Improving Voice Spoofing Detection Through Extensive Analysis of Multicepstral Feature Reduction

Area of Science:

• Biometrics and Security
• Artificial Intelligence and Machine Learning
• Signal Processing

Background:

• Voice biometric systems are crucial for security but vulnerable to AI-powered spoofing attacks.
• Realistic synthetic speech poses a significant threat to current voice authentication methods.
• Developing robust defenses against sophisticated spoofing is essential for maintaining system integrity.

Purpose of the Study:

• To explore and evaluate various dimensionality reduction strategies for identifying spoofed voice signals.
• To enhance the effectiveness of supervised machine learning models in voice anti-spoofing.
• To assess the performance gains achieved through dimensionality reduction in voice biometric security.

Main Methods:

• Extraction of multi-cepstral features from voice signals.
• Application of diverse dimensionality reduction techniques: PCA, SVD, ANOVA F-value, Mutual Information, RFE, LASSO, Random Forest importance, Permutation Importance.
• Empirical evaluation using the ASVSpoof 2017 v2.0 dataset and Equal Error Rate (EER) metric.

Main Results:

• Dimensionality reduction methods significantly improved the performance of spoof detection.
• Achieved an Equal Error Rate (EER) of approximately 10% on the ASVSpoof 2017 v2.0 dataset.
• Demonstrated the effectiveness of feature selection and reduction in combating voice spoofing.

Conclusions:

• Dimensionality reduction is a valuable approach for enhancing voice biometric security against AI-driven spoofing.
• The proposed framework offers improved accuracy and robustness for voice authentication systems.
• Further research into advanced feature engineering and reduction techniques can bolster defenses against evolving threats.