Identification and Exploitation of Inadvertent Spectral Artifacts in Digital Audio

Abstract

We show that modulation products from local oscillators in a variety of commercial camcorders are coupled into the recorded audio track, creating narrow band time invariant spectral features. These spectral features, left largely intact by transcoding, compression and other forms of audiovisual post processing, can encode characteristics of specific camcorders used to capture the audio files, including the make and model. Using data sets both downloaded from YouTube and collected under controlled laboratory conditions we demonstrate an average probability of detection (Pd) approaching 0.95 for identification of a specific camcorder in a population of thousands of similar recordings, with a probability of false alarm (Pfa) of about 0.11. We also demonstrate an average Pd of about 0.93 for correct association of make and model of camcorder based on comparison of audio spectral features extracted from random YouTube downloads compared to a reference library of spectral features captured from known makes and models of camcorders, with a Pfa of 0.06. The method described can be used independently or synergistically with image plane-based techniques such as those based upon Photo Response Non-Uniformity.