Novel transfer learning based acoustic feature engineering for scene fake audio detection.

Area of Science:

• Digital Security
• Audio Forensics
• Machine Learning

Background:

• Audio forensics is crucial for legal and security investigations.
• Sophisticated audio fake attacks pose a significant challenge to fake audio detection systems.
• The growing threat of manipulated audio content impacts media, legal evidence, and cybersecurity.

Purpose of the Study:

• To propose a novel transfer learning approach for enhanced fake audio detection.
• To improve the accuracy and efficiency of identifying manipulated audio content.
• To contribute to the integrity and reliability of digital communications.

Main Methods:

• Utilized the SceneFake benchmark dataset containing 12,668 audio files.
• Extracted Mel-Frequency Cepstral Coefficients (MFCC) and class prediction probability features.
• Developed a novel MfC-RF (MFCC-Random Forest) transfer learning method for feature extraction and classification.

Main Results:

• The proposed MfC-RF method achieved a high accuracy of 0.98.
• The MfC-RF approach outperformed existing state-of-the-art methods in fake audio detection.
• Hyperparameter tuning and cross-validation were applied to ensure robust performance.

Conclusions:

• The novel MfC-RF transfer learning method significantly enhances fake audio detection accuracy.
• The research contributes to more reliable digital security and forensic analysis.
• The study demonstrates the effectiveness of transfer learning in combating sophisticated audio manipulation techniques.