Explainable Camera Model Identification Employing Log-Mel Spectrograms from Videos' Audio

The Camera Model Identification (CMI) aims to identify the specific camera model used to capture an image or video, making it a fundamental challenge in multimedia forensics. An innovative CMI method is proposed that leverages both videos and log-Mel spectrograms from video audio, applying late fusion strategies using Convolutional Neural Networks whose hyperparameters are optimized through Neural Architecture Search. The method is integrated with Gradient-weighted Class Activation Mapping, which offers insights into influential spectrogram regions. A strong emphasis on high-frequency components within the audio data is observed, motivating band-pass filtering of the input spectrograms to retain high-frequency information selectively. By doing so, the classification accuracy of camera models is increased. Experiments conducted on the VISION dataset, comprising data from 35 different camera devices, demonstrate the effectiveness of the proposed model in achieving accurate and interpretable CMI.