Abstract:
The electrical network frequency (ENF) is the supply frequency of power grids (50 or 60 Hz). It fluctuates randomly over time around the nominal value. It has become a powerful tool in forensic applications such as multimedia authentication, tamper detection, time-of-recording validation, and region-of-recording identifica- tion. The temporal variation of ENF is uniform for a particular grid and separable from grid-to-grid observations due to different degrees of controls used to regulate the grid. ENF can potentially be used as the fingerprint of a particular grid lo- cation. Since the ENF signal is mostly used to authenticate the time of recording due to the high correlation between audio and power data, location identification from ENF data is a very challenging task, especially when time information for audio and power data is unavailable. This thesis mainly focuses on developing a novel algorithm to identify the geographic location in terms of grid from power and audio data, which will be helpful in location forensic and law enforcement applications. In this thesis, we have developed a robust scheme to extract ENF from power grid recordings and its harmonics. In this method, we have generated a sinusoidally time-frequency-distributed signal by applying the Kernel function and then extracted the instantaneous frequencies (ENFs) by applying the root MUSIC algorithm. Finally, we have investigated both a machine learning-based SVM clas- sifier and a deep neural network-based classifier that will differentiate between the grids based on media file location without relying on concurrent power references.