Source Acquisition Device Identification from Recorded Audio Based on Spatiotemporal Representation Learning with...

Area of Science:

• Digital Forensics
• Signal Processing
• Machine Learning

Background:

• Source acquisition device identification is crucial in audio forensics.
• Analyzing intrinsic audio characteristics for device identification presents significant challenges.
• Existing methods often struggle with accurate and efficient device attribution.

Purpose of the Study:

• To propose a spatiotemporal representation learning framework for source acquisition device identification.
• To enhance the accuracy and efficiency of identifying recording devices from audio signals.
• To develop a robust method for audio forensics using deep learning and attention mechanisms.

Main Methods:

• A two-branch deep feature extraction network combining residual dense temporal convolution networks (RD-TCNs) and convolutional neural networks (CNNs).
• Utilizing spatial probability distribution features for CNN-based spatial representation and temporal spectral features for RD-TCN-based temporal representation.
• Implementing multi-attention mechanisms (temporal, spatial, and branch attention) for effective spatiotemporal feature fusion.

Main Results:

• Achieved state-of-the-art performance on the CCNU_Mobile dataset.
• Reached an accuracy of 97.6% for identifying 45 different recording devices.
• Demonstrated a significant reduction in training time compared to existing models.

Conclusions:

• The proposed spatiotemporal framework effectively learns and fuses deep audio features for device identification.
• The multi-attention mechanism enhances the model's ability to capture relevant spatiotemporal information.
• This approach offers a promising solution for accurate and efficient source acquisition device identification in audio forensics.