Detecting fraud via statistical anomalies

Abstract

With the increasing reliance on mobile device locational data, ensuring data authenticity is essential. However, structured patterns in GPS pings suggest possible manipulation, highlighting the need for robust fraud detection techniques. This study aims to detect fraudulent locational data by identifying statistical anomalies in mobile device pings using machine learning and statistical methods. We analyze mobile GPS pings from CITYDATA.ai for Singapore in September 2020, applying 2-D Kernel Density Estimation (KDE) and chi-square goodness-of-fit tests to compare the distribution of real and synthetic data. Additionally, we employ statistical methods such as isolation forests to identify abnormal patterns. Our results indicate that approximately 96.5% of devices exhibit structured, non-random patterns, suggesting potential manipulation. The KDE analysis reveals a significant divergence between actual and synthetic data distributions, confirming the presence of anomalies. We also explore possible explanations for such manipulation, considering that it may not always be driven by ill intent but rather by attempts to improve data accuracy. These findings lay the foundation for more accurate fraud detection in locational data, with implications for urban mobility research and government planning. Ultimately, our analysis suggests that the CITYDATA.ai locational ping dataset may not be a fully accurate representation of real-world locational data, whether due to intentional manipulation or unintentional data adjustments.