Algorithm of face anti-spoofing based on pseudo-negative features generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics Security

Background:

• Face anti-spoofing technology faces challenges from evolving attacker methods, diverse presentation modes, and limited negative sample data.
• Scarcity of negative samples leads to domain shift issues, hindering robust generalization in face anti-spoofing models.
• Existing methods struggle with cross-domain generalization, necessitating more effective approaches.

Purpose of the Study:

• To enhance face anti-spoofing system effectiveness through pseudo-negative sample analysis.
• To expand training datasets and improve cross-domain generalization capabilities.
• To develop a more robust feature classifier for broader attack defense.

Main Methods:

• A novel algorithm generates pseudo-negative features to enrich the training dataset.
• KL divergence loss is utilized to align generated pseudo-negative features.
• The enriched dataset aids in training a more robust feature classifier.

Main Results:

• Experiments on four public datasets (MSU-MFSD, OULU-NPU, Replay-Attack, CASIA-FASD) demonstrate positive results.
• The method shows effectiveness in both intra-dataset and cross-dataset evaluations.
• Superior performance was observed on smaller datasets, indicating strong generalization.

Conclusions:

• The proposed method effectively enhances face anti-spoofing capabilities by leveraging pseudo-negative samples.
• The approach demonstrates excellent generalization across different datasets and scenarios.
• This technique offers a promising solution for robust face anti-spoofing in real-world applications.