Taming Self-Supervised Learning for Presentation Attack Detection: De-Folding and De-Mixing.

Area of Science:

• Computer Science
• Biometrics
• Machine Learning

Background:

• Biometric systems face security risks from presentation attacks (PAs) using diverse instruments (PAIs).
• Existing presentation attack detection (PAD) methods, while advanced, often lack generalization to unseen PAIs.
• Model initialization's impact on PAD generalization is a critical, underexplored factor.

Purpose of the Study:

• To address the challenge of poor generalization in PAD for unknown PAIs.
• To propose a novel self-supervised learning method, DF-DM, that enhances PAD model generalization.
• To demonstrate the significance of model initialization in achieving robust PAD performance.

Main Methods:

• Introduced DF-DM, a self-supervised learning approach utilizing a global-local view.
• Implemented a 'de-folding' strategy to learn region-specific features and minimize generative loss.
• Employed a 'de-mixing' strategy to extract instance-specific features with global context via interpolation-based consistency.

Main Results:

• DF-DM significantly improves PAD performance on complex, hybrid datasets for both face and fingerprint modalities.
• Achieved an 18.60% equal error rate (EER) on OULU-NPU and MSU-MFSD datasets when trained on CASIA-FASD and Idiap Replay-Attack.
• Outperformed baseline methods by 9.54%, demonstrating superior generalization capabilities.

Conclusions:

• Model initialization is a critical factor for achieving generalization in PAD.
• The proposed DF-DM method effectively enhances PAD generalization by learning task-specific representations.
• DF-DM offers a promising direction for developing more robust and secure biometric systems against presentation attacks.