1D-CNN-based audio tampering detection using ENF signals.

Area of Science:

• Digital Forensics
• Audio Analysis
• Signal Processing

Background:

• Digital audio recording is prevalent in legal proceedings.
• Electric Network Frequency (ENF) analysis is crucial for audio forensics.
• Lack of ENF reference datasets hinders current audio tampering detection.

Purpose of the Study:

• To develop a blind audio forensics framework (ATD) for detecting audio tampering.
• To address the limitations of existing ENF-based methods by eliminating the need for reference databases.
• To improve the accuracy and robustness of audio tampering detection.

Main Methods:

• A one-dimensional convolutional neural network (1D-CNN) forms the core of the ATD framework.
• Feature extraction is enhanced by incorporating fundamental harmonic characteristics of ENF signals.
• A novel denoising method, ENF noise reduction (ENR), combines Variational Mode Decomposition (VMD) and Robust Filtering Algorithm (RFA).

Main Results:

• ATD achieves over 93% accuracy in detecting four types of audio tampering (deletion, insertion, replacement).
• In binary-class scenarios, ATD's tampering detection accuracy exceeds 96%.
• Experiments confirm ATD's effectiveness, efficiency, adaptability, and robustness across different detection scenarios.

Conclusions:

• The proposed ATD framework offers a reliable solution for blind audio forensics.
• ATD effectively identifies phase mutations and waveform discontinuities in tampered ENF signals.
• The integrated ENR method enhances the resilience of ENF signals against noise interference.