Geographic Location Estimation from ENF Signals with High Accuracy

The Electric Network Signals (ENF) is the frequency characteristic of power distribution system, and can be captured in the form of power or audio recordings supplied by corresponding power system. The multi-dimensional diversity of the ENF between different regions allows it become a criterion of geographic localization. This triggered a lot of research on forensic authentication of digital and audio recordings. In this paper, we present a ENF-based location estimation system that integrates extraction and classification of power and audio signals from power grids. We uses several methods including autoregressive model and wavelet analysis to extract the statistical diversity of ENF signals as a whole to investigate the region-of-recording. The data is then classified by Support Vector Machine (SVM). We also built local sensing circuit to measure the local power grid data, and compare with the training data. The proposed scheme not only inherits high accuracy for power recordings, but also achieves improvements in audio recordings by pre-processing and utilization of higher harmonics. Extensive analytical and experimental results are presented which shows the accuracy and flexibility of our proposed system. We achieve accuracy rate of 93.3% for power signals and 80% for audio signals.