The goal of this project is to investigate the reliability, robustness, and computationally efficiency of digital audio forensic methods under various adversarial conditions, e.g., lossy compression attack. We aim to identify and develop mathematical tools for modeling and characterizing of microphone nonlinearities (fingerprints), statistical methods for acoustic environment estimation, and system identification based framework for linking an acquisition device to the audio recording. More specifically, the project uses statistical modeling and extraction of microphone fingerprints by developing computationally efficient algorithms for nonlinear system identification and acoustic environment modeling and extraction using nonlinear filtering, and use them for linking a given recording to the acquisition device and to the acoustic environment. The algorithms developed through this project holds the potential for immediate effect in the area of digital audio forensic analysis, particularly in forensic analysis in compressed domain. The developed techniques will be evaluated for robustness, reliability, and computational complexity using datasets collected during this exploratory investigation and datasets available in the public domain. The expected outcomes also include datasets for performance evaluation of audio forensic methods and audio forensics tools robust to targeted attacks. The findings of this research, resulting audio forensic tools, and datasets will be made available to the research community via project webpage.
EAGER: Model Driven Framework for Audio Forensics