Reliable and Balanced Transfer Learning for Generalized Multimodal Face Anti-Spoofing

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face Anti-Spoofing (FAS) is critical for secure face recognition systems.
• Multimodal FAS systems leverage diverse sensors but face challenges like modality unreliability and imbalance.
• Existing methods struggle with generalization to unseen attacks and varied environments.

Purpose of the Study:

• To propose MMDG++, a multimodal domain-generalized FAS framework.
• To enhance the reliability and balance of multimodal fusion in FAS.
• To improve the generalization capability of FAS systems against diverse presentation attacks.

Main Methods:

• Developed MMDG++, a framework utilizing the CLIP vision-language model.
• Introduced Uncertainty-Guided Cross-Adapter++ (U-Adapter++) to filter unreliable modality regions.
• Implemented Rebalanced Modality Gradient Modulation (ReGrad) for adaptive gradient balancing.
• Designed Asymmetric Domain Prompts (ADPs) to leverage CLIP's language priors for generalized decision boundaries.

Main Results:

• MMDG++ demonstrates superior generalization capability on a novel multimodal FAS benchmark.
• The proposed methods effectively address modality unreliability and imbalance.
• Experimental results show outperformance compared to state-of-the-art FAS methods.

Conclusions:

• MMDG++ offers a robust solution for multimodal domain-generalized face anti-spoofing.
• The framework enhances the reliability and balance of multimodal fusion for improved security.
• The study highlights the potential of vision-language models in advancing FAS research.