Presentation Attack Detection on Limited-Resource Devices Using Deep Neural Classifiers Trained on Consistent...

Area of Science:

• Speech Processing
• Biometrics
• Machine Learning

Background:

• Speaker verification systems face presentation attacks.
• Challenge-response schemes are used for detection.
• Resource-limited hardware poses challenges for complex models.

Purpose of the Study:

• To develop a novel approach for training convolutional phoneme classifiers.
• To enable efficient presentation attack detection on resource-limited devices.
• To improve phoneme recognition accuracy in simplified neural networks.

Main Methods:

• Utilized Deep Convolutional Neural Networks on Mel-Spectrograms.
• Developed a new training set construction method focusing on central phoneme articulation intervals.
• Employed bagging for ensembling simple classifiers.

Main Results:

• Achieved up to 76% accuracy in a 39-phoneme recognition task with simplified networks.
• Reduced within-class data scatter by optimizing training data selection.
• Ensembling improved accuracy by 2-3%, reaching 23% Phoneme Error Rate.

Conclusions:

• The proposed method allows reliable presentation attack detection on resource-limited hardware.
• Simplified yet effective neural architectures can achieve competitive performance.
• This approach is suitable for mobile and embedded biometric systems.