Enhancing Ensemble Learning Using Explainable CNN for Spoof Fingerprints.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in classification but lack interpretability, raising concerns about reliability with limited data.
• Ensemble learning with multiple CNNs can enhance robustness, but often at the cost of accuracy.
• Limited training data poses challenges for CNN reliability and robustness.

Purpose of the Study:

• To propose a novel CNN training method enhancing both accuracy and robustness.
• To address the interpretability and reliability issues in CNN-based classification systems.
• To improve CNN performance on limited training datasets through a new ensemble approach.

Main Methods:

• Utilized Class Activation Maps (CAMs) to identify influential regions in previously trained CNNs.
• Concealed identified 'fingerprint' regions during the training of new CNNs with identical architectures.
• Ensembled resultant networks to ensure comprehensive feature consideration for classification.

Main Results:

• Achieved significant enhancement in classification accuracy and robustness across multiple sensors.
• Demonstrated state-of-the-art accuracy on LivDet datasets.
• The novel training method improved reliability and consistency in CNN predictions.

Conclusions:

• The proposed CAM-based training method effectively improves CNN accuracy and robustness.
• This approach mitigates the accuracy-robustness trade-off often seen in ensemble methods.
• The technique offers a reliable solution for CNN-based classification, especially with limited data.