Fusion Methods for Face Presentation Attack Detection.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Face presentation attacks (PA) pose a significant threat to face recognition (FR) systems, as they are easily executed and challenging to detect.
• Existing methods, including deep learning and traditional feature engineering, show effectiveness but may not fully capture all relevant discriminative information.
• The optimal approach for PA detection remains an open question regarding feature representation.

Purpose of the Study:

• To investigate whether deep neural networks adequately learn traditional low-level features for optimal presentation attack detection.
• To propose and evaluate a simple feature-fusion method combining deep learning features with traditional color and texture features.
• To demonstrate the benefits of an enriched feature space for improving PA detection rates.

Main Methods:

• A feature-fusion strategy was developed to integrate features from pre-trained deep learning models with traditional color and texture features.
• The proposed method was evaluated on three public datasets: CASIA, Replay Attack, and SiW.
• Performance was assessed based on detection rates achieved by the fused feature space.

Main Results:

• Extensive experiments demonstrated that enriching the feature space by fusing deep learning and traditional features significantly improves detection rates.
• The proposed simple feature-fusion method proved effective in enhancing the performance of presentation attack detection.
• Results confirm the benefit of combining diverse feature types for robust PA detection.

Conclusions:

• Combining features from deep learning models with traditional color and texture features is beneficial for improving face presentation attack detection.
• The proposed feature-fusion method offers a practical approach to enhance the security of face recognition applications against spoofing.
• Future research should explore novel characterizing features and advanced fusion strategies for even more robust PA detection.