Meta-Teacher For Face Anti-Spoofing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face anti-spoofing (FAS) is crucial for securing face recognition systems against presentation attacks (PAs).
• Current FAS methods often rely on handcrafted labels, which may not fully capture intrinsic spoofing cues.
• Existing teacher-student models prioritize student accuracy over the teacher's ability to impart knowledge.

Purpose of the Study:

• To develop a more effective supervision method for presentation attack detectors in FAS.
• To train a meta-teacher model capable of providing superior guidance compared to traditional labels and existing models.
• To enhance the learning of intrinsic spoofing cues by presentation attack detectors.

Main Methods:

• Proposed a novel Meta-Teacher FAS (MT-FAS) method utilizing bi-level optimization.
• Trained a meta-teacher to supervise presentation attack detectors (students) through a lower-level training process.
• Optimized the meta-teacher's teaching performance by minimizing the detector's validation loss in a higher-level training process.

Main Results:

• The trained meta-teacher provided more suitable supervision than handcrafted labels and existing teacher-student models.
• The MT-FAS method significantly improved the performance of presentation attack detectors.
• Experiments on five FAS benchmarks validated the effectiveness of the proposed approach.

Conclusions:

• The MT-FAS method offers a superior approach to supervising presentation attack detectors.
• Bi-level optimization enables the meta-teacher to learn effective teaching strategies for FAS.
• This work advances the field of face anti-spoofing by enhancing detector performance and security.