Efficient Attention Branch Network with Combined Loss Function for Automatic Speaker Verification Spoof Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Automatic Speaker Verification (ASV) systems are vulnerable to spoofing attacks.
• Current ASV countermeasures lack generalization to unseen attacks.
• Improving ASV spoof detection requires advancements in feature extraction, classification, and loss functions.

Purpose of the Study:

• To propose an Efficient Attention Branch Network (EABN) architecture for improved ASV spoof detection.
• To enhance model generalization to unseen spoofing attacks.
• To develop a robust countermeasure for speaker verification systems.

Main Methods:

• Developed the EABN architecture with attention and perception branches.
• Optimized EfficientNet-A0 for the perception branch, reducing parameters and computations.
• Employed a combined loss function for improved classification and detection.
• Utilized logPowSpec and LFCC features for feature extraction.

Main Results:

• Achieved EER = 0.86% and t-DCF = 0.0239 in the Physical Access (PA) scenario using logPowSpec.
• Achieved EER = 1.89% and t-DCF = 0.507 in the Logical Access (LA) scenario using LFCC and SE-Res2Net50.
• Demonstrated superior performance compared to existing single-system ASV spoofing countermeasures.

Conclusions:

• The EABN architecture significantly improves generalization to unseen spoofing attacks in ASV systems.
• The proposed method offers a highly effective and efficient solution for ASV spoof detection.
• This work represents a state-of-the-art single-system countermeasure for speaker verification security.