LSTM autoencoder based parallel architecture for deepfake audio detection with dynamic residual encoding and feature...

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Deepfake audio detection is crucial due to advancements in synthetic speech technology.
• Existing models face challenges like temporal inconsistencies and weak contextual representation.

Purpose of the Study:

• To enhance deepfake audio detection performance.
• To address limitations in current detection models.

Main Methods:

• Proposed a novel framework: Long Short-Term Memory Auto-Encoder with Dynamic Residual Difference Encoding (LSTM-AE-DRDE).
• Utilized attention-enhanced LSTM with contrastive learning for temporal cues.
• Employed residual differences across audio variants for real-vs-fake separability.
• Integrated diverse speech features: MFCC, temporal, prosodic, wavelet packet, and glottal parameters.

Main Results:

• Achieved high classification accuracies on five benchmark datasets (90%-97%).
• Demonstrated superior performance compared to eleven state-of-the-art methods with an ROC-AUC of ~98%.
• Feature-wise ablation study confirmed the framework's robustness.

Conclusions:

• The proposed LSTM-AE-DRDE framework significantly improves deepfake audio detection.
• Integration of diverse features and novel encoding methods enhances model reliability and performance.