Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• 3D mask presentation attacks threaten face recognition systems.
• Existing detection methods using multimodal features or remote photoplethysmography (rPPG) are costly and have limited generalization.
• Detection-related text descriptions are cost-effective but underutilized for this task.

Purpose of the Study:

• To explore the potential of vision-language multimodal features for 3D mask presentation attack detection.
• To propose a novel knowledge-based prompt learning framework to enhance detection performance and generalization.
• To address the limitations of current methods by leveraging cost-effective text descriptions and advanced AI techniques.

Main Methods:

• Developed a knowledge-based prompt learning framework incorporating knowledge graph entities and triples.
• Introduced a visual-specific knowledge filter using attention mechanisms to refine knowledge graph elements based on visual context.
• Leveraged causal graph theory and a spurious correlation elimination paradigm during training to enhance generalization.

Main Results:

• The proposed framework effectively harnesses knowledge from pre-trained vision-language models.
• The method demonstrates strong generalization capabilities for 3D mask presentation attack detection.
• Achieved state-of-the-art performance in both intra- and cross-scenario detection on benchmark datasets.

Conclusions:

• Vision-language models, enhanced with knowledge graphs and causal reasoning, offer a powerful approach for 3D mask presentation attack detection.
• The proposed framework provides a cost-effective and highly generalizable solution compared to existing methods.
• This research opens new avenues for securing face recognition systems against sophisticated spoofing attacks.